for
ESTIMATION and CONTROL
A.B.Kurzhanski
I.V´alyi
Contents
Preface 1
Part I. Evolution and Control.
The Exact Theory 6
Introduction 6
1.1 The System 9
1.2 Attainability and the Solution Tubes 12
1.3 The Evolution Equation 15
1.4 The Problem of Control Synthesis:
a Solution Through SetValued
Techniques 22
1.5 Control Synthesis Through
Dynamic Programming Techniques 31
1.6 Uncertain Systems.
Attainability Under Uncertainty 40
1.7 Uncertain Systems :
the Solvability Tubes 47
1.8 Control Synthesis Under Uncertainty 53
1.9 State Constraints and Viability 62
1.10 Control Synthesis
Under State Constraints 68
i
1.11 State Constrained Uncertain Systems.
Viability Under Counteraction. 73
1.12 Guaranteed State Estimation :
the Bounding Approach 76
1.13 Synopsis 83
1.14 Why Ellipsoids ? 90
Part II. THE ELLIPSOIDAL CALCULUS 93
Introduction 93
2.1 Basic Notions : the Ellipsoids 94
2.2 External Approximations: the Sums.
Internal Approximations: the
Diﬀerences 105
2.3 Internal Approximations: the Sums.
External Approximations: the
Diﬀerences 119
2.4 Sums and Diﬀerences:
the Exact Representation 124
2.5 The Selection of Optimal Ellipsoids 127
2.6 Intersections of Ellipsoids 136
2.7 Finite Sums and Integrals:
External Approximations 149
2.8 Finite Sums and Integrals:
Internal Approximations 158
ii
Part III. ELLIPSOIDAL DYNAMICS:
EVOLUTION and CONTROL SYNTHESIS 164
Introduction 164
3.1 EllipsoidalValued Constraints 165
3.2 Attainability Sets and Attainability Tubes:
the External and Internal Approximations 168
3.3 Evolution Equations
with Ellipsoidal  Valued Solutions 177
3.4 Solvability in Absence of Uncertainty 181
3.5 Solvability Under Uncertainty 185
3.6 Control Synthesis
Through Ellipsoidal Techniques 194
3.7 Control Synthesis: Numerical Examples 199
3.8 ”Ellipsoidal” Control Synthesis
for Uncertain Systems 206
3.9 Control Synthesis for Uncertain Systems: Numerical Examples 211
3.10 Target Control Synthesis within
Free Time Interval 217
Part IV. ELLIPSOIDAL DYNAMICS:
STATE ESTIMATION and VIABILITY
PROBLEMS 221
Introduction 221
iii
4.1 Guaranteed State Estimation:
a Dynamic Programming Perspective 224
4.2 From Dynamic Programming to
Ellipsoidal State Estimates 232
4.3 The State Estimates, Error Bounds
and Error Sets 237
4.4 Attainability Revisited.
Viability Through Ellipsoids 240
4.5 The Dynamics of Information Domains.
State Estimation as a Tracking Problem 246
4.6 Discontinuous Measurements and
the Singular Perturbation Technique 254
References 258
iv
Preface
It is well known that the emphasis of mathematical modelling on the basis of available
observations is at ﬁrst  to use the data to specify or reﬁne the mathematical model, then 
to analyse the model through available or new mathematical tools and further on  to use
this analysis in order to predict or prescribe ( control ) the future course of the modelled
process. This is particularly done by specifying feedback control strategies (policies) that
realize the desired goals. An important component of the overall process is to verify the
model and its performance over the actual course of events.
The given principles are also among the objectives of modern control theory , whether
directed at traditional ( aerospace, mechanics, regulation, technology) or relatively new
applications ( environment, population , ﬁnances and economics, biomedical issues, com
munication and transport).
Among the speciﬁc features of the controlled processes in the mentioned areas are usually
their dynamic nature and the uncertainty in their description.
Following this reasoning, one may claim that control theory is a science of assigning feedback
control or regulation laws for dynamic processes on the basis of available information on
the system model and its performance goals (given both through online observations and
through data known in advance). It is on how to construct, in some appropriate sense, the
best or the better control laws. It is also to indicate how the level of uncertainty and the
” amount” of information used for designing the feedback control laws aﬀects the result
of the controlled process, particularly , the values of the cost functions or the aspired
”guaranteed” performance levels.
But it is not any type of theory that is desired. Not the least objective is to develop,
among the possible approaches, a solution theory that allows analytical designs relatively
simple for practical implementations or leads, at least, to eﬀective numerical algorithms.
These, desirably, should match the abilities of modern computer technology  allow parallel
calculations and graphic animation, for example.
So, what is this book about? The present book is devoted to an array of selected key
problems in dynamic modelling,  state estimation , viability and feedback control under
uncertainty. Its aim is to present a uniﬁed framework for eﬀectively solving these problems
and their generalizations to the end through modern computer tools.
The model of uncertainty considered here is deterministic, with setmembership descrip
tion of the uncertain items. These are taken to be unknown but bounded with preassigned
bounds and no statistical information whatever. The setmembership model of uncer
tainty reﬂects many actual informational situations in applied problems. Particularly, it
appears relevant in estimating nonrepetitive processes, processes with limited numbers of
1
observations , incomplete knowledge of the problem data and no available statistics. It is
a common approach in pursuitevasion diﬀerential games, in robust stabilization, control
and disturbance attenuation, particularly, under unmodelled dynamics. Needless to say ,
it also reﬂects the research preferences, interests and experiences of the authors.
1
The problems treated here are described through setvalued functions and are thus to be
treated through setvalued analysis. However, the aim of this book is not to produce any
type of setvalued technique, but a calculus that would allow eﬀective solutions of the
selected problems and their generalizations with fairly simple control designs and possi
bility of graphic animation. The attempt is based on introducing an ”ellipsoidal calculus”
that allows to represent the exact setvalued solutions of the respective problems through
ellipsoidalvalued functions. The solutions are thus constructed of elements that involve
only ellipsoidal sets or ellipsoidalvalued functions and operations over such sets or func
tions. This further allows to parallelize the calculations and to animate the solutions
through computer graphic tools.
It is necessary to indicate that the ellipsoidal techniques of this particular book are not
conﬁned to approximation of convex sets by one or several ellipsoids only as done in other
publications, but to indicate ellipsoidal representations of the exact solutions. Namely, each
convex set or convex setvalued function considered here is represented by a parametrized
variety of ellipsoids or ellipsoidalvalued functions, which, while their number increases,
jointly allow ( through their sums , unions or intersections), a more and more accurate
representation with exact one in the limit. The scheme includes approximations by single
ellipsoids as a particular element of the overall approach.
A particular emphasis of this book is on the possibility of computergraphic representations.
The animation of the problems in estimation , feedback control and gametype dynamics
not only allows to present the rather sophisticated mathematical solutions in ”visible”
forms ( and literally, to peer into the multidimensional spaces through computer windows
or more sophisticated computer tools). The authors believe that it may also give new
insights in to the mathematical structure of the solutions. (Thus, some assertions of general
nature proved in this book were ﬁrst ” noticed” during the animation experiments). The
authors also hope that though applying their techniques to a specially selected array of
problems, they also demonstrate an approach applicable to many other situations that
spread quite beyond the topics addressed here. These are certainly not conﬁned only to
control applications , but cover a broad variety of problems in systems modelling.
The book is divided into four parts designed along the following lines. The ﬁrst part gives
exact solutions to the problems of evolution  attainability (reachability) and solvability,
1
The references and some historical comments are given in the introductions to each part and throughout
the text. The authors apologize that among the enormous literature on the subject they were able to
mention only a very limited, representative, rather than exhaustive number of publications available to
them, with an emphasis on those directly related to the topics of this publication and those that would
allow, as we hope, to pursue the further directions indicated here.
2
as well as of estimation , viability and feedback terminal target control. The exact theory
( both known and new ) is rewritten within a uniﬁed framework that involves trajectory
tubes and their setvalued descriptions either through evolution equations of the ”funnel
type”, or through the evolution of support functions or through level sets of appropriate
HJB ( HamiltonJacobiBellman) equations. The feedback control designs are based on
using setvalued solvability tubes which may be interpreted as ”bridges” introduced by
N.N.Krasovski as a basis for further ”aiming rules”, as well as on Dynamic Programming
considerations. The principal schemes for this framework are speciﬁcally directed towards
the desired transition, through ellipsoidalvalued representations , to parallelizable compu
tation schemes.
The second part of the book describes the ellipsoidal calculus itself. It covers exter
nal and internal ellipsoidal representations for basic setvalued operationsgeometrical
(”Minkowski”) sums and diﬀerences as well as intersections of ellipsoids and integrals of
ellipsoidalvalued functions. Though written for applications to Problems of Part I, the
text of this chapter may be also considered as a separate theory with motivations and
applications coming from topics other than discussed here. Particularly, from optimiza
tion under uncertainty and multiobjective optimization, experiment planning, problems in
probability and statistics, interval analysis and its generalizations, adaptive systems and
robotics, image processing, mathematical morphology and related areas of theoretical and
applied research.
The third and fourth parts indicate the applications of ellipsoidal calculus to problems of
Part I. Thus, the third part describes ( both in forward and backward time) the internal
and external ellipsoidal representations of attainability (reachability) tubes for systems
without and with uncertainty. In the latter case these are more complicated, of course,
being related to reachability or solvability under uncertainty or counteraction and allowing,
particularly, a direct interpretation in terms of the above mentioned ”bridges”  the key
elements of gametheoretic feedback control. The third part also deals with feedback
control. The respective control designs are based on applying ”ellipsoidal” versions of
the exact solutions. This leads to a nonlinear control synthesis in the form of analytical
designs , except for a scalar parameter whose dependence on the state space vector may be
calculated in advance, through the solution of a simple algebraic equation. These analytical
designs are possible due to the fact that the internal ellipsoidal tubes that approximate the
solvability domains under uncertainty are precisely such, that they posses the property of
being an ”ellipsoidalvalued bridge”. The latter property arrives due to two basic features
: the fact that the respective ellipsoidalvalued mappings posses a generalized semigroup
property and the fact that the internal tubes are inclusionmaximal among all other internal
ellipsoidal tubes.
The fourth part deals with stateestimation under unknown but bounded errors, with at
tainability under state constraints and viability problems. It also indicates the applicability
of the suggested schemes to problems posed within the socalled H
∞
approaches, when the
3
value of the error bound is not speciﬁed. This is due to the involvement of Dynamic Pro
gramming techniques, particularly, of one and the same HJB equation for the treatment
of uncertain dynamics in both of the settings investigated here. Other topics include new
types of dynamic relations for the treatment of information sets and vectorvalued ”guar
anteed estimators” as well as an interpretation of the state estimation problem as one of
tracking an unknown motion under unspeciﬁed but bounded errors. This links the problem
with those of viability under counteraction. The ﬁnal Section 4.6 deals with problems
of state estimation under ”bad noise”, which are approached through the incorporation
of singular perturbation techniques. This approach allows to treat discontinuous observa
tions described by measurable functions and also to deal with viability problems under
”measurable constraints”. Numerical examples complement the theoretical parts.
The narrative stops just short of control under measurement feedback and adaptive control.
These are areas which require separate serious consideration and explanation. However,
the application of ellipsoidal techniques would be especially useful in these areas , as we
believe. The respective challenges are beyond the material of the present book.
As already mentioned , this book indicates a uniﬁed concise framework for problems of state
estimation, viability and feedback control under setmembership uncertainty for systems
with linear dynamics. It introduces an ellipsoidal calculus  to develop the solutions from
theory to algorithms and computer animation, and thus to solve the problem to the end.
This book is not a collection of numerous facts or artifacts in setvalued analysis or control
theory. It is rather a book on basic problems and principles for calculating their solutions
through setvalued models. Whether reached or not, our aim was also to stimulate and
encourage further investigation in the spirit of the present approach as well as implementa
tions in reallife modelling. (The latter issue could be the topic of a separate monograph).
In this text we are conﬁned to ”linearconvex” systems and problems. However, the con
trol synthesis given here is nonlinear and the synthesized systems are nonlinear systems.
Moreover, the Dynamic Programming approaches applied here open the routes to fur
ther penetration into generically nonlinear classes of systems. The algorithmization and
animation in these cases is certainly a worthy challenge.
2
Another important aspect hardly discussed here is the accuracy and computational com
plexity of the underlying algorithms.
The conceptual approaches to controlled dynamics that served as a background for this
work were inﬂuenced by the research of N.N.Krasovski and his associates at in Ekaterinburg
(Sverdlovsk), where the ﬁrst of the authors had earlier worked for many years. In the
2
The ”nondiﬀerentiable” version of Dynamic Programming has been substantially developed in the
recent years (see references [82], citecranish, [289]), becoming an eﬀective tool in nonlinear control theory,
particularly. For covering the needs of this book we use its simple versions that do not extend beyond the
use of subdiﬀerential calculus.
4
necessity of studying and applying setvalued analysis we believe to share the views of
JP.Aubin and his colleagues.
The principal parts of this book and the underlying ellipsoidal ”representation” approach
for the setvalued functions were worked out throughout the authors participation at the
SDS (Systems and Decision Sciences) Program of IIASA  the International Institute of
Applies Systems Analysis at Laxenburg, Austria. The serious but friendly atmosphere,
pleasant working conditions and possibility of regular contacts with a broad spectrum of
researchers certainly stimulated our work at the Institute and the direction of our eﬀorts.
The authors are grateful to the Directors of IIASA  Thomas Lee, Robert Pry and Peter
De J´anosi and to the Chairman of the IIASA Council in 19871992 , the late Vladimir
Sergeevich Michalevich, for their support of methodological research at IIASA, particularly
of our own investigations.
We wish to thank our colleagues at IIASA and its Advisory Committee on Methodol
ogy — J.P. Aubin, H.Frankowska, A.Gaivoronski, P.Kenderov, G.Pﬂug, R.T.Rockafellar,
W.Runggaldier, K.Sigmund, M.Thoma, V.Veliov, R.Wets, A.Wierzbicki for their stimulat
ing discussions and support, K.Fedra and M.Makowski for their help in computer graph
ics, the SDS secretaries E.Gruber and C.EnzlbergerVaughan for preparing papers and
manuscripts used in this book. We thank T.Filippova, O.Nikonov, M.Tanaka, K.Sugimoto
who have coauthored some of our IIASA Working Papers used here and O.A.Schepunova
for her help in arranging the ﬁnal version of the manuscript.
Throughout the last years we had the pleasant opportunity to discuss the topics
treated in this book with Z.Artstein, J.Baras, F.Chernousko, M.Gusev, A.Isidori, P.Kall,
R.Kalman, H.Knobloch, A.Krener, G.Leitmann, C.Martin, M.Milanese, E.Mischenko,
S.Mitter, J.Norton, Yu.Osipov, B.Pschenichnyi, S.Robinson, A.Rusczynski, P.SaintPierre,
A.Subbotin, T.Tanino, V.Tikhomirov, P.Varaiya, S.Veres, J.Willems. Their valuable com
ments certainly helped to shape the contents.
Our special thanks go to C.Byrnes  the Editor of the Birkhauser Series on Systems and
Control: Foundations and Applications, to E.Beschler, the Publisher and to the Birkhauser
staﬀ for their support, patience and understanding of the problems faced by the authors
in preparing the manuscript.
5
Part I. Evolution and Control.
The Exact Theory
Introduction
The present ﬁrst part of the book is a narrative on constructive techniques for modeling
and analyzing an array of key problems in uncertain dynamics, estimation and control. It
presents a uniﬁed approach to these topics based on descriptions involving the notions of
trajectory tubes and evolution equations for these tubes. The class of systems treated here
are linear timevariant systems
˙ x = A(t)x + u +f(t), x(t
0
) = x
0
,
with magnitude bounds on the controls u and the uncertain items f(t), x
0
. The target
control processes and the estimation problems are considered within ﬁnite time intervals
: t ∈ [t
0
, t
1
]. This requires a rather detailed investigation of the system dynamics. The
present topic thus diﬀers from problems where the objective lies only in the achievement of
an appropriate asymptotic behaviour of the trajectories with perhaps some desired quality
of the transient process.
The ﬁrst step is the description of attainability (reachability) domains A[t] for systems
without uncertainty. The evolution of these in time is naturally described by setvalued
(”Aumann”) integrals with variable upper time limit. This immediately leads to set
valued ”attainability tubes” whose crossections are the attainability domains. However,
it is not unimportant to introduce some sort of evolution equation with setvalued state
space variable that would describe the dynamics of sets A[t] in time. The tubes A[t] could
then be interpreted as ”trajectories ” of some generalized dynamic systems. (Among the
ﬁrst investigations with this emphasis are the works of E.Barbashin and E.Roxin ,see [34],
[270]). The serious obstacle for deriving such an equation in a diﬀerential form is the
diﬃculty in deﬁning an appropriate derivative for setvalued functions. The objective is
nevertheless reached through evolution equations of the funnel type that do not involve
such derivatives. Though somewhat cumbersome at ﬁrst glance, these equations indicate
setvalued discretetime schemes important for calculations.
3
.
The attainability tubes may be also constructed in backward time in which case they are
referred to as solvability tubes . The solvability tubes are used here in synthesizing feedback
control strategies for problems of terminal target control. Namely, if the solvability tube
ends at the target set /, then the synthesized control strategy should be designed to
keep the trajectory within this tube (or ”bridge”) throughout the process. This idea is
3
Among the recent investigations on evolution equations for trajectory tubes are papers [246], [298],
[193], [17]
6
the essence of the ”extremal aiming rule” introduced by N.N.Krasovski, [168], [169] and
used in Section 1.4. A key element that allows to use the solvability tubes for the control
synthesis problem is that the respective multivalued maps satisfy a semigroup property and
therefore generate a generalized dynamic system with setvalued trajectories.
A similar type of strategy may be derived through a Dynamic Programming technique
with cost function being the square of the Euclid distance d
2
(x[t
1
], /), from endpoint
x[t
1
] to the target set / , for example (see Section 1.5). Selecting a starting position
¦t, x¦, x = x(t), we may minimize the cost function by selecting an appropriate optimal
control ( in the class of either openloop or closedloop controls). Finding the optimal
value of the cost function for any position ¦t, x¦ we come to the value function V (t, x). One
should note that in the absence of uncertainty the value function V (t, x) is the same both for
openloop (programmed) control and for closedloop (positional) feedback control. For the
”linearconvex” problems of these Sections the function V (t, x) may be therefore calculated
through standard methods of convex analysis used traditionally for solving related problems
of openloop control, [167], [266], [181]. The function V (t, x) then satisﬁes a corresponding
generalized HJB (HamiltonJacobiBellman) equation.
The next stage is the treatment of systems with input uncertainty f(t) ( unknown but
bounded, with magnitude bounds). In this case the attainability set under counteraction
(in forward time ) and solvability set under uncertainty ( in backward time) are in general
far more complicated than in the absence of uncertainty (see Sections 1.61.8). One should
now distinguish, for example,the openloop solvability tubes from the closedloop solvability
tubes (Section 1.6). Under some nondegeneracy conditions, the latter ones may be again
interpreted as Krasovski’s ”bridges” ( now for uncertain systems) and may be used for
designing feedback strategies through the extremal aiming rules (Sections 1.7, 1.8). The
backward procedure for solvability tubes is also similar in nature to the schemes introduced
by P.Varaiya et al. [307] and B.Pschenichnyi [259]. In the linearconvex case considered
here the constructive description of solvability tubes may be given by a setvalued integral
known as L.S.Pontryagin’s alternated integral, [257], (Section 1.7). It is indicated here that
they also satisfy some special evolution equations of the funnel type. There is a particular
case, however, when the openloop and closedloop solvability tubes coincide. This is when
the system satisﬁes the socalled matching condition which mean that the bounds on the
controls u and the disturbances f are similar in some sense (Section 1.6). The calculation
of the solvability tubes is then as simple as in the absence of uncertainty.
One may also apply Dynamic Programming to the mentioned uncertain systems. Taking
the cost function d
2
(x(t
1
), /), for example, we note that now it should be minimaximized
over the control u and the input disturbance f respectively. But the value of this minmax,
when calculated over closed loop controls is diﬀerent, in general, from its value calculated
for open loop controls. It is the former value that may be described through a respective
HJBI ( HamiltonJacobiBellmanIsaacs) equation. (see [109], [171], [219], [289].) There
is an exception again, nevertheless. Namely, if the matching conditions are satisﬁed , then
7
the minmax of the cost function or, in other words, the value function V (t, x) is the same,
whether calculated over openloop or closedloop controls.
Having in mind the previous remarks, one may observe that the value functions V (t, x)
used in this book play the role of Liapunov functions used in respective approaches to the
design of feedback controllers for uncertain systems ( see [214], [215]). We should also
emphasize that the problem treated here is to reach the goal in ﬁnite time attenuating the
unknown disturbances. Namely, it is to ensure the system to be steered by control u to the
target set / ( at given time t
1
) under persistent disturbances f rather than to ﬁgure out
the saddle points of a positional (feedback) dynamic game between two ”equal” players u
and f which is the emphasis of the theory of diﬀerential games (see [37], [171], [50], [119]
).
The further problems are similar to the previous ones but complicated by state constraints
(”viability restrictions”). Evolution ” funnel” equations are introduced for the dynamics
of attainability sets under state constraints (in forward time) and respective solvability
sets ( in backward time). The latter ones then coincide with viability kernels introduced
by J.P.Aubin, ([15]), within the framework of viability theory (Section 1.9). The control
synthesis problem is now to ensure viability (Section 1.10) or viability under counterac
tion (or under persistent disturbances, in another interpretation), while also reaching the
terminal set ( Section 1.11).
The last problem of the ﬁrst part is the one of state estimation under unknown but bounded
errors and disturbances (Section 1.12).
4
The main objects of investigation here are the information sets consistent with the system
dynamics, the available measurement and the constraints on the uncertain items. The
information sets are actually the attainability domains under a state constraint that is
induced by the measurement equation and therefore arrives online, together with the result
of the measurement. The evolution equation for the information set acts a ”guaranteed
ﬁltering equation” and the guaranteed state estimate is then the ”Chebyshev center” of
this set ( namely, the center of the smallest ball that includes the information set ). This
ﬁrst part of the book gives but a general introduction into the problem, whilst constructive
techniques are introduced in parts III, IV, where one may also ﬁnd some connections with
other approaches to deterministic ﬁltering (particularly, the H
∞
approach, [94], in the
interpretation of J.Baras and M.James [30]).
A synopsis of the results and some suggestions on why ellipsoids were taken to be studied
ﬁnalize this part.
We now proceed with the main text, commencing with the basic notions.
4
The ﬁrst investigations of state estimation problems under unknown but bounded inputs date to papers
[166], [317], [178]. A systematic investigation of the setvalued approach in continuous time seems to have
started with [54], [276], [179], [181].
8
1.1 The System
In this book we consider dynamic models described in general by a linear timevariant
system
˙ x(t) = A(t)x(t) + u +f(t) (1.1.1)
with ﬁnitedimensional state space vector x(t) ∈ IR
n
and inputs u (the control) and f(t)
(the disturbance or external forcing term). The n n matrix function A(t) is taken to be
continuous on a preassigned interval of time T = ¦t ∈ IR : t
0
≤ t ≤ t
1
¦ within which we
consider the forthcoming problems, with u(t), f(t) assumed Lebesguemeasurable in t ∈ T
.
The values u of the controls are assumed to be restricted for almost all t by a magnitude
or ”geometrical ” constraint
u ∈ T(t) , (1.1.2)
where T(t) is a multivalued function T : T → convIR
n
, continuous in t.
We shall further consider two types of controls which are:
— open loop, when u = u(t) is a function of time t, measurable on T (a “measurable
control”), and
— closed loop, when u = (t, x(t)) is a multivalued map, namely
 : T IR
n
−→ convIR
n
measurable in t and upper semicontinuous in x, being a function of the position ¦t, x¦
of the system.
In the ﬁrst case we come to a linear diﬀerential equation
˙ x(t) = A(t)x(t) + u(t) + f(t) (1.1.3)
with u(t) ∈ T(t), t ∈ T, being an openloop control. The class of functions u() = u(t), t ∈
T,measurable in t ∈ T and restricted as in (1.1.2) is further denoted as U
O
P
.
In the second case we come to a nonlinear diﬀerential inclusion
˙ x(t) ∈ A(t)x(t) +(t, x(t)) + f(t), (1.1.4)
where
(t, x) ⊆ T(t), t ∈ T, (1.1.5)
9
is a feedback (closedloop) control strategy .
The class U
c
P
= ¦(t, x)¦ of feasible control strategies consists of all convex compactvalued
multifunctions that are measurable in t, upper semicontinuous in x, being restricted by
(1.1.5) and such that equation (1.1.4) does have a solution extensible to any ﬁnite time
interval T for any x
0
= x(t
0
) ∈ IR
n
. The latter means that there exists an absolutely
continuous function x(t), t ∈ T, that yields the inclusion
˙ x(t) ∈ A(t)x(t) +(t, x(t)) + f(t)
for almost all t ∈ T.
The existence of solution for system (1.1.3) is a standard property of linear diﬀerential
equations, ([61], [142], [167], [248]).
Systems (1.1.3), (1.1.4) may be transformed into simpler relations.
Let S(t, τ) stand for the matrix solution to the equation
∂
∂t
S(t, τ) = −S(t, τ)A(t), S(τ, τ) = I, (1.1.6)
which also satisﬁes the equation
∂
∂τ
S(t, τ) = A(τ)S(t, τ), S(t, t) = I.
As it is well known, the solution to (1.1.3) with initial value
x(t
0
) = x
0
(1.1.7)
is given by the formula
x(t) = S(t
0
, t)x
0
+
_
t
t
0
S(τ, t)(u(τ) + f(τ))dτ.
Taking the transformation
z(t) = S(t, t
1
)x(t) (1.1.8)
and substituting x for z in (1.1.3) we come to the equation
˙ z(t) = S(t, t
1
)u(t) + S(t, t
1
)f(t) (1.1.9)
z
0
= z(t
0
) = S(t
0
, t
1
)x
0
. (1.1.10)
Clearly, there is a onetoone correspondence of type (1.1.6) between the solutions x(t) and
z(t) to equations (1.1.3) and (1.1.9) respectively. The initial values for these are related
10
through (1.1.10). Therefore, instead of the systems (1.1.3), (1.1.4), constraint (1.1.2) and
initial condition (1.1.7), we come to
˙ z(t) = w(t) + g(t), z(t
0
) = S(t
0
, t
1
)x
0
, (1.1.11)
˙ z(t) ∈ J(t, z) + g(t) (1.1.12)
with constraint
w(t) ∈ T
0
(t). (1.1.13)
Here obviously
w(t) = S(t, t
1
)u(t)
g(t) = S(t, t
1
)f(t)
J(t, z) = S(t, t
1
)(t, S
−1
(t
0
, t)z)
T
0
(t) = S(t, t
1
)T(t)
and the setvalued function T
0
(t) remains continuous. The “new” feedback strategies
J(t, z) belong to the class deﬁned by the constraint T
0
(t), but otherwise the same as
before:
J(t, z) ∈ U
c
P
0
.
Without loss of generality we may therefore further treat systems (1.1.11)–(1.1.13) rather
than (1.1.2)–(1.1.4). It is compulsory however that the constraint function T
0
(t) would be
timevariant. In other terms, without loss of generality we may further follow the notations
of (1.1.2)–(1.1.4) with A(t) ≡ 0.
One should realize , however, that the described substitution (1.1.8) allows to consider the
forthcoming problems for A(t) ≡ 0 within the time range ¦t ≤ t
1
¦.A similar result may be
also obtained by substitution
z(t) = S(t, t
0
) x(t). (1.1.14)
Then the original system may be again , without loss of generality, taken with A(t) ≡ 0 ,
but the time range for which the respective substitution is true will be ¦t ≥ t
0
¦.
We shall often make use of the indicated facts in the sequel in the hope that this will
enable to demonstrate the basic techniques without overloading the text with unessential
but cumbersome procedures. The reader will always be able to return to A(t) ,≡ 0 as a
healthy exercise.
The ﬁrst issue to discuss is the description of the set of states that can be reached in ﬁnite
time due to systems (1.1.3), (1.1.4) under restriction (1.1.2) and (1.1.5).
11
1.2 Attainability and the Solution Tubes
Taking system (1.1.3), (1.1.2) for A(t) ≡ 0, we have
˙ x(t) = u(t) + f(t), t ∈ T, (1.2.1)
with constraint (1.1.2) u(t) ∈ T(t). We also presume that the initial state x
0
= x(t
0
) is
restricted by the inclusion
x
0
∈ A
0
, A
0
∈ comp IR
n
. (1.2.2)
One of the ﬁrst questions that arise in control theory is to describe the variety of all states
x = x(t) that can be reached by the system trajectories that start at a prescribed set A
0
.
Let x[t] = x(t, t
0
, x
0
) denote an isolated trajectory of system (1.2.1) that starts at instant
t
0
from state x
0
, being driven by a certain control u(t). We will be further interested
in the union of all such isolated trajectories over all possible initial states x
0
∈ X
0
and
measurable controls u(t) ∈ T(t). Therefore we denote
A[t] = A(t, t
0
, A
0
) =
_
¦x(t, t
0
, x
0
) : x
0
∈ A
0
, u(t) ∈ T(t), t ∈ T¦.
For the mapping A(t, t
0
, ) : compIR
n
→ compIR
n
it is not diﬃcult to check that it satisﬁes
the following semigroup property:
A(t, t
0
, A
0
) = A(t, τ, A(τ, t
0
, A
0
)),
whatever are the values t, τ with t
1
≥ t ≥ τ ≥ t
0
.
Deﬁniton 1.2.1 The set A[t] = A(t, t
0
, A
0
) is referred to as the attainability domain
for system (1.2.3) (or (1.2.1), (1.1.2.) at time t, from set A
0
.
The attainability domain A[t] is often said to be the reachability domain .
The setvalued map
A[t] = A(t, t
0
, A
0
) , t ∈ T,
deﬁnes a solution tube to the diﬀerential inclusion
˙ x(t) ∈ T(t) + f(t), (1.2.3)
12
that starts from set A
0
. In other words, the set A[t] = ¦x
∗
¦ consists of all those vectors
x
∗
for each of which there exists an isolated trajectory x[τ] = x(τ, t
0
, x
0
), t
0
≤ τ ≤ t, of
(1.2.3) that satisﬁes the boundary conditions x[t
0
] ∈ A
0
, x[t] = x
∗
.
It is clear that the control
u(t) = ˙ x[t] −f(t)
is the one that corresponds to x[t], so that we could also indicate
A[t] = ¦x[t] : ˙ x[τ] −f(τ) ∈ T(τ), t
0
≤ τ ≤ t, x[t] ∈ X
0
¦. (1.2.4)
The multivalued function A[t], t ∈ T, A[t
0
] = A
0
is also known as the solution tube to
system (1.2.1), under restriction (1.1.2), from set A
0
, for the interval t ∈ [t
0
, t
1
] = T.
5
As a preliminary exercise it is not diﬃcult to prove the following
Lemma 1.2.1 The multifunction A[t] is convex compactvalued (A[t] ∈ convIR
n
) and
continuous on the interval T.
Remark 1.2.1 One of the popular problems studied on the subject of attainability is the
following: given A
0
= ¦0¦, will the set
A
0
= ∪¦A[t], t ∈ [t
0
, ∞)¦,
coincide with the whole space IR
n
?.
An aﬃrmative answer will indicate that any point in IR
n
may be reached in ﬁnite time
through a bounded control u() ∈ U
O
P
.Otherwise one is to specify A as a subset of IR
n
.
Exercise 1.2.1. Investigate the problem of Remark 1.2.1.
Passing to the diﬀerential inclusion
˙ x(t) ∈ (t, x) + f(t), (1.2.5)
(, ) ∈ U
c
P
, (1.2.6)
5
Other terminology says that A[t] is the trajectory assembly generated by system (1.2.3) and set A[t
0
] =
A
0
, [181].
13
where the class of feasible feedback strategies U
c
P
is as deﬁned in Section 1.1, we come to
the following questions.
Let A
0
U
[t] = A
U
(t, t
0
, x
0
) be the crossection of the set of all isolated trajectories x[t] that
satisfy the relation
˙ x[t] ∈ (t, x[t]) + f(t), x[t
0
] = x
0
,
for a given multivalued map (, ) ∈ U
c
P
. A particular element of U
c
P
is the setvalued map
T(t) itself, so that (1.2.3) could be viewed upon as a particular case of equation (1.2.5),
when (t, x) ≡ T(t).
Denote for a ﬁxed  of (1.2.6):
A
U
[t] = A
U
(t, t
0
, A
0
) =
_
¦A
U
(t, t
0
, x
0
) : x
0
∈ A
0
¦,
and further
A
∗
[t] = A
∗
(t, t
0
, A
0
) =
_
¦A
U
(t, t
0
, A
0
) : (, ) ∈ U
c
P
¦.
Then one may want to know what is the relation between the tubes A
∗
[t] obtained for the
closedloop system (1.2.5), (1.2.6) and A[t] obtained for the openloop system (1.2.3)?
Theorem 1.2.1 With A
∗
[t
0
] = A[t
0
] = A
0
the following relation is true:
A[t] ≡ A
∗
[t], t ∈ T.
To prove this assertion we observe that every singlevalued function u(t) can be treated as
an element of U
c
P
. Therefore A[t] ⊆ A
∗
[t], t ∈ T. To show the opposite assume there exists
a trajectory x
∗
[t] = x
∗
(t, t
0
, x
0
), x
0
∈ A, which satisﬁes the inclusion x
∗
[t] ∈ A
∗
[t], t ∈ T,
namely
˙ x
∗
[τ] ∈ 
∗
(t, x
∗
[τ]) + f(τ).
for some (, ) = 
∗
(, ) ∈ U
c
P
. Then obviously
˙ x
∗
[τ] ∈ 
∗
(τ, x
∗
[τ]) + f(τ) ⊆ P(τ) + f(τ)
and due to (1.2.4) this yields x
∗
[t] ∈ A[t], t ∈ T.
The main conclusion given by Theorem 1.2.1 is such that with function f(t) given (there
is no uncertainty in system (1.1.1), (1.1.2)), the solvability tube A[t] for system (1.1.1),
(1.1.2) taken in the class U
0
P
of open loop controls u[t] is the same as the solvability tube
A
∗
[t] taken in the class U
c
P
of closedloop controls (t, x). This conclusion is also true when
the closed loop controls are selected among appropriate classes of singlevalued functions
u = u(t, x) ∈ T(t) that allow the existence and prolongation of solutions of (1.1.1) with
u = u(t, x), t ∈ T.
The next question is whether it would be possible to describe the evolution of sets A[t] in
time t through some type of evolution equation with setvalued states A = A[t].
14
1.3 The Evolution Equation
We shall now introduce an evolution equation with state space variable A ∈ convIR
n
, whose
solution will be precisely the tube A[t] of Section 1.1.2.
Obviously
A[t] = A
0
+
t
_
t
0
T(τ)dτ +
t
_
t
0
f(τ)dτ, (1.3.1)
where the second term in the right hand is the setvalued Lebesgue integral (“the Aumann
integral”, [25]) for the function T. The question is therefore whether one could construct
an evolution equation for describing A[t].
Deﬁniton 1.3.1 The Hausdorﬀ semidistance h
+
(A, ¸) between sets A, ¸ ∈ conv IR
n
is
introduced as
h
+
(A, ¸) = min¦γ ≥ 0 : A ⊆ ¸ +γo¦
or equivalently
h
+
(A, ¸) = max
x
min
y
¦(x −y, x −y)
1
2
[x ∈ A, y ∈ ¸¦.
Similarly
h
−
(A, ¸) = h
+
(¸, A).
The following properties are true for A, ¸, Z ∈ convIR
n
:
(i) h
+
(A, ¸) = 0 implies A ⊆ ¸
(and h
−
(A, ¸) = 0 implies ¸ ⊆ A).
(ii) h
+
(A, Z) + h
+
(Z, ¸) ≥ h
+
(A, ¸).
Deﬁniton 1.3.2 The Hausdorﬀ distance h(A, ¸) between sets A, ¸ ∈ convIR
n
is intro
duced as
h(A, ¸) = max¦h
+
(A, ¸), h
−
(A, ¸)¦.
15
Obviously
(iii). h(A, ¸) = 0 implies A = ¸; for A, ¸ ∈ convIR
n
As it is well known, a closed convex set A ∈ convIR
n
may be described by its support
function
ρ(l[A) = sup¦(l, x) [ x ∈ A¦
which is a positively homogeneous convex function of l, namely
ρ(αl[A) = αρ(l[A) for α ≥ 0
and
ρ(α, l
1
+ α
2
l
2
[A) ≤ α
1
ρ(l
1
[A) + α
2
ρ(l
2
[A),
where α
1
> 0, α
2
≥ 0, α
1
+ α
2
= 1.
For A ∈ compIR
n
we have ρ(l[A) < ∞, ∀l ∈ IR
n
. A well known property is given by
Lemma 1.3.1 The inclusion x ∈ A, A ∈ convIR
n
is equivalent to the inequality
(l, x) ≤ ρ(l[A), ∀l ∈ IR
n
.
Direct calculation gives us the following formulae:
h
+
(A, ¸) = max ¦ ρ(l[A) −ρ(l[¸) : l ≤ 1¦
h(A, ¸) = max ¦ [ρ(l[A) −ρ(l[¸)[ : l ≤ 1¦ (1.3.2)
16
Deﬁniton 1.3.3 A function A : T → convIR
n
is said to be absolutely hcontinuous on T
if for any ε > 0 there exists a δ > 0 such that condition
i
(t
i
−t
i
) < δ
yields
i
h(A[t
i
], A[t
i
]) ≤ ε.
The deﬁnition of absolute h
+
continuity is given by mere substitution of h by h
+
in Deﬁ
nition 1.3.3.
Lemma 1.3.2 A function A : T → IR
n
is absolutely hcontinuous if the support function
ρ(l[A[t]) = f(l, t) is absolutely continuous in t ∈ T uniformly in l ∈ S, S = ¦l : (l, l) ≤ 1¦.
Now we may consider the “equation”
limσ
−1
σ→0
h(A[t + σ], A[t] + σT(t) + σf(t)) = 0 (1.3.3)
with “initial value”
A[t
0
] = A
0
. (1.3.4)
Deﬁniton 1.3.4 A multivalued function Z : T → convIR
n
is said to be a solution of
(1.3.3), (1.3.4) if it is absolutely hcontinuous and satisﬁes (1.3.3) for almost all t ∈ T,
together with (1.3.4).
Let us see whether A[t] is a solution to (1.3.3) in the sense of the last deﬁnition.
Rewriting (1.3.1) in terms of support functions, we come to
ρ(l[A[t]) = ρ(l[A
0
) +
t
_
t
0
ρ(l[T(τ))dτ +
t
_
t
0
(l, f(τ))dτ. (1.3.5)
17
Here we made use of the fact that for a continuous map T : T → convIR
n
, the following is
true
ρ(l[
t
_
t
0
T(τ)dτ) =
t
_
t
0
ρ(l[T(τ))dτ.
To calculate
h(A[t + σ], A[t] + σT(t) + σf(t)) = H(σ, t),
due to (1.3.3), (1.3.5) at ﬁrst we have
R(l, σ, t) = ρ(l[A[t + σ]) − ρ(l[A[t]) − σρ(l[T(t)) −σ(l, f(t)),
R(l, σ, t) =
t+σ
_
t
[(ρ(l[T(τ)) + (l, f(τ))]dτ −σρ(l[T(t)) − σ(l, f(t))
In case of continuous f(t) and T(t) we further have
σ
−1
t+σ
_
t
f(τ)dτ → f(t), σ → 0 (1.3.6)
for all t and
σ
−1
t+σ
_
t
ρ(l[T(τ))dτ → ρ(l[T(t)), σ → 0 (1.3.7)
if f(t) is not continuous, being only measurable, relation (1.3.6) is still true, but now only
for almost all of the values of t, which are the “points of density” of f(t), [19]. A similar
remark is true for (1.3.7) with ρ(l[T(t)) measurable in t, [232].
Taking into account the equality
H(σ, t) = max¦[R(l, σ, t)[ : l = 1¦
and the relation
18
lim
σ→0
σ
−1
R(l, σ, t) = 0,
that follows from (1.3.6), (1.3.7), and is uniform in l ∈ S, being true for almost all t ∈ T,
we observe
limσ
−1
σ→0
H(σ, t) = 0
for almost all t ∈ T. This proves the following assertion:
Theorem 1.3.1 The map A : T → convIR
n
, is a solution to the evolution equation (1.3.3).
Theorem 1.2.1 implies the following
Corollary 1.3.1 The map A
∗
[t] of Section 1.2.1 is a solution to the evolution equation
(1.3.3).
It is not uninteresting to write down a formal analogy of equation (1.3.3) when A(t) ,≡ 0.
This is as follows:
lim
σ→0
σ
−1
h(A[t + σ], (I + σA(t))A[t] + σT(t)
+ σf(t)) = 0. (1.3.8)
A solution A[t] to (1.3.7) with given initial state A[t
0
] = A
0
, A
0
∈ convIR
n
is one that
satisﬁes Deﬁnition 1.3.4, but with equation (1.3.3) substituted by (1.3.8).
Let us now have a look at what would equation (1.3.8) turn to be when A[t] = ¦x[t]¦ and
T(t) = ¦p(t)¦ are singlevalued. Then, clearly,
h(x
, x
) = d(x
, x
) = (x
−x
, x
−x
)
1/2
and
x[t +σ] = (I +σA(t))x[t] + σp(t) + σf(t) + o(σ), (1.3.9)
19
where σ
−1
o(σ) → 0 with σ → 0. This yields
σ
−1
(x[t + σ] −x[t]) = A(t)x[t] + p(t) + f(t) + σ
−1
o(σ)
for almost all t ∈ T or after a limit transition σ → 0:
˙ x[t] = A(t)x[t] + p(t) + f(t), x[t
0
] = x
0
(1.3.10)
for almost all t ∈ T.
Thus, equation (1.3.8) is clearly a setvalued analogy of the ordinary diﬀerential equation
(ODE) (1.3.10) that may also be presented in a form similar to (1.3.8), which is (1.3.9).
There is no special point, however, in presenting an ODE in the form (1.3.9). It is not so
for setvalued maps, where equation (1.3.8) may be quite convenient, particularly because
here we avoid the unpleasant operation of subtraction of sets or setvalued functions.
Equation (1.3.3) may be integrated. The integral form for its solution A[t] is given by
(1.3.1) and is described by a multivalued Lebesgue integral,[25].
The support function for A[t] satisﬁes the partial diﬀerential equation
∂ρ
∂t
(l[A[t]) = ρ(l[T(t)) + (l, f(t)), (1.3.11)
ρ(l[A[t
0
]) = ρ(l[A
0
), t ∈ T, l ∈ IR
n
(1.3.12)
for almost all t ∈ T and all l ∈ IR
n
, that follows from (1.3.5). The derivative ∂ρ/∂t exists
for almost all t due to the properties of the integrals in (1.3.5). From (1.3.11), (1.3.12) its
is not diﬃcult to observe that A[t] is the only solution to (1.3.3), (1.3.4):
Lemma 1.3.3 The solution A[t] to equation (1.3.3), (1.3.4) is unique.
To conclude this paragraph we shall introduce another version of the evolution equation
(1.3.3), namely by substituting the Hausdorﬀ distance h() for a semidistance h
+
(). This
gives
lim
σ→0
σ
−1
h
+
(Z[t + σ], Z[t] + σT(t) + σf(t)) = 0 , (1.3.13)
20
Z[t
0
] ⊆ A
0
. (1.3.14)
A solution Z[t] to (1.3.12), is speciﬁed as in Deﬁnition 1.3.4 but with equation (1.3.3)
substituted by (1.3.13). Here a solution Z[t] to (1.3.12) satisﬁes an inclusion
Z[t + σ] ⊆ Z[t] + σT(t) + σf(t) + o(σ)S,
rather than an equality, (⊆ instead of =,) which would be the case for (1.3.3). This directly
yields a partial diﬀerential inequality
∂ρ
∂t
(l[Z[t]) ≤ ρ(l[T(t)) + (l, f(t)) (1.3.15)
for the solution Z[t]. The initial condition also satisﬁes an inequality
ρ(l[Z[t
0
]) ≤ ρ(l[A
0
). (1.3.16)
It is not diﬃcult to observe that (1.3.13) has a nonunique solution. Particularly, any single
valued trajectory x(t) driven by a control u(t) ∈ T(t) with x
0
∈ A
0
will be one of these.
Integrating (1.3.13), we come to
ρ(l[Z[t]) ≤ ρ(l[A[t
0
]) +
t
_
t
0
(ρ(l[T(τ)) + (l, f(τ)))dτ = ρ(l[A[t]),
for all t ∈ T, l ∈ IR
n
, in view of (1.3.15) and (1.3.16). This leads us to the assertion
Lemma 1.3.4 The solutions Z[t], A[t] to the evolution equations (1.3.3) (1.3.4) and
(1.3.13), (1.3.4) respectively, satisfy the inclusion
Z(t) ⊆ A[t],
for all t ∈ T.
We emphasize again that (1.3.3), (1.3.4) has a unique solution, while, in general, the
solution to (1.3.13), (1.3.4) is nonunique.
21
Deﬁniton 1.3.5 A solution Z
0
[t] to (1.3.13) is maximal if
Z[t] ⊆ Z
0
[t], ∀t ∈ T,
for any solution Z[t] to (1.3.13) with the same initial condition (1.3.4).
As an exercise the reader may prove the following
Lemma 1.3.5 The maximal solution Z
0
[t] to (1.3.13), (1.3.14) exists and coincides with
the unique solution A[t] to (1.3.3), (1.3.4).
We will further use the evolution equations (1.3.3), (1.3.8), (1.3.18) and their generaliza
tions as an essential tool for describing the topics of this book. Among the ﬁrst of these
in the problem of Control Synthesis.
We shall ﬁrst present a constructive technique for Control Synthesis based on setvalued
calculus and further used here in the Sections devoted to ellipsoidalvalued dynamics. A
still further Section 1.1.5 is intended to indicate that the technique of Section 1.1.4 is not
an isolated approach, but allows an equivalent representation in conventional terms terms
of Dynamic Programming as applied in either a standard or a ”nondiﬀerentiable” version.
1.4 The Problem of Control Synthesis:
a Solution Through SetValued
Techniques
Consider system (1.1.1)  (1.1.2) and a “terminal set” / ∈ convIR
n
.
Deﬁniton 1.4.1 The problem of control synthesis consists in specifying a solvability set
J
∗
(τ, t
1
, /) and a feedback control strategy u = (t, x), (, ) ∈ U
c
P
such that all
the solutions to the diﬀerential inclusion
˙ x(t) ∈ (t, x) + f(t) (1.4.1)
that start from any given position ¦τ, x
τ
¦, x
τ
= x[τ], x
τ
∈ J
∗
(τ, t
1
, /), τ ∈ [t
0
, t
1
), would
reach the terminal set / at time t
1
: x[t
1
] ∈ /.
22
The deﬁnition is nonredundant provided x
τ
∈ J
∗
(τ, t
1
, /) ,= ∅, where the solvability set
J
∗
(τ, t
1
, /) = J
∗
[τ] is the ”largest” set of states from which the solution to the problem
of control synthesis does exist at all.(More precisely this will be speciﬁed below).
Taking J
∗
(t, t
1
, /) for any instant t ∈ [t
0
, t
1
], we come to a setvalued map J
∗
[t] =
J
∗
(t, t
1
, /), t ∈ T, (the “solvability tube”) where J
∗
[t
1
] = /.
To describe the tube J
∗
[t] we ﬁrst start from the following
Deﬁniton 1.4.2 The “openloop” solvability set J(τ, t
1
, /) is the set of all states x
τ
∈
IR
n
such that there exists a control u(t) ∈ T(t), τ ≤ t ≤ t
1
that steers the system from x
τ
to / due to a respective trajectory x[t], τ ≤ t ≤ t
1
, so that x[τ] = x
τ
, and x[t
1
] ∈ /.
The set J[τ] = J(τ, t
1
, /) is nothing more than the attainability domain at instant τ for
system (1.1.1), (1.1.2), from set /, but calculated in backward time, namely, from t
1
to τ.
The respective map J[t], t ∈ T, J(t
1
) = / is deﬁned as the “open loop” solvability tube
for set /, on the interval T.
A direct consequence of Theorem 1.3.1 and the deﬁnition of J[t] is the following
Theorem 1.4.1 The setvalued function J[t] satisﬁes the evolution equation
lim
σ→0
σ
−1
h(J[t −σ], J[t] −σT(t) −σf(t)) = 0 (1.4.2)
J[t
1
] = / (1.4.3)
and the semigroup property
J(τ, t
1
, /) = J(τ, t, J(t, t
1
, /))
for all t
0
≤ τ ≤ t ≤ t
1
.
Its solution is obviously
J[t] = /−
t
1
_
t
T(τ)dτ −
t
1
_
t
f(τ)dτ. (1.4.4)
Equation (1.4.2) is the same as (1.3.3), but is treated in backward time. The deﬁnition of
the solution is naturally, also the same.
23
Deﬁniton 1.4.3 The “closed loop” solvability set J
∗
(τ, t
1
, /) is the set of all states
x
τ
∈ IR
n
such that there exists a control strategy u = (t, x), (, ) ∈ U
c
P
that ensures
every trajectory x[t] of the diﬀerential inclusion (1.4.1) that starts at τ, x[τ] = x
τ
, to end
in set / : x[t
1
] ∈ /.
The respective map J
∗
[t] = J
∗
(t, t
1
, /), t ∈ T, J
∗
[t
1
] = / deﬁnes the “closedloop”
solvability tube J[] for set /.
From Theorem 1.2.1 we come to
Lemma 1.4.1 With J[t
1
] = J
∗
[t
1
] = / the open loop and closed loop solvability tubes,
which are J[t] and J
∗
[t], do coincide, namely
J[t] ≡ J
∗
[t], ; t ∈ T.
Tube J
∗
[t] therefore satisﬁes the evolution equation (1.4.2), (1.4.3). With A(t) ,≡ 0 we
have
(1.4.5)
lim
σ→0
σ
−1
h (J[t −σ], (I −σA(t)) J[t] −σT(t) −σf(t)) ≡ 0.
The solutions to (1.4.2), (1.4.3), or (1.4.5), (1.4.3) are unique and given by convex compact
valued maps.
Substituting Hausdorﬀ distance h() for semidistance h
+
() we come to the equation
lim
σ→0
σ
−1
h
+
(Z[t −σ], Z[t] −σT(t) −σf(t)) = 0 (1.4.6)
Z[t
1
] ⊆ / (1.4.7)
which is the same as (1.3.13), (1.3.14) but taken in backward time. The deﬁnition of its
solution and maximal solution are analogies of those given in the previous section for direct
(”forward”) time. By analogy with Lemma 1.3.5 we also come to
Lemma 1.4.2 With J[t
1
] = /, the map J[t], t ∈ T, is the maximal solution to (1.4.6),
(1.4.7).
24
This is a consequence of the deﬁnition of the solvability sets and it is certainly important
to emphasize that the inclusion
x
τ
∈ J(τ, t
1
, /)
is necessary and suﬃcient for the existence of a synthesizing strategy (, ) ∈ U
c
P
. An
essential element in constructing the strategy  = (t, x) is the tube J[t].
Assume x ∈ IR
n
and set J[τ] to be given. Let us introduce a ”synthesizing function”
V (t, x) = d
2
[τ, x], where
d[τ, x] = h
+
(x, J[τ]),
h
+
(x, J[τ]) = min¦x −w [ w ∈ J[τ]¦.
Clearly,
V [τ, x] = 0 implies x ∈ J[τ],
and
V [τ, x] > 0, implies x ,∈ J[τ].
(One may observe that
J[τ] = ¦x : V (t, x) ≤ 0¦
is the level set for V (t, x)).
We may now investigate the derivative
d
dt
V (t, x)
along the trajectories of system (1.2.1). The control set 
0
(t, x) will then consist, as we
shall see, of all the values u(t) ∈ T(t) that minimize this derivative, namely,

0
(t, x) = arg min¦
d
dt
V (t, x)
¸
¸
¸
¸
(1.2.1)
u ∈ T(t)¦.
25
Let us specify this in detail. A direct diﬀerentiation yields
d
dt
V (t, x)
¸
¸
¸
¸
(1.2.1)
= 2d[t, x](
d
dt
d[t, x]
¸
¸
¸
¸
(1.2.1)
), (1.4.8)
where
d[τ, x] = max¦(l, x) −ρ(l[J[τ]) : l ≤ 1¦ = (1.4.9)
= (l
0
, x) −ρ(l
0
[J[t]),
and where l
0
,= 0, l
0
 = 1, is a unique maximizer for d[τ, x] > 0. We will always choose
the maximizer to be l
0
= ¦0¦ if d[τ, x] = 0.
Since J[t] is absolutely continuous, it is not diﬃcult to prove the following property.
Lemma 1.4.3 Let x
∗
[t] be an absolutely continuous function on an interval e where d
∗
(t) =
h
+
(x
∗
[t], J[t]) > 0. Then the function d
∗
(t) is absolutely continuous on the same interval.
We further need the derivative
dd[t, x])/dt when d[t, x] > 0, due to the system (1.2.1), which is
˙ x(t) = u(t) + f(t).
For this we obtain:
(1.4.10)
d
dt
d[t, x] =
∂
∂t
d[t, x] +
_
∂
∂x
d[t, x], ˙ x(t)
_
= (l
0
, ˙ x(t)) −
∂
∂t
ρ(l
0
[J[t])
= (l
0
, u(t) + f(t)) + ρ(−l
0
[T(t)) − (l
0
, f(t)),
and therefore
d
dt
d[t, x(t)] = (l
0
, u(t)) + ρ(−l
0
[T(t)). (1.4.11)
26
Here we have used the formula
ρ(l[J[t]) = ρ(l[/) +
t
1
_
t
ρ(−l[T(τ))dτ −
t
1
_
t
(l, f(τ))dτ
that follows from (1.4.4) and the fact that in calculating the derivative (1.4.10) for d[t, x]
which is the maximum over l in (1.4.9), we should avoid diﬀerentiation in l
0
.
Indeed, following [86], [265], [261], we observe, that for a diﬀerentiable function of type
h(t, x(t)) = max¦H(t, x(t), l) : l = 1¦,
with unique maximizer l
0
we have
dh(t, x)
dt
=
∂H(t, x, l
0
)
∂t
+
_
∂H(t, x, l
0
)
∂x
, ˙ x
_
,
where l
0
= arg max ¦H(t, x, l) : l = 1¦.
Remark 1.4.1 The direct calculation of ∂ρ(l[J[t])/∂t introduced in (1.4.9) also indicates
that with Z(t) = J[t] the inequality (1.3.15) turns to be an equality.
We now proceed with specifying the feedback strategy 
0
(t, x). Since l
0
depends on t and
x we further use the notation l
0
= l
0
(t, x).
The strategy 
0
(t, x) will have to be speciﬁed both in the domain ¦x ,∈ J[t]¦ (or V (t, x) >
0) and in ¦x ∈ J[t]¦ (or V (t, x) = 0).
Assume V (t, x) > 0. Then 
0
(t, x) will be deﬁned as

0
(t, x) = arg min
_
d
dt
d[t, x]
¸
¸
¸
¸
u(t) ∈ T(t)
_
=
= arg min
_
d
dt
V (t, x)
¸
¸
¸
¸
u(t) ∈ T(t)
_
.
(We further omit the index (1.2.1) that indicates the system along whose solutions we
calculate the derivative).
27
Due to (1.4.11), this turns into

0
(t, x) = arg min¦(l
0
(t, x), u)
¸
¸
¸
¸
u(t) ∈ T(t)¦
or, what is the same,

0
(t, x) = arg max¦(−l
0
(t, x), u)
¸
¸
¸
¸
u(t) ∈ T(t)¦. (1.4.12)
(One should observe, that with V (t, x) = 0 we have l
0
= 0 and therefore 
0
(t, x) = T(t)).
Relations (1.4.9), (1.4.10) yield the following assertion
Lemma 1.4.4 With d[t, x] > 0 the derivative
d
dt
d[t, x] ≥ 0 for any u ∈ T, (1.4.13)
d
dt
d[t, x] = 0 for u(t) ∈ 
0
(t, x), (1.4.14)
where 
0
(t, x) is deﬁned by relation (1.4.11).
Combining this with (1.4.8),(1.4.11),(1.4.12), we come to
Lemma 1.4.5 For any position ¦t, x¦the derivative
d
dt
V (t, x)
¸
¸
¸
¸
(1.2.1)
≤ 0 , (1.4.15)
provided u ∈ 
0
(t, x)
The latter relations allow to prove
Theorem 1.4.2 The strategy u(t) = 
0
(t, x) deﬁned by equation (1.4.11), does solve the
problem of control synthesis speciﬁed in Deﬁnition 1.4.1.
28
Assume that x
0
∈ J[t
0
] and that the inclusion (1.4.1) is run by strategy u(t) = 
0
(t, x),
which, in general, generates a tube
A[t, 
0
] = A(t, t
0
, x
0
[
0
) = ¦x
0
(t, t
0
, x
0
¦, t ∈ T,
of isolated trajectories x
0
[t] = x
0
(t, t
0
, x
0
) to (1.4.1), ( = 
0
) .
The proof of Theorem 1.4.2 is based on the following
Lemma 1.4.6 The tube A[t[
0
], t ∈ T; A[t
0
[
0
] = x
0
, x
0
∈ J[t
0
], satisﬁes the inclusion
A[t[
0
] ⊆ J[t], t ∈ T,
therefore
A[t
1
[
0
] ⊆ /.
Proof. Assume x[t] = x(t, t
0
, x
0
) is a trajectory of inclusion (1.4.1),  = 
0
, with
x
0
∈ J[t
0
] or equivalently, with V (t
0
, x
0
) ≤ 0 and x[t] ∈ A[t[
0
], t ∈ T. We shall prove
that x[t] ∈ J[t], or equivalently, that V (t, x[t]) ≤ 0, for all t ∈ T. Then , for any value of
t ∈ (t
0
, t
1
], we observe that the integral
t
_
t
0
dV (τ, x[τ])
dτ
dτ = V (t, x[t]) −V (t
0
, x[t
0
]) ≤ 0
due to (1.4.15).Since x[t
0
] ∈ J[t
0
] , this yields
V (t, x[t]) ≤ V (t
0
, x[t
0
]) ≤ 0
for any t ∈ (t
0
, t
1
] and thus proves Lemma 1.4.6 from which Theorem 1.4.1 follows directly.
The same property is true if x
0
is substituted by a set A
0
∈ J[t
0
].
Corollary 1.4.1 With A
0
= A(t
0
[
0
) ⊆ J[t
0
] the respective tube of solutions A[t[
0
] =
A(t, t
0
, A
0
[
0
) to the diﬀerential inclusion (1.4.1) generated by strategy (t, x) = 
0
(t, x),
satisﬁes the relation
A[t[
0
] ⊆ J[t], t ∈ T,
and therefore A[t
1
[
0
] ⊆ /.
29
We thus observe that if for an instant τ ∈ T the inclusion A
τ
= A[τ[
0
] ⊆ J[τ] is true,
then
A[t[
0
] ⊆ J[t] (1.4.16)
will hold for all t ≥ τ,i.e., for the whole trajectory tube A[t[
0
], t ∈ T
τ
, T
τ
= [τ, t
1
]
generated by system (1.4.1),  = 
0
.
The tube A[t[
0
] therefore satisﬁes (1.4.16) as a state constraint. According to the ter
minology of , [17] A
τ
is strongly invariant relative to tube J[t], t ∈ T
τ
, for time τ. The
latter means that every trajectory of the inclusion (1.4.1),  = 
0
that evolves from set
A
τ
remains within J[t]. It is now obvious that the largest strongly invariant set for time
τ, relative to J[t], t ∈ T
τ
, is J[τ] itself.
The feedback strategy 
0
(t, x) may be rewritten in terms of the notion of subdiﬀerential,
[265], [267], [261]. We recall that a subdiﬀerential ∂
l
f(t, l
0
) (in the variable l, at point
l = l
0
) of a function f(t, l) convex in l, is the set of all vectors q such that
f(t, l) − f(t, l
0
) ≥ (q, l −l
0
), ∀l ∈ IR
n
. (1.4.17)
Assume f(t, l) = ρ(l[T(t)). Then, due to the deﬁnition, a vector q ∈ ∂
l
f(t, l
0
) if and only
if
ρ(l[T(t)) − ρ(l
0
[T(t)) ≥ (q, l −l
0
), ∀l ∈ IR
n
.
From here it follows (taking l = 2l
0
), that
ρ(l
0
[T(t)) − (l
0
, q) ≥ 0, (1.4.18)
and therefore,
ρ(l[T(t)) − (q, l) ≥ (l
0
[T(t)) − (l
0
, q) ≥ 0, ∀l ∈ IR
n
,
whence q ∈ T(t).
On the other hand, with l = 0, we come to
(l
0
, q) ≥ ρ(l
0
[T(t)). (1.4.19)
A comparison of (1.4.18), (1.4.19) and a substitution of q for u yields
30
Lemma 1.4.7 With f(t, l) = ρ(l[T(t)), the respective subdiﬀerential ∂
l
f(t, l) is given by:
∂
l
f(t, l
0
) = ¦u ∈ IR
n
[ (l
0
, u) = ρ(l
0
[T(t))¦.
Clearly, for l
0
= 0 we have ∂
l
f(t, l
0
) = T(t), and therefore

0
(t, x) = ∂
l
f(t, −l
0
(t, x)),
where l
0
(t, x) is the maximizer for (1.4.9). l Summarizing the reasoning of the above, we
conclude the following.
Theorem 1.4.3 The feedback strategy 
0
(t, x) that solves the problem of control synthesis
may be deﬁned as

0
(t, x) = ∂
l
f(t, −l
0
(t, x)), (1.4.20)
where f(t, l) = ρ(l[T(t)) and l
0
(t, x) is the maximizer for problem (1.4.9).
One may check, as an exercise, that the strategy 
0
(t, x) ∈ U
c
P
and belongs to the class of
feasible strategies introduced in Section 1.2.
1.5 Control Synthesis Through
Dynamic Programming Techniques
For a controltheorist experienced in the conventional methods of this theory the geomet
rical techniques of setvalued calculus, as introduced in the above and further used in the
sequel, may seem, at ﬁrst glance , to be somewhat unusual. It may be demonstrated , how
ever, that they are quite in line with the wellknown fundamentals of control theory. We
therefore feel obliged to indicate, in a very concise form, a ”conventional” way of looking
at the problems under discussion.
Assume a position ¦τ, x¦ due to system (1.1.1) to be given together with a terminal set
/ ∈ convIR
n
.(Although matrix A(t) ,≡ 0 is present in the ﬁrst part of this Section,
one may always take A(t) ≡ 0 as shown in Section 1.2). Let us indicate a cost criterion
1(τ, x) for the problem of control synthesis, presuming that our objective will be to ﬁnd
31
an ”optimal” control strategy u = u(t, x) ∈ U
c
P
that minimizes this criterion. More
speciﬁcally, let us assume
1(τ, x
∗
) = h
2
+
(x[t
1
], /]), (1.5.1)
where x[t
1
] = x(t
1
, τ, x
∗
)
The optimal value
1
0
(τ, x) = min ¦1(τ, x)[ u(, ) ∈ U
c
P
¦
when taken for any position ¦τ, x¦ will be further referred to as the value function
1(τ, x) = 1
0
(τ, x).
It is obvious that 1(τ, x) = 0 , if x[t
1
] ∈ / and 1(τ, x) > 0 if x[t
1
] ,∈ /
Therefore, the solvability domain W[τ] of Section 1.4 is actually the level set
W[τ] = ¦x : 1(τ, x) ≤ 0¦.
Let us now calculate function 1(τ, x) by formally writing down the H−J −B (Hamilton
JacobiBellman) equation for the problem of minimizing cost criterion (1.5.1) along the
trajectories of system (1.1.1) with u = u(t, x) ∈ U
c
P
. (The respective theory may be
found in [109], [53]). This is
(1.5.2)
∂1(t, x)
∂t
+ min
__
∂1(t, x)
∂x
, A(t)x + u +f(t)
_¸
¸
¸
¸
u ∈ T(t)
_
= 0
with boundary condition
1(t
1
, x) = h
2
+
(x, /) (1.5.3)
or, more precisely,
(1.5.4)
∂1(t, x)
∂t
+
_
A
(t)
∂1(t, x)
∂x
, x
_
− ρ
_
∂(−1(t, x))
∂x
¸
¸
¸
¸
T(t) + f(t)
_
= 0
with same boundary condition (1.5.3).
32
We have to check, however, whether these formal operations are justiﬁed. We shall do that
by calculating the value 1(τ, x) directly, through the technique of convex analysis.
Obviously, the function
φ(x) = h
2
+
(x, /) = min¦(x −q, x −q)[q ∈ /¦
has a conjugate
φ
∗
(l) = max¦(l, x) −φ(x)[x ∈ IR
n
¦ =
= max¦max¦(l, x) − (x −q, x −q)[x ∈ IR
n
¦[q ∈ /¦ = max¦(l, q) +
1
4
(l, l)[q ∈ /¦
which is
φ
∗
(l) = ρ(l[/) +
1
4
(l, l). (1.5.5)
Our problem is to ﬁnd
1
0
(τ, x) = min¦φ(x(t
1
))[u(t) ∈ T(t), τ ≤ t ≤ t
1
¦
over the trajectories of system (1.2.1), (1.1.2) with given initial position ¦τ, x¦.We have
min
u(·)
φ(x(t
1
)) = min
u(·)
max
l
¦(l, x(t
1
)) −φ
∗
(l)¦ =
= max
l
¦min
u(·)
¦(l, x(t
1
)) − φ
∗
(l)¦¦
The function in the brackets in the righthand side is linear in u() and concave in l ,
with φ(l) → ∞ as l → ∞. This indicates that the operations of min and max are
interchangeable ( [101] ).
Denoting s(t, t
1
, l) to be the solution of the adjoint equation
˙ s = −sA(t) , s(t
1
) = l, t ≤ t
1
,
33
( s is a vectorrow) and using the notations of (1.1.6),(1.1.7) , we may rewrite
(l, x(t
1
)) = (l, S(τ, t
1
)x) +
t
1
_
τ
(l, S(t, t
1
)u(t) + f(t)) dt =
= s(τ, t
1
, l)x +
t
1
_
τ
s(t, t
1
, l)(u(t) + f(t)) dt.
Hence
1
0
(τ, x) = max¦Φ(τ, x, l)[l ∈ IR
n
¦, (1.5.6)
where
Φ(τ, x, l) = s(τ, t
1
, l)x −
−
t
1
_
τ
ρ(−s(t, t
1
, l)[T(t) + f(t)) dt −φ
∗
(l).
The function Φ(τ, x, l) is concave in l (moreover, even strictly concave, due to the quadratic
term). The maximum in (1.5.6) is therefore attained at a single vector l
0
= l
0
(τ, x),
whatever is the position ¦τ, x¦.
Lemma 1.5.1 The maximizer l
0
(τ, x) of (1.5.6) is continuous in ¦τ, x¦.
Indeed, suppose the sequence
¦τ
k
, x
(k)
¦ → ¦τ, x¦, k → ∞,
Then, due to the properties of Φ(τ, x, l) in l, the respective sequence of maximisers
¦l
0
(τ
k
, x
(k)
)¦ will be equibounded, so that there exists a limit
lim
k→∞
l
0
(τ
k
, x
(k)
) = l
∗
Taking into account the obvious inequalities,
Φ(τ
k
, x
(k)
, l
0
(τ
k
, x
(k)
)) ≥ Φ(τ
k
, x, l
0
(τ, x))
Φ(τ, x, l
0
(τ, x)) ≥ Φ(τ, x, l
0
(τ
k
, x
(k)
))
34
and passing to the limits (k → ∞), we observe (due to the continuity of Φ(τ, x, l) in all of
its variables) ,
Φ(τ, x, l
∗
) ≥ Φ(τ, x, l
0
(τ, x)),
Φ(τ, x, l
0
(τ, x)) ≥ Φ(τ, x, l
∗
),
so that Φ(τ, x, l
∗
) = Φ(τ, x, l
0
(τ, x)). Due to the unicity of l
0
(τ, x) this yields l
∗
= l
0
(τ, x)
and the Lemma is thus proved.
Denote A
0
[τ] = ¦x : 1
0
(τ, x) ≤ 0¦
Lemma 1.5.2 For any x ∈ A
0
[τ] we have
l
0
(τ, x) = arg max¦Φ(τ, x, l)[l ∈ IR
n
¦ = ¦0¦.
This follows from the explicit expression for Φ(τ, x, l).
A direct diﬀerentiation of 1
0
(τ, x) in x now gives
∂1
0
(τ, x)
∂x
= s(τ, t
1
, l
0
(τ, x)). (1.5.7)
(Recall that since l
0
(τ, x) is unique, the respective formula is as follows
∂
∂x
_
max¦Φ(τ, x, l)[l¦
_
=
∂
∂x
Φ(τ, x, l)
¸
¸
¸
¸
l=l
0
(τ.x)
).
Similarly
∂1
0
(τ, x)
∂τ
= ρ(−s(τ, t
1
, l)[T(t) + f(t)) −s(τ, t
1
, l)A(τ)x. (1.5.8)
Taking 1(t, x) = 1
0
(t, x) and substituting into (1.5.2), we have, in view of (1.5.7),(1.5.8),
ρ(−s(t, t
1
, l
0
(t, x))[T(t)) −s(t, t
1
, l
0
(t, x))(A(t)x + f(t))+ (1.5.9)
min¦s(t, t
1
, l
0
(t, x))(A(t)x + u + f(t))[u ∈ T(t)¦ = 0.
In order to check the boundary condition (1.5.3) , we observe from formula (1.5.6)
1
0
(t
1
, x) = max¦(l, x) −ρ(l[/) −
1
4
(l, l)[l ∈ IR
n
¦ = ψ
∗
(x), (1.5.10)
where
ψ(l) = ρ(l[/) +
1
4
(l, l) = φ
∗
(l)
due to (1.5.5.). Hence ψ
∗
(x) = (φ
∗
)
∗
(x) = φ(x) = h
2
+
(x, /).
Therefore, the following assertion turns to be true
35
Theorem 1.5.1 The value function 1(t, x) = 1
0
(t, x) given by formula (1.5.6) satisﬁes the
Dynamic Programming (HJB) equation (1.5.2) ((1.5.4)) with boundary condition (1.5.3).
The respective control u = u(t, x) is then formally determined from (1.5.2),(1.5.9) as
u
0
(t, x) = arg max
__
∂(−1(t, x))
∂x
, u
_¸
¸
¸
¸
u ∈ T(t)
_
. (1.5.11)
Particularly, with 1(t, x) = 0 , this gives
u
0
(t, x) ≡ T(t). (1.5.12)
(This reﬂects that ∂1(t, x)/∂t = 0 if 1(t, x) = 0, in view of Lemma 1.5.2).
The control u
0
(t, x) is thus similar to 
0
(t, x) deﬁned in Section 1.4. , while (−∂1(t, x)/∂x)
plays the role of vector l
0
(t, x) in (1.4.12),(1.4.20).
We therefore come to an equivalent of Theorem 1.4.2.
Theorem 1.5.2 The solution strategy u
0
(t, x) is given by by relations (1.5.11), (1.5.12),
where 1(t, x) is the value function 1
0
(t, x).
Let us now calculate the value
r[τ, x] = min¦h
+
(x[t
1
], /)[u(t) ∈ P(t), τ ≤ t ≤ t
1
¦
assuming A(t) ≡ 0. Following the scheme for calculating (1.5.6), we have
r[τ, x] = max¦Ψ(τ, x, l)[l ≤ 1¦,
where
Ψ(τ, x, l) = (l, x) −
t
1
_
τ
ρ(−l[T(t) + f(t))dt −ρ(l[/) =
= (l, x) −ρ(l[J[t])
and J[t] is deﬁned by (1.4.4). This yields r[τ, x] = d[τ, x] and therefore leads us to
Lemma 1.5.3 With A(t) ≡ 0 the value function
1
0
(τ, x) = 1(τ, x) = d
2
[τ, x]
36
Thus, under condition A(t) ≡ 0 ( which does not imply any loss of generality, as we have
seen), the solution given in Section 1.4 through setvalued techniques is precisely the one
derived in this Section through Dynamic Programming.
We shall further continue to indicate the Dynamic Programming interpretations of the
outcoming relations which, of course, shall be somewhat more complicated in the case of
uncertain systems and state constraints. Nevertheless, in the problems of this book, aimed
particularly at the applicability of ellipsoidal calculus , the value functions will turn to be
convex in x. They will be directionally diﬀerentiable and therefore allow a more or less
clear propagation of the notions of Dynamic Programming (DP).
In the more general case of nonlinear systems and an arbitrary terminal cost φ(x) the
main inconvenience is that there may be no diﬀerentiable function 1(t, x) that solves the
DP ( HJB ) equation (a nonlinear analogy of equation (1.5.2)), whereas if we look for
nondiﬀerentiable functions, then the partial derivatives of 1 may not be continuous or may
not even exist at all. The solution to the DP equation should then be interpreted in some
generalized sense. Particularly, it may be interpreted as a viscosity solution , [82], [109], or
its equivalent  the minmax solution , [289].
Looking at the solution (1.5.10),(1.5.11), one may observe that for deﬁning u
0
(t, x) through
DP techniques, one needs to know the following elements:
• the level sets
W
0
[t] = ¦x : 1(t, x) ≤ 0¦
• the partial derivatives ∂1/∂x in the domain ¦x : 1(t, x) > 0¦.
For the problems treated in this book these elements may be determined without integrating
equation (1.5.4) but through direct constructive techniques which , particularly, are those
formulated in Section 1.4.One just has to recognise that d
2
[τ, x] = 1(τ, x) and therefore that
the level set W
0
[t] is the solvability set J[t] , (W
0
[t] ≡ J[t](!)), while the antigradient
(−∂1/∂x) is collinear with l
0
(t, x) in (1.4.12).
Needless to say, the elements 1(t, x), ∂1(t, x)/∂x, may be, of course, calculated by inte
grating equation (1.5.4) or its analogies ( in a generalized sense, perhaps) . This integration
will be an essential tool for the treatment of those nonlinear problems for which the tech
niques of this book cease to be eﬀective.
Example 1.5.1. Let us write down equation (1.5.2), with boundary condition (1.5.3) for
the particular case when the system is autonomous , A ≡ 0,and /, T(t) are nondegenerate
ellipsoids,namely,
/ = ¦x : (x −m, M
−
1(x −m)) ≤ 1¦ = c(m, M),
37
T(t) = ¦u : (u −p, P
−
1(u −p)) ≤ 1¦ = c(p, P),
M, P > 0.
We have
∂1(t, x)
∂t
+
_
∂1(t, x)
∂x
, f(t) + p
_
+ (1.5.13)
+
_
∂1(t, x)
∂x
, P
−1
∂1(t, x)
∂x
_1
2
= 0
with
(1.5.14)
1(t, x) = (x −m, M(x −m))(1 − (x −m, M(x −m))
−
1
2
)
2
, x ,∈ c(m, M),
1(t, x) = 0 , x ∈ c(m, M).
Relation (1.5.14) follows from (1.5.10) by direct calculation.
Exercise 1.5.1: with A(t) ,≡ 0 indicate the cost criterion 1
∗
for which the value function
1
∗
would be
1
∗
(t, x) = h
2
+
(x, J[t]),
where J[t] is the solvability set of Section 1.4.
Let us now indicate another relation for the solvability set W[τ]. Taking system (1.2.1) and
position ¦τ, x¦, τ ∈ [t
0
, t
1
], x = x(τ), solve the following problem: minimize the functional
Ψ(τ, x, u()) = max¦I
T
, I
1
¦, (1.5.15)
where
I
T
= (x(t
1
) −m, M(x(t
1
) −m)),
I
1
= esssup
t
(u(t) −p(t), P(t)(u(t) −p(t))), t ∈ [τ, t
1
].
38
Introduce value function
V (τ, x) = min¦Ψ(τ, x, u())[u()¦
Then, clearly
W[τ] = ¦x : V (τ, x) ≤ 1¦. (1.5.16)
We shall now indicate an explicit relation for V (τ, x).
First – consider set
W
µ
[τ] = m + µc(0, M) −
_
t
1
τ
(p(t) + f(t) + µc(0, P(t)))dt.
This set is similar to set J[τ] of (1.4.4) with / = m+µc(0, M), T(t) = p(t)+µc(0, P(t)).
Its support function
ρ(l[W
µ
[τ]) = (l, x
∗
(τ)) + µ
_
(l, Ml)
1/2
+
_
t
1
τ
(l, P(t)l)
1/2
dt
_
,
where x
∗
(t) is the solution to system
˙
x
∗
= p(t) + f(t), x
∗
(t
1
) = m.
Second – for a given position ¦τ, x¦ ﬁnd smallest µ for which x ∈ W
µ
[τ]. We have x ∈ W
µ
[τ]
iﬀ
(l, x) ≤ ρ(l[W
µ
[τ]), ∀l ∈ IR
n
,
or otherwise
(l, x −x
∗
(τ))H
−1
(t, l) ≤ µ, ∀l,
where
H(t, l) = (l, Ml)
1/2
+
_
t
1
τ
(l, P(t)l)
1/2
dt.
This immediately yields
Lemma 1.5.4 The value function
V (τ, x) = max¦(l, x −x
∗
(τ))H
−1
(t, l)[l ∈ IR
n
¦. (1.5.17)
39
( check that here, with M > 0, the maximum is attained).
Exercise 1.5.2. Try to write a formal HJB equation for cost criterion Ψ(τ, x, u()). Check
whether this equation does have a classical solution. In what sense could function V (τ, x)
be considered a solution to this equation? Would it be a viscosity solution, [82], [109] ?
Later, in Part IV, sections 4.1  4.3, we shall indicate an approach for approximating the
solution of the HJB equation of Exercise 1.5.2, rather than solving it explicitly.
Naturally, the description of attainability domains also allows an application of DP. Indeed,
since J[t] is similar to the attainability domain A[t], if the latter would be calculated in
backward time, it is possible to formulate an optimization problem , such that A[t] would be
the level set for the respective value function.(We oﬀer the reader to specify the formulation
of such a problem). Later, in Sections 4.14.4, we shall discuss this issue in conjunction
with ellipsoidal techniques.
Our next subject will be the issue of uncertainty in the knowledge of the system inputs.
1.6 Uncertain Systems.
Attainability Under Uncertainty
We are returning to systems (1.1.1), (1.1.2) and (1.1.4), but now the disturbance (or
“forcing term”) f(t) will be taken to be unknown but bounded, namely, the information on
f(t) will be restricted to the inclusion
f(t) ∈ Q(t), (1.6.1)
where Q(t) is a given multivalued map Q : T → convIR
n
, continuous in t.
We therefore come to the following systems:
(i) the linear diﬀerential equation
˙ x = u(t) + f(t), x(t
0
) ∈ A
0
, t ∈ T, (1.6.2)
that reﬂects the availability of only openloop controls u() ∈ U
0
P
and also has an
unknown disturbance f(t) subject to a given constraint (1.6.1),
40
(ii) the nonlinear diﬀerential inclusion
˙ x ∈ (t, x) + f(t), x
0
∈ A
0
, (1.6.3)
where (, ) ∈ U
c
P
and f(t) is unknown, but bounded by constraint (1.6.1). This
reﬂects the availability of closedloop (feedback) controls.
What would be the notion of attainability now, that the input f(t) is unknown? It is quite
obvious that the respective deﬁnitions for both openloop and closed loop controls could
be presented in several ways. We shall start with the following open loop construction that
will be used in the sequel.
Deﬁniton 1.6.1 An ”openloop” domain of attainability under counteraction for system
(1.6.2), (1.6.1) from set A
0
= A[t
0
] at time t
1
is deﬁned as the set A(t
1
, t
0
, A
0
) = A[τ]
of all states x
∗
, such that for any f
∗
() ∈ Q() there exists a pair ¦x
0∗
, u
∗
()¦, u
∗
() ∈ U
0
P
,
that generates an isolated trajectory x
∗
[t] of system
˙ x = u
∗
+ f
∗
, x
∗
[t
0
] = x
0∗
, t ∈ T,
that satisﬁes the boundary conditions
x
∗
[t
0
] ∈ A
0
, x
∗
[t
1
] = x
∗
.
Let us further add the symbol f(), to the notation of the attainability domain A[t] =
A(t, t
0
, A
0
) of Section 1.2 emphasizing its dependence on a given input f(), namely
A[t] = A(t, t
0
, A
0
, f).
In other terms A[t] = A(t, t
0
, A
0
, f) is the crosssection at instant t of the solution tube
to the linearconvex diﬀerential inclusion
˙ x ∈ T(t) + f(t), A(t
0
) = A
0
.
The set X[t
1
] of Deﬁnition 1.6.1 may then be presented as
A[t
1
] = A(t
1
, t
0
, A
0
) = (1.6.4)
41
=
f
_
x
0
_
A(t
1
, t
0
, ¦x
0
¦, f)
¸
¸
¸
¸
x
0
∈ A
0
, f() ∈ Q()
_
or
A[t
1
] =
_
A(t
1
, t
0
, A
0
, f)
¸
¸
¸
¸
f() ∈ Q()
_
Remark 1.6.1 Other types of attainability domains, than in Deﬁnition 1.6.1, may be
deﬁned by introducing operations of either intersection ∩ or union ∪ over x
0
, f(), in an
order other than (1.6.4). We invite the reader to investigate this issue.
Returning to (1.6.4) and taking A[t], for any t ∈ T, we come to the open loop solution tube
(under counteraction).
Let us see, whether it is possible to derive an evolution equation for the tube A[t]. Obviously
x
∗
∈ A[t] if and only if
(l, x
∗
) ≤ ρ(l[A(t, t
0
, A
0
, f)) ∀l ∈ IR
n
, ∀f() ∈ Q()
or
(l, x
∗
) ≤ h(t, l),
with
h(t, l) = inf
_
ρ(l[A(t, t
0
, A
0
, f) : f() ∈ Q()
_
.
Here direct calculation gives
h(t, l) = ρ(l[A
0
) +
t
_
t
0
ρ(l[T(τ))dτ + inf
_
t
_
t
0
(l, f(τ))dτ : f() ∈ Q()
_
h(t, l) = ρ(l[A
0
) +
t
_
t
0
(ρ(l[T(τ)) − ρ(−l[Q(τ)))dτ. (1.6.5)
The function h(t, l) is positively homogeneous in l, namely h(t, αl) = αh(t, l), for all
α > 0.
42
Assumption 1.6.1 The function
g(τ, l) = ρ(l[T(τ)) −ρ(−l[Q(τ))
is convex in l (and ﬁnite valued: g(τ, l) > −∞, ∀l, ∀τ ∈ T ).
Under Assumption 1.6.1 the function g(τ, l) is convex in l. It is also positively homogeneous
in l and therefore turns to be the support function for a certain set 1(τ) that evolves
continuously in time. A standard result in convex analysis consists in the following, [265],
Lemma 1.6.1 Under Assumption 1.6.1 the function
g(τ, l) = ρ(l[T(τ)) −ρ(−l[Q(τ))
is the support function for set
1(τ) = T(τ)
˙
−(−Q(τ)),
namely,
ρ(l[1(τ)) = ρ(l[T(τ)
˙
−(−Q(τ))) = ρ(l[T(τ)) − ρ(l[Q(τ)) = ρ(l[T(τ)) − ρ(l[Q(τ))
and 1(τ) ,= ∅. The setvalued map 1(τ) is continuous.
Here T
˙
−Q stands for the geometrical (“Minkowski”) diﬀerence between T and Q:
T
˙
−Q = ¦x : x +Q ⊆ T¦.
Following the proof of Theorem 1.4.1 we come to
Theorem 1.6.1 Under Assumption 1.6.1 the tube A[t] satisﬁes the following evolution
equation
lim
σ→0
σ
−1
h(A[t + σ], A[t] + σ1(t)) = 0, A[t
0
] = A
0
. (1.6.6)
43
A consequence of Theorem 1.6.1 is
Corollary 1.6.1 The solution A[t] = A(t, t
0
, A
0
) to equation (1.6.6) satisﬁes the semi
group property
A(t, τ, A(τ, t
0
, A
0
)) = A(t, t
0
, A
0
),
for t
0
≤ τ ≤ t ≤ t
1
.
Remark 1.6.2 A typical example for Assumption 1.6.1 is when 0 ∈ T, −Q = αT + c,
0 < α < 1, c ∈ IR
n
, so that
T
˙
−(−Q) = (1 −α)T −c.
This case is known as the matching condition for the constraints on u, f in equation (1.2.1).
It is not diﬃcult to formulate the necessary and suﬃcient condition for A[t] to be nonvoid.
This is given by
Lemma 1.6.2 The set A[t] is nonvoid (A[t] ,= ∅) if and only if there exists a vector
c ∈ IR
n
such that the function h(t, l) −(l, c) ≥ 0, ∀l ∈ IR
n
.
The proof of this assertion is a standard exercise in convex analysis.
Once A
0
= ¦x
0
¦ is a singleton, the function h(t, l) satisﬁes the condition of Lemma 1.6.2
if the following assumption is fulﬁlled.
Assumption 1.6.2 The geometrical diﬀerence of the following two integrals is not empty:
t
_
t
0
T(τ)dτ
˙
−
t
_
t
0
(−Q(τ))dτ ,= ∅.
44
Thus the set A[t] ,= ∅ if and only if (co
l
h)(t, l) ,= −∞, for all l, where (co
l
h) (t, l) is the
closed convex hull of h(t, l) in the variable l, [265], [100].
Then x
∗
∈ A[t] if and only if
(l, x
∗
) ≤ (co
l
h)(t, l), ∀l ∈ IR
n
or
(co
l
h)(t, l) = ρ(l[A[t]). (1.6.7)
Therefore
A[t] = (A[t
0
] +
t
_
t
0
T(τ)dτ)
˙
−
t
_
t
0
Q(τ)dτ (1.6.8)
It follows under Assumption 1.6.1, that the set A[t] ,= ∅ for any convex compact set A[t
0
]
and any t ∈ T, since
t
_
t
0
T(τ)dτ)
˙
−
t
_
t
0
Q(τ)dτ ⊇
t
_
t
0
(T(τ)
˙
− Q(τ)) dτ
(prove this inclusion).
Remark 1.6.3 The results of convex analysis imply a formal calculation for determining
the closed convex hull (co
l
h)(t, l). This is given by the relation
(co
l
h)(t, l) = h
∗∗
l
(t, l),
where h
∗∗
l
(t, l) is the second conjugate of h(t, l) in the variable l. Recall that
k
∗∗
(l) = (k
∗
)
∗
(l),
where
k
∗
(p) = sup¦(l, p) −k(l)[ ∈ IR
n
¦.
45
Remark 1.6.4 It is not diﬃcult to check that the tube A[t] of (1.6.8) may not satisfy the
semigroup property.
Exercise:Construct an example for this remark.
Let us now deﬁne an attainability domain under counteraction in the class of feedback
(closedloop) control strategies.
Given a strategy u = 
∗
(t, x) , 
∗
(t, x) ∈ U
c
P
, and f() ∈ Q(t), we shall deﬁne the
respective solution tube to system
˙ x ∈ 
∗
(t, x) + f(t), (1.6.9)
x(t
0
) = x
0
as A(t, t
0
, x
0
, f[
∗
), so that
A(t, t
0
, A
0
, f[
∗
) =
_
_
A(t, t
0
, x
0
, f[
∗
) : x
0
∈ A
0
_
The union of such tubes will be
A(t, t
0
, A
0
, f) =
_
_
A(t, t
0
, A
0
, f[) :  ∈ U
c
P
_
Deﬁniton 1.6.2 A closedloop domain of attainability under counteraction for system
(1.6.3), (1.6.1) from set A
0
= A
∗
[t
0
] at time t
1
is deﬁned as the set X[t
1
] = X(t
1
, t
0
, A
0
)
of all states x
∗
, such that for any f() ∈ Q() there exists a vector x
0∗
∈ A
0
, and a strat
egy 
∗
∈ U
c
P
, such that the pair ¦x
0∗
, 
∗
¦, generates a solution tube A(t
1
, t
0
, x
0∗
, f[
∗
) to
system (1.6.2) that satisﬁes the boundary condition
x
∗
∈ A(t
1
, t
0
, x
0∗
, f[
∗
).
In other words, X(t
1
, t
0
, A
0
) can be described as
X(t
1
, t
0
, A
0
[) = (1.6.10)
46
=
f
_
U
_
x
0
_
A
∗
(t
1
, t
0
, x
0
, f[) : x
0
∈ A
0
,  ∈ U
c
P
, f() ∈ Q()
_
=
=
_
A(t
1
, t
0
, A
0
, f[
∗
) : f() ∈ Q()
_
This also means that set X[t] = X(t, t
0
, A
0
[
∗
) consists, for any ﬁxed t ∈ T, of all those
states x
∗
such that for any f() ∈ Q() there exists a solution x[τ], τ ∈ [t
0
, t] to (1.6.3)
generated by some x[t
0
] ∈ A
0
,  ∈ U
c
P
, such that x[t] = x
∗
.
Other types of attainability domains under feedback and counteraction could be deﬁned
by introducing operations of intersection or union over x
0
, f,  in an order other than in
(1.6.10). This is left to the reader.
1.7 Uncertain Systems :
the Solvability Tubes
The notion further important for control synthesis is that of the solvability set. We shall
start with a respective deﬁnition for the case of openloop controls.
Deﬁniton 1.7.1 The ”openloop” solvability set under counteraction at time t, t < t
1
, for
terminal set / is the set W[t] = W(t, t
1
, /) of all states x
∗
∈ IR
n
, such that for every
function f(τ) ∈ Q(τ), t ≤ τ ≤ t
1
, there exists an openloop control u = u(τ), u() ∈ U
O
P
that steers system
˙ x = u(τ) + f(τ) (1.7.1)
from x
∗
= x(t) to set /, so that x(t
1
) ∈ /.
A direct calculation similar to that of Section 1.6 (see (1.6.5)) gives:
x
∗
∈ J[t]
if and only if
(l, x
∗
) ≤ k(t, l), ∀l ∈ IR
n
,
47
where
k(t, l) = ρ(l[/) +
t
1
_
t
(ρ(l[ −T(τ)) −ρ(l[Q(τ)))dτ.
In terms of setvalued maps this allows the relation
J[t] =
_
/+
t
1
_
t
(−T(τ))dτ
_
˙
−
t
1
_
t
Q(τ)dτ. (1.7.2)
The function J[t] = J(t, t
1
, /) generates a multivalued map with convex compact values
(provided / ∈ conv IR
n
). For the openloop case, considered here, it is clear that the
inverse problem of ﬁnding W[t] is precisely the one of constructing the attainability domain
of Deﬁnition 1.6.1, but when the latter is taken in backward time. This does not mean,
however, that J[t] would immediately satisfy a semigroup property and therefore, an
evolution “funnel equation”, since an additional assumption is required here.
Lemma 1.7.1 Under Assumption 1.6.1 the map J[t] satisﬁes the evolution equation
lim
σ→0
σ
−1
h(J[t −σ], J[t] −σ(T(t)
˙
−(−Q(t))) = 0 (1.7.3)
J[t
1
] = /.
This lemma follows directly from Theorem 1.6.1. Particularly, under the conditions of
Remark 1.6.2 (namely, 0 ∈ T(t), −Q(t) = αT(t) + c, α ∈ (0, 1)), we have
T(t)
˙
−(−Q(t)) = (1 −α)T(t) + c = 1(t). (1.7.4)
The evolution equation (1.7.3) in this case is precisely the equation (1.6.4), but its solution
is evolving in backward time, starting at t
1
and ”moving” towards given instant t < t
1
.
Our aim, however, is to devise a feedback control strategy for an uncertain system that
operates under unknown but bounded input disturbances. A precise deﬁnition of the problem
as well as its solution will be given in the next section. This solution will require some
preparatory work. Let us formally construct a set J
∗
(t, t
1
, /) which shall be a certain
superposition of the “openloop” sets J(t, t
1
, /) deﬁned above.
48
Taking the interval t ≤ τ ≤ t
1
, introduce a subdivision Σ = ¦σ
1
, . . . , σ
k
¦,
t = t
1
−
k
i=1
σ
i
, . . . , t
1
−σ
1
, t
1
,
where
σ
i
> 0,
k
i=1
σ
i
= t
1
−t.
As a ﬁrst step, starting at instant t
1
, ﬁnd the openloop solvability set J[t
1
− σ
1
] =
J(t
1
−σ
1
, t
1
, /). Due to (1.7.2) this gives
J[t
1
−σ
1
] =
_
/+
t
1
_
t
1
−σ
1
(−T(τ))dτ
_
˙
−
t
1
_
t
1
−σ
1
Q(τ)dτ. (1.7.5)
Following the procedure, we come to
J(t −σ
1
−σ
2
, t
1
−σ
1
, J[t
1
−σ
1
]) = (1.7.6)
=
_
J[t
1
−σ
1
] +
t
1
−σ
1
_
t
1
−σ
1
−σ
2
(−T(τ))dτ
_
˙
−
t
1
−σ
1
_
t
1
−σ
1
−σ
2
Q(τ)dτ
and may ﬁnally calculate the value
J
_
t, t + σ
k
, J(t + σ
k
, t + σ
k
+ σ
k−1
, . . . , J(t
1
−σ
1
, t
1
, /) . . .)
_
=
= ¸(t, t
1
, /, Σ). (1.7.7)
The formal procedure described here presumes all the sets J() of type (1.7.5)–(1.7.7)
involved in the construction to be nonvoid.
Assumption 1.7.1 There exists a continuous function β(t) > 0, t ∈ T, such that all the
sets
J(t
1
−
j
i=1
σ
i
, t
1
−
j−1
i=1
σ
i
, J(t
1
−
j−1
i=1
σ
1
, t
1
−
j−2
i=1
σ
i
, . . . ,
49
..., J(t
1
−σ
1
, t
1
, /)) . . .) −β(t
1
−
j
i=1
σ
i
)o
are nonvoid with j = 1, . . . , k, whatever is the subdivision
Σ = ¦σ
1
, . . . , σ
k
¦,
k
i=1
σ
i
= t
1
−t, σ
i
> 0.
Assumption 1.7.1 clearly ensures ¸(t, t
1
, /, Σ) ,= ∅ for any subdivision Σ.
Following (1.7.2), (1.7.5)–(1.7.7), we come to the analytical expression
¸(t, t
1
, /, Σ) =
= (. . . ((/+
t
1
_
t
1
−σ
1
(−T(τ)dτ)
˙
−
t
1
_
t
1
−σ
1
Q(τ)dτ) . . .
˙
−
t+σ
k
_
t
Q(τ)dτ)).
The set ¸(t, t
1
, /, Σ) is convex and compact for any subdivision Σ. We may consider the
limit of these sets with max ¦σ
i
: i = 1, . . . , k¦ → 0.
Lemma 1.7.2 Under Assumption 1.7.1 there exists a Hausdorﬀ limit
limh
_
¸(t, t
1
, /, Σ) , ¸(t, t
1
, /)
_
= 0
with
max¦σ
i
: i = 1, . . . , k¦ → 0, k → ∞,
k
i=1
σ
i
= t
1
−t.
We shall refer to
¸(t, t
1
, /) = J
∗
(t, t
1
, /) = J
∗
[t] (1.7.8)
as the “alternated” solvability domain and denote
50
¸(t, t
1
, /) =
t
_
t
1
,M
((−T(τ))dτ
˙
−Q(τ)dτ).
The set ¸(t, t
1
, /) is actually the value of a certain type of setvalued integral that is
known as the “Alternated Integral of L.S. Pontryagin.” The integral was introduced and
described in detail in papers [256], [257].
Deﬁniton 1.7.2 With t varying, the setvalued function J
∗
[t] of (1.7.8) will be referred
to as the “alternated solvability tube”.
Lemma 1.7.3 Once J
∗
[t] ,= ∅, t ∈ T, the setvalued function J
∗
[t] satisﬁes for all t ∈ T
the evolution equation
lim
σ→0
σ
−1
h
+
(J[t −σ] + σQ(t), J[t] −σT(t)) = 0, (1.7.9)
J[t
1
] = /. (1.7.10)
Obviously J
∗
[t
1
] = /. Taking J
∗
at an arbitrary instant of time t and also J
∗
[t − σ],
and following the deﬁnitions of these sets we observe that there exists a function γ(σ), such
that
J
∗
[t −σ] ⊆ J(t −σ, t, J
∗
[t]) + γ(σ)o, (1.7.11)
where
γ(σ) > 0, σ > 0; σ
−1
γ(σ) → 0 with σ → 0, (see [257]) .
Due to
J(t −σ, t, J
∗
[t]) = (1.7.12)
= (J
∗
[t] +
t
_
t−σ
(−T(τ))dτ)
˙
−
t
_
t−σ
Q(τ)dτ,
and to the deﬁnition of geometric diﬀerence, relation (1.7.11) yields the following:
51
_
J
∗
[t −σ] +
t
_
t−σ
Q(τ)dτ
_
⊆ (1.7.13)
⊆ J
∗
[t] +
t
_
t−σ
(−T(τ))dτ + γ(σ)o.
The continuity of T(τ), and Q(τ) implies
lim
σ→0
σ
−1
h
_
t
_
t−σ
T(τ)dτ, σT(τ)
_
= 0,
lim
σ→0
σ
−1
h
_
t
_
t−σ
Q(τ)dτ, σQ(t)
_
= 0.
The latter relations, together with (1.7.1) give the inclusion
J
∗
[t −σ] + σQ(t) ⊆ J
∗
[t] −σT(t) + α(σ)o, (1.7.14)
where σ
−1
α(σ) → 0 with σ → 0.
Relation (1.7.14) is equivalent to the existence of a solution to (1.7.9) at any given instant
t, particularly, at t = t
1
. The prolongability of the solution J
∗
[t] towards time t
0
follows
from the condition J
∗
[t] ,= ∅, t ∈ T and from the boundedness of the tube J
∗
[t]. This
justiﬁes the assertion of Lemma 1.7.3. Studying equation (1.7.9) it is possible to observe
that its solution in nonunique (devise an example) and moreover, that J
∗
[t] satisﬁes the
following properties.
Lemma 1.7.4 The setvalued function J
∗
[t] is a maximal solution to equation (1.7.9).
The proof of this assertion is left to the reader. It also follows from Lemma 1.8.3 of the
next Section.
Lemma 1.7.5 The setvalued map J
∗
(t, t
1
, /) satisﬁes the semigroup property (in back
ward time). Namely
J
∗
(t, t
1
, /) = J
∗
(t, τ, J
∗
(τ, t
1
, /)), (1.7.15)
with t ≤ τ ≤ t
1
.
52
The proof of the relation (1.7.14) follows from the additivity properties of the alternated
integral ¸(t, t
1
, /),[257].
Remark 1.7.1 The assertions of this Section concerning the Alternated Solvability
Tube W
∗
[t] have been all derived under Assumption 1.7.1. Hence, all the propositions
that follow in the sequel and involve the tube W
∗
[t] are true only under this assumption.
For future operation it may be sometimes more convenient to use an assumption of equiv
alent type
Assumption 1.7.2 The alternated solvability tube J
∗
[t] is nondegenerate. Namely, there
exists an absolutely continuous function x(t) and a function β(t) > 0, t
0
≤ t ≤ t
1
, such
that
x(t) + β(t)o ⊂ J ∗ [t], t
0
≤ t ≤ t
1
,
where o is a unit ball in IR
n
.
6
As we shall see in the next section, the tube J
∗
[t] will coincide with the solvability tube
for the problem of control synthesis under uncertainty.
1.8 Control Synthesis Under Uncertainty
Consider system (1.7.1) and terminal set /.
Deﬁniton 1.8.1 The problem of control synthesis under uncertainty consists in
specifying a solvability set J
∗
(τ, t
1
, /) and a setvalued feedback control strategy
u = (t, x), (, ) ∈ U
c
P
such that all the solutions to the diﬀerential inclusion
˙ x ∈ (t, x) +Q(t) (1.8.1)
that start from any given position ¦τ, x
τ
¦, x
τ
= x[τ] ∈ J
∗
(τ, t
1
, /), τ ∈ [t
0
, t
1
), would
reach the terminal set / at time t
1
: x(t
1
) ∈ /.
6
Once Assumptions 1.7.1 or 1.7.2 are not fulﬁlled, there is a degenerate situation which has to be
approached separately, by means of a regularization procedure that would allow to keep up with the basic
solution scheme. Such situations are not discussed in this book and are left for additional treatment.
53
Deﬁniton 1.8.2 The solvability set J
∗
[t] = J
∗
(t, t
1
, /) for the problem of control syn
thesis under uncertainty is the set of all states x
τ
∈ IR
n
, such that there exists a control
strategy (t, x) that solves the problem of control synthesis of Deﬁnition 1.7.1, provided
x
τ
∈ J
∗
[t].
Deﬁnition 1.8.1 is nonredundant if J
∗
(τ, t
1
, /) ,= ∅, where, as we have seen, J
∗
(τ, t
1
, /)
is the solvability set, which is the ”largest” set of states from which the solution to the
problem does exist at all.
Taking J
∗
(t, t
1
, /) = J
∗
[t], we come to a setvalued map (the “solvability tube”). We
shall prove that the “alternated solvability tube” J
∗
[t] of Section 1.6 does coincide with
J
∗
[t].
Let us ﬁrst try to ﬁnd a set and a tube that would provide an appropriate solution for the
problem of control synthesis, but would not necessarily be the largest solvability set and
solvability tube as required by Deﬁnitions 1.8.1, 1.8.2. Assume Z(t) to be a solution to
the evolution equation (1.7.9) with boundary condition
Z(t
1
) ⊆ /
and therefore an absolutely continuous setvalued map with convex compact values.
For every solution Z(t) let us assign a feedback strategy 
Z
(t, x) constructed similar to
the one in Section 1.4 (see (1.4.11), (1.4.14) and (1.4.15)). Thus

Z
(t, x) = ∂
l
f(t, −l
0
Z
(t, x)), (1.8.2)
where f(t, l) = ρ(l[T(t)) and l
0
= l
0
Z
(t, x) is the maximizer of the expression for calculating
d
Z
[τ, x] = h
+
(x, Z[τ]), which is
d
Z
[t, x] = max¦(l, x) −ρ(l[Z[t]) : l ≤ 1¦ =
= (l
0
, x) −ρ(l
0
[Z[t]),
(l
0
= ¦0¦ for d
Z
[t, x] = 0).
Relation (1.8.3) is formally similar to the deﬁnition of the “extremal strategy” (1.4.19).
54
Consider the derivative
d
dt
d
2
Z
[t, x] = d[t, x]
d
dt
d
Z
[t, x]
due to system (1.7.1).At a point ¦t, x¦ that gives d
Z
[t, x] > 0, a direct calculation yields
d
dt
d
Z
[t, x] = (l
0
, ˙ x) −
∂ρ(l
0
[Z(t))
∂t
. (1.8.3)
Lemma 1.8.1 The following inequality is true
∂ρ(l[Z(t))
∂t
≥ ρ(l[Q(t)) −ρ(−l[ −T(t)), l ∈ IR
n
. (1.8.4)
The evolution equation (1.7.9) leads to the inclusion
Z(t −σ) + σQ(t) ⊆ Z(t) −σT(t) + o(σ)o
with
σ
−1
o(σ) → 0, σ → 0,
and further on,
ρ(l[Z(t −σ)) + σρ(l[Q(t)) ≤ ρ(l[Z(t)) + σρ(−l[T(t)) + o(σ)(l, l)
1/2
or otherwise, the inequality
σ
−1
(ρ(l[Z(t) −ρ(l[Z(t −σ)) ≥ ρ(l[Q(t)) −ρ(−l[T(t)) + σ
−1
o(σ)(l, l)
1/2
which gives, after a limit transition σ → 0, the result (1.8.4) of the Lemma.
A consequence of Lemma 1.8.1 (see (1.8.2), (1.8.3)) is that with d
Z
[t, x] > 0 for the
derivative the following inequality holds:
55
d
dt
d
Z
[t, x] ≤ (l
0
, u(t) + f(t)) −ρ(l
0
[Q(t)) + ρ(−l
0
[T(t)), (1.8.5)
where
u(t) ∈ T(t), f(t) ∈ Q(t)
Therefore, with u = u
0
, where
−(l
0
, u
0
) = ρ(−l
0
[T(t)),
we will have
d
dt
d
Z
[t, x] ≤ 0, ∀f(t) ∈ Q(t). (1.8.6)
This leads us to
Lemma 1.8.2 The derivative
d
dt
d
2
Z
[t, x] calculated due to the system
˙ x ∈ 
Z
(t, x) + f(t) (1.8.7)
satisﬁes the inequality
d
dt
d
2
Z
[t, x] ≤ 0, ∀f(t) ∈ Q(t).
Some further reasoning yields the next assertion
Lemma 1.8.3 With x
τ
∈ Z[τ], τ < t
1
, the solution tube A
Z
[t] = A
Z
(t, τ, x
τ
[f()) of
system (1.8.6), x[τ] = x
τ
, τ ≤ t ≤ t
1
, satisﬁes the inclusion
A
Z
[t] ⊆ Z[t], f(t) ∈ Q(t), τ ≤ t ≤ t
1
, (1.8.8)
and therefore, the boundary condition Z(t
1
) ⊆ /.
56
The proof of Lemma 1.8.3 is similar to that of Lemma 1.3.3 as the main relation used in
the proof is the inequality (1.8.4).
Lemma 1.8.3 thus indicates that 
Z
(t, x) is a synthesizing strategy that solves the problem
of control synthesis of Deﬁnition 1.8.1 with Z[τ] being the solvability domain (but not
necessarily the largest one). It is therefore possible, in principle, to solve the problem of
control synthesis through any solution Z[t] of equation (1.7.9) with boundary condition
Z[t
1
] ⊆ /.
The set of states for which the problem of Deﬁnition 1.8.1 is solvable will then be restricted
to Z[t]. Our problem, however, is to ﬁnd the maximal solvability domain J
∗
[t] for the
problem of Deﬁnition 1.8.1 and the respective strategy (t, x).
Referring to Lemmas 1.7.3, 1.7.4, we observe that tube J
∗
[t] is the maximal solution to
equation (1.7.9) with an equality in the boundary condition (J
∗
[t
1
] = /). The tube
J
∗
[t] generates a strategy

0
(t, x) = ∂
l
f(t, −l
0
W
∗(t, x)), (1.8.9)
where l
0
= l
0
W
∗(t, x) is the maximizer for the problem
d
W
∗[τ, x] = max¦(l, x) −ρ(l[J
∗
[τ][ l ≤ 1¦
d
W
∗[τ, x] = (l
0
, x) −ρ(l
0
[J
∗
[τ]). (1.8.10)
The results of Lemmas 1.8.2, 1.8.3 imply
Lemma 1.8.4 Strategy 
0
(t, x) ensures the inclusion
X
W
∗(t, τ, x
τ
) ⊆ J
∗
[t], τ ≤ t ≤ t
1
, (1.8.11)
provided x
τ
∈ J
∗
[τ].
Here X
W
∗(t, τ, x
τ
) is the solution tube for system (1.8.1), x[τ] = x
τ
, with (t, x) = 
0
(t, x).
The results of the above may be summarized into
Theorem 1.8.1 The synthesizing strategy 
0
(t, x) of (1.8.9) resolves the problem of con
trol synthesis under uncertainty of Deﬁnition 1.8.1.
57
Remark 1.8.1 It is necessary to emphasize that the last theorem is true in the absence of
matching conditions of the Assumption 1.6.1 type. The result presumes however that the
solution J
∗
[t] to the evolution equation (1.7.9),(1.7.10) exists and that intJ∗ ,≡ ∅. The
latter is ensured by Assumption 1.7.1.
Finally, to make the ends meet, we have to answer the following question: is the maximal
solution J
∗
[t] to equation (1.7.9), (1.7.10) also the maximal solution tube for the problem
of control synthesis of Deﬁnition 1.8.1? As we shall see, the answer to the question is
aﬃrmative. This may be proved due to the inequality (1.8.4). Namely, once x
∗
,∈ J
∗
[τ],
it is possible to select in the domain d[x
∗
, J
∗
[τ]] > 0, a strategy
1
0
(t, x) = ¦v [ (l
0
, v) = ρ(l
0
[Q(t))¦. (1.8.12)
This strategy will aﬀect the sign of the derivative dd
W
∗(t, x
∗
)/dt due to system (1.7.1).
Let us calculate this derivative solving the extremal problem
d
W
∗(t, x
∗
) = max¦(l, x
∗
) −ρ(l[J
∗
[t]) : l ≤ 1¦
d
W
∗(t, x
∗
) = (l
0
∗
, x
∗
) −ρ(l
0
∗
[J
∗
[t]),
and
l
0
∗
= 0 for d
W
∗(t, x
∗
) = 0.
This gives
d
dt
d
W
∗(t, x
∗
) = (l
0
∗
, ˙ x
∗
) −∂ρ(l
0
∗
[J
∗
[t])/∂t.
The calculation of the derivative ∂ρ(l
0
∗
[J
∗
[t])/∂t can be done using the representations
(1.7.12)(1.7.14). Thus, in view of a relation of type (1.7.13) this further gives
ρ(l[J
∗
[t + σ]) −ρ(l[J
∗
[t]) ≥ − ρ(l[
t+σ
_
t
(−T(τ)dτ −Q(τ)dτ)),
58
∂ρ(l[J
∗
[t])
∂t
≥ ρ(l[Q(τ)) −ρ(l[ −T(τ)), (1.8.13)
or, under Assumption 1.6.1,
∂ρ(l[J
∗
[t])
∂t
= ρ(l[Q(τ)) −ρ(l[ −T(τ)) = (1.8.14)
= ρ(l[ −T(τ)
˙
−Q(τ))
We shall further ﬁrst continue under this assumption so that, with Z(t) = J
∗
(t), the
inequality (1.8.5) turns to an equality. Diﬀerentiating d
W
∗(t, x
∗
) for x
∗
,∈ J
∗
[t], using
(1.8.13) and also the respective rule indicated in Remark 1.4.1, we come to the following
relation.
Lemma 1.8.5 Under Assumption 1.6.1 the derivative d
W
∗(t, x
∗
)/dt due to system (1.7.1)
is given by the relation
d
dt
d
W
∗(t, x
∗
) = (l
0
∗
, u + f) + ρ(−l
0
∗
[T(t)) −ρ(l
0
∗
[Q(t)) (1.8.15)
Selecting f ∈ 1
0
(t, x), (1.8.12) and observing that
u ∈ T(t) implies −(l
0
∗
, u) ≤ ρ(l
0
∗
[T(t)), ∀u ∈ T(t),
we arrive, due to (1.8.15), to the following relations
d
dt
d
W
∗(t, x
∗
)
¸
¸
¸
¸
f∈V
0
≥ 0, ∀u ∈ T(t),
d
dt
d
W
∗(t, x
∗
)
¸
¸
¸
¸
f∈V
0
= 0, ∀u ∈ 
0
(t, x).
This implies that any solution x[t] to the diﬀerential inclusion
dx
dt
∈ (t, x) + f; f ∈ 1
0
(t, x), x[τ] = x
∗
,
that starts at a point x
∗
,∈ J
∗
[t] or in other words, with d(x
∗
, J
∗
[τ]) = r
τ
> 0, does satisfy
the inequality d(x[t], J
∗
[t]) ≥ r
τ
, t ∈ [τ, t
1
] whatever is the strategy (, ) ∈ U
c
P
. Under
Assumption 1.6.1 we have therefore proved
59
Theorem 1.8.2 (i) The alternated solvability tube J
∗
[t] coincides with the solvability
tube J
∗
[t] of the problem of control synthesis under uncertainty; namely
J
∗
[t] = J
∗
[t], t
0
≤ t ≤ t
1
.
(ii) The set J
∗
[τ], τ ∈ [t
0
, t
1
) is the largest solvability domain for this problem.
It should be emphasized that this theorem remains true without the Assumption 1.6.1.
To prove (i),(ii) in the general case, one has to substitute strategy 1
0
(t, x) of (1.8.12) by
another one 1
∗
that would in some sense ensure a relation similar to the following
d
dt
d
W
∗[t, x] = (l
0
(t), u + f(t)) + ρ(l
0
[(−T(t))
˙
−Q(t)) ≥ 0
f ∈ 1
ast
(t, x), ∀u ∈ T(t).
Since in general we have
ρ(l[(−T)
˙
−Q) = co(f
1
−f
2
)(l), (1.8.16)
f
1
(l) = ρ(l[ −T), f
2
(l) = ρ(l[Q),
the desired strategy 1
∗
may not exist in the explicit form of (1.8.12). It exists, however, in
the class of mixed strategies (also known as relaxed controls), where 1
∗
has to be speciﬁed as
a probabilistic measure concentrated on Q. Loosely speaking, the value v may be required
to run around a variety v
(i)
of some extremal points of the set Q, throughout any minor
interval of time. We shall not specify the rigorous deﬁnition and precise construction of
such strategies 1
∗
as this would require to discuss notions that are quite beyond the scope
of this book, addressing the reader to monographs [169], [170], [171], on diﬀerential games,
where these topics are discussed in detail.
Let us now pass to the DP interpretation for this Section. Consider equation (1.7.1) and
target set /.Introduce the value function
V
∗
(t, x) = min
u
max
f
¦1(t, x)[(, ) ∈ U
c
P
, f() ∈ Q()¦
where 1(t, x) is the same as in (1.5.1). our aim is to minimaxize the cost 1(t, x) over all
the strategies (, ) ∈ U
c
P
and disturbances f() ∈ Q().Here the formal HJB equation
for the value V
∗
(t, x) looks as follows
∂V
∂t
+ min
u
max
f
¦
_
∂V
∂x
, u + f
_
¦ = 0 (1.8.17)
60
with boundary condition
V (t
1
, x) = h
2
+
(x, /) (1.8.18)
(Having had a minmax operation involved, the latter HJB equation is often referred to
as the the HJBI equation with letter I being a reference to R.Isaacs and his contribution
to diﬀerential games).
Presuming J
∗
[t] ,= ∅ , consider the function V (t, x) = d
2
W
∗[t, x].Obviously
V (t, x) = h
2
+
(x, /) (1.8.19)
Then in view of Lemma 1.8.2 one may observe that V (t, x) satisﬁes the inequality
∂V (t, x)
∂t
+ max
f
_
∂V (t, x)
∂x
, u + f
_
≤ 0 (1.8.20)
and boundary condition (1.8.18), provided
u ∈ 
W
∗(t, x) = ∂
l
f(t, −l
0
W
∗(t, x))
where f(t, l) = ρ(l[T(t)) and l
0
W
∗(t, x) = ∂d
2
W
∗[t, x]/∂x with d
W
∗[t, x] > 0 ( l
0
W
∗(t, x) = 0
with d
W
∗[t, x] = 0).
Denoting 
W
∗(t, x) = 
∗
(t, x), we may rewrite

∗
(t, x) = arg min¦
_
∂V (t, x)
∂x
, u
_¸
¸
¸
¸
u ∈ T(t)¦ = (1.8.21)
= ∂
l
f
_
t,
_
−
∂V (t, x)
∂x
__
.
Relations (1.8.20),(1.8.17) and (1.8.21),(1.8.9) then imply
Lemma 1.8.6 Suppose J
∗
[t] ,= ∅. Then the value function V
∗
(t, x) ≤ V (t, x) and the
strategy 
∗
(t, x), (1.8.21), solves the problem of Control Synthesis of Deﬁnition 1.8.1.
Indeed, inequality (1.8.20) ensures that once V (t, x) ≤ 0, then V (τ, x[τ]) ≤ 0 for any
trajectory x[τ] = x(τ, t, x), τ ∈ [t, t
1
], of the diﬀerential inclusion
dx(τ)
dτ
= 
∗
(τ, x) + f(τ), x[t] = x, (1.8.22)
whatever is the disturbance f(τ) that satisﬁes (1.6.1).The continuity of J
∗
[τ] in τ implies
the upper semicontinuity of 
∗
(τ, x) in its variables and therefore, the existence of solutions
to the diﬀerential inclusion (1.8.22).
61
Under Assumption 1.6.1 relation (1.8.20) turns into an equality and 1
∗
(t, x) = V (t, x). In
order to achieve this equality without such an assumption one has to allow the disturbance
f to be selected as indicated in the comments after Theorem 1.8.2, namely, in the class
of functions generated by a mixed strategy which may result in sliding modes or socalled
chattering functions f. Then V (t, x) will be the value of a respective diﬀerential game (
see [171], [290] ).
Our next issue is to deal with state constraints.
1.9 State Constraints and Viability
Let us return to system
˙ x(t) ∈ T(t) + f(t), t
0
≤ t ≤ t
1
, (1.9.1)
x
0
∈ A
0
, (1.9.2)
with a ﬁxed disturbance f(t), taking it here to be continuous. We shall now introduce an
additional state constraint
Gx(t) ∈ ¸(t), t
0
≤ t ≤ t
1
, (1.9.3)
where G is a given matrix of dimensions mn (m ≤ n) and ¸(t) is a multivalued function
continuous in t with convex compact values (¸(t) ∈ comp R
n
, ∀t).
We shall start from
Deﬁniton 1.9.1 A trajectory x[t] = x(t, t
0
, x
0
) of system (1.8.1), (1.8.2) is said to be
viable relative to constraint (1.9.3) if it satisﬁes the state constraint (1.9.3).
Our interest is in describing the tube of such trajectories.
7
A detailed theory of viable
trajectory tubes for diﬀerential inclusions may be found in [17],[193].
7
This section gives a very concise description of the subject, being only an introduction to other parts
of the book.
62
Deﬁniton 1.9.2 A viability tube
A[t] = A(t, t
0
, A
0
)
is the union
A[t] =
_
¦x(t, t
0
, x
0
) : x
0
∈ A
0
¦
over A
0
of all viable trajectories of system (1.9.1), (1.9.2) relative to constraint (1.9.3).
It is obvious that A[t
0
] ∈ ¸(t
0
).
Let us ﬁrst calculate the support function ρ([A[t]) of the crossection A[t] of the tube A[]
at time t. This is also the attainability domain of system (1.9.1), (1.9.2) under a state
constraint (1.9.3). The set A[t] is generated through relations (1.9.1)(1.9.3). For a certain
instant t = ϑ these relations yield
x(ϑ) = x
0
+
_
ϑ
t
0
u(τ)dτ +
_
ϑ
t
0
f(τ)dτ, (1.9.4)
Gx(t) = Gx
0
+
_
t
t
0
Gu(τ)dτ +
_
t
t
0
Gf(τ)dτ, (1.9.5)
t
0
≤ t ≤ ϑ,
with restrictions (1.9.3) and
x
0
∈ A
0
, u(t) ∈ T(t), t
0
≤ t ≤ ϑ. (1.9.6)
It is not diﬃcult to observe that (1.9.4) is equivalent to the equality
x(θ) =
x
0
+
_
ϑ
t
0
u(τ)dτ +
_
ϑ
t
0
f(τ)dτ (1.9.7)
that should be true for any vector ∈ IR
n
, while (1.9.5) is equivalent to the equality
_
ϑ
t
0
λ
(t)Gx(t)dt =
_
ϑ
t
0
λ
(t)Gx
0
dt + (1.9.8)
_
ϑ
t
0
_
_
ϑ
τ
λ
(t)Gdt
_
u(τ)dτ +
_
ϑ
t
0
_
_
ϑ
τ
λ
(t)Gdt
_
f(τ)dτ
63
that should be true for any continuous vector function λ(t) ∈ C
m
[t
0
, ϑ].
On the other hand, the inclusions (1.9.6) are equivalent to the following inequalities
x
0
≤ ρ([A
0
), ∀ ∈ IR
n
, (1.9.9)
u(t) ≤ ρ([T(t)), ∀ ∈ IR
n
, t
0
≤ t ≤ ϑ, (1.9.10)
while (1.9.3) is equivalent to
0 ≤ −
_
ϑ
t
0
λ
(t)Gx(t)dt +
_
ϑ
t
0
ρ(λ(t)[¸(t))dt, (1.9.11)
∀λ() ∈ C
n
[t
0
, ϑ].
The set A[ϑ] will now consist of all those vectors x[θ] that satisfy (1.9.7), (1.9.8) under
restrictions (1.9.9)–(1.9.11). In other terms, collecting relations (1.9.7)–(1.9.11), we observe
that
x(θ) ∈ A[ϑ] (1.9.12)
if and only if there exists a vector x
0
and a function u(t) that respectively satisfy (1.9.9)
and (1.9.10) and also the inequality
x(θ) ≤ (
−
_
ϑ
t
0
λ
(t)Gdt)x
0
+
_
ϑ
t
0
_
−
_
ϑ
τ
λ
(t)Gdt
_
u(τ)dτ +
_
ϑ
t
0
(
−
_
ϑ
τ
λ
(t)G dt)f(τ)dτ +
_
ϑ
t
0
ρ(λ(τ)[¸(τ))dτ,
whatever are the elements ∈ IR
n
, λ() ∈ C
n
[t
0
, ϑ].
Following the theory of convex analysis, [100], it was proved in [181], that the latter
requirement will be fulﬁlled if and only if
x(θ) ≤ Φ
ϑ
(, λ())
for any ∈ IR
n
, λ() ∈ C
m
[t
0
, ϑ], where
Φ
ϑ
(, λ()) = ρ(
−
_
ϑ
t
0
λ
(t)G dt[A
0
) +
_
ϑ
t
0
ρ(
−
_
ϑ
τ
λ
(t)G dt[T(τ) + f(τ))dτ +
_
ϑ
t
0
ρ(λ(τ)[¸(τ))dτ.
64
This in its turn will be true if and only if
x(θ) ≤ inf ¦Φ
ϑ
(, λ()) [ λ() ∈ C
m
[t
0
, ϑ]¦ = Φ
ϑ
[]. (1.9.13)
Function Φ
ϑ
[] happens to be convex and positively homogenous (this may be veriﬁed as
an exercise) and therefore, due to Lemma 1.3.1 (b), is a support function of some set A
ϑ
.
Since (1.9.13) is necessary and suﬃcient for (1.9.12) we come to the equality A
ϑ
= A[ϑ],
having proved
Theorem 1.9.1 The support function for A[ϑ] is given by
ρ([A[ϑ]) = Φ
ϑ
[]. (1.9.14)
A more detailed version of these calculations could be also found in paper [181]. Let us
now introduce an evolution equation which will prove to be an appropriate description for
A[t]. This will be
lim
σ→0
σ
−1
h
+
(Z[t + σ], Z[t] ∩ ¸(t) + σT(t) + σf(t)) = 0 (1.9.15)
Z[t
0
] ⊆ A
0
.
A solution to (1.9.15) is a multifunction Z[t] that satisﬁes (1.9.15) almost everywhere and
is also h
+
– absolutely continuous in the sense of Deﬁnition 1.3.2.
Let A[t] = ¦x[t]¦ be the union of all trajectories of (1.9.1), (1.9.2) viable relative to
constraint (1.9.3). Then, obviously
A[t] ⊂ ¸[t], t
0
≤ t ≤ t
1
, (1.9.16)
and
A[t
0
] = A
0
∩ ¸[t
0
].
At the same time A[t] = ¦x[t]¦ is a collection of all solutions to (1.9.1), (1.9.2) and
therefore, due to Lemma 1.3.5, is a solution to the evolution equation (1.3.13). For any
t ∈ [t
0
, t
1
] , σ > 0, σ < t
1
−t
0
this yields the inclusion
A[t + σ] ⊆ A[t] +
_
t+σ
t
T(τ)dτ +
_
t+σ
t
f(τ)dτ (1.9.17)
⊆ A[t] + σT(t) + σf(t) + o(σ)S
that follows from the deﬁnition of the Hausdorﬀ semidistance h
+
as well as from relations
(1.3.6), (1.3.7). The inclusion (1.9.17) may be rewritten due to (1.9.16) as
A[t +σ] ⊆ A[t] ∩ ¸[t] + σT(t) + σf(t) + o(σ)S. (1.9.18)
65
The latter relation indicates that (1.9.15) is true for t ∈ [t
0
, t
1
]. From the relations of the
above it also follows that A[t] is absolutely h
+
continuous in the sense of Deﬁnition 1.3.2.
We thus come to the proposition
Theorem 1.9.2 The setvalued function A[t] is a solution to the evolution equation
(1.9.15).
It is not diﬃcult to observe that an isolated trajectory x[t] = x(t, t
0
, x
0
) that is viable
relative to constraint (1.9.3) is also a solution to (1.9.15). Given A[t] and any other
solution Z[t] to (1.9.15), the following assertion is true.
Lemma 1.9.1 The setvalued function A[t] is a maximal solution to (1.9.15), namely
Z[t] ⊆ A[t]
for any other solution Z[t] to (1.9.15).
We leave to the reader to verify both Lemma 1.9.1 as well as the following assertion
Lemma 1.9.2 The mapping A[t] = A(t, t
0
, A
0
) satisﬁes the semigroup property
A(t, t
0
, A
0
) = A(t, τ, A(τ, t
0
, A
0
)). (1.9.19)
There is another , stronger form of an evolution equation which should be mentioned in
this context. We will precede this with a deﬁnition
Deﬁniton 1.9.3 A convex compact multivalued function ¸(t) ∈ compIR
n
is said to be
absolutely continuous on an interval T = [t
0
, t
1
] if its support function
ρ([¸(t)) = f(t, )
is absolutely continuous on the interval T, for any ∈ S.
This deﬁnition ensures that f(t, ) is absolutely continuous on T uniformly in ∈ S.
We may now formulate
66
Theorem 1.9.3 Assume the multifunction ¸(t) to be absolutely continuous on the interval
T. Also assume that there exists a trajectory x(t) of system (1.8.1),(1.8.2) such that
x(t) ∈ int¸(t). Then the multifunction A[t] is the unique solution to the evolution equation
(1.9.20)
lim
σ→0
σ
−1
h(A[t + σ], (A[t] + σT(t) + σf(t)) ∩ ¸(t + σ)) = 0
A[t
0
] = A
0
.
Equation (1.9.19) is somewhat diﬀerent from (1.9.15), particularly in having involved the
Hausdorﬀ distance h rather than the semidistance h
+
. The proof of Theorem 1.9.3 is given
in papers [190], [193].
Let us now formally write the equation (1.9.19) with A(t) ,≡ 0. This gives
(1.9.21)
lim
σ→0
σ
−1
h(A[t + σ], ((I + σA(t))A[t] + σT(t) + σf(t)) ∩ ¸(t + σ)) = 0.
Equation (1.9.19) can also be treated in backward time, namely in the following form
(1.9.22)
lim
σ→0
σ
−1
h(J[t −σ], (J[t] −σT(t) −σf(t)) ∩ ¸(t −σ)) = 0,
J[t
1
] = /.
Theorem (1.9.3) obviously yields
Lemma 1.9.3 With ¸(t) absolutely continuous, equation (1.9.21) has a unique solution
deﬁned on the interval T.
Set J[τ] allows the following interpretation.
Deﬁniton 1.9.4 A solvability set J
0
[τ] under state constraints (1.9.3), is the set of all
states ¦x
τ
¦ = J
0
[τ] such that there exists a measurable function u(t) that generates a
trajectory x(t, τ, x
τ
) = x[t], τ ≤ t ≤ t
1
of system (1.9.1) that satisﬁes the inclusion x[t
1
] ∈
/ together with restriction (1.9.3).
67
Here J
0
[τ] is obviously the same as the attainability set, but is taken in backward time
and is clearly a solution to equation (1.9.21), so that J
0
[τ] ≡ J[τ]. With / = IR
n
, the
set J
0
[τ] is also known as the viability kernel (relative to constraint (1.9.3)), [17].
An alternative version of (1.9.21) is given by the equation
lim
σ→0
σ
−1
h
+
(Z[t −σ], Z[t] ∩ ¸[t] −σT(t) −σf(t)) = 0 (1.9.23)
Z[t
1
] ⊆ /.
A solution to (1.9.23) exists under weaker assumptions than those for (1.9.22) (¸[t] may
be assumed to be merely continuous and even upper semicontinuous). Its solution is
nonunique, however. Thus any viable , single  valued trajectory of (1.8.1),(1.8.2), if taken
in backward time, satisﬁes (1.9.23) and the proof of the respective existence theorem is
similar to that of Theorem 1.9.2.
We ﬁnally emphasize the following property that could be proved through standard pro
cedures.
Lemma 1.9.4 The multifunction J[t] is the maximal solution to equation (1.9.22).
Function J
0
[t], t
0
≤ t ≤ t
1
generates a solvability tube that is a crucial element for solving
the problem of control synthesis under state constraints.
Remark 1.9.1 With / = IR
n
, A
0
= IR
n
, the viability set J[t] is the collection of all
positions ¦t, x¦ , from each of which there exists a control u() ∈ 
0
P
that keeps the respective
trajectory x(t) within the state constraint (1.9.3.).A set J[t] with such a property is referred
to as weakly invariant relative to constraint (1.8.3),(see [17]).
1.10 Control Synthesis
Under State Constraints
Given the solvability tube J[t] of the previous paragraph we may construct a multival
ued synthesizing strategy (t, x) that solves the problem of control synthesis under state
constraints.
68
Deﬁniton 1.10.1 Given a terminal set / ∈ compIR
n
, the problem of control synthesis
under state constraints consists in specifying a solvability set J(τ, t
1
, /) = J
0
[τ]
and a setvalued feedback control strategy u = (t, x), (, ) ∈ U
c
P
, such that all the
solutions to the diﬀerential inclusion
˙ x(t) ∈ (t, x) + f(t) (1.10.1)
that start from any given position ¦τ, x
τ
¦, x
τ
= x[τ], x
τ
∈ J(τ, t
1
, /), τ ∈ [t
0
, t
1
] would
satisfy the restrictions
x(t) ∈ ¸(t), τ ≤ t ≤ t
1
, (1.10.2)
x(t
1
) ∈ /. (1.10.3)
Deﬁnition 1.9.1 is nonredundant provided J
0
[τ] = J(τ, t
1
, /) ,= ∅, where, J
0
[τ] is the
”largest” set of states x
τ
, from which the solution to the problem of Deﬁnition 1.9.1 does
exist at all.
Following the same reasoning as in the absence of state constraints (see Sections 1.3,1.4 )
, it may be shown that the set J
0
[τ] will coincide with set J[τ] of Section 1.1.9, so that
J
0
[τ] ≡ J
0
[τ] ≡ J[τ], τ ∈ T.We shall further use notation J[τ] for this set and for the
respective tube ( τ ∈ T).
Let us now consider the tube J[τ], τ ≤ t ≤ t
1
, and deﬁne a feedback strategy
(t, x) = ∂
l
f(t, −l
0
W
(t, x)) (1.10.4)
similar to that of (1.4.20), (1.8.9). Here, as before, f(t, l) = ρ(l[T(t)) and l = l
0
W
(t, x) is
the maximizer for the expression
d
W
[t, x] = max¦(l, x) −ρ(l[J[t])
¸
¸
¸
¸
l ≤ 1¦ (1.10.5)
or
d
W
[t, x] = (l
0
, x) −ρ(l
0
[J[t])
if d
W
[t, x] > 0 (otherwise l
0
= 0).
Here
69
d
2
W
[t, x] = min¦(x −z, x −z)[z ∈ J[t]¦ = V (t, x) (1.10.6)
To prove that (t, x) is a solution to our problem we have to calculate the derivative
d
dt
V (t, x) = d
W
[t, x]
d
dt
d
W
[t, x] = d
W
[t, x]
d
+
dt
d
W
[t, x] (1.10.7)
due to the inclusion
˙ x ∈ (t, x) + f(t). (1.10.8)
We assume in this section that the support function ρ(l[¸(t)) of the multifunction ¸(t) is
absolutely continuous.
In order to do that, let us ﬁrst calculate the left partial derivative in t of the support
function ρ(l[W[t]), namely
∂
−
ρ(l[W[τ])
∂τ
¸
¸
¸
¸
τ=t+0
where
∂
−
ρ(l[W[τ])
∂τ
= lim
σ→0
σ
−1
(ρ(l[J[τ −σ]) − (ρ(l[J[τ])
for a given direction l ∈ IR
n
. We will further use the relation (1.9.21), particularly to
calculate the increment
ρ(l[J[τ −σ]) −ρ(l[J[τ])
through the relation
J[τ −σ] = (J[τ] −σT(τ) −f(τ)) ∩ ¸(τ −σ) + r(σ)
where σ
−1
h(r(σ), 0) → 0 with σ → 0.
Since
h(J
, J
) = max¦ρ(l[J
) −ρ(l[J
)
¸
¸
¸
¸
l = 1¦
we observe that the increments
∆
1
(σ) = σ
−1
(ρ(l[J[τ −σ] −ρ(l[J[τ]))
and
∆
2
(σ) = σ
−1
(ρ(l[(J[τ] −σT(τ) −σf(τ)) ∩ ¸(τ −σ)) −ρ(l[J[τ]))
are such that
lim
σ→0
[∆
1
(σ) −∆
2
(σ)[ = 0.
70
Therefore it suﬃces to calculate the left derivative
dg(σ)
dσ
¸
¸
¸
¸
σ=0
for the function
g(σ) = ρ(l[(J[τ] −σT(τ) −σf(τ)) ∩ ¸(τ −σ))
= min¦ρ(p[(J[τ] −σT(τ) −σf(τ))) + ρ(l −p[¸(τ −σ))[p ∈ IR
n
¦
since
dg(σ)
dσ
¸
¸
¸
¸
σ=0
=
∂
−
ρ(l[J[τ])
∂τ
The calculation then follows the techniques of directional diﬀerentiation given, for example,
in [89]. This ﬁnally yields
Lemma 1.10.1 The following relation holds
∂
−
ρ(l[J[τ])
∂τ
¸
¸
¸
¸
τ=t
=
min¦ρ(−p[T(t)) −(p, f(t)) −
∂
∂t
(ρ(l −p[¸(t))
¸
¸
¸
¸
p ∈ T(t, l)¦, (1.10.9)
where
T(t, l) = ¦p ∈ IR
n
: k(t, l) −k(t, p) = ρ(l −p[¸(t))¦ (1.10.10)
and
k(t, l) = ρ(l[J[t]).
Here the relations (1.10.9), (.10.10) reﬂect the fact that the “inﬁmal convolution” that
deﬁnes g(0), t = τ, is exact, namely the minimum in (1.0.9) is taken over all p ∈ IR
n
that
satisfy the equality
ρ(p[J[t]) + ρ(l −p[¸(t)) = ρ(l[J[t]).
Let us elaborate on this result. Since the properties of J[t] imply J[t] ⊆ ¸(t), we have
ρ(l[J[t] ∩ ¸(t)) = min¦ρ(l[J[t]), ρ(l[¸(t))¦
and therefore the minimum of g(0) over p is attained at either p = 0 (which is when
ρ(l[J[t]) ≤ ρ(l[¸(t)) or p = l (which is when ρ(l[¸(t) = ρ(l[J[t])). Formula (1.10.9)
actually yields
∂
−
ρ(l[J[τ])
∂τ
¸
¸
¸
¸
τ=t
=
min¦ρ(−l[T(t)) −(l, f(t)) , −
∂
∂t
(ρ(l[¸(t))¦ (1.10.11)
71
Relation (1.10.9) allows to calculate the right directional derivative
d
+
d
W
[t, x]/dt
due to system (1.9.1) through formula (1.9.5).
In view of the equality
∂
−
ρ(l[J[τ])
∂τ
¸
¸
¸
¸
τ=t
= −
∂ρ(l[J[τ])
∂τ
we come to
d
+
d
W
[t, x]
dt
= (l
0
, u + f(t)) +
∂ρ(l
0
[J[t])
∂t
=
= (l
0
, u + f(t)) +
min¦ρ(−l
0
[T(t)) −(l
0
, f(t)), −
∂
∂t
ρ(l
0
[¸(t))¦
≤ (l
0
, u + f(t)) + ρ(−l
0
[T(t)) + (−l
0
, f(t)),
which is true for almost all t. The last relation turns into an equality d
+
d
W
[t, x]/dt = 0,
if u ∈ (t, x), where (t, x) is given by (1.10.4). This conclusion produces
Lemma 1.10.2 Once u ∈ (t, x), where (t, x) is deﬁned by to (1.10.4), then almost
everywhere the derivative
d
+
d
W
[t, x]/dt ≤ 0,
and therefore
d
+
dt
V (t, x)
¸
¸
¸
¸
u∈U(t,x)
≤ 0 a.e. (1.10.12)
Similar to how it was before, in Sections 1.1.4 and 1.1.8, inequality (1.10.12) suﬃces to
prove that strategy (t, x) of (1.10.4) does solve the problem 1.10.1 of control synthesis
under state constraints. We thus come to the proposition
Theorem 1.10.1 The problem of control synthesis under state constraints of Deﬁnition
1.10.1 is solved by strategy (t, x) of (1.10.4).
The problem is obviously solvable if the starting position ¦t, x¦ is such that x ∈ J[t],
where J[t] is the solvability set given by the unique solution to equation (1.9.22) or by
the unique maximal solution to equation (1.9.23). It is not diﬃcult to prove though that
J[t] is the “largest” set from which the solution does exist at all. The respective proof is
similar to the one given in the last part of Section 1.8, so that the last theorem may be
complemented by
72
Lemma 1.10.3 In order that the problem of Deﬁnition 1.10.1 would be solvable it is nec
essary and suﬃcient that x
τ
∈ J[τ].
The results of this section may be again explained through DP techniques .
Exercise 1.10.1.
(a)Introducing the value function
V
0
(t, x) = min
_
t
1
_
t
h
2
+
(x(τ, ¸(τ))dτ + h
2
+
(x(t
1
), /)
¸
¸
¸
¸
u(τ) ∈ T(τ)
_
,
check, whether it satisﬁes the corresponding HJB equation for system (1.10.8) and what
would be the relations between the solutions to the problem of Deﬁnition 1.10.1 achieved
through V
0
(t, x) and through function V (t, x) of (1.10.6).
(b) Taking Example 1.5.1, complement it by a state constraint
(x(t) −n(t), N(t)(x(t) −n(t)) ≤ 1, N(t) > 0,
and ﬁnd the solvability set J
0
[τ] of Deﬁnition 1.10.1 by following the schemes of (1.5.15)
(1.5.17). Calculate the analogy of formula (1.5.17) for the given state constraint.
We ﬁnally come to the next topic which incorporates all the diﬃculties speciﬁc for the
previous Sections.This is the problem of control synthesis under both uncertainty and
state constraints.
1.11 State Constrained Uncertain Systems.
Viability Under Counteraction.
Consider system (1.7.1) with terminal set /, state constraint (1.10.2) and constraints
(1.1.2), (1.6.1) on the control u and the uncertain input f.
Deﬁniton 1.11.1 The problem of control synthesis under uncertainty and state
constraint consists in specifying a solvability set J[τ] = J
∗
(τ, t
1
, /) and a setvalued
control strategy (t, x) such that all the solutions x[t] = x(t, τ, x
τ
) to the diﬀerential in
clusion (1.8.1) that start at a given position ¦τ, x
τ
¦, x
τ
= x[τ] ∈ J
∗
(τ, t
1
, /), τ ∈ [t
0
, t
1
),
would reach the terminal set / at time t
1
, so that x(t
1
) ∈ /, and would also satisfy the
state constraint (1.10.2), namely
x[t] ∈ ¸(t), τ ≤ t ≤ t
1
.
73
Here the multivalued function ¸(t) with values in compIR
n
is again taken to be absolutely
continuous.
It is clearly the strategy (t, x) that is responsible for the solution x[t] to satisfy the state
constraint (1.10.2), no matter what is the disturbance f(t).
In this section we will have to combine the schemes of sections 1.1.7, 1.1.8 and sections
1.1.9, 1.1.10. The technicalities of this combination require a more or less sophisticated
mathematical treatment, the details of which are not directly relevant to the topics of
this book. They are the subject of other publications ( see [194], [195]). We will however
give a concise presentation of the solution to this problem emphasizing the substantial
interrelations important for the results.
The solution strategy (t, x) will again be determined by a relation of type (1.10.4),
(1.10.5), where J[t] has to be substituted by J
∗
[t] — the solvability set of Deﬁnition
1.11.1. The basic evolution equation for J
∗
[t] now has the form
(1.11.1)
lim
σ→0
σ
−1
h
+
(Z[t −σ] + σQ(t), Z[t] ∩ ¸(t) −σT(t)) = 0,
Z[t
1
] ⊆ /,
so that the following assertion holds.
Lemma 1.11.1 The solvability set J
∗
[t] for the problem of control synthesis under both
uncertainty and state constraints as formulated in Deﬁnition 1.11.1 is the maximal solution
to equation (1.11.1) with boundary condition Z[t
1
] = /.
The proof of this assertion follows the lines of sections 1.1.7 – 1.1.10 . As we have seen
above, the property important for Control Synthesis is the behavior of the directional
derivative
d
dt
V (t, x)
V (t, x) = d
2
W∗
[t, x] = h
+
(x, J
∗
[t])
along the solutions to the diﬀerential inclusion (1.10.1) with f(t) unknown, but bounded:
74
f(t) ∈ Q(t), t ∈ [t
0
, t
1
]. (1.11.2)
Combining the calculations of sections 1.1.8 and 1.1.10 under Assumption 1.6.1 we come
to
Lemma 1.11.2 The derivative dd(x, J
∗
[t])/dt is given by
d
dt
d(x, J
∗
[t]) = (l
0
, u + f) + min¦ρ(l
0
[(−T(t))
˙
− Q(t)) , −
∂
∂t
(ρ(l
0
[¸(t))¦ ≤
≤ (l
0
, u + f) + ρ(l
0
[ −T(t)) − ρ(l
0
[Q(t)). (1.11.3)
The synthesizing strategy 
0
∗
(t, x) may now be deﬁned in the same way as 
0
(t, x) of Section
1.1.8, that is according to (1.8.9), but with J
∗
[t] substituted by J
∗
[t] of Deﬁnition 1.11.1,
(the notation 
0
(t, x) is also substituted by 
0
∗
(t, x) ).
Similarly to Section 1.1.8 (Theorem 1.8.1), the previous lemma implies
Theorem 1.11.1 The strategy 
0
∗
(t, x) deﬁned by (1.8.9) (with J
∗
[t] substituted by J
∗
[t])
resolves the problem of control synthesis under uncertainty and state constraints of Deﬁni
tion 1.11.1.
Therefore every solution x[t] to the system
˙ x ∈ 
0
∗
(t, x) + f(t) (1.11.4)
x(t
0
) ∈ J
∗
[t
0
] (1.11.5)
satisﬁes the constraint
x[t] ∈ J
∗
[t], t
0
≤ t ≤ t
1
(1.11.6)
and therefore the inclusion
x[t
1
] ∈ /.
In this case we will say that system (1.11.3) is viable relative to constraint (1.10.2) under
counteraction (1.11.2), provided x(t
0
) satisﬁes (1.11.6).
75
In other terms we may say that J
∗
[t], t
0
≤ t ≤ t
1
, is a tube of strongly invariant sets for
system (1.11.3) under counteraction (1.11.2).( The latter term indicates that all the solu
tions to the diﬀerential inclusion (1.11.3) that start in J[t
0
], do satisfy the state constraint
(1.10.2)).
The assertions of this section ﬁnalize the concise description of the solutions to the problems
of evolution and control synthesis in the presence of uncertainty and state constraints. A
topic for further discussion is the application of set valued calculus to the problem of
stateestimation.
1.12 Guaranteed State Estimation :
the Bounding Approach
One of the basic problems of modelling and control is to estimate the state of an uncertain
or incompletely deﬁned dynamic system on the basis of on or oﬀline observations cor
rupted by noise. Leaving aside the welldeveloped stochastic approach to these problems,
and following the emphasis of the present book, we shall again assume the ”setvalued”
interpretations of the respective problems.
Namely, as in Section 1.6, an uncertain system is understood to be one of the following
type
˙ x(t) ∈ A(t)x(t) + u(t) + f(t), (1.12.1)
t
0
≤ t ≤ t
1
, x(t
0
) = x
0
,
where A(t) ∈ L(IR
n
IR
n
), u(t) is a given function ( a preselected control) and f(t) ∈ IR
n
is the unknown but bounded input (disturbance).
It is presumed that the initial state x
0
∈ IR
n
is also unknown but bounded, so that
f(t) ∈ Q(t), t
0
≤ t ≤ t
1
, (1.12.2)
x
0
∈ A
0
, (1.12.3)
where the set A
0
⊂ conv1
n
and the continuous setvalued function Q(t) ∈ compIR
n
, are
given in advance.
Equation (1.12.1) may be complemented, as we have seen earlier, in Section 1.9, by a state
constraint
G(t)x(t) ∈ ¸(t), t
0
≤ t ≤ t
1
(1.12.4)
76
with G(t) ∈ L(IR
n
IR
m
) and ¸(t) ∈ conv IR
m
, m ≤ n. The constraint (1.12.3) may be
particularly generated by a measurement equation
y(t) = G(t)x(t) + v(t), t
0
≤ t ≤ t
1
, (1.12.5)
with an unknown but bounded error
v(t) ∈ /(t), t
0
≤ t ≤ t
1
(1.12.6)
where /(t) ∈ conv IR
m
, t
0
≤ t
1
is an absolutely continuous setvalued map, ( recall Section
1.9).
With the realization y() being known, restriction (1.12.4), (1.12.5) turns into
G(t)x(t) ∈ y(t) −/(t), t
0
≤ t ≤ t
1
, (1.12.7)
so that
¸(t) = ¦x : G(t)x ∈ y(t) −/(t)¦,
( however, the whole function y() may not be known in advance, arriving online).
Our objective will be to estimate the system output
z(t) = Hx(t), z ∈ IR
r
, r ≤ n, t
0
≤ t ≤ t
1
(1.12.8)
at any prescribed instant of time t.
More precisely, the problem is to specify the range of the output z(t) that is consistent
with relations (1.12.1)(1.12.4) (the Attainability Problem under State Constraints), or the
set fo all outputs z(t) consistent with system (1.12.1)(1.12.3), and measurement equation
(1.12.5), (1.12.6) ,with realization y(t) of the measurement being given (the Guaranteed
State Estimation Problem).
The solution to both problems is therefore given in the form of a set representing thus the
bounding approach to state estimation.
Our aim here is not to repeat the wellknown information ( [276], [181], [225]). on these
issues, but to rewrite some theoretical results focusing them on the main objective, which is
further to devise in Part IV some constructive algorithmic procedures based on ellipsoidal
techniques that would allow a computer simulation with graphical representations.
8
8
The ﬁrst descriptions of state estimation ( observation ) problems under unknown but bounded errors
may be traced to papers [166], [317], [276], [177]. The setvalued approach to such problems in continuous
time appears to have started from publications [54], [178], [277], [181].
77
Let us specify the problems considered here. starting with the Attainability Problem. As
indicated in Section 1.9, the attainability domain A(, t
0
, x
0
) for (1.12.1),(1.12.2) under state
constraint (1.12.4) at time t ∈ [t
0
, t
1
] from point x
0
∈ IR
n
is the crosssection at t ∈ [t
0
, t
1
]
of the tube of all trajectories x[] = x(, t
0
, x
0
) that satisfy (1.12.1),(1.2.2),(1.12.4), namely,
A(, t
0
, x
0
) =
_
¦x(, t
0
, x
0
) [ x
0
∈ A
0
¦ (1.12.9)
Deﬁne the map A[t] = A(t, t
0
, A
0
) as
A(t, t
0
, A
0
) =
_
¦A(t, t
0
, x
0
)[x
0
∈ A
0
¦.
The multivalued map A[] generates a generalized dynamic system. Namely the mapping
A : [t
0
, t
1
] [t
0
, t
1
] convIR
n
→ convIR
n
possesses a semigroup property, that is, whatever are the values t
0
≤ t ≤ τ ≤ θ ≤ t
1
, we
have
A(θ, t, A[t]) = A(θ, τ, A(τ, t, A[t])).
Also, the setvalued map A, or in other words, the tube A[t], (t
0
≤ t ≤ t
1
) satisﬁes an
evolution equation — the ”funnel” equation of type (1.9.20), ([190], [193]) – which is
(1.12.10)
lim
σ→+0
σ
−1
h(A[t + σ], ((I + A(t)σ)A[t] + σT(t)) ∩ ¸(t + σ)) = 0,
t
0
≤ t ≤ t
1
,
A[t
0
] = A
0
,
Equation (1.12.10) is correctly posed and has a unique solution that deﬁnes the tube
A[] = A(, t
0
, A
0
) for system (1.12.1)(1.12.4) if the map ¸() is such that the support
function
ρ( [ ¸()) = max¦(, p) [ p ∈ /(t)¦
is absolutely continuous in t,[193].
Using only one of the Hausdorﬀ semidistances in (1.12.10) leads to the loss of uniqueness
of the solutions, but allows to relax the requirements on the multivalued function ¸(t).
Consider the evolution equation of type (1.9.23) which, in our case transforms into
lim
σ→+0
σ
−1
h
+
(Z[t + σ], ((I + A(t)σ)Z[t] ∩ ¸(t)) + σT(t)) = 0, (1.12.11)
78
with
t
0
≤ t ≤ t
1
,
and
Z[t
0
] = A
0
.
As we have observed earlier, the solution to this equation is nonunique. By complementing
it with an extremality condition, we obtain alternative descriptions for the multivalued
map A[].
A setvalued map A
+
[] will be deﬁned as a maximal solution to (1.12.11) if it satisﬁes
(1.12.11) for almost all t ∈ [t
0
, t
1
] and if there exists no other solution Z[] , such that
A
+
[t] ⊂ Z[t] for all t ∈ [t
0
, t
1
] and A
+
[] ,= J[]. Equation (1.12.11) has a unique maximal
solution under relatively mild conditions ( for example, if ¸(t) is only upper semicontinuous
in t), ( [194]). Particularly, it allows to treat a reasonably large class of discontinuous set
valued functions ¸(t).
Under the conditions required for the existence and uniqueness of the solutions to (1.12.10)
one may also observe, that A[] = A
+
[]. As mentioned in Section 1.9, equation (1.12.11)
is an alternative version relative to (1.12.10).
The Guaranteed State Estimation Problem may now be formulated more precisely. Namely,
suppose that the measurement y() = y
∗
(), due to system (1.12.1),(1.12.4) is given, and is
generated by an unknown triplet
ζ
∗
(t) = ¦x
∗
0
, f
∗
(t), v
∗
(t)¦, (1.12.12)
t
0
≤ t ≤ t
1
,
that complies with the constraints (1.12.2),(1.12.3),(1.12.5),(1.12.6), that is:
˙ x
∗
[t] = A(t)x
∗
[t] + u(t) + f
∗
(t) x
∗
0
∈ A
∗
0
, (1.12.13)
y
∗
(t) = G(t)x
∗
[t] + v
∗
(t), t
0
≤ t ≤ t
1
. (1.12.14)
Then the tube At = A
∗
[] of domains A
∗
[t] = A[t] = A(t, t
0
, A
0
) generated by (1.12.1)
(1.12.3), (1.12.5), (1.12.6) and calculated due to the knowledge of the measurement y[] =
y
∗
[], does always contain the unknown actual trajectory x
∗
[], generated by ζ
∗
(). Each set
A
∗
[t] therefore gives a guaranteed estimate of the state x
∗
[t] of system (1.12.1) on the basis
of the available measurement y
∗
(τ), t
0
≤ τ ≤ t under the constraints (1.12.2), (1.12.3),
(1.12.7).
Deﬁniton 1.12.1 The set A[t] = A(t, t
0
, A
0
), of states x = x(t) of system (1.12.1) that,
with given y(τ), t
0
≤ τ ≤ t , are consistent with the constraints (1.12.2),(1.12.3),(1.12.7)
is referred to as the information domain relative to measurement y()).
79
The information domain A[t] , [181], is also referred to as the ”domain of consistency”,or
the ”feasibility domain”, ( [56], [277], [225]). As mentioned above, it is the attainability
domain A[t] for system (1.12.1),(1.12.2),(1.12.7).
The solution of the guaranteed estimation problem is to specify the tube A[t] = A
∗
[t],
t
0
≤ t ≤ t
1
, deﬁned for a given measurement y(t) = y
∗
(t).The results of Section 1.9 allow
the following assertion.
Theorem 1.12.1 (i) With ¸(t) upper semicontinuous, the tube A
∗
[t] is the unique max
imal solution to the evolution equation (1.2.11).
(ii) With ¸(t) absolutely continuous, the tube (setvalued function)A
∗
[t] satisﬁes the evo
lution equation (1.12.10).It is also the unique maximal solution to (1.2.11).
(iii) Once A
∗
[t] is known, the estimate of the output z(t) is the set Z(t) = HA
∗
[t].
To conform with the assumptions on ¸(t) of the above, one ought to presume, for exam
ple, that y(t) is piecewisecontinuous (”from the right”), if we use equation (1.12.11), or
absolutely continuous, if we use (1.12.10).
It is important to emphasize that in many applied problems the observed measurement
output y(t) is not obliged to be continuous. We shall therefore further allow it to be only
Lebesguemeasurable. In order to imbed this situation in the given schemes, we shall apply
the idea of singular perturbation technique. But one must of course realize that this time
the object of application is a diﬀerential inclusion and that the propagation of wellknown
results [294], [295] in singular perturbation theory to trajectory tubes would require speciﬁc
treatment.
Consider the system of diﬀerential inclusions ( > 0):
˙ x ∈ A(t)x +T(t), (1.12.15)
˙ w ∈ −G(t)x +¸(t), (1.12.16)
¦x(t
0
), w(t
0
)¦ ∈ Z
0
t
0
≤ t ≤ τ. (1.12.17)
Here w ∈ IR
m
, z
0
∈ conv(IR
n
IR
m
)). As in the earlier Sections by A[t] = A(t, t
0
, A
0
), we
denote the trajectory tube of system (1.1.16) that consists of all those trajectories that start
at A[t
0
] = A
0
and satisfy the state constraint (1.12.4) for all t ∈ [t
0
, τ] A[t] is obviously
the tube of all viable trajectories relative to constraints (1.12.4) and (1.12.3).
80
Following this notation symbol Z[t] = Z(τ, t
0
, Z
0
, ) witll denote the tube of solutions
z(t) = ¦x(t), y(t)¦ to the system (1.12.15)(1.12.17) on the interval [t
0
, τ]. We will also use
the notation Π
x
W for the projection of set W ⊂ IR
n
IR
m
onto the space IR
n
of variables
x.
Here the constraint (1.12.4) may particularly be generated by a measurement equation, as
in (1.12.7) or (1.12.5),(1.12.6), where function y(t)  the realization of the observations  is
allowed to be Lebesguemeasurable. The ”bad” properties of y(t) are then clearly due to
the ”bad” measurement ”noise” v(t) in (1.12.5).
Our aim is still to describe the tube A[t]. However, in order to achieve this, we shall not
study system (1.12.16),(1.12.4) directly, but shall rather deal with the perturbed system
(1.12.16)  (1.12.18). The latter system may then be fully treated within the ”standard”
framework of Sections 1.9, 1.10 and papers [192], [193]. The following assertion is true.
Theorem 1.12.2 Assume that
A
0
⊆ Π
x
Z
0
. (1.12.18)
Then for every trajectory x() ∈ A[] of (1.12.15),(1.12.3),(1.12.4) there exists a vector
w
0
∈ IR
m
such that ¦x(t
0
), w
0
¦ ∈ Z
0
and for every τ ∈ [t
0
, t
1
]
z(τ) = ¦x(τ), w
(
τ)¦ ∈ Z(τ, t
0
, Z
0
, )
for all > 0.
Corollary 1.12.1 Assume (1.12.18) to be true. Then
A[τ] ⊆ Π
x
(∩¦Z(τ, t
0
, Z
0
, )[ > 0¦)
Let us now introduce another system of diﬀerential inclusions of type (1.12.15), (1.12.16),
but with a timedependent matrix L(t) instead of the scalar > 0 :
˙ x ∈ A(t)x +T(t), (1.12.19)
L(t) ˙ y ∈ −G(t)x +¸(t) (1.12.20)
z
0
= ¦x(t
0
), w(t
0
)¦ ∈ Z
0
, t
0
≤ t ≤ τ. (1.12.21)
The class of all continuous invertible matrix functions L(t) ∈ L(IR
n
, IR
n
), t ∈ [t
0
, t
1
] will
be denoted as L and the solution tube to system (1.12.19)(1.12.21) will be denoted as
Z[t, L] = Z(t, t
0
, X
0
, L).
The following analogy of Theorem 12.2 is true.
81
Theorem 1.12.3 Assume relation (1.12.18) to be true. Then for every x() ∈ A[] there
exists a vector w
0
∈ IR
m
such that
¦x(t
0
), w
0
¦ ∈ Z
0
,
and for every τ ∈ [t
0
, t
1
]
z(τ) = ¦x(τ), y(τ)¦ ∈ Z[τ, L],
whatever is the function L() ∈ L.
Corollary 1.12.2 Assume relation (1.12.18) to be true. Then
A[τ] ⊆ Π
x
(∩¦Z[τ, L][L() ∈ L¦). (1.12.22)
The principal result of the singular perturbations method applied to the guaranteed esti
mation problem discussed here is formulated as follows
Theorem 1.12.4 Let us assume
Π
x
Z
0
⊆ X
0
.
Then for every τ ∈ [t
0
, t
1
] the following inclusion is true
Π
x
(∩¦Z[τ, L][L() ∈ L¦) ⊆ A[τ]. (1.12.23)
This result may be proved within the techniques of Sections 1.9, 1.10. Its details may be
found in [193].
Relations (1.12.22), (1.12.23) yield an exact description of the set X[τ] through the solu
tion of the perturbed diﬀerential inclusions (1.12.19)(1.12.21) that are without any state
constraints:
Theorem 1.12.5 Under the assumption
Π
x
Z
0
= A
0
the following formula is true
A[τ] = Π
x
(∩¦Z[τ, L][L() ∈ L¦) (1.12.24)
for any τ ∈ [t
0
, t
1
].
82
The application of this theorem to the calculation of information sets will be illustrated in
Section 4.6, where it will be further modiﬁed to suit the related ellipsoidal techniques.
The conventional theory of guaranteed state estimation as introduced in [181], [225], may
require to ﬁnd the worstcase estimate of x(t) as a vector x
0
(t), which is usually taken to
be the ”Chebyshev center” of set A(t), namely , as the solution to the problem
max
z
¦ x
0
(t) −z  [z ∈ A(t)¦ = (1.12.25)
= min
x
max
z
¦ x −z  [x ∈ A(t), z ∈ A(t)¦.
The Chebyshev center of a set A is the center of the smallest Euclidean ball that includes
A. Its calculation leads to mathematical programming problems of special type, [139],
[86], [88], [69], [209]. The approximate calculation of Chebyshev centers is generating an
increasing literature,[225]. A less investigated problem is to ﬁnd the Steiner center, [275]
of set A[t].
The interested reader, who has managed to reach these lines, may be curious to know ,
whether the results of the last few Sections could again be interpreted in some conventional
way, in terms of DP , as in Section 5 , for example. These questions are discussed further,
in the ﬁrst Sections of Part 1V.
1.13 Synopsis
We shall now summarize the results of the previous Sections. Namely, recall that we have
considered the system
˙ x(t) = u + f(t) , t
0
≤ t ≤ t
1
, (1.13.1)
with constraints on the controls
u ∈ T(t) , (1.13.2)
the unknown inputs
f(t) ∈ Q(t) , (1.13.3)
the initial state
83
x
0
∈ X
0
, x(t
0
) = x
0
, (1.13.4)
and the state space variables
G(t)x ∈ ¸(t) , G(t) ∈ L(IR
n
, IR
m
), (1.13.5)
and with continuous in time set–valued functions
T(t) ∈ convIR
n
, Q(t) ∈ convIR
n
, ¸(t) ∈ convIR
m
and matrixvalued function G(t) taken to be continuous in t. ( Recall that the presumed
property of Q(t) being continuous is translated into the presumption that f(t) is continuous,
whenever Q(t) = f(t) is reduced to a singleton f(t), otherwise, f(t) is allowed to be
measurable in t).
Among the problems of control and estimation for this system we have singled out ﬁve for
detailed treatment, to demonstrate the suggested approach. These are the following
I System with no input uncertainty and no state constraints ( Sections 1.2 , 1.3 ):
f(t)  given ;Q(t) = f(t)  singlevalued, ¸(t) ≡ IR
m
.
II System with input uncertainty and no state constraints (Sections 1.6,1.7 ):
f(t)  unknown, but bounded, due to (1.14.3), ¸(t) ≡ IR
m
.
III System with state constraint but no uncertainty (Section 1.9):
f(t)  given; Q(t) ≡ f(t); ¸(t) ∈ convIR
m
 absolutely continuous in t,
IV System with uncertainty and with state constraints (Section 1.11) :
f(t)  unknown but bounded, due to (1.13.3); ¸(t) ∈ convIR
m
 same as in III.
IV’ System with measurement output (Section 1.12), with uncertainty in the inputs, initial
states and measurement noise : control u(t)  given, input f(t)  unknown, but bounded,
due to (1.13.3), state constraint given in the form
y(t) ∈ G(t)x +/(t)
or, equivalently,
84
y(t) ∈ ¸(t) , ¸(t) = y(t) −/(t),
where y(t) is the available measurement, /(t) is the bound on the measurement error.
The ﬁrst issue discussed was the calculation of the attainability domains and the attain
ability tubes . These were given through the solutions of the following evolution ”funnel”
equations with setvalued solutions, namely,
for case I
lim
σ→0
σ
−1
h(A[t + σ], A[t] + σT(t) + σf(t)) = 0,
A[t
0
] = A
0
; (1.13.6)
for case II
lim
σ→0
σ
−1
h
+
(A[t + σ] −σQ(t), A[t] + σT(t)) = 0,
or equivalently,
lim
σ→0
σ
−1
h
+
(A[t + σ], A[t] + σ(T(t)
˙
−(−Q(t)))) = 0,
under condition (1.13.6) ;
for case III
lim
σ→0
σ
−1
h(A[t + σ], (A[t] + σT(t) + f(t)) ∩ ¸(t + σ),
or
lim
σ→0
σ
−1
h
+
(A[t + σ], (A[t] ∩ ¸(t) + σT(t) + f(t)) + 0,
under (1.13.6) ;
for case IV
lim
σ→0
σ
−1
h
+
(A[t +σ] −σQ(t), A[t] ∩ ¸(t) +T(t) + f(t)) = 0,
85
under condition (1.13.6) ;
for case IV’
lim
σ→0
σ
−1
h(A[t + σ], (A[t] + σ(u(t) + y(t)) + σQ(t)) ∩ /(t)) = 0,
(if the function ¸(t) = y(t) −/(t) is absolutely continuous in t ) or
lim
σ→0
σ
−1
h
+
(A[t] + σ, A[t] ∩ ((y(t) −/(t)) + σ(u(t) +Q(t))) = 0,
(if the function ¸(t) is upper semicontinuous in t, particularly, if /(t) is continuous and
y(t) is piecewise continuous from the right).
Both equations are considered under conditions (1.13.6) .
The respective attainability domains are given through the respective unique solutions to
the evolution equations when these are written in terms of the Hausdorﬀ distance h(.)
and through the maximal solutions ( with respect to inclusion) for the equations written
in terms of the Hausdorﬀ semidistance h
+
(, ).The attainability domain for case IV’ is
the informational domain for the guaranteed state estimation problem of Section 1.12.
The second group of issues consists of problems of goal  oriented nonlinear control syn
thesis. Here the objective is to reach a preassigned terminal target set / at given time
t = t
1
by selecting a feedback control strategy (t, x) ∈ U
c
P
which in general turns to be
nonlinear, as the controls are bounded here by magnitude bounds. The overall synthesized
system is then described by a nonlinear diﬀerential inclusion.
For each of the system types I  IV this strategy  = 
0
(t, x) is selected in a standard
way by minimizing the derivative
d
dt
V (t, x)
¸
¸
¸
¸
u=U
0
= min
_
d
dt
V (t, x)
¸
¸
¸
¸
u=U
:  ∈ U
c
P
_
, (1.13.7)
where
V (t, x) = d
2
(x, J[t])
and W[t] is the crossection of the respective solvability tube.
9
9
This type of solution was introduced by N.N.Krasovski under the name of extremal aiming strategy
with solution tubes W[t] being referred to as ”bridges”, see [168], [169] and also [171].
86
The strategy (t, x) may also be calculated directly, without introducing the tube J[t],
but, as indicated in Sections 1.5,1.8,1.10,1.11, by solving for the respective problems the
respective HJB equations with value functions
V
∗
(t, x) = min¦h
2
+
(x(t
1
, t
0
, x), /)[(, ) ∈ U
c
P
¦
for case I,
V
∗
(t, x) = min
U
max
f
¦h
2
+
(x(t
1
, t
0
, x), /)[(, ) ∈ U
c
P
, f() ∈ Q()¦
for case II,
V
0
(t, x) = min
_
t
1
_
t
h
2
+
(x(τ, ¸(τ))dτ + h
2
+
(x(t
1
, t
0
, x)
¸
¸
¸
¸
u() ∈ T()
_
for case III,
V
0
(t, x) = min
U
max
f
_
t
1
_
t
h
2
+
(x(τ, ¸(τ))dτ + h
2
+
(x(t
1
, t
0
, x)
¸
¸
¸
¸
(, ) ∈ U
c
P
, f() ∈ Q()
_
for case IV
and with further application of (1.14.6) to V = V
∗
(t, x) or V = V
0
(t, x).
For systems of type (1.14.1) the respective solvability sets J[t] will turn to be level sets
for the corresponding value functions, namely,
J[t] = ¦x : V
∗
(t, x) ≤ 0¦
for cases I, II and
J[t] = ¦x : V
0
(t, x) ≤ 0¦
for cases III, IV.
The control strategies are then determined from relation (1.13.6) taken for the correspond
ing value functions.
87
The ability to calculate the solvability tubes J() eliminates the necessity to solve the
HJB equation . The speciﬁc emphasis is that these tubes may be calculated through
evolution equations which are precisely the ones introduced for the attainability domains
, but should be now taken in backward time. Namely, we have introduced the following
equations:
for case I
lim
σ→0
σ
−1
h(W[t −σ], W[t] −σT(t) −σf(t)) = 0, (1.13.8)
for case II
lim
σ→0
σ
−1
h
+
(W[t −σ] + σQ(t), W[t] −σT(t)) = 0, (1.13.9)
or
lim
σ→0
σ
−1
h
+
(W[t −σ], W[t] −σ(T(t)
˙
−(−Q(t))) = 0, (1.13.10)
for case III
lim
σ→0
σ
−1
h(W[t −σ], (W[t] −σT(t) −σf(t)) ∩ ¸(t −σ)) = 0, (1.13.11)
or
lim
σ→0
σ
−1
h
+
(W[t −σ], W[t] ∩ ¸(t) −σT(t) −σf(t)) = 0, (1.13.12)
for case IV
lim
σ→0
σ
−1
h
+
(W[t −σ] + σQ(t), W[t] ∩ ¸(t) −σT(t)) = 0, (1.13.13)
All of these equations have to be solved with boundary condition
W[t
1
] = /. (1.13.14)
The unique solutions to equations ( 1.13.7), (1.13.10) with boundary condition (1.13.13)
and the maximal solutions to equations (1.13.8), (1.13.9), (1.13.11),(1.13.12) with same
boundary condition give us the respective solvability tubes J[] that produce the the
crucial elements J[t] for calculating the required control strategies (t, x).
88
Needless to say, equations (1.13.8),(1.13.9), (1.13.11) are particular cases of equation
(1.13.12) and equation (1.13.7) is a particular case of (1.13.10).( The solution to (1.13.7) is
also the maximal solution to a modiﬁcation of this equation, where distance h is substituted
by semidistance h
+
).
Now it should be probably clear, that equations (1.13.7)(1.13.12) may serve to be the
motivation and the basis for introducing discretized schemes with setvalued elements. In
other words, we may loosely assume:
for case I
W[t −σ] · W[t −σ] −σT(t) −σf(t), (1.13.15)
for case II
W[t −σ] · W[t] −σ(T(t)
˙
−(−Q(t))), (1.13.16)
for case III
W[t −σ] · (W[t] −σT(t) −σf(t)) ∩ ¸(t −σ) (1.13.17)
or
W[t −σ] · W[t] ∩ ¸(t) −σT(t) −σf(t), (1.13.18)
for case IV
W[t −σ] · W[t] ∩ ¸(t) −σ(T(t)
˙
−(−Q(t))) −σf(t). (1.13.19)
The ”equalities” (1.13.14)  (1.13.18) are true relative to an error of order γ(σ), where
σ
−1
γ(σ) → 0, σ → 0.
10
.
The last relations indicate that the basic setvalued operations for the topics of this book
are the geometrical ( ”Minkowski”) sums (+,) and diﬀerences(
˙
−) of convex compact sets
as well as their intersections (∩).Since an arbitrary convex compact set is an inﬁnite
dimensional element ( that may be identiﬁed with its support function, for example), the
respective numerical calculations require ﬁnite dimensional approximations.This book
indicates ellipsoidal approximations as an appropriate technique.
10
The indicated relations describe ﬁrst order approximations to the exact set  valued solutions of the
above. The theory of secondorder approximations to solution tubes for diﬀerential inclusions was discussed
in paper [308]
89
1.14 Why Ellipsoids ?
The aim of this book is to indicate some constructive techniques for solving problems of es
timation and feedback control under setmembership uncertainty and state constraints with
an aspiration that these techniques would allow eﬀective algorithmization and computer
animation. As we have seen in the above , the basic mathematical tool for describing the
class of problems raised here is setvalued calculus. It is probably not unnatural , therefore,
that the speciﬁc methods selected in the sequel are based on an ellipsoidal technique that
would allow to approximate the setvalued solutions of the above by ellipsoidalvalued solu
tions. Particularly, the setvalued attainability and solution tubes of the previous Sections
will be further approximated by ellipsoidalvalued functions.
Technically one of the basic justiﬁcations for such an approach is that crossections
A[t], J[t] of the convex compact valued functions A[], J[] may be presented ( for all
the cases I  IV under consideration) in the form of intersections
A[t] = ∩c
(α)
+
(t) , J[t] = ∩c
(β)
+
(t) (1.14.1)
over a parametrized inﬁnite variety of ellipsoidalvalued functions c
(α)
+
(t), c
(β)
+
(t) ( which
may even be assumed denumerable). Each of these , in its turn, may be calculated by
solving a system of ordinary diﬀerential equations ( ODE’s). The calculation of A[t], J[t]
would thus be parallelized into an array of identical problems each of which would consist
in solving an ODE that describes an ellipsoidalvalued function c
(α
∗
)
+
(t) or c
(β
∗
)
+
(t).
The ellipsoidal representations (1.14.1) for A[t], J[t] are exact and are true for the solutions
of each of the evolution equations indicated in Section 1.14, (that is, for all the cases I 
IV indicated in this Section). Moreover , in the absence of state constraints, ( cases I, II),
the ellipsoidal calculus used here also allows eﬀective internal ellipsoidal approximations
in the form of
A[t] = ∪c
(γ)
−
(t), J[t] = ∪c
(δ)
−
(t), (1.14.2)
where as before the dash stands for the closure of the respective set and where
c
(γ)
−
(t), c
(δ)
−
(t), again stands for the elements of an inﬁnite denumerable variety of
ellipsoidalvalued functions described by ordinary diﬀerential equations.
The ellipsoidal calculus suggested further yields, among others, the ability to address the
following issues:
(i) The exact representation and approximation of attainability domains for linear systems
with or without state constraints through both external and internal ellipsoids.
90
(ii) The treatment of attainability and solvability tubes A[t], J[t] under setmembership
uncertainty ( counteraction ) in the inputs . These, in general , my be particularly de
scribed by alternated integrals of L.S.Pontryagin  an object far more complicated than the
standard setvalued ( ” Aumann” ) integral that represents similar tubes in the absence
of input uncertainty. Nevertheless, the respective tubes, given by alternated integrals or
by corresponding evolution equations of the ” funnel type” , still allow exact internal and
external ellipsoidal representations.
(iii) The exact ellipsoidal representation or external approximation of the informational
domains for guaranteed ( setmembership) state estimation under unknown but bounded
errors.
(iv) The possibility to single out ”individual” external or internal approximating ellipsoids
that are optimal relative to some given optimality criterion ( trace, volume, diameter, etc.
) or a combination of such criteria. They also allow to apply vector  valued criteria to the
approximation problem.
11
Loosely speaking, the representations of type (1.14.1), (1.14.2) mean that the more ellip
soids are allowed to approximate A[t], J[t] ( in practice this depends on the number of
available processors ), the more accurate will be the approximation, so that , in theory, an
inﬁnite ( denumerable ) variety of ellipsoids would produce the exact relations (1.14.1) ,
(1.14.2). Thus, each ellipsoidalvalued function could be treated through a single processor
which solves an ODE of ﬁxed dimension. The number of available processors would then
determine the accuracy of the solution.
The application of ellipsoidal techniques will further allow to devise relatively simple con
trol strategies for control synthesis that would ensure guaranteed results for the related
problems. The strategies will then be given in the form of analytic designs rather than algo
rithms as it is in the exact case. The important feature that allows such ”ellipsoidalbased”
analytical designs is that the multivalued mappings that generate the internal ellipsoidal
tubes c
δ
−
(t) for the solvability sets J[t] satisfy a generalized semigroup property on one
hand, and on the other, the tubes are nondominated ( ” inclusionmaximal”) among all
such ellipsoidal valued functions. These two properties allow to demonstrate that the
tubes c
δ
−
(t) possess the property of being an ”ellipsoidal bridge” similar to the ”Krasovski
bridge” of the exact solution, (see Sections 3.5, 3.8).
It is obvious , of course, that one of the options is to approximate the setvalued func
tions A[t], J[t] by using boxes or , more generally, by polyhedralvalued functions. This
approach, which, of course, has its advantages and disadvantages, lies beyond the scope
of the present book. ( We address the reader to [187], [175], [225], [75]). However, it
appears that the main diﬃculty lies in the fact that computational complexity is such
11
the possibility of exact representations and of vectorvalued criteria for approximating attainability
domains under state constraints by external ellipsoids was indicated in monograph [181]. The minimal
volume criteria for these problems was thoroughly studied in [277], [73].
91
that the number of elementary operations here increases exponentially with the number of
steps in the sampled problem. A natural desire will then be to parallelize the polyhedral
approximation into problems of smaller dimensions.
(v) Another motivation for using ellipsoids comes from Section 4.2. There the nondiﬀeren
tiable solution V (t, x) to the HJB equation ( its level set is the attainability domain) is
approximated by quadratic functions whose level sets are nondegenerate ellipsoids. At the
same time, these functions are precisely the test functions used in deﬁning the generalized
”viscosity” solutions of the HJB equation.
It also appears useful to remark that any convex compact set Q in IR
n
may be presented
as an intersection of ellipsoids
Q = ∩c
(σ)
(This fact is a consequence of an ellipsoidal separation theorem  the property that every
point x ,∈ Q may be separated from Q by an ellipsoidal surface ).
The last fact justiﬁes that we further take all the sets A
0
, T(t), Q(t), ¸(t), X
0
that deﬁne
the preassigned constraints (1.2.1),(1.1.2),(1.6.1),(1.9.3) on the system to be ellipsoidal 
valued.
92
Part II. THE ELLIPSOIDAL CALCULUS
Introduction
This part is a separate text on ellipsoidal calculus  a technique of representing basic
operations on ellipsoids. The operations treated here are motivated by the requirements of
part I of the present book. However, the results given here may be applied, of course, to a
substantially broader class of problems that arise in mathematical modeling, particularly,
in optimization and approximation, identiﬁcation and experiment planning, probability
and statistics, stabilization, adaptive control, mathematical morphology and other areas.
The operations on ellipsoids are discussed in the following order. First of all , these are
the geometrical (Minkowski) sums and diﬀerences of two nondegenerate ellipsoids with
the diﬀerence having a nonvoid interior. Each sum and diﬀerence is approximated  both
externally and internally  by a corresponding parametrized variety of ellipsoids. With the
number of approximating ellipsoids increasing to inﬁnity, the approximations converge, in
the limit, to exact representations. The external representations are given by intersections
and the internal by unions ( or their closures) over the respective varieties each of which is
inﬁnite, denumerable, at least. Taking intersections or unions over some ﬁnite subsets of
these varieties, we come to external and internal approximations of the sums and diﬀerences
(Sections 2.2,2.3). Particularly, we make take only one element of the respective variety,
which is optimal in some sense ( an array of possible optimality criteria is discussed at the
end of Section 2.1). Then the sums and diﬀerences will be approximated (internally or
externally) by an optimal ellipsoid (Section 2.5). These criteria may include its diameter,
the sum of its axes or their squares (the trace of the matrix deﬁning the ellipsoid), ([111],
[181], [263], [225]) etc. A widely studied criterion is the volume of the ellipsoid ( see [277],
[73]).
A certain reciprocity consists in the fact that the external ellipsoids for the sums and the
internal ones for the diﬀerences are given by the same type of parametrization which diﬀers
in both cases only in some signs in the representation formula. A similar fact is true for
the internal ellipsoidal approximations of the sums and the external ones of the diﬀerences
(Sections 2.2, 2.3).
The obtained representations are then propagated to ﬁnite sums of nondegenerate ellipsoids
and to setvalued integrals of ellipsoidalvalued functions c[t] ( which are not obliged to
be ellipsoids). These sums and integrals are again approximated externally and internally.
Moreover, if the upper limit of the setvalued integral
A[t] =
_
t
t
0
c[τ]dτ
93
varies, then the parameters of the ellipsoidal functions that approximate A[t] may be
described by ordinary diﬀerential equations ( Sections 2.7, 2.8). The important element
here is that these ellipsoidalvalued functions that approximate A[t] are ”nondominated”
with respect to inclusion. Namely, they are inclusion minimal for the external ellipsoids
and inclusionmaximal for the internals.
Intersections of ellipsoids are the topic of Section 2.6 . Several types of external ellipsoidal
approximations are described here with exact representations in the limit. An indication
is ﬁnally given on how to construct varieties of internal ellipsoidal approximations of inter
sections of ellipsoids. The construction of internal ellipsoidal approximations to polyhedral
and other types of convex sets are important particularly in algorithmic problems of math
ematical programming ( [152], [281]). The solution to these is usually given in the form of
an algorithm. (Also note the theory of analytical centers , [286]). However, to be consis
tent with the approach presented here, one must be able to indicate a variety of internal
ellipsoids whose union (or its closure) would approximate a nondegenerate intersection of
ellipsoids ( ” from inside ”) with any desired degree of accuracy.
In Section 2.4 we also mention a direction towards the calculation of approximation errors (
depending on the number of approximating ellipsoids). Eﬀective algorithms for estimating
the errors as well as the computational complexity of these problems are among the issues
that present a further challenge. We also believe that one should not drop the problem of
ﬁnding perhaps rough, but simple error estimates.
2.1 Basic Notions : the Ellipsoids
As we have seen in Part I, the basic setvalued operations involved in the calculation of
solutions to the control problems of the above are the following:
• the geometrical (Minkowski) sum of convex sets,
• the geometrical (Minkowski) diﬀerence of convex sets,
• the intersection of convex sets,
• the aﬃne transformations of convex sets.
Let us elaborate on the ﬁrst two of these operations presuming that the sets involved are
convex and compact.
Deﬁniton 2.1.1 Given sets H
1
, H
2
∈ comp IR
n
, the geometrical (Minkowski) sum H
1
+H
2
is deﬁned as
H
1
+H
2
=
_
h
1
∈H
1
_
h
2
∈H
2
¦h
1
+h
2
¦.
94
Obviously, the support function
ρ([H
1
+H
2
) = ρ(l[H
1
) + ρ(l[H
2
).
Deﬁniton 2.1.2 Given sets H
1
, H
2
∈ comp IR
n
, the geometrical
( or Minkowski or also “internal”) diﬀerence H
1
˙
−H
2
is deﬁned as
H
1
˙
−H
2
= ¦h ∈ IR
n
: h +H
2
⊆ H
1
¦.
This means H
1
˙
−H
2
,= ∅ if there is an element h ∈ IR
n
, such that
h +H
2
⊆ H
1
.
Clearly
H
1
˙
−H
2
= ¦h ∈ IR
n
: h ∈ H
1
−h
2
, for all h
2
∈ H
2
¦.
What follows from here are the assertions
Lemma 2.1.1 The set H
1
˙
−H
2
may be presented as
H
1
˙
−H
2
=
h
2
∈H
2
_
h
1
∈H
1
¦h
1
−h
2
¦. (2.1.1)
Lemma 2.1.2 The geometrical diﬀerence H = H
1
˙
−H
2
is the maximal convex set (with
respect to inclusion) among those that satisfy the relation
H+H
2
⊆ H
1
, (2.1.2)
namely,
H = H
1
˙
−H
2
if and only if
H+H
2
⊆ H
1
and H
+H
2
⊆ H
1
imply H
⊂ H.
In terms of support functions the inclusion (2.1.2) yields
ρ(l[H) ≤ ρ(l[H
1
) −ρ(l[H
2
) = f(l), ∀l ∈ IR
n
, (2.1.3)
so that, if f(l) were convex, we would have
ρ(l[H) = ρ(l[H
1
) −ρ(l[H
2
)
(since H is the inclusionmaximal set that satisﬁes (2.1.3)).
Otherwise,
ρ(l[H) = ( co f)(l) = f
∗∗
(l)
where co f is the “lower envelope” of f(l), [100].
The following properties may happen to be useful
95
Lemma 2.1.3 With H
1
, H
2
, H
3
∈ compIR
n
we have
H
1
˙
−(H
2
+H
3
) = (H
1
˙
−H
2
)
˙
−H
3
(2.1.4)
(H
1
+H
2
)
˙
−H
3
⊇ H
1
+ (H
2
˙
−H
3
). (2.1.5)
Let us indicate an example which would illustrate that in the last relation both an equality
and a strict inclusion are possible.
96
Example 2.1.1
Assume
S
r
(0) = ¦(x, y) ∈ IR
2
: x
2
+ y
2
≤ r
2
¦
H
1
= S
1
(0)
H
2
= S
1
(0)
H
3
= ¦(x, y) ∈ IR
2
: −1 ≤ x ≤ 1, y = 0¦
Then
H
1
+H
2
= S
2
(0)
(H
1
+H
2
)
˙
−H
3
= S
2
(0)
˙
−H
3
= H,
where, obviously
H = ¦(x, y) ∈ IR
2
: (x + z)
2
+ y
2
≤ 4, [z[ ≤ 1¦.
In other words, the set H is the intersection of the sets
¦(x, y) ∈ IR
2
: (x + 1)
2
+ y
2
≤ 4¦
¦(x, y) ∈ IR
2
: (x −1)
2
+ y
2
≤ 4¦.
On the other hand, clearly, H
2
˙
−H
3
= ¦0¦, according to the deﬁnition of the geometric
diﬀerence. Therefore, H
1
+ (H
2
˙
−H
3
) = H
1
+¦0¦ = S
1
(0) and
S
1
(0) ⊂ H.
The relation (2.1.5) is therefore a strict inclusion.
Example 2.1.2
Take
H
1
= ¦(x, y) ∈ IR
2
: x = 0, [y[ ≤ 1¦
H
2
= S
1
(0)
H
3
= ¦(x, y) ∈ IR
2
: [x[ ≤ 1, y = 0¦
Then
H
1
+H
2
=
_
¦(x, y) ∈ IR
2
: x
2
+ (y + z)
2
≤ 1 [ [z[ ≤ 1¦
and one may observe
(H
1
+H
2
)
˙
−H
3
= H
1
.
On the other hand, clearly,
H
2
˙
−H
3
= ¦0¦,
97
and
H
1
+ (H
2
˙
−H
3
) = H
1
.
In this case the inclusion (2.1.5) is an equality.
The convex compact set H = H
1
˙
−H
2
is deﬁned to be inclusionmaximal relative to those
convex compact sets that satisfy the relation
H
2
+H ⊆ H
1
,
being therefore an internal diﬀerence. It is not diﬃcult to prove that the internal diﬀerence
is unique.
Deﬁniton 2.1.3 A set H
e
will be deﬁned as an external diﬀerence H
e
= H
1
÷H
2
if it
is inclusionminimal relative to those convex compact sets that satisfy the relation
H
2
+H ⊇ H
1
. (2.1.6)
The external diﬀerence may also be deﬁned as the class H = ¦H¦ of all sets H ∈ conv (IR
n
)
that satisfy the inclusion (2.1.6).
Then
ρ(l[H
2
) + ρ(l[H) ≥ ρ(l[H
1
)
∀H ∈ H, ∀l ∈ IR
n
.
and it is possible to demonstrate that
inf¦ρ(l[H)[H ∈ H¦ = ρ(l[H
1
) −ρ(l[H
2
) ≥ ρ([H
e
)
(since H
1
, H
2
∈ comp (IR
n
) and l ∈ IR
n
is ﬁnitedimensional).
Our further aim is to introduce an ellipsoidal calculus that would allow to approximate the
above relations for convex compact sets through ellipsoidalvalued relations.
Deﬁniton 2.1.4 An ellipsoid c(a, Q) with center a ∈ IR
n
and “conﬁguration” matrix Q
(symmetric and nonnegative) is deﬁned as the set
c(a, Q) = ¦x ∈ IR
n
: (l, x) ≤ (l, a) + (l, Ql)
1
2
, ∀l ∈ IR
n
¦,
where its support function ρ(l[c(a, Q) is deﬁned by the equation
ρ(l[c(a, Q)) = (l, a) + (l, Ql)
1
2
.
98
With Q nondegenerate, the ellipsoid c(a, Q) could also be presented otherwise, in terms of
the inequality
c(a, Q) = ¦x ∈ IR
n
: (x −a)
Q
−1
(x −a) ≤ 1¦.
which gives a direct, conventional description.
12
We will now proceed with the following basic setvalued operations, applying them to
ellipsoids
The geometrical (Minkowski) sum. Given two ellipsoids c
1
= c(a
1
, Q
1
), c
2
= c(a
2
, Q
2
),
their sum c
1
+ c
2
may obviously not be an ellipsoid (ﬁnd an example). We will
therefore be interested in ellipsoidal approximations c
(+)
+
, c
(−)
+
of the sum c
1
+ c
2
where
c
(+)
+
⊇ c
1
+c
2
is an external approximation and
c
(−)
+
⊆ c
1
+c
2
is an internal approximation.
As we shall see, it is not diﬃcult to observe that c
(+)
+
, c
(−)
+
are not unique. Therefore we
shall further describe a rather ”complete” parametrised variety of such ellipsoids.
Deﬁniton 2.1.5 Let E
0
denote a certain variety of ellipsoids c. An ellipsoid c
0
will be
inclusionmaximal relative to the variety E
0
, iftheinclusions c
0
∈ E
0
, c
0
⊆ c imply
the equality c
0
= c. Inclusionminimality is deﬁned similarly.
We will further indicate the inclusionminimal external approximations and the inclusion
maximal internal approximations of c
1
+c
2
.
The geometrical (Minkowski) diﬀerence. For two given ellipsoids c
1
, c
2
∈ comp IR
n
the
geometrical (internal) diﬀerence
c
1
˙
−c
2
= ¦x : x +c
2
⊆ c
1
¦
is unique. However it may not be an ellipsoid (give an example).
12
This representation is also true for degenerate matrices Q but then Q
−1
does not exist and has to be
substituted by the MoorePenrose pseudoinverse for Q, [120].
99
Deﬁniton 2.1.6 An external ellipsoidal estimate of the diﬀerence c
1
˙
−c
2
will be deﬁned
as an ellipsoid c
(+)
−
that satisﬁes the inclusion
c
(+)
−
⊇ c
1
˙
−c
2
,
while an internal ellipsoidal approximation of the diﬀerence c
1
˙
−c
2
is an ellipsoid c
(−)
−
that
satisﬁes the inclusion
c
(−)
−
+c
2
⊆ c
1
.
Obviously
c
(−)
−
⊆ c
1
˙
−c
2
.
We will further be interested in the inclusionmaximal internal and the inclusionminimal
external ellipsoidal approximations c
(−)
−
, c
˙
−c
2
of the diﬀerence c
1
˙
−c
2
.
As we shall observe in the sequel, the inclusionmaximal approximations c
(−)
−
and the
inclusionminimal approximations c
(+)
−
are not unique. They could also be interpreted as
nondominated elements of a (partially) ordered family. (It is now the family of sets in dR
n
and the ordering is inclusion). This interpretation naturally follows from Deﬁnition 2.1.5.
The Intersections. Given c
1
, c
2
∈ comp IR
n
, its intersection c
1
∩ c
2
in general is not an
ellipsoid. We will be interested ﬁrst of all in its external approximations
c
+
⊃ c
1
∩ c
2
seeking for example the inclusionminimal (nondominated) ellipsoids. We will again
discover that these are not unique and will try to describe a rather complete variety
of such ellipsoids.
A more diﬃcult problem is to ﬁnd an internal ellipsoidal approximation
c
−
⊂ c
1
∩ c
2
to the intersection. Indeed the intersection may easily turn to be a convex set of rather
general nature, namely a nonsymmetrical set relative to any point or plane or even a
degenerate convex set in the sense that if taken in IR
n
it will have no interior point.
A relatively simpler situation occurs when the centers of c
1
and c
2
coincide.
100
Aﬃne Transformations As an exercise one may easily check that the inclusion
x ∈ c(a, Q)
is equivalent to the following
Ax +b ∈ c(Aa + b, AQA
)
Let us now indicate two useful properties of symmetrical sets.
Lemma 2.1.4 Suppose that a set H ∈ comp IR
n
is symmetrical, namely H = −H. Then
(a) the inclusion H ⊂ c(a, Q) implies H ⊂ c(0, Q) and
(b) H ⊃ c(a, Q) implies H ⊃ c(0, Q).
Proof. Suppose H ⊂ c(a, Q), but the inclusion H ⊂ c(0, Q) does not hold. Then there
exists a set L ,= ∅ such that
L = ¦l ∈ IR
n
: ρ(l[H) > (l, Ql)
1
2
, l = 1¦.
Since H, c(0, Q) are symmetrical, the property
ρ(l[H) > (l, Ql)
1
2
will also hold for l ∈ −L.
Since H ⊂ c(a, Q), we have
ρ(l[H) < (a, l) + (l, Ql)
1
2
, l ∈ IR
n
.
Due to the previous supposition this implies
(a, l) > 0, l ∈ L (2.1.7)
and also the inequality
(a, l) > 0, l ∈ −L
which contradicts with (2.1.7). The assertion (a) is thus proved.
To prove assertion (b) we observe
ρ(l[H) ≥ (l, a) + (l, Ql)
1
2
, l ∈ IR
n
. (2.1.8)
101
Since H = −H we have ρ(l[H) = ρ(−l[H) for all l ∈ IR
n
. Therefore
ρ(l[H) ≥ (−l, a) + (l, Ql)
1
2
, l ∈ IR
n
.
Together with (2.1.8) the latter inequality yields
ρ(l[H) ≥ (l, Ql)
1
2
, l ∈ IR
n
.
The assertion is thus proved.
As we have already mentioned, our objective is to add and subtract ellipsoids ( in the
“geometrical” sense) and also to intersect them and to apply aﬃne transformations. The
results of these operations are convex sets which may either again be ellipsoids or what is
more common, may not be ellipsoidal at all. In the latter case we will introduce internal
and external ellipsoidal approximations of these sets. Out of all the possible approximating
ellipsoids we will prefer to select the inclusion minimal or maximal ellipsoids observing
that these “extremal” ellipsoids are the nondominated elements (relative to inclusion) of
the respective varieties.
Among the nondominated varieties of inclusionminimal or inclusionmaximal ellipsoids we
may then want to single out some individual elements that would be optimal relative to
some prescribed optimality criterion. We will therefore indicate a class Ψ = ¦ψ(c(a, Q))¦
of criteria functions ψ(c(a, Q)) that would be
(a) deﬁned on the set of all nondegenerate ellipsoids ¦c(a, Q)¦ and nonnegativevalued,
(b) monotonous by increasing with respect to inclusion:
ψ(c
1
) ≤ ψ(c
2
) if c
1
⊆ c
2
.
(We shall generally also require the monotonicity property (b) to be invariant relative to
aﬃne transformations of ellipsoids.)
Let Q stand for a symmetric positive matrix and ϕ
k
(Q), k = 1, . . . , n, for the coeﬃcient of
the (n −k)th degree term of its characteristic polynomial,
χ(λ) =
n
k=0
χ
k
λ
k
,
so that χ
k
= ϕ
n−k
(Q).
Let σ(Q) = ¦λ
1
, . . . , λ
n
¦ denote the set of eigenvalues of Q.
Lemma 2.1.5 Suppose c(a
1
, Q
1
) ⊇ c(a
2
, Q
2
).Then for all m ∈ N we have
102
(i) ϕ
k
(Q
m
1
) ≥ ϕ
k
(Q
m
2
), (k = 1, . . . , n),
(ii) max¦σ(Q
m
1
)¦ ≥ max¦σ(Q
m
2
)¦ ,
(iii) min¦σ(Q
m
1
)¦ ≥ min¦σ(Q
m
2
)¦ .
Proof. From Lemma 2.1.4 it follows that if c(a
1
, Q
1
) ⊇ c(a
2
, Q
2
), then
c(a
1
, Q
1
) ⊇ c(a
1
, Q
2
), c(0, Q
1
) ⊇ c(0, Q
2
).
The latter inclusion yields Q
1
≥ Q
2
.
Having two positive n n matrices Q
1
, Q
2
with respective eigenvalues
λ
(1)
1
≤ λ
(1)
2
≤ . . . ≤ λ
(1)
n
and
λ
(2)
1
≤ λ
(2)
2
≤ . . . ≤ λ
(2)
n
,
we observe that if Q
1
≥ Q
2
, then, [120],
λ
(1)
i
≥ λ
(2)
i
, i = 1, . . . , n.
The latter property yields the assertions of the Lemma.
Let us now indicate some common type measures
ψ[Q] = ψ(c(0, Q))
for the “size” of an ellipsoid c(0, Q)
(a) The trace:
ψ[Q] = tr(Q) = ϕ
n−1
(Q) = λ
1
+ . . . + λ
n
,
is actually the sum of the squares of the semiaxes of c(0, Q).
Given c(0, Q) = ¦x ∈ IR
n
: (Q
−1
x, x) ≤ 1¦, with support function ρ(l[c(0, Q)) =
(l, Ql)
1
2
, a canonic orthogonal transformation Tx = z ([T[ ,= 0) transforms c(0, Q)
into c(0, TQT
), where TQT
is diagonal, with diagonal elements λ
i
. (Here the
transformation Q → TQT
keeps the eigenvalues of TQT
the same as of Q and the
lengths of the semiaxes of c(0, TQT
) the same as of c(0, Q)). Thus
ρ(l[c(0, TQT
)) = (
n
i=1
λ
i
l
2
i
)
1
2
,
so that the length of the ith semiaxis of c(0, Q) is
ρ(e
(i)
[c(0, TQT
)) =
_
λ
i
,
where e
(i)
= (e
(i)
1
, . . . , e
(i)
n
), e
(i)
j
= δ
ij
, is the ith orth in the orthogonal coordinate
space of IR
n
.
Therefore tr(Q) is equal to the sum of the squares of the semiaxes of c(0, Q).
103
(b) The “trace of the square” yields a criteria
ψ[Q] = tr(Q
2
) = ϕ
n−1
(Q
2
).
(c) The product ψ[Q] = λ
1
λ
2
. . . λ
n
= ϕ
0
(Q) is proportional to the volume vol (c(0, Q))
of c(0, Q).
Indeed a direct calculation yields , [213],
vol (c(0, Q)) = Π
n
2
( det Q)(Γ(
n
2
+ 1)
−1
where Γ stands for the gammafunction. One just has to recall that the determinant
det Q of Q is equal to the product ϕ
0
(Q) = λ
1
. . . λ
n
.
(d) The diameter : ψ[Q] = d(c(0, Q)). Here the value
max¦λ
i
∈ IR : i = 1, . . . , n¦ =
_
d
2
_
2
,
where d = d(c(0, Q)) is the diameter of c(0, Q) ,so that d/2 is the radius of the
“smallest” ndimensional ball that includes c(0, Q).
This follows from the fact that d/2 is equal to the length of the largest semiaxis of
c(0, Q).
It is obvious that monotonous functions of these appearing in Lemma 2.1.5 as well as their
positive combinations are also monotonous with respect to inclusion. This indicates the
range of cases that we are able to handle. However we shall formulate our results primarily
for vol c(0, Q), tr(Q) and tr(Q
2
).
We shall now specify some parametrized varieties of ellipsoids that allow to approximate
the geometrical sums and diﬀerences of ellipsoids and even to give an exact representation
of these.
104
2.2 External Approximations: the Sums.
Internal Approximations: the
Diﬀerences
In this section we will deal only with nondegenerate ellipsoids. Given two such ellipsoids
c
1
= c(a, Q
1
) and c
2
= c(a, Q
2
), denote the roots of the equation
det(Q
1
−λQ
2
) = 0
as
λ
min
= λ
1
≤ λ
2
≤ . . . ≤ λ
n
= λ
max
, (λ
1
> 0, λ
n
< ∞).
These roots are also said to be the “relative eigenvalues” of the matrices Q
1
, Q
2
∈
L(IR
n
, IR
n
).
Consider a parametric family of matrices
Q(p) = (1 + p
−1
)Q
1
+ (1 + p)Q
2
.
We will also be interested in the family Q(−p).
Denote
+
= [λ
1/2
min
, λ
1/2
max
]
−
=
+
(1, λ
min
).
Lemma 2.2.1 (a) The ellipsoid c = c(a
1
+ a
2
, Q(p)), p > 0 is properly deﬁned and is
an external approximation of the sum c
1
+c
2
, i.e.
c
1
+c
2
⊂ c(a
1
+ a
2
, Q(p)) (2.2.1)
for any p > 0.
(b) With vector l ∈ IR
n
, l = 1, given, the equality
p = (Q
1
l, l)
1/2
(Q
2
l, l)
−1/2
(2.2.2)
deﬁnes a scalar parameter p ∈
+
, such that
ρ(l[c(a
1
+ a
2
, Q(p))) = ρ(l[c(a
1
, Q
1
) +c(a
2
, Q
2
)). (2.2.3)
Conversely, with parameter p ∈
+
given, there exists a vector l ∈ IR
n
, l = 1, such that
equalities (2.2.2) and (2.2.3) are true.
105
Proof. The inequality
p
−1
(Q
1
l, l) + p(Q
2
l, l) ≥ 2(Q
1
l, l)
1/2
(Q
2
l, l)
1/2
is obviously true for any p > 0.
Adding (Q
1
l, l) + (Q
2
l, l) to both sides, we obtain
(Q(p)l, l)
1/2
≥ (Q
1
l, l)
1/2
+ (Q
2
l, l)
1/2
(2.2.4)
where Q(p) > 0 for any p > 0.
With a further addition of (l, a
1
+a
2
) to both sides this implies
ρ(l[c(a
1
+a
2
, Q(p))) ≥ ρ(l[c(a
1
, Q
1
)) + ρ(l[c(a
2
, Q
2
))
for any l ∈ IR
n
and therefore, implies the inclusion (2.1.2).
To prove the assertion (b), with l ∈ IR
n
given, we select the parameter p due to (2.2.2),
observing that p ∈
+
(check the latter inclusion as an exercise, using the extremal prop
erties of matrix eigenvalues, see e.g. [120]). After a substitution of (2.2.2) into (2.2.4), the
latter turns into an equality for the given l (this can be veriﬁed through direct calculation).
The equality (2.2.3) is therefore true for any given l with p and l related through (2.2.2).
Conversely, with p ∈
+
given, there exists a vector l ∈ IR
n
, l = 1, such that (2.2.2) and
therefore (2.2.3) do hold. This follows from Theorem 7.10, Chapter X of reference [120],
due to the continuity in l of the righthand side of (2.2.2).
Q.E.D.
A similar reasoning passes through for geometrical diﬀerences.
Lemma 2.2.2 Suppose int c(0, Q
1
) ⊃ c(0, Q
2
). Then
(a) c = c(a
1
−a
2
, Q(−p)) is a nondegenerate ellipsoid if and only if
p ∈ (1, λ
min
). (2.2.5)
For these values of p the ellipsoid c is an internal approximation of the diﬀerence
c
1
˙
− c
2
, i.e.
(b)
c(a
1
−a
2
, Q(−p)) ⊆ c
1
˙
− c
2
.
106
Proof. Consider the inclusion
int c(0, Q
1
) ⊃ c(0, Q
2
),
which implies
(l, Q
1
l) > (l, Q
2
l)
for any l ∈ IR
n
and therefore implies the condition
p =
(l, Q
1
l)
1/2
(l, Q
2
l)
1/2
> 1.
Taking
Q(−p) = (1 −p
−1
)Q
1
+ (1 −p)Q
2
= (p −1)(Q
1
p
−1
−Q
2
),
we observe, in view of the condition p > 1 that Q(−p) is positive deﬁnite if and only if
Q
1
p
−1
−Q
2
> 0, which means p < λ
min
. This yields p ∈ (1, λ
min
).
Following the proof along a scheme similar to that of Lemma 2.2.1 with p substituted by
(−p), we come to the inequality
¸Q(−p)l, l)
1/2
≤ ¸Q
1
l, l)
1/2
−¸Q
2
l, l)
1/2
, (2.2.6)
which is true for any l ∈ IR
n
and equivalent to
ρ(l[c(a
1
−a
2
, Q(−p))) ≤ ρ(l[c(a
1
, Q
1
)) −ρ(l[c(a
2
, Q
2
)).
The latter inequality is further equivalent to the inclusion
c(a
1
−a
2
, Q(−p)) +c(a
2
, Q
2
) ⊆ c(a
1
, Q
1
)
which, due to the deﬁnition of the geometrical diﬀerence and the conditions of the lemma,
implies
c(a
1
−a
2
, Q(−p)) ⊆ c(a
1
, Q
1
)
˙
− c(a
2
, Q
2
) (2.2.7)
for any p ∈ (1, λ
min
). As shown above, the latter condition ensures that Q(−p) > 0.
To prove assertion (b) with l ∈ IR
n
given, we suppose the parameter p to be deﬁned due
to (2.2.2) and such that p ∈
−
. Under these conditions a direct substitution of p into
(2.2.6) turns the latter into an equality. The inclusion (2.2.8) together with the relation
ρ(l[c(a
1
, Q
1
)
˙
− c(a
2
, Q
2
)) ≤ ρ(l[c(a
1
, Q
1
)) −ρ(l[c(a
2
, Q
2
))
then yields equality (2.2.6) for the given values of l and p.
On the other hand, once p ∈
−
is given, there exists a vector l ∈ IR
n
, such that equality
(2.2.2) is fulﬁlled, (due to Theorem 7.10, Chapter X of [120], and the continuity of the
righthand side of (2.2.2.) in l).
107
This also yields (2.2.6) for the given p and l.
Q.E.D.
Consider a positive deﬁnite, symmetric matrix C with elements ¦c
ij
¦ where i stands for
the row and j for the column of C. Also assume the symbol 1, 0 = ∅.
Lemma 2.2.3 Fix a vector l ∈ IR
n
, l = 1 and suppose that for some m ∈ [0, n], we have
lj = 0 if j ∈ 1, m (2.2.8)
lj ,= 0 if j = m + 1, n
Suppose in addition that c
1
= c(0, Q
1
), c
2
= c(0, Q
2
) and that the matrices Q
1
, Q
2
, Q are
diagonal. Then the following implications hold:
(a) If
c(0, Q) ⊆ c(0, C) ⊆ c
1
+c
2
and
ρ(l[c(0, Q)) = ρ(l[c
1
+c
2
)
then
c
ij
= 0 for all i ,= j, i ∈ m + 1, n
(b) If
c(0, Q) ⊆ c(0, C) ⊆ c
1
˙
− c
2
and
ρ(l[c(0, Q)) = ρ(l[c
1
˙
− c
2
)
then
c
ij
= 0 for all i ,= j, i ∈ m + 1, n
.
Proof. We will now prove the assertion (a). Assertion (b) will be left as an exercise. Its
proof is similar to that of (a).
For the given vector l deﬁne an array of vectors h
(k)
, k ∈ m + 1, n, where
h
(k)
j
=
_
l
j
if j ,= k
−l
j
if j = k.
It can be checked that vectors h
(k)
are linearly independent.
108
By diagonality of the respective matrices and the equality of supports we have
(Qh
(k)
, h
(k)
)
1/2
= (Q
1
h
(k)
, h
(k)
)
1/2
+ (Q
2
h
(k)
, h
(k)
)
1/2
. (2.2.9)
Combined with the inclusion relations this implies
(Qh
(k)
, h
(k)
)
1/2
= (C h
(k)
, h
(k)
)
1/2
. (2.2.10)
Assuming
Q = diag ¦q
11
, . . . , q
nn
¦,
Q
j
= diag ¦q
(j)
11
, . . . , q
(j)
nn
¦, j = 1, 2.
deﬁne the function χ : IR
n
→ IR by the equality
χ(z) =
n
i=1
(q
ii
−c
ii
)z
2
i
.
Here we have for all k ∈ m + 1, n
χ(l) = χ(h
(k)
) (2.2.11)
and due to (2.2.11) we come to
χ(l) =
n
i,j=m+1;i=j
c
ij
l
i
l
j
. (2.2.12)
Substituting l for l
(k)
with a ﬁxed value of k into (2.2.12), (2.2.13) and taking into account
the symmetry of C, we have
2
n
j=1;j=k
c
kj
l
k
l
j
= 0.
The respective terms may now be cancelled out from the righthand side of (2.2.12) for the
respective value of k. Taking equation (2.2.12) in its reduced form we can cancel out similar
terms for a new value k
∗
,= k. Repeating this procedure for all the values k
∗
∈ m + 1, n
except for a previously ﬁxed pair r, s, r ,= s, r, s ∈ m + 1, n, we ﬁnally come to
χ(l) = 2c
rs
l
r
l
s
,
so that the “last” cancellation yields χ(l) = 0. This directly implies c
rs
,= 0. Since r, s
were chosen arbitrarily, what follows is that c
rs
= 0, for any r, s ∈ m + 1, n; r ,= s. The
proof is therefore complete for m = 0.
If m > 0, then take
φ(l) =
1
2
(ρ
2
(l[c(0, Q)) −ρ
2
(l[c(0, C))). (2.2.13)
109
The function ϕ(l) has a local minimum at l = h
(k)
for any k ∈ m + 1, n.
By diﬀerentiability we necessarily have
∂ϕ(l)
∂l
¸
¸
¸
¸
l=h
(k)
= 0.
For all the values of i ∈ 1, m, k ∈ m + 1, n this yields
n
j=m+1
c
ij
h
(k)
j
= 0
and since the vectors h
(k)
, k ∈ m + 1, n are linearly independent we conclude that c
ij
= 0
for any i ∈ 1, m, j ∈ m + 1, n.
Q.E.D.
Lemma 2.2.4 Consider an ellipsoid c(0, C) together with ellipsoids c
1
= c(0, Q
1
), c
2
=
c(0, Q
2
), assuming that Q
1
, Q
2
are diagonal. Further assume the vector l ∈ IR
n
, l = 1 to
be given and the parameter p to be deﬁned due to relation (2.2.2) with given l.
Then the following implications hold:
If
c(0, Q(p)) ⊇ c(0, C) ⊇ c
1
+c
2
and
ρ(l[c(0, Q(p))) = ρ(l[c
1
+c
2
),
then
c(0, Q(p)) = c(0, C) and p ∈
+
.
Proof. Denote Q
j
= diag ¦q
(j)
11
, . . . , q
(j)
nn
¦, j = 1, 2, Q(p) = diag ¦q
n
, . . . , q
nn
¦, C =
¦c
ij
¦. Also keep the notation of (2.2.9) and the deﬁnition of h
(k)
of the previous theorem.
Due to the previous Lemma 2.2.3 (a) and the inclusion c(0, Q(p)) ⊆ c(0, C), we have
c
ii
≤ q
ii
, i ∈ m + 1, n. (2.2.14)
The equality (2.2.3) with value of p from (2.2.2) yields
n
i=m+1
q
ii
l
2
i
=
n
i=m+1
c
ii
l
2
i
,
where l
i
,= 0. Together with (2.2.14) this implies c
ii
= q
i
for i ∈ m + 1, n.
110
By the conditions of the lemma both nonnegative functions
ξ(l) = ρ(l[c(0, Q(p))) −ρ(l[c(0, C))
and
ζ(l) = ρ(l[c(0, C)) −ρ(l[c
1
+c
2
)
have local minima at l = l
(k)
, k ∈ m + 1, n with ξ(l
(k)
) = 0, ζ(l
(k)
) = 0. Due to the
diﬀerentiability of these functions the second order necessary conditions of optimality imply
that the matrices of the second order partial derivatives are nonnegative, namely
_
∂
∂l
_
2
ξ(l)
¸
¸
¸
¸
l=l
(k)
≥ 0
and
_
∂
∂l
_
2
ζ(l)
¸
¸
¸
¸
l=l
(k)
≥ 0.
In particular, this implies that the diagonal elements of the respective matrices are non
negative, or, after a direct calculation,
q
ii
(Q(p)l
(k)
, l
(k)
)
1/2
≥
c
ii
(Cl
(k)
, l
(k)
)
1/2
−
_
n
j=m+1
c
ij
l
(k)
j
_
2
(Cl
(k)
, l
(k)
)
3/2
for all l = 1, m, k ∈ m + 1, n. Here the second term on the righthand side disappears due
to Lemma 2.2.3. If we now observe that the denominators in the resulting inequality are
equal (due to the condition of the lemma and being the values of the support functions at
l = l
(k)
) we may conclude that q
ii
≥ c
ii
for all i ∈ 1, m.
Due to Lemma 2.2.3 and the deﬁnition of l
(k)
a similar condition for the matrix of second
derivatives of ζ yields the following inequality for the diagonal elements i = 1, m:
c
ii
(Cl
(k)
, l
(k)
)
1/2
≥
q
(1)
ii
(Q
1
l
(k)
, l
(k)
)
1/2
+
q
(2)
ii
(Q
2
l
(k)
, l
(k)
)
1/2
(2.2.15)
Take the righthand side of (2.2.16), multiply and divide it by (Q
1
l
(k)
, l
(k)
)
1/2
+
(Q
2
l
(k)
, l
(k)
)
1/2
, recalling that
Q(p) = (1 + p
−1
)Q
(1)
+ (1 + p)Q
(2)
,
where
p = (l, Q
(1)
l)
1/2
(l, Q
(2)
l)
−1/2
, l = l
(k)
,
substituting the obtained relations into (2.2.15) and using the equality relation for the
support functions at l = l
(k)
, we come to the condition q
ii
≤ c
ii
and therefore to the
equality q
ii
= c
ii
, i ∈ 1, m.
111
The inclusion c(0, Q(p)) ⊇ c(0, C) implies that the matrix Q(p) −C is nonnegative. Since
it was just established that its diagonal elements are all equal to zero, what follows is that
all the rest of the elements must also be zero. The theorem is thus proved.
However, before proving an assertion similar to Lemma 2.2.4, but for the diﬀerences c
1
˙
−c
2
,
we will ﬁrst prove the following essential
Theorem 2.2.1 Suppose that for the ellipsoids c
1
= c(a
1
, Q
1
), c
2
= c(a
2
, Q
2
) the matri
ces Q
1
, Q
2
are positive deﬁnite and that Q(p) is deﬁned due to formula (2.2.1).
Then the set of inclusionminimal external estimates of the sum c
1
+c
2
will consist of the
ellipsoids of the form c(a
1
+ a
2
, Q(p)), with p ∈
+
.
Proof. Without loss of generality, referring also to Lemma 2.2.4, we may assume all the
centers of the ellipsoids considered here to be zero, particularly a
1
= a
2
= 0.
Given an ellipsoid c(0, Q) ⊇ c
1
+ c
2
let us indicate that there exists a value p such that
the ellipsoid c(0, Q(p)) could be “squeezed” in between c(0, Q) and c
1
+ c
2
, so that we
would have
c
1
+c
2
⊆ c(0, Q(p)) ⊆ c(0, Q).
We may obviously consider c(0, Q) to be tangential to c
1
+ c
2
, assuming the existence of
a vector l =
¯
l ∈ IR
n
, 
¯
l = 1, such that
ρ(
¯
l[c(0, Q)) = ρ(
¯
l[c
1
+c
2
). (2.2.16)
Let us now select an invertible matrix T such that the matrices Q
∗
1
= T
Q
1
T, Q
∗
2
= T
Q
2
T
would both be diagonal. The existence of such a transformation T follows from results in
Linear Algebra and the theory of matrices (see e.g. [120] ).
The transformation T obviously does not violate the inclusion c(0, Q) ⊇ c
1
+ c
2
, so that
with Q
∗
= T
QT we still have
c(0, Q
∗
) ⊇ c(0, Q
∗
1
) +c(0, Q
∗
2
).
Taking the mapping
¯
l = Tz one may transform the equality (2.2.17), which is
(
¯
l, Q
¯
l)
1/2
= (
¯
l, Q
∗
1
¯
l)
1/2
+ (
¯
l, Q
∗
2
¯
l)
1/2
into
(¯ z, Q
∗
¯ z)
1/2
= (z, Q
∗
1
z)
1/2
+ (z, Q
∗
2
z)
1/2
112
where ¯ z = T
−1
¯
l. Following (2.2.2) we may now select
¯ p = (¯ z, Q
∗
1
¯ z)
1/2
(¯ z, Q
∗
2
¯ z)
−1/2
and further take
Q
∗
(¯ p) = (1 + ¯ p
−1
)Q
∗
1
+ (1 + ¯ p)Q
∗
2
.
We then come to the relations
c(0, Q
∗
1
) +c(0, Q
∗
2
) ⊆ c(0, Q
∗
(¯ p))
ρ(¯ z[c(0, Q
∗
1
)) + ρ(¯ z[c(0, Q
∗
2
)) (2.2.17)
= ρ(¯ z[c(0, Q
∗
(¯ p))) = ρ(¯ z[c(0, Q
∗
)).
From Lemma 2.1.4, part (a) it now follows that c(0, Q
∗
(¯ p)) ⊆ c(0, Q
∗
).
Indeed with if this inclusion being false, there would have existed a vector z
∗
such that
ρ(z
∗
[c(0, Q
∗
(p))) > ρ(z
∗
[c(0, Q
∗
)). (2.2.18)
The vector z
∗
is obviously noncollinear with ¯ z. Deﬁne Z to be the 2dimensional space
generated by ¯ z, z
∗
and c
z
(0, Q) to be the projection of the ellipsoid c(0, Q) on space Z. In
view of (2.2.18) and the inequality (2.2.19) we have
c
z
(0, Q
∗
(¯ p)) ⊇ c
z
(0, Q
∗
)
and
ρ(¯ z[c
z
(0, Q
∗
)) = ρ(¯ z[c
z
(0, Q
∗
(p)))
= ρ(¯ z[c
z
(0, Q
∗
1
) +c
z
(0, Q
∗
2
)).
From Lemma 2.2.4 it then follows that c
z
(0, Q
∗
) = c
z
(0, Q
∗
(p)) is in contradiction with
(2.2.18).
Q.E.D.
The following proposition is similar to Lemma 2.2.4, but is applied to geometrical diﬀer
ences.
Lemma 2.2.5 Consider a nondegenerate ellipsoid c(0, C) together with ellipsoids c
1
=
c(0, Q
1
), c
2
= c(0, Q
2
), assuming that Q
1
, Q
2
are diagonal. Further assume the vector
l =
¯
l, l = 1 to be given and the parameter p = ¯ p to be deﬁned due to relation (2.2.2),
l =
¯
l.
113
Then, if the ellipsoid c(0, Q(−¯ p)) is properly deﬁned (Q(−¯ p) > 0), the relations
c(0, Q(−¯ p)) ⊆ c(0, C) ⊆ c
1
˙
−c
2
(2.2.19)
ρ(
¯
l[c(0, Q(−¯ p)) = ρ(
¯
l[c
1
˙
−c
2
) (2.2.20)
imply
c(0, Q(−¯ p)) = c(0, C) and ¯ p ∈
−
.
Proof. Let us start with the indication that the inclusion c(0, C) ⊆ c
1
˙
−c
2
implies the
existence of an ε > 0 such that
cl ≤ (l, Cl)
1/2
≤ (l, Q
1
l)
1/2
−(l, Q
2
l)
1/2
and therefore
¯ p = (
¯
l, Q
1
¯
l)
1/2
(
¯
l, Q
2
¯
l)
−1/2
> 1.
Let us further proceed with all the formal procedures, presuming also
¯ p < λ
min
= min¦(l, Q
1
l)(l, Q
2
l)
−1
∈ IR : l = 1¦
so that altogether ¯ p ∈ (1, λ
min
) and therefore Q(−¯ p) > 0.
Suppose that ¯ p,
¯
l is the pair given in the formulation of the theorem. Then for this pair the
relations (2.2.20), (2.2.21) are fulﬁlled and the ﬁrst of them is equivalent to the inequalities
ρ(l[Q(−¯ p)) ≤ ρ(l[c(0, C)) ≤ ρ(l[c
1
) −ρ(l[c
2
) ≤ ρ(l[c
1
˙
−c
2
)
for any l ∈ IR
n
, and consequently
ρ(l[Q(−¯ p)) ≤ ρ(l[c(0, C)) ≤ ρ(l[c
1
˙
−c
2
).
Due to (2.2.21) this yields
ρ(
¯
l[c(0, Q(−¯ p)) + ρ(
¯
l[c
2
) = ρ(
¯
l[c
1
)
and
ρ(
¯
l[c(0, C)) + ρ(
¯
l[c
2
) = ρ(
¯
l, c
1
). (2.2.21)
Further on, the inclusion c(0, C) ⊆ c
1
˙
−c
2
implies
c(0, C) +c
2
⊆ c
1
.
By Lemma 2.2.1 there exists an ellipsoid c(0, C(p)) with
C(p) = (1 + p
−1
)C + (1 + p)Q
2
(2.2.22)
114
which satisﬁes the inclusion
c(0, C) +c
2
⊆ c(0, C(p))
for any p > 0. With l =
¯
l given and p = p
∗
taken as
p
∗
= (C
¯
l,
¯
l)
1/2
(Q
2
¯
l,
¯
l)
−1/2
(2.2.23)
it also satisﬁes the equality
ρ(
¯
l[c(0, C(p
∗
))) = ρ(
¯
l[c(0, C)) + ρ(
¯
l[c
2
). (2.2.24)
According to Theorem 2.2.1 we then have
c(0, C) +c
2
⊆ c(0, C(¯ p)) ⊆ c
1
(2.2.25)
and therefore
c(0, C) ⊆ c(0, C(¯ p))
˙
−c
2
⊆ c
1
˙
−c
2
.
Rearranging (2.2.23) and taking p
0
= 1 + p
∗
, we obtain
C = (1 + (p
0
)
−1
)C(¯ p) + (1 −p
0
)Q
2
. (2.2.26)
Being deﬁned through (2.2.24) the value p
∗
is positive (as C > 0) and p
0
> 1. From
(2.2.26) we also observe C(¯ p) ≤ Q
1
and due to (2.2.22), (2.2.25) we have
p
0
= p
∗
+ 1 = ((C
¯
l,
¯
l)
1/2
+ (Q
2
¯
l,
¯
l)
1/2
)(Q
2
¯
l,
¯
l)
−1/2
= ((C(¯ p),
¯
l,
¯
l)
1/2
)(Q
2
¯
l,
¯
l)
1/2
≤ (Q
1
¯
l,
¯
l)
1/2
(Q
2
¯
l,
¯
l)
−1/2
= ¯ p
so that p
0
= ¯ p.
Combining (2.2.26), (2.2.27) we come to the inclusion
C ⊆ Q(−¯ p) = (1 + ¯ p
−1
)Q
1
+ (1 − ¯ p)Q
2
which, together with (2.2.20) produces
C = Q(−¯ p).
Since C is nondegenerate, we have Q(−¯ p) > 0 and therefore ¯ p indeed lies within the domain
p ∈ (1, λ
min
).
Q.E.D.
Let us now prove the analogy of Theorem 2.2.1 for geometrical diﬀerences.
115
Theorem 2.2.2 Suppose that int c(0, Q
1
) ⊇ c(0, Q
2
) holds. Then the set of maximal
internal estimates of the diﬀerence c
1
˙
−c
2
consists of ellipsoids of the form
c(a
1
−a
2
, Q(−p)), p ∈
−
.
.
Proof. Without loss of generality suppose again that a
1
= 0, a
2
= 0. Given a nondegen
erate ellipsoid c(0, Q) ⊆ c
1
˙
−c
2
let us indicate that there exists a value p such that the
ellipsoid c(0, Q(−p)) could be “squeezed” in between c(0, Q) and c
1
÷c
2
, so that we would
have
c(0, Q) ⊆ c(0, Q(−p)) ⊆ c
1
÷c
2
. (2.2.27)
We may consider c(0, Q) to be tangential to c
1
˙
−c
2
, assuming the existence of a vector
¯
l ∈ IR
n
, 
¯
l = 1, such that
ρ(
¯
l[c(0, Q)) = ρ(
¯
l[c
1
˙
−c
2
). (2.2.28)
Let us ﬁrst deﬁne c(0, Q(−¯ p)) with
p = ¯ p = (Q,
¯
l,
¯
l)
1/2
(Q
2
¯
l,
¯
l)
−1/2
and prove that Q(−¯ p) is positive deﬁnite. To do this, deﬁne the matrix
D(p
∗
) = (1 + (p
∗
)
−1
)Q + (1 + p
∗
)Q
2
or
Q = (1 −p
−1
D(p
∗
) + (1 −p)Q
2
(2.2.29)
where
p
∗
= (
¯
l, Q
¯
l)
1/2
(
¯
l, Q
2
¯
l)
−1/2
and
p = p
∗
+ 1.
By (2.2.29) we have
ρ(
¯
l[c(0, Q)) + ρ(
¯
l[c
2
) = ρ(
¯
l[c
1
) (2.2.30)
or
(
¯
l, Q
¯
l)
1/2
+ (
¯
l, Q
2
¯
l)
1/2
= (
¯
l, Q
1
¯
l)
1/2
so that
p = p
∗
+ 1 = ((
¯
l, Q
¯
l)
1/2
+ (
¯
l, Q
2
¯
l)
1/2
)(
¯
l, Q
2
¯
l)
−1/2
= (
¯
l, Q
1
¯
l)
1/2
(
¯
l, Q
2
¯
l)
−1/2
= ¯ p.
116
Due to Theorem 2.2.1, we further observe
c(0, Q) +c
2
⊆ c(0, D(p
∗
)) ⊆ c
1
or, in other words, that D(p
∗
) ≤ Q
1
.
Together with (2.2.30) and the equality p = ¯ p this yields Q = Q(−¯ p), also proving that
Q(−¯ p) > 0, which means ¯ p ∈ (1, λ
min
). Following Lemma 2.2.2(a) we come to the desired
inclusion (2.2.8). Q.E.D.
To conclude this section we shall summarize its results in the following
Theorem 2.2.3 Given nondegenerate ellipsoids c
1
, c
2
, the following relations are true
c
1
+c
2
= ∩¦c(a
1
+a
2
, Q(p)) : p ∈
+
¦, (2.2.31)
and with int c
1
⊇ c
2
,
c
1
˙
−c
2
=
_
¦c(a
1
−a
2
, Q(−p)) : p ∈
−
¦, (2.2.32)
where
¯
Q stands for the closure of set Q.
Proof. It is clearly suﬃcient to prove the theorem for a
1
= a
2
= 0.
From Lemma 2.2.1 it follows
c
1
+c
2
⊆ ∩¦c(0, Q(p)) : p ∈
+
¦.
To prove the exact equality, assume the existence of a point x
∗
such that
x
∗
∈ ∩¦c(0, Q(p)) : p ∈
+
¦, (2.2.33)
x ,∈ c
1
+c
2
. (2.2.34)
The last condition ensures the existence of a vector l = l
∗
that yields
(l
∗
, x
∗
) > ρ(l
∗
[c
1
+c
2
). (2.2.35)
Selecting
p = p
∗
= (l
∗
, Q
1
l
∗
)
1/2
(l
∗
, Q
2
l
∗
)
−1/2
and following Lemma 2.2.1 (a), we come to
ρ(l
∗
[c
1
+c
2
) = ρ(l
∗
[c(0, Q(p
∗
)).
117
Together with (2.2.36) this implies x
∗
,∈ c(0, Q(p
∗
)) in contradiction with (2.2.34). The
equality (2.2.32) is thus proved.
To prove (2.2.33), we recall that int (c
1
˙
−c
2
) ,= ∅ and follow Lemma 2.2.2 which immedi
ately yields
_
¦c(0, Q(−p)) : p ∈
−
¦ ⊆ c
1
˙
−c
2
.
To indicate that there is actually an exact equality, assume the existence of such a vector
x
∗
that
x
∗
∈ int (c
1
˙
−c
2
), (2.2.36)
x
∗
,∈
_
¦c(0, Q(−p)) : p ∈
−
¦. (2.2.37)
Since x
∗
∈ int (c
1
˙
−c
2
) there exists an ε > 0 for which
o
ε
(x
∗
) = ¦(x −x
∗
, x −x
∗
) ≤ ε
2
¦ ⊆ int (c
1
˙
−c
2
).
As c
1
= −c
1
, c
2
= −c
2
(these sets are symmetrical around the origin), we obviously have
c
1
˙
−c
2
= −(c
1
˙
−c
2
) and therefore the whole set
o = ¦x : x ∈ o
ε
(z), z = −x
∗
+ 2αx
∗
, α ∈ [0, 1]¦
satisﬁes o ⊆ int (c
1
˙
−c
2
). What follows is that there exists a nondegenerate ellipsoid
c(0, C
∗
) ⊆ o ⊆ int (c
1
˙
−c
2
). (Give an example by explicit calculation of C
∗
, assuming set
A
∗
= ¦z : z = −x
∗
+ 2αx
∗
, α ∈ [0, 1]¦
to be its largest axis.) From Theorem 2.2.2 it now follows that for some p
∗
∈
−
there
exists an ellipsoid c(0, Q(−p
∗
)) that satisﬁes x
∗
⊆ c(0, C) ⊆ c(0, Q(−p)) ∈ c
1
˙
−c
2
in
contradiction with (2.2.37), (2.2.38).
Q.E.D.
Theorem 2.2.3 may be illustrated on a 2dimensional example. In the center of Fig.2.2
(a) we see two ellipsoids whose sum is the nonellipsoidal set that is the intersection of
the ”nondominated” (inclusionminimal) ellipsoids that approximate it externally and are
constructed due to formula (2.2.31). Fig.2.2 (b) shows a nondegenerate geometric diﬀerence
of two ellipsoids ( the set with two kinks) that also arrives as the ( closure of the) union
of the nondominated (inclusionmaximal) ellipsoids that approximate it internally and are
constructed due to formula (2.2.32). In both examples the parameters p ∈ Π
+
, S ∈ Σ are
chosen randomly but give a good illustration of the nature of the approximations.
118
———— !!! insert Figures 2.2.1 (a) and (b) !!! ——————–
2.3 Internal Approximations: the Sums.
External Approximations: the
Diﬀerences
We shall now introduce a representation that will allow an internal approximation of the
sum of two nondegenerate ellipsoids by a parametrized variety of ellipsoids, and an external
approximation of the geometrical diﬀerence of these. It will be demonstrated that this
approximation may be exact.
Given c
1
= c(a
1
, Q
1
), c
2
= c(a
2
, Q
2
), where Q
1
> 0, Q
2
> 0, we introduce a parametric
family of matrices Q
+
[S], where
Q
+
[S] = S
−1
[(SQ
1
S
)
1/2
+ (SQ
2
S
)
1/2
]
2
S
−1
(2.3.1)
and S ∈ Σ with
Σ = ¦S ∈ L(IR
n
, IR
n
) : S
= S, [S[ , = 0¦.
The matrix S is therefore selected from the set Σ of symmetrical nondegenerate matrices.
In a similar way we deﬁne the variety
Q
−
[S] = S
−1
[(SQ
1
S
)
1/2
−(SQ
2
S)
1/2
]
2
S
−1
(2.3.2)
with S ∈ Σ.
The variety Q
+
[S] will be used for approximating the sums c
1
+c
2
(internally), while the
variety Q
−
[S] for approximating the diﬀerences c
1
÷c
2
(externally).
Let us start from the ﬁrst case
119
Lemma 2.3.1 The ellipsoid c = c(a
1
+ a
2
, Q
+
[S]) is an internal approximation of the
(Minkowski) sum c
1
+c
2
, namely, for any S ∈ Σ, one has
c[S] = c(a
1
+ a
2
, Q
+
[S]) ⊆ c
1
+c
2
. (2.3.3)
For each S ∈ Σ there exists a vector l = l
∗
, l = 1, such that the equality
ρ (l[c(a
1
+ a
2
, Q
+
[S])) = ρ(l[c
1
) + ρ(l[c
2
) (2.3.4)
is true with l = l
∗
. Conversely, for any l ∈ IR
n
, l = 1, there exists a matrix S
∗
∈ Σ such
that (2.3.4) is true with S = S
∗
.
Proof. As in the previous sections it is clearly suﬃcient to consider the case, when
a
1
= 0, a
2
= 0.
For any matrix S ∈ L(IR
n
, IR
n
) we have
(ρ(l[c[S]))
2
= (l, Q
+
[S]l) =
= (l, Q
1
l) + (l, Q
2
l) +
+ 2
_
(SQ
1
S
)
1/2
S
−1
l, (SQ
2
S
)
1/2
S
−1
l)
_
.
By H¨older’s inequality
(ρ(l[c[S]))
2
≤ (l, Q
1
l) + (l, Q
2
l) +
+ 2
_
(SQ
1
S
)
1/2
S
−1
l, (SQ
1
S
)
1/2
S
−1
_
1/2
_
(SQ
2
S
)
1/2
S
−1
l, (SQ
2
S
)
1/2
S
−1
l
_
1/2
or
(ρ(l[c[S]))
2
≤ (l, Q
1
l) + (l, Q
2
l) + 2(l, Q
1
l)
1/2
(l, Q
2
l)
1/2
which proves the inclusion (2.3.3).
To prove that for a given S there exists an l = l
∗
that ensures the equality (2.3.4) we
observe, by direct substitution, that this would be possible if there existed a number λ > 0
and a vector l = l
∗
,
l
∗
 = 1, that would ensure the relation
_
(SQ
1
S
)
1/2
−λ(SQ
2
S
)
1/2
_
S
−1
l
∗
= 0. (2.3.5)
Denote
D = (Q
−1/2
2
Q
1
Q
−1/2
2
)
1/2
, z = Q
−1/2
2
l
∗
, T = SQ
1/2
2
120
then (2.3.5) reduces to
(TDDT
)
1/2
T
−1
z = λ(TT
)
1/2
T
−1
z.
Suppose, in addition that the matrix T is symmetrical. Then the last relation takes the
form
(TD DT)
1/2
T
−1
z = λz. (2.3.6)
Taking the polar decomposition ( [120]), of the matrix TD, we obtain an orthogonal matrix
and a symmetrical (here also nonsingular) matrix H such that
TD = UH. (2.3.7)
The condition of symmetricity for H means that
TDU
−1
= UDT. (2.3.8)
Substituting (2.3.7) into (2.3.6), we ﬁnally transform the original equation (2.3.4) to
U Dz = λz. (2.3.9)
We now have to solve the system (2.3.8), (2.3.9) for λ ∈ IR, U ∈ L(IR
n
, IR
n
) orthogonal
and T ∈ L(IR
n
, IR
n
) symmetrical and nonsingular, where the symmetrical, positive deﬁnite
matrix D ∈ L(IR
n
, IR
n
) and the nonzero vector z ∈ IR
n
are given in advance.
With vector Dz ,= 0 given, there obviously exists an orthogonal matrix U such that vector
UDz is directed along with z, hence there is a λ > 0 that ensures (2.3.9) with the U
selected as above. What remains is to ﬁnd a solution to the equation
TB
= BT
for symmetrical and nonsingular T where
B = UD.
This can be done by using a well known result of matrix theory. The proof given in Chapter
VIII of ( [120]) needs a slight modiﬁcation which we leave to the reader.
For matrix S deﬁned by the solution T obtained this way, equation (2.3.5) will hold. Using
the polar decomposition theorem, we ﬁnd an orthogonal matrix 0 such that S
∗
= OS is
symmetrical and therefore S
∗
∈
. From formula (2.3.1) it follows that
Q[S] = Q[OS]
and in such a way also that (2.3.5) is valid for S = S
∗
. Q.E.D.
The proof of Lemma 2.3.1 implies
121
Corollary 2.3.1 The following equality is true
c
1
+c
2
= ∪¦c(a
1
+a
2
, Q
+
[S][S ∈
¦.
The next step is to prove that the ellipsoids c(0, Q
+
[S]), c(0, Q
−
[S]) are the inclusion
maximal internal and the inclusionminimal external estimates for c
1
+ c
2
and c
1
÷ c
2
respectively.
Theorem 2.3.1 Consider the parametrized varieties of ellipsoids c(a
1
+a
2
, Q
+
[S]), c(a
1
−
a
2
, Q
−
[S]), S ∈ Σ, generated respectively by the varieties of matrices Q
+
[S], Q
−
[S], due to
(2.3.1), (2.3.2). Then the following assertions are true
(a) the set of inclusionmaximal internal estimates of the sum c
1
+c
2
consists of ellipsoids
of the form c(a
1
+ a
2
, Q
+
[S]) where S ∈ Σ.
(b) Assuming int c
1
⊇ c
2
, the set of inclusionminimal external estimates of the diﬀerence
c
1
÷c
2
consists of the ellipsoids of the form c(a
1
−a
2
, Q
−
[S]), S ∈ Σ.
Proof. As previously, we assume a
1
= a
2
.
In order to prove the maximality of c(0, Q
+
[S]) we shall demonstrate that for any ellipsoid
c(0, Q) the inclusions
c(0, Q
+
[S]) ⊆ c(0, Q) ⊆ c
1
+c
2
imply
Q
+
[S] = Q.
According to Lemma 2.3.1 there is a condition that there exists a vector ∈ IR
n
,  = 1,
such that (2.3.4) is true. This is actually (2.3.6) (using the notations of the lemma),
where the matrix (TDDT)
1/2
is positive deﬁnite and symmetrical and the matrix T
−1
is
symmetrical. It is left to the reader to prove as an exercise that, under the above conditions,
their product has simple structure, namely a complete set ¦z
i
∈ IR
n
: i = 1, n¦ of linearly
independent eigenvectors (that are not necessarily orthogonal). From this it follows that
there is an invertible matrix B ∈ L(IR
n
, IR
n
) that maps the ith unit vector e
i
∈ IR
n
into
z
i
∈ IR
n
, for all i ∈ 1, n.
This leads to the relation
ρ([c(0, B
Q
+
[S]B)) ≤ ρ([c(0, B
QB)) ≤ (2.3.10)
≤ ρ([c
1
) + ρ([c
2
)
122
for all ∈ IR
n
with equality holding for = e
i
, i ∈ 1, n. This implies that the diagonal
elements of B
Q
+
[S]B and B
QB coincide.
Substituting = e
i
+ e
j
, i ,= j into (2.3.10) we obtain
q
(+)
ii
+ 2q
(+)
ij
≤ q
ii
+ 2q
ij
+ q
jj
for arbitrary ﬁxed i and j, where q
(+)
kr
and q
kr
denote the element in the kth row and
rth column of the matrix B
Q
+
[S]B and B
QB
1
respectively the element in the kth row
and rth column of the matrix B
Q
+
[S]B and B
QB
1
respectively. By the equality of the
diagonal elements, this implies
q
(+)
ij
≤ q
ij
.
Carrying out the substitution of = e
i
−e
j
into (2.3.10), we arrive at the reverse inequality.
Taken together this means
B
Q
+
[S]B = B
QB.
and by the invertibility of B, the equality
Q
+
[S] = Q.
Part (a) is thus proved. The proof of part (b) is similar, and is left to the reader. However,
one should have in mind that part (b) is true only if the diﬀerence c
1
÷c
2
has a nonvoid
interior which implies that matrix Q
−
[S] > 0.
Q.E.D.
The second part of the Theorem implies the following assertion
Corollary 2.3.2 The following representation is true
c
1
÷c
2
= ∩¦c(a
1
−a
2
, Q
−
[S])[S ∈ Σ¦. (2.3.11)
A 2dimensional illustration of Theorem 2.3.1 is given in Figures 2.3.1 (a) and (b). The
ﬁrst one shows the nonellipsoidal sum of two ellipsoids and the variety of nondominated
(inclusionmaximal) ellipsoids that approximate it internally, due to formula (2.3.3). The
sum then arrives as the union of the internal ellipsoids ( over all the variety of these,
see Lemma 2.3.1). The second one shows a nondegenerated geometrical diﬀerence of two
ellipsoids ( the set with two kinks) and the variety of nondominated (inclusionminimal)
ellipsoids tha approximate it externally, due to formula (2.3.11). This diﬀerence then
appears as the intersection of the external ellipsoids ( over all the variety). The parameters
123
p ∈ Π
−
, S ∈ Σ are chosen randomly but give a good illustration of the nature of the
representations.
———–!!! insert ﬁgures 2. 3 (a) and (b) —————————
—
2.4 Sums and Diﬀerences:
the Exact Representation
The results of the previous Sections indicate that the sums c
1
+ c
2
and diﬀerences c
1
˙
−c
2
of ellipsoids could be exactly represented through the unions and intersections of the el
ements of certain parameterized families of ellipsoids. Let us once more indicate this
result collecting all the facts in one proposition. When calculating c
1
˙
−c
2
we also assume
int(c
1
˙
−c
2
) ,= ∅.
Theorem 2.4.1 (The Representation Theorem).
Let c
1
= c(a
1
, Q
1
), c
2
= c(a
2
, Q
2
) be a pair of nondegenerate ellipsoids. Let Q(p) be a
parameterized family of ellipsoids
Q(p) = (1 + p
−1
)Q
1
+ (1 + p)Q
2
,
p ∈
+
= [λ
1/2
min
, λ
1/2
max
],
where λ
min
> 0, λ
max
< ∞ are the roots of the equation
det(Q
1
−λQ
2
) = 0
(the relative eigenvalues of Q
1
, Q
2
).
Let
−
=
+
∪ (1, λ
min
).
124
Also let Q
+
[S], Q
−
[S] denote the following parametrized families of ellipsoids
Q
+
[S] = S
−1
[(SQ
1
S
)
1/2
+ (SQ
2
S
)
1/2
]
2
S
−1
Q
−
[S] = S
−1
[(SQ
1
S
)
1/2
−(SQ
2
S
)
1/2
]
2
S
−1
where
S ∈
= ¦S ∈ L(IR
n
, IR
n
) : S
= S, [S[ ,= 0¦.
Then the following inclusions are true
c
1
+c
2
⊆ c(a
1
+a
2
, Q(p)), ∀p ∈
+
, (2.4.1)
c
1
+c
2
⊇ c(a
1
+a
2
, Q
+
[S]), ∀S ∈
, (2.4.2)
c
1
÷c
2
⊆ c(a
1
−a
2
, Q
−
[S]), ∀S ∈
, (2.4.3)
c
1
÷c
2
⊇ c(a
1
−a
2
, Q(−p)), ∀p ∈
−
. (2.4.4)
Moreover, the following exact representations are valid:
c
1
+c
2
=
¦c(a
1
+a
2
, Q(p))[p ∈
+
¦ (2.4.5)
c
1
+c
2
=
_
¦c(a
1
+a
2
, Q
+
[S])[S ∈
¦ (2.4.6)
c
1
÷c
2
=
¦c(a
1
−a
2
, Q
−
[S])[S ∈
¦ (2.4.7)
c
1
÷c
2
=
_
¦c(a
1
−a
2
, Q(−p))[p ∈
−
¦ (2.4.8)
The facts given in this theorem may be treated as being related to integral geometry,
particularly, to the representations of ellipsoidal sets (bodies) in IR
n
. The speciﬁc prop
erties formulated in (2.4.1)(2.4.4) and (2.4.5)(2.4.8) reﬂect a certain type of geometrical
duality in treating the geometrical sums and diﬀerences of ellipsoids. Namely, the exter
nal representations (2.4.1) for approximating the sum yields, with a change in the sign
of the parameter p, the internal representation (2.4.4) for the diﬀerence and the internal
representation (2.4.2) for the sum yields (with a change of sign from (Q
+
[S] to Q
−
[S]),
the external representation (2.4.3) for the diﬀerence. As it was also demonstrated in the
previous sections, Theorem 2.4.1 also indicates that the parametrized varieties involved are
also the varieties of inclusionminimal external and inclusionmaximal internal estimates
for c
1
+c
2
, c
1
÷c
2
. This could be summarized in
Theorem 2.4.2 (i) Given c
1
+ c
2
and an ellipsoid c ⊇ c
1
+ c
2
, there exists a value
p ∈
+
such that
c
1
+c
2
⊆ c(a
1
+a
2
, Q(p)) ⊆ c (2.4.9)
125
(ii) Given c
1
+c
2
and an ellipsoid c ⊆ c
1
+c
2
there exists an S ∈
, such that
c ⊆ c(a
1
+a
2
, Q
+
[S]) ⊆ c
1
+c
2
. (2.4.10)
(iii) Given c
1
÷c
2
, (int (c
1
÷c
2
) ,= ∅) and an ellipsoid c ⊇ c
1
÷c
2
, there exists an S ∈
,
such that
c
1
÷c
2
⊆ c(a
1
−a
2
, Q
−
[S]) ⊆ c. (2.4.11)
(iv) Given c
1
÷ c
2
, (int (c
1
÷ c
2
) ,= ∅) and an ellipsoid c ⊇ c
1
÷ c
2
, there exists an
p ∈
−
, such that
c ⊆ c(a
1
−a
2
, Q(−p)) ⊆ c
1
÷c
2
. (2.4.12)
The variety ¦c(a
1
+ a
2
, Q(p)), p ∈
+
¦ is therefore the set of nondominated (inclusion
minimal) upper ellipsoidal estimates for c
1
+c
2
. The variety of nondominated (inclusion
maximal) internal estimates for c
1
+c
2
is therefore ¦c(a
1
+ a
2
, Q
+
[S], S ∈
¦.
Similarly, the varieties of nondominated (inclusionminimal) external and nondomi
nated (inclusionmaximal) external estimates for c
1
÷ c
2
are c(a
1
− a
2
, Q
−
[S]) and
c(a
1
−a
2
, Q(−p)).
The mentioned relations allow to say that the nondominated ellipsoids, as described above,
posses a certain type of ”Pareto” property. The important fact is that the mentioned
Pareto property is invariant under linear transformations. This means that after a linear
transformation the nondominated ellipsoids remain nondominated. It is precisely this
fact that allows us to propagate the static schemes of this Section to systems with linear
dynamics.
One of the further problems in ellipsoidal approximations to be discussed would be to to
estimate the number of ellipsoids that would give a desired accuracy of approximation.
Without going into the details of this problem, we shall brieﬂy discuss it for the case of
external approximation of the sum of two ellipsoids c
1
= c(a
1
, Q
1
) and c
2
= c(a
2
, Q
2
).
Taking k arbitrary external ellipsoids c(a
1
+a
2
, Q(p
k
)) of the type given in (1.4.1), we have
c
1
+c
2
⊂ ∩¦c(a
1
+ a
2
, Q(p
i
))[i = 1, ...k¦,
where p
i
∈ Π
+
. Without loss of generality we further set a
1
+ a
2
= 0.
Calculating the Hausdorﬀ semidistance (i = 1, ..., k)
h
+
(∩c(0, Q(p
i
)), c
1
+c
2
) = ζ(p[k])
where p[k] = ¦p
1
, ..., p
k
¦ is a kdimensional vector, we come to
126
ζ(p[k]) = max¦ρ(l[ ∩
i
c(0, Q(p
i
)) −ρ(l[c
1
+c
2
)[(l, l) ≤ 1¦
where, in its turn,
ρ(l[ ∩
i
c(0, Q(p
i
))) = min¦
k
i=1
(l
(i)
, Q(p
i
)l
(i)
)
1/2
[Σ
k
i=1
l
(i)
= l¦
Presuming that the best k ellipsoids are those that give the smallest Hausdorﬀ semidistance
ζ(p[k]), we may specify them by solving the problem
ζ
0
(k) = min¦ζ(p[k])[p[k] : p
i
∈ Π
+
, i = 1, ...k¦
The best ellipsoids are those that are generated by the optimalizing vector p[k] = p
0
[k] of
the previous problem. It is not diﬃcult to observe that ζ
0
(k) ≥ ζ
0
(k + 1) and ζ
0
(k) → 0
if k → ∞. We thus assert the following
Lemma 2.4.1 The minimal number of ellipsoids that approximate the sum c
1
+ c
2
with
given accuracy > 0 may be determined as the smallest integer k = k() that satisﬁes the
inequality ζ
0
(k) ≤ .
The respective optimal ellipsoids are those that are generated by the vector p[k()] = p
0
[k()].
It would be also important, of course, to obtain the estimate in a more explicit form or
to obtain estimates that are perhaps less precise, but simpler than the exact one, speciﬁed
by the previous Lemma. An appropriate issue is also to describe eﬀective algorithms for
calculating ζ
0
(k) or its estimates.
A similar reasoning may be applied to the other approximation problems. However, speak
ing in general, we should note that the nonsimple problem of the accuracy and computa
tional complexity of the ellipsoidal approximation requires special treatment and in its full
detail spreads beyond the scope of this book.
It may be also interesting to single out, among the variety of approximating ellipsoids, a
single ellipsoid that is optimal in some sense.
2.5 The Selection of Optimal Ellipsoids
We shall no proceed with describing the optimal ellipsoidal estimates (external or internal)
for c
1
+ c
2
and c
1
÷c
2
, selected through some cost function (optimality criterion). If the
127
cost function ψ(c) is monotonous bby increasing with respect to inclusion (ψ(c
) ≥ ψ(c
)
or ψ(c
) ≤ ψ(c
) if c
⊇ c
), then, due to Theorem 2.4.2, the solution may be sought
for only from the parametrized varieties of this theorem. Some rather simple necessary
conditions of optimality could then be applied for this purpose.
Lemma 2.5.1 Suppose the function ψ(c) is continuous and monotonously increasing with
respect to inclusion. Then
(a) The ψminimal external ellipsoidal estimate of the sum c
1
+c
2
is c(a
1
+ a
2
, Q(p
∗
)),
where p
∗
is the value for which the minimum of the function f : ¦0, ∞¦ −→ 1,
f(p) = ψ(Q(p)), p ∈
+
is attained,
(b) Suppose that int(c
1
) ⊃ c
2
holds. Then the ψmaximal internal estimate of the diﬀer
ence c
1
÷ c
2
is c(a
1
− a
2
, Q(−p
∗∗
)), where p
∗∗
is the value for which the maximum
of the function g : (0, λ
min
) −→ 1
g(p) = ψ(Q(−p)), p ∈
−
is attained.
Proof. By the monotonicity of ψ, follows from Theorem 2.4.2. Q.E.D.
We shall now apply Lemma 2.5.1 to ﬁnd the minimal external estimates of the sum of two
ellipsoids with respect to three important parameters of ellipsoids: the volume, the sum
of squares of semiaxes (or trace) and Tr(Q
2
), and do the same for the maximal external
estimates of the diﬀerence. Using our technique, analogous results can be obtained for
other functions, like Tr(Q
3
), etc.
Although the case of the diﬀerence is similar to the above, in our test cases we are not
in such a favourable position, because the existence of a unique stationary point cannot
always be guaranteed. However, as we shall see later, this still can be done up to o(ε) in
the important special case when Q
2
= ε
2
Q
0
.
Lemma 2.5.2 (a) There exists a unique ellipsoid with minimal sum of squares of
semiaxes, (which is Tr(Q)) that contains the sum c
1
+c
2
. It is of the form
c(a
1
+a
2
, Q(p
∗
)), where
p
∗
=
(TrQ
1
)
1/2
(TrQ
2
)
1/2
. (2.5.1)
128
(b) Suppose that int(c
1
) ⊃ c
2
holds, and also that there exists an internally estimating
ellipsoid for the diﬀerence c
1
÷c
2
such that its sum of squares of semiaxes (which is
Tr(Q)) is maximal. Then it is of the form c(a
1
−a
2
, Q(−p
∗
)). Here p
∗
is deﬁned by
the equality (2.4.13) and p
∗
∈ Π
−
.
Proof. The function f : (0, ∞) −→ 1
f(p) =
d
dp
Tr[Q(p)] =
d
dp
Tr[(1 + p
−1
)Q
1
+ (1 + p)Q
2
]
= Tr
d
dp
[(1 + p
−1
)Q
1
+ (1 + p)Q
2
]
has one single root at p = p
∗
, therefore it has to be an element of Π
+
.
In case (b), considering the function deﬁned by g(p) = f(−p), we have g(p) = f(p), and
therefore the root is the same as in the above case and unique. The matrix Q(−p
∗
) being
positive deﬁnite implies p
∗
< λ
min
. Q.E.D.
Lemma 2.5.3 (a) There exists a unique ellipsoid of minimal volume that contains the
sum c
1
+c
2
. It is of the form c(a
1
+ a
2
, Q(p
∗
)) where p
∗
∈ (0, ∞) is the unique
solution of the equation
n
i=1
1
λ
i
+ p
=
n
p(p + 1)
. (2.5.2)
For it we also have p
∗
∈ Π
+
.
(b) Suppose that c(a
2
, Q
2
) ⊂ int(c(a
1
, Q
1
)), then there exists a unique ellipsoid c of
maximal volume contained in the diﬀerence c
1
÷ c
2
. It is of the form c = c(a
1
−
a
2
, Q(−p
∗∗
)) where p
∗∗
is the single root of equation (2.5.6) falling into the set Π
−
.
Proof. By Lemma 2.5.1 we only have to ﬁnd p
∗
∈ (0, ∞) that minimizes the volume of
c(a, Q()). This is obviously equivalent to the minimization of log det(Q()), and det(Q())
depends on the product of the eigenvalues of Q. Since the eigenvalues of Q do not change
with nondegenerate linear transformations, the minimal property of the estimating ellipsoid
is not changed if a nondegenerateaﬃne transformation is applied. Because of the way Q(p)
is derived from Q
1
and Q
2
, so is the value of p
∗
∈ (0, ∞). Therefore, we are allowed to
suppose that
Q
j
= diag¦q
(j)
1
, . . . q
(j)
n
¦, j ∈ 1, n
and hence
Q(p) = diag¦q
1
(p), . . . q
n
(p)¦.
129
All this means that we have to ﬁnd the roots of the function f : ¦0, ∞¦ −→ IR
f(p) =
d
dp
log det(Q(p)).
Using the relationship
d
dp
log det(Q(p)) = Tr
_
Q(p)
−1
d
dp
Q(p)
_
we obtain
Tr[(p
−1
Q
2
−1
Q
1
+ I)
−1
] =
n
p + 1
.
By diagonality this means that
n
i=1
p
λ
i
+ p
=
n
p + 1
,
where λ
i
, i ∈ 1, n are the eigenvalues of the pencil Q
1
−λQ
2
— and they are again invariant
with respect to the aﬃne transformation used for diagonalization.
The left hand side of this equality strictly increases from 0 to n and the right hand side
strictly decreases from n to 0 while p increases from 0 to ∞. Therefore it has one single
root corresponding to a minimum, as we have
lim
p−→0
det(Q(p)) = lim
p−→∞
det(Q(p)) = ∞.
The proof of (a) now is complete.
In the case of (b) we need the roots in (−∞, 0). If here the parameter p corresponding to
the maximal volume fell onto any of the endpoints of the interval (0, λ
min
) then the matrix
would be semideﬁnite, i. e. the ”ellipsoid” would have zero volume. But this is excluded
by the condition we set. After having established this, a similar argument shows that there
is one single local maximum. Q.E.D.
Lemma 2.5.4 (a) There exists a unique ellipsoid with minimal Tr(Q
2
) that contains the
sum c
1
+c
2
. It is of the form c(a
1
+a
2
, Q(p
∗
)), where p
∗
is the unique positive root
of the polynomial
f(p) = γ
22
p
3
+γ
12
p
2
−γ
12
p −γ
11
(2.5.3)
γ
ij
= Tr(Q
i
Q
j
),
for i, j ∈ 1, 2. The value p
∗
∈
+
.
130
(b) Suppose that int(c
1
) ⊃ c
2
holds, and also that there exists an internally estimating
ellipsoid for the diﬀerence c
1
÷c
2
with maximal Tr(Q
2
). Then it is of the form
c(a
1
− a
2
, Q(−p
∗∗
)). Here p
∗∗
is a root of the polynomial deﬁned by the equality
g(p) = f(−p) that falls into Π
−
.
Proof. Direct calculations indicate that
f(p) =
d
dp
Tr[Q(p)
2
].
All the coeﬃcients γ
ij
s are positive, as we have the equality Tr(Q
i
Q
j
) = Tr(R
j
Q
i
R
j
)
where R
j
is the square root of Q
j
. The right hand side is positive here, because we take
the trace of a positive deﬁnite matrix. Hence γ
22
> 0 and γ
11
> 0, implying the existence
of a positive root.
The equality f(p) = 0 is equivalent to
p
2
=
γ
12
p + γ
11
γ
22
p + γ
12
.
Here the left hand side is a convex strictly increasing function, while the right hand side
is a continuous function with value γ
1
, γ
−1
12
at p = 0, which tends to γ
11
γ
−1
22
with p →
∞. Therefore their graphs may have no more than one intersection. The proof of (a) is
complete.
The proof of the statement of part (b) is obvious. Q.E.D.
For the sake of approximating the solutions to diﬀerential inclusions we may need to treat
a particular case, when the parameters of c
2
= c(a
2
, Q
2
) may be presented as
a
2
= εa
0
, Q
2
= ε
2
Q
0
so that
ρ(l[c(a
2
, Q
2
)) = (l, a
2
) + (l, Q
2
l)
1/2
= ε(l, a
0
) + ε(l, Q
0
l)
1/2
= ερ(l[c(a
0
, Q
0
)),
or in other terms
c(a
2
, Q
2
) = εc(a
0
, Q
0
)
i.e.
ε
2
= εc
0
, c
0
= c(a
0
, Q
0
). (2.5.4)
131
Lemma 2.5.5 Let us have for ε > 0 the relation (2.5.4).
(a) Then the ellipsoid with minimal sum of squares of semiaxes, (which is Tr(Q)) that
contains the c
1
+c
2
is of the form c(a
1
+ a
2
, Q
ε
(q
∗
)), where
Q
ε
(q) = Q
1
+ ε(q
−1
Q
1
+ qQ
0
) + ε
2
Q
0
(2.5.5)
and
q
∗
=
Tr
1/2
(Q
1
)
Tr
1/2
(Q
0
))
. (2.5.6)
(b) Suppose that int(c
2
) ,= ∅. Then the ellipsoid with the maximal sum of squares of
semiaxes, (which is Tr(Q)) that is contained in the diﬀerence c
1
÷c
2
is of the form
c(a
1
−εa
0
, Q
ε
(−q
∗
)).
Proof. Writing formally
Q
0
(p) = (1 + p
−1
)Q
1
+ ε
2
(1 + p)Q
0
, (2.5.7)
and calculating the optimal p = p
∗
that gives Tr(Q(p)) = min, we obtain
ε
2
(p
∗
) =
TrQ
1
TrQ
0
)
. (2.5.8)
Introducing the notation
Q
ε
(q) = Q(qε
−1
), q = pε, (2.5.9)
(2.5.5) follows. By (2.5.8) now the proof of part (a) is complete.
The proof of case (b) is similar to the above, however here we also need the inequality
ε
−1
q
∗
< λ
max
= ε
−2
µ
max
that will automatically hold for small ε. Here µ
max
is the maximal relative eigenvalue of
Q
1
and Q
0
. The lemma is now proved for both cases. Q.E.D.
Lemma 2.5.6 Let us have for ε > 0 the relation (2.5.4).
(a) Then the ellipsoid with minimal volume that contains the sum c
0
+c
2
is of the form
c(a
1
+εa
0
, Q
ε
(q
∗
)), with Q
ε
(q) given by (2.5.6) and
q
∗
=
n
1/2
Tr
1/2
(Q
0
Q
−1
1
)
. (2.5.10)
132
(b) Suppose that int(c
2
) ,= ∅. Then the ellipsoid with maximal volume that is contained
in the diﬀerence c
1
÷c
2
is of the form
c(a
1
−εa
0
, Q
ε
(−q
∗
)), where q
∗
is deﬁned by (2.5.11).
Proof. Denote by µ
1
≤ µ
2
≤ . . . ≤ µ
n
the relative eigenvalues of the matrices c
1
and
c
0
. Comparing them with the relative eigenvalues λ
1
≤ λ
2
≤ . . . ≤ λ
n
of c
1
and c
2
we
obviously have for all i ∈ 1, n
ε
2
λ
i
= µ
i
. (2.5.11)
We look for the root of equation (2.5.6) where we have to use (2.5.8) and according to the
equality q = pε. Rewriting equation (2.5.6), into the form
n
i=1
λ
i
λ
i
+ p
=
np
p + 1
, (2.5.12)
and carrying out the substitutions, by the analytic dependence of the roots on the param
eters, we obtain for the Taylorseries expansion in ε of equation (2.4.20)
n
i=1
1
1 + εqµ
−1
i
=
n
1 + εq
−1
.
and
n −εq
∗
n
i=1
µ
−1
i
= n(1 −ε(q
∗
)
−1
) + o(ε).
From the comparison of the coeﬃcients of ε,
(q
∗
)
2
=
n
n
i=1
µ
−1
i
,
and then (a) follows. For part (b) we have to take the negative sign, and the condition of
positive deﬁniteness follows in the same way as for Lemma 2.5.2. Q.E.D.
Lemma 2.5.7 Assume the relation the relation (2.5.4) to be true.
(a) Then the ellipsoid with minimal Tr(Q
2
) that contains the sum c
1
+c
2
is of the form
c(a
1
+εa
0
, Q
ε
(q
∗
)), with formula Q
ε
(q) given by (2.5.6)
q
∗
=
Tr
1/2
(Q
2
1
)
Tr
1/2
(Q
1
Q
0
)
. (2.5.13)
133
(b) Suppose that int(c
2
) ,= ∅. Then the ellipsoid with maximal Tr(Q
2
) that is contained
in the diﬀerence c
2
÷c
1
is of the form c(a
1
− εa
0
, Q
ε
(−q
∗
)), where q
∗
is deﬁned by
(2.5.14) 2.5.13.
Proof. The notation of equation (2.5.3)
γ
ij
= Tr(Q
i
Q
j
),
is now used for i, j = 1 or 0. Using the above scheme, we look for the root again in the
form of q = pε. Substituting this into equation (2.5.4)
εγ
00
q
3
+ γ
10
q
2
−εγ
10
q −γ
11
= 0 (2.5.14)
and from the comparison of coeﬃcients in the Taylorseries expansion in ε for this equation
(q
∗
)
2
=
Tr(Q
2
1
)
Tr(Q
1
Q
0
)
. (2.5.15)
The rest of the proof is analogous. Q.E.D.
Corresponding to the above, we formulate the converse statements concerning the internal
estimates of maximal volume of the Minkowskisum, and the external estimates of minimal
volume of the diﬀerence. Part (a) of the following theorem is given in ( [73]), but this proof
is diﬀerent, and appears to be more general. On the other hand, part (b) is a new result,
which is proved by the technique used there.
Lemma 2.5.8 (a) There exists a unique ellipsoid of maximal volume contained in the
sum c
1
+c
2
. It is of the form c(a
1
+a
2
,
˜
Q
+
) where
Q
∗
+
= Q
1
+ Q
2
+ 2Q
2
1/2
[Q
2
−1/2
Q
1
Q
2
−1/2
]
1/2
Q
2
1/2
. (2.5.16)
(b) There exists a unique ellipsoid of minimal volume containing the diﬀerence c
1
÷c
2
.
It is of the form c(a
1
−a
2
, Q
∗
−
) where
Q
∗
−
= Q
1
+ Q
2
−2Q
2
1/2
[Q
2
−1/2
Q
1
Q
2
−1/2
]
1/2
Q
2
1/2
. (2.5.17)
Proof. It is not diﬃcult to prove, as in [Chern], that (a) is valid with
Q
∗
+
= S
−1
[(SQ
1
S
)
1/2
+ (SQ
2
S
)
1/2
]
2
S
−1
134
where S is the matrix diagonalizing both Q
1
and Q
2
. It is also possible to observe that
although S is not unique, this expression is independent of the choice of S. Let us select
S = NQ
2
−1/2
where N is orthogonal. Then for a suitable N the matrix S will meet the requirements. A
substitution now yields (a).
Let us now consider (b). By the aﬃne invariance of the volume function, we can use again
the matrix S to diagonalize Q
1
and Q
2
, i. e.
to get SQ
1
S
= D
1
= diag¦r
1
, . . . r
n
¦. and SQ
2
S
= P, where P = diag¦p
1
, . . . p
n
¦ is a
partial identity. Our aim is to ﬁnd a minimal volume ellipsoid c(a, D) among those with
the property:
¸a, y) +¸Dy, y)
1/2
≥ ¸D
1
y, y)
1/2
−¸Py, y)
1/2
By the argument of the proof of Part 2 in [84], the existence and uniqueness of such an
ellipsoid follows, and the same argument implies that a = 0 and D is diagonal. This is
justiﬁed because of the inclusion int(c
1
) ⊃ c
2
holds. Substituting the unit coordinate
vectors into the previous inequality we obtain
d
1/2
i
≥ r
i
1/2
−p
i
1/2
i ∈ 1, n.
If we deﬁne d
i
, i ∈ 1, n with d
i
= (r
1/2
i
− p
i
1/2
)
2
, then it can be established by direct
calculation that D is an external estimate. The statement is thus true. Q.E.D.
Lemma 2.5.9 Let us have the relation (2.5.9) for ε > 0. Then
(a) the ellipsoid with maximal volume that is contained in the sum c
1
+c
2
is of the form
c(a
1
+εa
0
, Q
ε+
), where
Q
ε+
= Q
1
+ 2ε(Q
1/2
o
[Q
−1/2
o
Q
1
Q
−1/2
0
]
1/2
Q
1/2
0
) + o(ε)I (2.5.18)
(b) the ellipsoid with minimal volume that contains the diﬀerence c
1
÷c
2
is of the form
c(a
1
−εa
0
, Q
ε−
)), where
Q
ε−
= Q
1
−2ε(Q
1/2
o
[Q
−1/2
o
Q
1
Q
−1/2
0
]
1/2
Q
1/2
0
) + o(ε)I (2.5.19)
The proof follows the lines of the above through expansions of the respective relations in ε
and the solution is found within the terms with ε of power 1. The reader may verify this
as an exercise.
To illustrate the material of this Section, we introduce several ﬁgures. Thus Fig.2.5.1(a)
is the same as 2.3(a), except that in addition to the internal estimates that indicate the
135
sum of two ellipsoids, we also indicate the external estimates that are of minimal trace (
TrQ), minimal Tr(Q
2
) ( both drawn in continuous lines) and minimal volume (drawn in
dotted line). Fig.2.5.1(b) is the same as 2.2.(a), but in addition to the external estimates
that indicate the sum of two ellipsoids we also indicate the internal ellipsoid of maximal
volume ( shown with dotted line). Fig. 2.5.2(a) is the same as 2.2.(b), but in addition
the dotted line shows the volumeminimal ellipsoid for the Minkowskidiﬀerence of two
ellipsoids. Finally, Fig.2.5.2(b) is the same as 2.3.(b), but in addition the dotted line shows
the volumemaximal internal ellipsoid for the Minkowski diﬀerence.
——–!!! insert Figures 2.5.1(a),(b), 2.5.2(a),(b) !!! —————–
The next issue is to consider intersections of ellipsoids. The description of this operation
is more complicated than before and does not reach the same degree of detailization being
conﬁned mostly to the external ellipsoidal estimates of these intersections.
2.6 Intersections of Ellipsoids
Let us consider m nondegenerate ellipsoids c
i
= c(a
(i)
, Q
i
),
i = 1, . . . , m. Their intersection
m
i=1
c(a
(i)
, Q
i
) = T[m]
consists of all the vectors x ∈ IR
m
that simultaneously satisfy the inequalities
(x −a
(i)
, Q
−1
i
(x −a
(i)
)) ≤ 1, (i = 1, . . . , m). (2.6.1)
Assuming
/ = ¦α ∈ IR
m
:
α
i
= 1, α
i
≥ 0, i = 1, . . . , m¦,
take the inequality
m
i=1
α
i
(x −a
(i)
, Q
−1
i
(x −a
(i)
)) ≤ 1. (2.6.2)
The following assertion is obvious
136
Lemma 2.6.1 If x
∗
∈ IR
m
is a solution to the system (2.6.1), then x
∗
satisﬁes (2.6.2) for
any α ∈ / (and vice versa).
By direct calculation we may observe that for a given α ∈ / the inequality (2.5.2) deﬁnes
an ellipsoid
c[α] = ¦x : (x −a[α], Q[α](x −a[α])) ≤ 1 −h
2
[α]¦ (2.6.3)
where
a[α] =
_
m
i=1
α
i
Q
−1
i
_
−1
_
m
i=1
α
i
Q
−1
i
a
(i)
_
Q[α] =
m
i=1
α
i
Q
−1
i
h
2
[α] =
m
i=1
α
i
_
a
(i)
, Q
−1
i
a
(i)
_
−
_
_
m
i=1
α
i
Q
−1
i
a
(i)
,
_
m
i=1
α
i
Q
−1
i
_
−1
_
m
i=1
α
i
Q
−1
i
a
(i)
_
_
_
_
_
m
i=1
α
i
Q
−1
i
a
(i)
,
_
m
i=1
α
i
Q
−1
i
_
−1
_
m
i=1
α
i
Q
−1
i
a
(i)
_
_
_
.
It is not diﬃcult to check that h
2
[α] ∈ [0, 1]. In other terms
c[α] = c(a[α], (1 −h
2
[α])
−1
Q[α]). (2.6.4)
A direct consequence of Lemma 2.6.1 is
Lemma 2.6.2 The following assertion is true. The set
T[m] = ¦∩c[α][α ∈ /¦.
Each of the ellipsoids c[α] is therefore an external estimate for T[m], so that P
m
⊆ c[α],
for any α ∈ /.
The intersection T[m] of m ellipsoids c
i
is now presented as an intersection of a
parametrized family of ellipsoids c[α], α ∈ /. Among these we may select, if necessary, an
optimal external ellipsoidal estimate for T[m].
It is clear that the variety c[α] α ∈ / contains each of the ellipsoids c
i
, so that
c
i
= c[e
(i)
]; i = 1, . . . , m,
137
where e
(i)
∈ IR
m
is a unit vector (an orth) along the ith coordinate axis and its coordinates
are deﬁned as
e
(i)
j
= δ
ij
;
δ
ij
= 1 if i = j, δ
ij
= 0 if i ,= j j = 1, . . . , m.
In ﬁgure 2.6.1(a) one may observe an illustration of Lemma 2.6.2. Here numbers 1,2
indicate the intersecting ellipsoids whilst the unmarked ones are the external estimates
c[α] calculated due to formula (2.6.4). The intersecting ellipsoids 1,2 correspond to values
α
1
= 1, α
2
= 0 and α
1
= 0, α
2
= 1 in (2.6.4). Figure 2.6.1(b) shows the two intersecting
ellipsoids and with dotted line the volumeminimal external estimate obtained by a one
dimensional search in α
1
∈ [0, 1],, ( α
2
≥ 0, α
1
+ α
2
= 1).
——————!!! Insert ﬁg 2.6.1(a),(b) !!! ——————
However, in general, the optimal external ellipsoidal estimate c[α
0
] for T[m] (taken, for
example, for one of the criteria of the above) may not be among the ellipsoids c
i
. One
of the questions that arise here is whether the variety c[α] is suﬃciently complete in the
sense of the following question: will the optimal external estimate c[α
0
] (with respect to
a function Ψ(c)) chosen only among the ellipsoids c[α], α ∈ / be the same as the optimal
ellipsoid c (also with respect to Ψ(c)) chosen among some other class of external ellipsoids
or, particularly, among all the possible external estimates?
In the sequel of this section we shall produce some examples (see Examples 2.6.2, 2.6.3 )
that give an answer to this question.
Meanwhile let us introduce another formula for the intersection
m
i=1
c(a
(i)
, Q
i
) = T[m]
of nondegenerate ellipsoids c(a
(i)
, Q
i
).
138
Assumption 2.6.1 The intersection T[m] has an interior point : intT[m] ,= ∅.
We shall further proceed under this assumption. Taking the support functions
ρ([c(a
(i)
, Q
i
)) for the ellipsoids c(a
(i)
, Q
i
), we may apply the convolution formula ( [265])
ρ([Q) = ¦ inf
m
i=1
ρ(
(i)
[Q
i
)[
m
i=1
(i)
= ¦ (2.6.5)
that relates the support function ρ([Q) of an intersection Q = ∩
m
i=1
Q
i
with the support
functions ρ([Q
i
) for each set Q
i
.
Then, assuming
(i)
= M
(i)
(2.6.6)
where the matrix M
(i)
∈ IR
n×n
does exist for any ,
(i)
∈ IR
n
, ,= ¦0¦, we may substitute
(2.6.6) into (2.6.5), coming to
ρ([Q) =
_
inf
m
i=1
ρ([M
(i)
Q
i
) [M
(i)
:
m
i=1
(M
(i)
−I) = 0
_
(2.6.7)
or to the relations
ρ([Q) ≤ ρ([
m
i=1
M
(i)
Q
i
), (
m
i=1
M
(i)
−I) = 0
which should be true for any ∈ IR
n
and any array of matrices M
(i)
that satisfy the last
equality.
Otherwise this means
Q ⊆
n
i=1
M
(i)
Q
i
, (2.6.8)
whatever are the matrices M
(i)
that satisfy the equality
m
i=1
M
(i)
= I. (2.6.9)
Moreover, (2.6.7) implies
Q = ∩
m
i=1
M
(i)
Q
i
over all the matrices M
(i)
that satisfy (2.6.9) (we may omit the transpose, since M
(i)
are
chosen arbitrarily, provided only that (2.6.9.) does hold.)
139
In terms of ellipsoids and in view of the formula
Mc(a, Q) = c(Ma, MQM
)
this gives
T[m] ⊆
m
i=1
c(M
(i)
a
(i)
, M
(i)
Q
i
M
(i)
),
m
i=1
M
(i)
= I (2.6.10)
and for the same class of matrices (2.6.9) we have
T[m] = ∩
m
i=1
c(M
i
a
(i)
M
(i)
Q
i
M
(i)
),
m
i=1
M
(i)
= I. (2.6.11)
We therefore come to the assertion
Lemma 2.6.3 The intersection T[m] of m nondegenerate ellipsoids c(a
(i)
, Q
i
) satisﬁes
the inclusion (2.6.10), whatever are the matrices M
(i)
of (2.6.9). Moreover, the equality
(2.6.11) is true with the intersection taken over all M
(i)
of (2.6.9).
Particularly, for m = 2 we have
T[2] ⊆ (2.6.12)
⊆ c(M
a
(1)
, MQ
1
M
) +c((I −M)
a
(2)
, (I −M)Q
2
(I −M)
))
for any n n matrix M ∈ L(IR
n
, IR
n
).
The intersection T[m] of n ellipsoids c(a
(i)
, Q
i
) is therefore approximated from above in
(2.6.10) by the sum of m ellipsoids c(a
(i)
, M
(i)
Q
i
M
(i)
) restricted only by the equality
(2.6.9). As we have seen earlier, the sum of m ellipsoids may, however, be approximated
from above by one ellipsoid. Namely,
m
i=1
c(a
(i)
, M
(i)
Q
i
M
(i)
) ⊆ c(a[m], Q[m, p, /])
where
a[m] =
m
i=1
M
(i)
a
(i)
Q[m, p, /] = (
m
i=1
p
i
)
m
i=1
p
−1
i
M
(i)
Q
i
M
(i)
, p
i
≥ 0
/ = ¦M
(1)
, . . . , M
(m)
¦, p = ¦p
1
, . . . , p
m
¦
140
Combining the results of Lemma 2.6.3 and Theorem 2.4.1 ( formula (2.4.5)) , we conclude
that the intersection T[m] may be presented through the inclusion
T[m] ⊆ c(a[m, /], Q[m, p, /]) (2.6.13)
which is true for any /, p > 0, provided / satisﬁes (2.6.9), or the equality
T[m] =
p
M
c(a[m, /], Q[m, p, /]) (2.6.14)
with / of (2.6.9), p > 0.
Lemma 2.6.4 The intersection T[m] satisﬁes the inclusion (2.6.13) over all / of (2.6.9)
and the equality (2.6.14) over all / of (2.6.9).
Among the ellipsoids c(a[m], p, /) we may now select those that are optimal relative to
some criteria, taking perhaps one of the above, deﬁned at the end of Section 2.1.
Let us ﬁrst consider two ellipsoids with centers a
(1)
= a
(2)
= 0 so that
T[2] = (2.6.15)
= c(0, Q
1
) ∩ c(0, Q
2
) ⊆ c(0, MQ
1
M
) +c(0, (I −M)Q
2
(I −M)
)
The external bounding ellipsoid may be now designed through the following schemes.
Scheme A
For a matrix Q positive symmetrical we may rewrite
MQM
= (Q
1/2
M
)
(Q
1/2
M
)
and introduce the norm
MQM

2
= ¸Q
1/2
M
, Q
1/2
M
) = trMQM
,
where the scalar product ¸K, L) of two n nmatrices K, L ∈ IR
n×n
is deﬁned as
¸K, L) = trK
L.
The present scheme is now deﬁned through minimizing
MQ
1
M

2
+ (I −M)Q
2
(I −M)

2
=
= ¸Q
1/2
1
M
, Q
1/2
1
M
) +¸Q
1/2
2
(I −M)
, Q
1/2
2
(I −M)
) (2.6.16)
141
over M which leads to the optimal M = M
0
:
M
0
= M
0
= (Q
1
+Q
2
)
−1
Q
2
.
Further on, following (2.6.13) we have
c(0, M
0
Q
1
M
0
) +c(0, (I −M
0
)Q
2
(I −M
0
)
) (2.6.17)
⊆ c(0, (1 + p
−1
)M
0
Q
1
M
0
+ (1 + p)(I −M
0
)Q
2
(I −M
0
))
whatever is the p > 0. The bounding ellipsoid may now be optimalized over p due to one
of the criteria of the above (see Section 2.1).
Let us for example select an optimal p = p
0
, minimizing over p the trace
tr((1 + p
−1
)S
0
1
+ (1 + p)S
0
2
) = f
1
(p)
where
S
0
1
= M
0
Q
1
M
0
, S
0
2
= (I −M
0
)Q
2
(I −M
0
)
. (2.6.18)
Solving this problem through the equation f
1
(p) = 0, ( check here that what one gets is
precisely a minimum ) ,we observe
p
2
0
= trS
0
1
/trS
0
2
.
The ﬁnal calculation gives an upper bound for T[m], which is
T[m] ⊆ c(0, (1 + p
−1
0
)S
0
1
+ (1 + p
0
)S
0
2
) = c(0, S
0
) (2.6.19)
where
trS
0
= ((trS
0
1
)
1/2
+ (trS
0
2
)
1/2
)
2
. (2.6.20)
Consider a speciﬁc case
Example 2.6.1 Take the twodimensional ellipsoids c
1
= c(0, Q
1
), c
2
= c(0, Q
2
), where
Q
1
=
_
1 0
0 k
2
_
, Q
2
=
_
4k
2
0
0 1
_
Then
M
0
= (Q
1
+ Q
2
)
−1
Q
2
=
_
4k
2
(1 + 4k
2
)
−1
0,
0 (1 + k
2
)
−1
_
trM
0
Q
1
M
0
= 16k
2
(1 + 4k
2
)
−4
+ k
2
(1 + k
2
)
−2
= α
2
(k)
tr(I −M
0
)Q
2
(I −M
0
)
= 4k
2
(1 + 4k
2
)
−2
+ k
4
(1 + k
2
)
−2
= β
2
(k)
142
p
0
= α(k)β
−1
(k).
Following (2.6.19), (2.6.20), we have
trS
0
= (α(k) + β(k))
2
S
0
= (1 + p
0
)(p
0−1
S
0
1
+ S
0
2
).
Scheme B
The next option is not to minimize (2.6.16) ﬁrst over M, then over p, but to take the
bounding ellipsoid c(0, S[p, M]) given by the inclusion
c(0, MQM
) +c(0, (I −M)Q
2
(I −M)
) ⊆
⊆ c(0, (1 + p
−1
)MQM
+ (1 + p)(I −M)Q
2
(I −M)
) = c(0, S[p, M])
and to minimize c(0, S[p, M]) directly over the pair p, M(p > 0, M ∈ IR
n×n
) having
trc(0, S[p, M]) = min as the criterion. After a minimization of trc(0, S[p, M]) over p, this
leads to the problem of minimizing the function
f
2
(M) = ((trMQ
1
M
)
1/2
+ (tr(I −M)Q
2
(I −M)
)
1/2
)
2
over M. Since f
2
(M) is strictly convex and f
2
(M) → ∞ with M → ∞, there exists a
unique minimizer M
∗
.
We also gather that
p
∗
= (trM
∗
Q
1
M
∗
)
1/2
(tr(I −M
∗
)Q
2
(I −M
∗
)
)
−1/2
so that the optimal ellipsoid
c
∗
= c(0, S[p
∗
, M
∗
]).
We have thus indicated two options for the bounding ellipsoid
T[m] = c(0, S
0
)
The one of Scheme A is when S
0
is taken due to (2.6.18). The value M
0
for this case is
calculated through the minimum of (2.6.16) which is
¸M, Q
1
M
) +¸(I −M), Q
2
(I −M)
) =
= trMQ
1
M
+tr(I −M)Q
2
(I −M)
143
On the other hand, in Scheme B, we have
T[m] = c(0, S[p
∗
, M
∗
])
where M
∗
is calculated by minimizing f
2
[M] ,which is equivalent to the minimization of
(trMQ
1
M
)
1/2
+ (tr(I −M)Q
2
(I −M)
)
1/2
We shall now illustrate the given schemes on twodimensional examples, comparing on
them the results given by Schemes A and B. Apart from distinguishing these two cases, we
shall also distinguish for each case a minimization over diagonal matrices M only (cases
AD and BD respectively) from a minimization over all possible matrices M (cases AA and
BA ). In all the consecutive ﬁgures the intersecting ellipsoids are marked by numbers 1,2
, while the approximating ellipsoids are marked as A (AD,AA) and B(BD,BA).
Consider ﬁrst the case when the centers of c
1
and c
2
coincide.
13
Example 2.6.2
(a) The ellipsoids 1,2 are centerd at 0. Here both schemes AA and BA give the same
external ellipsoid (Fig.2.6.2 (a)). However, one may observe, that scheme AD gives a
larger one than AA . At the same time, scheme BD does not give any other ellipsoid
except 1,2.
(b) The ellipsoids 1,2 are centered at 0. Here schemes A and B give diﬀerent external
ellipsoids AA and BA ( Fig.2.6.2(b)) At the same time, for each of these schemes the
ellipsoids AA,BA are smaller (by inclusion) than AD,BD (which are not shown).
—!!! place ﬁgures 2.6.2 (a),(b) !!!—————————
The schemes A,B are now applied to ellipsoids c
1
, c
2
with diﬀerent centers.
13
Examples 2.6.2, 2.6.3 were calculated by S.Fefelov.
144
Example 2.6.3
(a) Here schemes AA,BA give the same external ellipsoid which clearly is not optimal by
either trace or volume. Note that scheme BD gives nothing more than ellipsoids 1,2 ( Fig.
2.6.3 (a)).
(b0 Here schemes A,B give diﬀerent external ellipsoids, but AA coincides with AD and
BA with BD (Fig. 2.6.3(b),(b1)).
(c) Here schemes AA,BA give the same external ellipsoid which is close to optimal by trace
or volume ( Fig.2.6.3(c)). Note that AD,BD give worse results in both cases.
——!!! place ﬁgures 2.6.3 (a), 2.6.3(b),(b1), 2.6.3(c) !!!——
Scheme C
This one is similar to Scheme 1, but instead of minimizing the trace f
1
(p), we minimize
f
3
(p) = tr(((1 + p
−1
)S
0
1
+ (1 + p)S
0
2
)((1 + p
−1
)S
0
1
+ (1 + p)S
0
2
)
)l
Equation f
3
(p) = 0 now yields a cubic polynomial
a
0
p
3
+ a
1
p
2
+a
2
p + a
3
= 0
where (S
0
1
= S
0
1
, S
0
2
= S
0
2
)
a
0
= trS
0
2
S
0
2
, a
1
= a
2
= trS
0
1
S
0
2
a
3
= trS
0
1
S
0
1
It may be checked, without diﬃculty, that the given polynomial has a unique positive root
p = p
∗
> 0, which turns out to be the optimalizer and therefore may be substituted into
f
3
(p) allowing to write
f
3
(p
∗
) = min
p
f(p), p > 0.
145
The optimal circumscribed (external) ellipsoid
c(0, S
0
) ⊇ c(0, S
0
1
) +c(0, S
0
2
)
is given by
S
0
= (1 + p
∗−1
)S
0
1
+ (1 + p
∗
)S
0
2
.
Let us now return to the case of an arbitrary ﬁnite number m of intersecting ellipsoids and
select the external circumscribed ellipsoid as a traceminimal set. We have
trQ[m, p] =
m
i=1
γ
i
c
2
i
= ϕ(p) (2.6.21)
where
γ
i
=
_
m
i=1
p
i
_
p
−1
i
, c
2
i
= trM
(i)
Q
i
M
(i)
and
m
i=1
M
(i)
= I. (2.6.22)
Minimizing trQ[m, p] over p = ¦p
i
, . . . , p
m
¦ and assuming
m
i=1
p
i
,= 0, p
i
> 0,
we come to the system
∂ϕ
∂p
i
= 0, i = 1, . . . , m,
or otherwise, to the equations
_
m
i=1
p
i
__
−c
2
i
γ
2
i
+
m
i=1
γ
i
c
2
i
_
= 0, i = 1, . . . , m,
the solution to which is given by
γ
0
i
=
_
m
i=1
c
i
_
c
−1
i
and therefore, by c
i
= p
i
(i = 1, . . . , m) so that the optimal value
ϕ(p
0
) = (
m
i=1
c
i
)
2
Further on we shall brieﬂy describe a possible approach to the calculation of internal el
lipsoidal approximations of an intersection of two nondegenerate ellipsoids c
1
= c(a
(1)
, Q
1
)
146
and c
2
= c(a
(2)
, Q
2
). We assume that this intersection has an interior point : intE
1
∩c
2
,= ∅
( Assumption 2.6.1 ).
Consider the direct product
c
1
⊗c
2
= c(a
(1)
∗
, Q
(1)
) +c(a
(2)
∗
, Q
(2)
) = H,
where
a
(1)
∗
=
_
a
(1)
0
_
, a
(2)
∗
=
_
0
a
(2)
_
and
Q
(1)
≡
_
Q
1
0
0 0
_
, Q
(2)
≡
_
0 0
0 Q
2
_
Clearly, a
(1)
∗
, a
(2)
∗
∈ IR
2n
, Q
(1)
, Q
(2)
, H ∈ L(IR
2n
, IR
2n
)
The set H is the sum of two degenerate ellipsoids in IR
2n
. Nevertheless, since c
1
, c
2
are
nondegenerate in IR
n
and the set H is assumed to have an interior point in IR
2n
, it may
still be approximated internally according to formula (2.3.3) and Corollary 2.3.3 ( where
one just has to take the closure of the approximating variety). We may therefore write
H ⊃ c(a
(1)
+ a
(2)
, Q[H]),
where
Q[S] = S
−1
[(SQ
(1)
S
)
1/2
+ (SQ
(2)
S
)
1/2
]
2
(S
)
−1
and S is a symmetrical matrix of dimension 2n 2n.
Let us denote a = a
(1)
+ a
(2)
and
Q
−1
[S] ≡
_
Q
−
11
Q
−
12
Q
−
21
Q
−
22
_
z =
_
x
(1)
x
(2)
_
where x
(i)
∈ IR
2n
, Q
−
i,j
∈ L(IR
n
, IR
n
), i, j = 1, 2.
Then
c(a, Q[S]) = ¦z : (z −a, Q
−1
[S](z −a)) ≤ 1¦ = (2.6.23)
¦z : Σ
i,j
(x
(i)
−a
(i)
, Q
−
i,j
(x
(i)
−a
(j)
))[i, j, = 1, 2¦
Let us now intersect sets H and c(a, Q[S]) with the hyperplane x
(1)
≡ x
(2)
= L. Then
H∩ L ⊃ c(a, Q[S]) ∩ L, ∀S ∈ Σ (2.6.24)
147
( Here Σ is the set of all symmetrical matrices in IR
2n
). Moreover, due to an extension of
Corollary 2.3.1 , we will have
H∩ L = ∪¦c(a, Q[S]) ∩ L[S ∈ Σ¦ (2.6.25)
The obtained relations may be now rewritten in IR
n
. Namely, taking x
(1)
= x
(2)
= x, x ∈
IR
n
, we may observe that then
H∩ L = c
1
∩ c
2
and
c(a, Q[S]) ∩ L = ¦x : Σ
i,j
((x −a
(i)
, Q
−
i,j
(x −a
(j)
)) ≤ 1¦
We may now rearrange the previous relation and rewrite (3.6.23) as
c
1
∩ c
2
⊃ c(q[S], (1 −h
2
[S])
−1
Q
∗
[S]), (2.6.26)
where
Q
∗
[S] = (Σ
2
i,j=1
Q
−
i,j
)
−1
, h
2
[S] = Σ
2
i,j=1
(a
(i)
, Q
−
i,j
a
(j)
−(q[S], q[S]),
and
q[S] =
1
2
(Q
∗
)
1/2
[S] b[S], b[S] = Σ
2
i,j=1
(Q
−
i,j
a
(i)
+Q
−
i,j
a
(j)
)
The previous reasoning results in
Lemma 2.6.5 Suppose Assumption 2.6.1 holds: the intersection c
1
∩ c
2
of two nonde
generate ellipsoids has an interior point ( intc
1
∩ c
2
,= ∅). Then the internal ellipsoidal
approximation of c
1
∩ c
2
may be described due to formula (2.6.25), where S is any sym
metrical matrix in IR
2n
.
The following relation is true
c
1
∩ c
2
= ∪¦c(q[S], (1 −h
2
[S])
−1
Q
∗
[S])[S ∈ Σ¦, (2.6.27)
The last relation follows from (3.6.24).
148
Remark 2.6.1 Under nondegeneracy conditions similar to those of Lemma 3.6.4 relations
analogous to ( 3.6.26), (3.6.27) are true for intersections of a ﬁnite number n of ellipsoids.
An interesting exercise here would be to specify some types of optimal or extremal internal
ellipsoids and also to describe some ”minimal ” variety of internals that would nevertheless
”wipe out” the set c
1
∩c
2
” from inside”. We leave this to the interested reader. However
we shall ﬁnalize this section with yet another illustration
Example 2.6.4 Here we demonstrate some internal ellipsoids for an intersection c
1
∩c
2
of
two ellipsoids where these are marked by numbers 1,2, as before. The internal ellipsoids,
calculated due to relations (2.6.25), are unmarked ( Fig.2.6.4 (a) and (b)).
14
——!!! insert ﬁgures 2.6.4(a) and (b) !!! ——————
2.7 Finite Sums and Integrals:
External Approximations
Consider m nondegenerate ellipsoids c
i
= c(q
i
, Q
i
), q
i
∈ IR
n
, Q
i
∈ L(IR
n
, IR
n
), Q
i
> 0,
i = 1, . . . , m. Let us ﬁnd the external estimates of their Minkowski sum
o[m] =
m
i=1
c
i
(2.7.1)
which is, by deﬁnition,
c =
_
p
(1)
∈E
1
. . .
_
p
(m)
∈Em
_
p
(1)
+ . . . p
(m)
_
.
14
This example was calculated by D.Potapov.
149
We shall try to get a hint at the type of formula required. Let us ﬁrst take three ellipsoids:
c
1
= c(0, Q
1
), c
2
= c(0, Q
2
), c
3
= c(0, Q
3
).
Applying formula (2.7.1) ﬁrst to c
1
+c
2
, one comes to
c
1
+c
2
⊆ c(p[2]) = c(0, Q(p[2])),
where
Q(p[2]) = (p
1
+ p
2
)(p
−1
1
Q
1
+ p
−1
2
Q
2
),
and parameter p = p[2] of (2.2.1) is presented in the form p = p
1
/p
2
, p
1
> 0, p
2
> 0.
Applying (2.2.1) once more (now to c(p[2]) and c
3
), one obtains
c(p[3]) = c(0, Q(p[3])),
where
Q(p[3]) = (1 + p
−1
)Q(p[2]) + (1 + p)Q
3
),
with parameter p > 0 taken as
p =
p
1
+ p
2
p
3
p
3
> 0.
This gives
Q(p[3]) = (p
1
+ p
2
+ p
3
)(p
−1
1
Q
1
+ p
−1
2
Q
2
+ p
−1
3
Q
3
). (2.7.2)
Now the general assertion is as follows:
Theorem 2.7.1 The external estimate
c(p[m]) (2.7.3)
of the Minkowski sum o[m] =
m
i=1
c
i
of m nondegenerate ellipsoids c
i
= c(q
i
, Q
i
) is given
by
c(p[m]) = c(q[m], Q(p[m])) (2.7.4)
where
q[m] =
m
i=1
q
i
(2.7.5)
and
Q(p[m]) =
m
i=1
p
i
_
m
i=1
p
−1
i
Q
i
_
(2.7.6)
for each set of p
i
> 0, i = 1, . . . m.
150
Proof. The proof is given by induction, For m = 2 the statement clearly follows from
(2.2.1). Assuming (2.7.4)  (2.7.6) to be true for given m and applying (2.2.1) to
c(p[m]) +c
m+1
one comes to
q
∗
[m + 1] = q
∗
[m] + q
m+1
(2.7.7)
Q(p[m + 1]) = (1 + p
−1
)Q[m] + (1 + p)Q
m+1
. (2.7.8)
After taking p > 0 as
p =
m
i=1
p
i
p
m+1
this gives
Q(p[m + 1]) =
_
m+1
i=1
p
i
_
m+1
i=1
p
−1
i
Q
i
.
Q.E.D.
In the form of recurrence relations, one has
Q(p[k + 1]) = (1 + p
k+1
p
−1
[k])Q[k] + (1 + p
−1
k+1
p[k])Q
k+1
(2.7.9)
p[k + 1] = p[k] + p
k+1
, p
k
> 0, k = 1, . . . m. (2.7.10)
Direct calculations yield the following
Lemma 2.7.1 If the parameter p[m] = ¦p
1
, . . . p
m
¦ of (2.7.10) is selected as
p
i
= (Q
i
∗
,
∗
)
1
2
, i = 1, . . . m (2.7.11)
with ∈ IR
n
, (
∗
,
∗
) = 1 ﬁxed, then
ρ(
∗
[c(q
∗
[m], Q(p[m]) = ρ(
∗
[o[m]). (2.7.12)
151
Formula (2.7.12) implies
Lemma 2.7.2 The following relation is true
o[m] =
¦ c(q
∗
[m], Q(p[m]) [ p[m] ∈ IR
m
¦ . (2.7.13)
As in the case of two ellipsoids, the ﬁnite sum o[m] may be presented as an intersection
of ellipsoids, which now belong to the parametrized variety c(q
∗
[m], Q(p[m]).
Although the equality (2.7.13) is true, this does not mean that the variety c(q
∗
[m], Q(p[m])
contains all the inclusionminimal ellipsoids circumscribed around o[m]
The following example illustrates that in the case of adding three, (or in general more than
two) ellipsoids, the family
¦c(0, Q[p
1
, p
2
, p
3
]) : p
i
> 0, i = 1, 2, 3 ¦ , Q[p
1
, p
2
, p
3
] = Q(p[3]),
does not contain the covering ellipsoid of minimal volume.
Example 2.7.1.
Consider the segments T
i
= [A
i
, B
i
] ⊂ IR
2
, i = 1, 2, 3 where
A
1
= (−1, 0) B
1
= (1, 0)
A
2
=
_
−
1
2
,
√
3
2
_
B
2
=
_
1
2
, −
√
3
2
_
(2.7.14)
A
3
=
_
1
2
,
√
3
2
_
B
3
=
_
−
1
2
, −
√
3
2
_
.
The Minkowski sum
T =
3
i=1
T
i
is the regular hexagon, that is covered by the ball of radius 2 around the origin, o(0, 2) ⊂ IR
2
,
with
1
π
2
Vol
2
(S(0, 2)) = 16.
On the other hand
min
_
1
π
2
Vol
2
(c(0, Q[p
1
, p
2
, p
3
]) : p
i
> 0, i = 1, 2, 3
_
=
81
4
.
152
Proof. We have
T
i
= c(0, Q
i
) i = 1, 2, 3
with
Q
1
=
_
1 0
0 0
_
Q
2
=
_
3
4
−
√
3
4
−
√
3
4
1
4
_
Q
3
=
_
3
4
√
3
4 √
3
4
1
4
_
.
Consider the matrix
Q[p
1
, p
2
, p
3
] = (p
1
+ p
2
+ p
3
)
_
_
1
p
1
+
1
4p
2
+
1
4p
3
√
3
4
_
1
p
2
−
1
p
3
_
√
3
4
_
1
p
2
−
1
p
3
_
3
4
_
1
p
2
+
1
p
3
_
_
_
.
Calculating the determinant, we obtain:
det(Q[p
1
, p
2
, p
3
]) =
1
π
2
Vol
2
(c(0, Q[p
1
, p
2
, p
3
])) =
3
4
_
p
1
+ p
2
+p
3
(p
1
p
2
p
3
)
1
3
_
3
.
The well known inequality between the arithmetic and geometric mean completes the proof.
Q.E.D.
Exercise 2.7.1 Consider the variety
E[m] = c(q
∗
[m], Q(p[m]))
by vectors p[m] > 0. Select an optimal ellipsoid among those of the form c ∈ E[m] relative
to the criterion
ψ(Q(p[m])) = min
where the function ψ is one of those given in Section 2.1.1.
A further step is to approximate setvalued integrals. Assume an ellipsoidal valued function
T(t) = c(q(t), Q(t)) t ∈ [t
0
, t
1
]
with the functions q : [t
0
, t
1
] −→ IR
n
Q : [t
0
, t
1
] −→ L(IR
n
, IR
n
) continuous and Q(t) > 0
for all t ∈ [t
0
, t
1
] given. What would be its setvalued integral
153
I[t
0
, t
1
] =
_
t
1
t
0
c(q(t), Q(t))dt ?
Since the sum of a ﬁnite number of ellipsoids is not obliged to be an ellipsoid, this, obviously
is all the more true for the integral of an ellipsoidal valued function T().
With the functions q(),Q() continuous, the integral I[t
0
, t
1
] can be treated as a setvalued
Riemannintegral with integral sums
I(Σ
N
) =
N
i=1
c(q(τ
i
), Q(τ
i
))σ
i
(2.7.15)
with
Σ
N
= ¦τ
0
= t
0
, τ
i
= τ
i−1
+σi −1, σ
i
>= 0, i = 1, . . . N¦
and
σ(N) = max¦σ
i
: i = 1, . . . N¦
converging to I[t
0
, t
1
] in the Hausdorﬀmetric h for any subdivision Σ
N
lim
σ
N
→0 N→∞
h(I(Σ
N
), I[t
0
, t
1
]) = 0. (2.7.16)
In the sequel assume σ
i
= (t
1
− t
0
)/N = σ(N) for i = 0, . . . N − 1. Applying Theorem
(2.7.1) to (2.7.15) we have
I(Σ
N
) ⊆ c(q
∗
(Σ
N
), Q(Σ
N
)) (2.7.17)
where
q
∗
(Σ
N
) =
N
i=1
q(τ
i
)σ
i
and
Q(Σ
N
) =
_
N
i=1
p
∗
i
(N)
_
N
i=1
σ
2
i
(p
∗
i
)
−1
(N)Q(τ
i
),
with p
∗
i
> 0.
154
After substitution p
i
(N) = σ
−1
(N)p
∗
i
(N) the last equality transforms into
Q(Σ
N
) =
_
N
i=1
p
i
(N)σ(N)
_
N
i=1
σ
i
p
−1
i
(N)Q(τ
i
).
Assuming p : [t
0
, t
1
] −→ IR to be a continuous function with positive values, taking
p
i
(N) = p(t
0
+ σ(N))
and having in view the continuity of Q we observe
lim
N→∞
Q(Σ
N
) =
__
t
1
t
0
p(τ)dτ
___
t
1
t
0
p
−1
(τ)Q(τ)dτ = Q
+
(t
1
[p(.))
_
(2.7.18)
while
lim
N→∞
q
∗
(Σ
N
) =
_
t
1
t
0
q(τ)dτ = q
t
0
(t
1
). (2.7.19)
Making a limit transition in (2.7.17) in view of (2.7.16), (2.7.18) and (2.7.19), we arrive at
the inclusion
I[t
0
, t
1
] ⊆ c(q
t
0
(t
1
), Q
+
(t
1
[p(.))) (2.7.20)
whatever is the function p() > 0. The last argument allows to formulate
Theorem 2.7.2 An external ellipsoidal estimate for the integral I[t
0
, t
1
] is given by rela
tion (2.7.20). Moreover, the following equality holds
I[t
0
, t
1
] =
_
c(q
t
0
(t
1
), Q
+
(t
1
[p()))[p() ∈ (
+
[t
0
, t
1
],
_
(2.7.21)
where (
+
[t
0
, t
1
] denotes the open cone of continuous, positive valued functions over the
interval [t
0
, t
1
].
Equality (2.7.21) follows from propositions similar to Lemmas (2.7.1), (2.7.13), namely
from
155
Lemma 2.7.3 If the function p() ∈ (
+
[t
0
, t
1
] of (2.7.20) is selected as
p(t) = (Q(t)
∗
,
∗
)
1
2
, t ∈ [t
0
, t
1
]
with
∗
∈ IR
n
, (
∗
,
∗
) = 1 ﬁxed, then the respective support function verify the equality:
ρ(
∗
[I[t
0
, t
1
]) = ρ(
∗
[c(q
t
0
(t
1
), Q
+
(t
1
[p()))) (2.7.22)
Proof. The proof follows from direct substitution. Q.E.D.
Let us ﬁnally indicate some diﬀerential relations for q
t
0
(t) and Q(t) = Q
+
(t[p()), taking
p() to be ﬁxed. Recalling (2.7.18) we have the representation
Q
+
(t) =
__
t
t
0
p(τ)dτ
___
t
t
0
p
−1
(τ)Q(τ)dτ
_
,
or, after diﬀerentiating both sides by t and introducing the notation
π(t) = p(t)
_
t
t
0
p(τ)dτ, (2.7.23)
the diﬀerential equation
˙
Q
+
(t) = π(t)Q
+
(t) + π
−1
(t)Q(t) (2.7.24)
Q
+
(t
0
) = 0 (2.7.25)
complemented by
˙ q
t
0
(t) = q(t) (2.7.26)
q
t
0
(t
0
) = q(t
0
). (2.7.27)
Exercise 2.7.2. Prove that for the sum
c(q
0
, Q
0
) +
_
t
1
t
0
c(q(t), Q(t))dt ⊆ c(q
t
0
(t
1
), Q
+
(t
1
)) (2.7.28)
the external ellipsoidal representation is still given by equations (2.7.24), and (2.7.26), the
change appearing only in the initial conditions (2.7.25) and (2.7.27), so that Q
0
and q
0
have to be added on the respective right hand side.
156
Before ending this Section, let us single out some ”individual” external ellipsoids. We shall
discuss two ways of selecting these. Integrating relation (2.7.4), in view of (2.7.25), we
have
Q
+
(τ) =
_
τ
t
0
T(π(t), Q
+
(t), Q(t))dt,
where
T(π(t), Q
+
(t), Q(t)) = π(t)Q
+
(t) + π
−1
(t)Q(t), (2.7.29)
Let us now minimize the matrix T[t] = T(π, Q
+
(t), Q(t)) over π ( at each instant t ∈ [t
0
, τ]
), taking, for example, the following ”local” optimality criteria ( see Section 2.1)
(a)
ψ[T[t]] = Tr(T[t]),
(b)
ψ[T[t]] = Tr(T
2
[t]),
(c)
ψ[T[t]] = detT[t],
Due to the results of Section 2.5 the respective optimalizers are
(a) π(t) =
Tr
1/2
(Q
+
(t))
Tr
1/2
(Q(t)
(2.7.30)
(b) π(t) =
Tr
1/2
((Q
+
(t))
2
)
Tr
1/2
(Q(t)Q
+
(t))
(2.7.31)
(c) π(t) =
n
1/2
Tr
1/2
(Q(t)(Q
+
(t))
−1
)
(2.7.32)
Summarizing these results, we come to
157
Lemma 2.7.4 (a)The parameters Q
+
(τ) of the external ellipsoids c(q
t
0
(τ, Q
+
(τ)) = c
+
[τ]
singled out through the local optimality criteria (a),(b),(c) taken for each t ∈ [t
0
, τ] , may
be calculated due to equation (2.7.24), where the function π(t), t ∈ [t
0
, τ] has to be selected
due to equalities (2.7.30)  (2.7.32) respectively.
(b) Each of the ellipsoidal ”tubes” c
+
[t], t
0
≤ t ≤ τ, generated by equations (2.7.24)
(2.7.27) is nondominated with respect to inclusion. (In the sense that for each t the re
spective set c
+
[t] is an inclusionminimal external ellipsoidal estimate of I[t
0
, t]).
One may observe that in equation (2.7.24) the functions π(t) may treated as ( positive
valued ) controls. The problem of selecting optimal ellipsoids may then be reduced to an
openloop terminal control problem, where the nonlocal optimality criteria to be minimized
over π(t), t ∈ [t
0
, τ], could be
TrQ
+
(τ), Tr(Q
+
(τ))
2
, det(Q
+
(τ)) (2.7.33)
accordingly.
15
Exercise 2.7.3 Compare the solutions of the optimal terminal control problem for system
(2.7.24) with control π(t), due to optimality criteria (2.7.33), with the solutions obtained
due to local criteria (a),(b),(c), as speciﬁed in Lemma 2.7.4.
We shall also calculate the internal ellipsoidal approximations for ﬁnite sums of ellipsoids
and for integrals of ellipsoidalvalued functions.
2.8 Finite Sums and Integrals:
Internal Approximations
Consider again the sum
o[m] =
m
i=1
c
i
of m nondegenerate ellipsoids c
i
= c(q
i
, Q
i
). We shall introduce the internal ellipsoidal
approximation of these, assuming again , without loss of generality, that q
i
= 0, i = 1, ..., m.
Applying formula (2.4.3) to c
0
, c
1
, we have,
c
0
+ c
1
= c(0, Q
0
) +c(0, Q
1
) ⊇ c(0, Q[S
1
]),
15
One should be aware, in view of Example 2.7.1, that these criteria would be minimized only in the
class of ellipsoids described by formula (2.7.24).
158
where S[1] = S
1
and
Q(S[1]) = S
−1
1
[(S
1
Q
0
S
1
)
1
2
+ (S
1
Q
1
S
1
)]
2
S
−1
. (2.8.1)
Moreover , the representation of Corollary 2.4.1 yields
c(0, Q
0
) + c(0, Q
1
) = ∪¦c(0, Q(S[1]))[S[1] ∈ Σ¦. (2.8.2)
Continuing this procedure, we have, due to the same representations
o[2] = c
0
+c
1
+c
2
⊇ c(0, Q(S[2])), (2.8.3)
∀S[2] = ¦S
1
, S
2
¦, S
i
∈ Σ,
where
Q(S[2]) = S
−1
2
[(S
2
Q(S[1])S
2
)
1
2
+ (S
2
Q
2
S
2
)
1
2
]
2
S
−1
2
.
Further on, assuming that the last relations are true for S[m−1], we have,
o[m] =
m
i=1
c
i
⊇ c(0, Q(S[m−1])) + c(0, Q
m
) ⊇ c(0, Q(S[m]), (2.8.4)
∀S[m] = ¦S
1
, .., S
m
¦,
where
Q(S[k]) = S
−1
k
[(S
k
Q(S[k −1])S
k
)
1
2
+ (S
k
Q
k
S
k
)
1
2
]
2
(S
)
−1
k
, (2.8.5)
S[k] = ¦S
1
, ..S
k
¦, Q(S[0]) = Q
0
.
Applying the representation of Corollary 2.4.1 to (2.8.3), we come to
c
0
+c
1
+c
2
= ∪¦c(0, Q(S[1]))[S
1
¦ + c(0, Q
2
),
c(0, Q(S[1])) + c(0, Q
2
) = ∪¦c(0, Q(S[2]))[S
2
¦,
159
which gives
o[2] = ∪ ∪ ¦c(0, Q(S[2])[S
1
, S
2
¦.
Similarly, by induction,
o[m] = ∪¦c(0, Q(S[m−1]))[S[m−1]¦ + c(0, Q
m
) = (2.8.6)
= ∪¦c(0, Q(S[m]))[S[m]¦.
Concluding the discussion, we are now able to formulate
Theorem 2.8.1 The internal ellipsoidal estimate
c
−
[m] ⊆
m
i=1
c(0, Q
i
) = o[m],
for the sum o[m] of m+1 nondegenerate ellipsoids c(0, Q
i
) is given by the inclusion (2.8.4)
with exact representation (2.8.6), where the union is taken over all the sequences S[m] of
symmetrical matrices S
i
∈ Σ, i = 1, ..., m.
The general case, with
o[m] =
m
i=0
c(q
i
, Q
i
),
is treated similarly. This allows
Corollary 2.8.1 The inclusion
o[m] ⊇ c
−
[m] = c(q[m], Q(S[m])), (2.8.7)
q[m] =
m
i=1
q
i
,
holds for any sequence S[m]. The following representation is true
o[m] = ∪¦c(q[m], Q[m])[S[m]¦. (2.8.8)
160
Remark 2.8.1 The last assertions were proved for the sum o[m] of m+1 nondegenerate
ellipsoids c
i
, i = 1, ...m. The basic relations turn to be also true if these are degenerate.
However, the union in the righthand side of (2.9.8) has to be substituted by its closure.
Exercise 2.8.1. Prove the assertion of the previous remark.
Let us now pass to the internal approximation of the setvalued integral
I[t
0
, t
1
] =
_
t
1
t
0
c(q(t), Q(t))dt.
Its Riemannean integral sum is the one given in (2.7.15) with convergence property (2.7.16).
Applying Theorem 2.8.1 to (2.7.15), we observe
I(Σ
k
) ⊇ c(
k−1
i=1
q
i
, Q
σ
(S[k −1]) + c(q
k
, σ
2
Q
k
) ⊇ (2.8.9)
⊇ c(
k
i=1
σ
i
q
i
, Q
σ
(S[k])
and
I(Σ
k
) = ∪¦c(
k
i=1
σ
i
q
i
, Q
σ
(S[k]))[S[k]¦ (2.8.10)
where S[k] is such that S
i
∈ Σ, i = 1, ..., k, and
Q
σ
(S[k]) = S
−1
k
(S
k
Q
σ
(S[k −1])S
k
)
1
2
+ σ
k
(S
k
Q
k
S
k
)
1
2
]
2
S
−1
k
The last relations are equivalent to
Q
σ
(S[k]) −Q
σ
(S[k −1]) =
= σ
k
S
−1
k
((S
k
Q
σ
(S[k −1])S
k
)
1
2
(S
k
Q
k
S
k
)
1
2
+ σ
k
(S
k
Q
k
S
k
)
1
2
(S
k
Q
σ
(S[k −1])S
k
)
1
2
)S
−1
k
+
+ σ
2
k
(S
k
Q
k
S
k
).
Denoting
τ
k
= t, Σ
k
= ¦τ
0
= t
0
, τ
i
= τ
i−1
+ σ
i−1
, σ
i
> 0, i = 1, ..., k¦
and
S[k] = S[t], S[i] = S[τ
i
] = ¦S(τ
j
); j = 0, ..., i¦, S(τ
j
) = S
j
,
161
q(i) = q(τ
i
), q
σ
(t) =
k
i=1
q(i)σ
i
we observe, that the previous relations may be rewritten as
(2.8.11)
Q
σ
(S[t]) − Q
σ
(S[t −σ]) = σ
k
S
−1
[t](S[t]Q
σ
(S[t −σ
k
])S
[t])
1
2
(S[t]Q(t)S
[t])
1
2
+
+ σ
k
(S[t]Q[t]S
[t])
1
2
(S[t]Q
σ
(S[t −σ
k
])S
[t])
1
2
)S
−1
[t] + σ
2
k
(S[t]Q[t]S
[t]),
q
σ
(t) − q
σ
(t −σ
k
) = q(t)σ
k
. (2.8.12)
Let us assume that, the values S[τ
i
] in the above are generated by a measurable, matrix
valued function S[τ], τ ∈ [t
0
, t] with values in Σ.
Passing to the limit in (2.8.11),(2.8.12) with
max¦σ
i
, i = 1, ..., k¦ → 0, k → ∞, (2.8.13)
( for an arbitrary t ) and denoting
limq
σ
(t) = q
0
(t), limQ
σ
(S[t]) = Q
−
(t)
we arrive at the diﬀerential equations
dQ
−
(t)/dt = S
−1
(t)((S[t]Q
−
(t)S
[t])
1
2
(S[t]Q(t)S
[t])
1
2
+ (2.8.14)
+ (S[t]Q(t)S
[t])
1
2
(S[t]Q
−
(t)S
[t])
1
2
)S
−1
(t),
dq
0
(t)/dt = q(t), (2.8.15)
Q
−
(t
0
) = 0 , q
0
(t
0
) = 0.
The inclusion (2.8.9) and the relation
lim
k→∞
I[Σ
k
] = I[t
0
, t] (2.8.16)
imply
162
Lemma 2.8.1 The inclusion
I[t
0
, t] ⊇ c(q
0
(t), Q
−
(t)), (2.8.17)
is true, whatever is the measurable function S[t] with values in Σ.
Further, since (2.8.10) is true for any value of k and since (2.8.16) is true with (2.8.13),
the limit transition in (2.8.10)(2.8.10) , (2.8.16) yields
I[t
0
, t] = cl¦∪¦c(q
0
(t), Q
−
(t))[S[t]¦¦ (2.8.18)
over all measurable functions S[] of the type considered above. The result may be sum
marized in
Theorem 2.8.2 The integral I[t
0
, t] allows an internal approximation (2.8.17) where
q
0
(t), Q
−
(t) satisfy the diﬀerential equations (2.8.14),(2.8.15) with zero initial conditions .
The representation (2.8.18) is true , where the union is taken over all measurable functions
S[t] with values in Σ.
An obvious consequence of this Theorem is
Corollary 2.8.2 The sum
c(q
0
, Q
0
) +
_
t
1
t
0
c(q(t), Q(t))dt = I[t
0
, t
1
]
allows an internal approximation (2.8.17) and a representation of type (2.8.18) , where
q
0
(t), Q(t) are the solutions to the diﬀerential equations (2.8.14),(2.8.15) with initial con
ditions
q
0
(t
0
) = q
0
, Q(t
0
) = Q
0
. (2.8.19)
We ﬁnally oﬀer the reader to formulate and solve a problem similar to Exercise 2.8.1, but
taken for internal ellipsoids.
This Section ﬁnalizes Part II. We shall now apply the results of this part to the problems
of Part I.
163
Part III. ELLIPSOIDAL DYNAMICS :
EVOLUTION and CONTROL SYNTHESIS
Introduction
In this part we apply the calculus of Part II to the problems of Part I. We start from systems
with no uncertainty, constructing external and internal ellipsoidal valued approximations
of the attainability (reachability) domains and tubes. In order to achieve these results
we introduce two corresponding types of evolution funnel equations with ellipsoidalvalued
solutions. Each of these evolution equations generates a respective variety of ellipsoidal
valued tubes that approximate the original attainability tubes externally or internally
and ﬁnally yield, through their intersections or unions, an exact representation of the
approximated tube. This result is similar to those achieved for static situations in Sections
2.2. 2.4, but is now given for a dynamic problem ( Sections 3.2, 3.3). The main point,
however, is that the timevarying coeﬃcients of the approximating ellipsoidal tubes are
further described through ordinary diﬀerential equations with righthand sides depending
on parameters. The same result is given in backward time (Section 3.4). This allows to
us the internal approximations for synthesizing the control strategies in the target control
problem. It is shown that the scheme of Section 1.4 remains true except that the the
solvability tube of Deﬁnition 1.4.3 is substituted for its internal ellipsoidal approximation
and the control strategy is constructed accordingly ( Section 3.6 ).
The speciﬁc advantage of such solutions is that the strategies are given ( relative to the
solution of a simple algebraic equation) in the form of an analytical design.
One should realize, however, that attainability domains for linear systems are among the
relatively simpler constructions in control theory. The problem is substantially more diﬃ
cult if the system is under the action of uncertain ( unknown but bounded) inputs. The
approximation of the domains of attainability under counteraction or of the solvability do
mains for uncertain systems requires, in its general setting, to incorporate both internal
and external approximations of sums or geometrical (Minkowski) diﬀerences of ellipsoids.
The external and internal ellipsoidal approximations of the solvability tubes for uncertain
systems are derived in Section 3.5 ( under conventional nondegeneracy conditions). The
important point is that these ellipsoidal approximations that reﬂect the evolution dynam
ics of uncertain or conﬂictcontrol systems are again described through the solutions of
ordinary diﬀerential equations. Once the internal approximation of the solvability tubes
are known, it is again possible, ( now following the schemes of Section 1.8), to implement
an ” ellipsoidal ” control synthesis in the form of an analytical design ( relative to the
solution of an algebraic equation). Moreover, the ellipsoidal solvability tubes constructed
here are such that they retain the property of being ”Krasovski bridges”. Namely, once
the starting position is in a speciﬁc internal ellipsoidal solvability tube, there exists an
164
analytical control design that keeps the trajectory within this tube despite the unknown
disturbances.
We should emphasize the key elements that allow to use the ellipsoidal tubes introduced
here for designing synthesizing control strategies ( both with and without uncertainty).
These are – ﬁrst – that the approximating ( internal) ellipsoidal tubes are nondominated
with respect to inclusion, their crossections being inclusionmaximal at each instant of
time and – second – that the respective ellipsoidalvalued mappings satisfy a generalized
semigroup property which we call the lower and upper semigroup property – for internal
and external tubes accordingly. It is these two elements that allow the internal ellipsoidal
approximations to retain the property of being bridges, speciﬁcally, to be the ”ellipsoidal
valued bridges”.
The techniques of this part are illustrated in Sections 3.7, 3.9, where one may observe some
examples on solvability tubes and ellipsoidal control synthesis for 4dimensional systems
animated through computer windows.
3.1 EllipsoidalValued Constraints
Let us again consider system (1.1.1) and pass to its transformed version (1.14.1), where
A(t) ≡ 0. Namely, taking
˙ x = u + f(t) (3.1.1)
x(t
0
) = x
0
, t
0
≤ t ≤ t
1
,
we shall further presume the constraints on u, f, x
0
to be ellipsoidal–valued:
(u −p(t), P
−1
(t)(u −p(t)) ≤ 1, (3.1.2)
(f −q(t), Q
−1
(t)(f −q(t)) ≤ 1, (3.1.3)
(x
0
−x
∗
, X
−1
0
(x
0
−x
∗
)) ≤ 1, (3.1.4)
where the continuous functions p(t), q(t) and the vector x
∗
are given together with contin
uous matrix functions P(t) > 0, Q(t) > 0 and matrix X
0
> 0.
In terms of inclusions we have,
u ∈ c(p(t), P(t)) , (3.1.5)
165
f ∈ c(q(t), Q(t)) , (3.1.6)
x
0
∈ c(x
∗
, X
0
) (3.1.7)
or, in terms of support functions, the inequalities
(l, u) ≤ (l, p(t)) + (l, P(t)l)
1
2
, (3.1.8)
(l, f) ≤ (l, q(t)) + (l, Q(t)l)
1
2
, (3.1.9)
(l, x
0
) ≤ (l, x
∗
) + (l, X
0
l)
1
2
(3.1.10)
∀l ∈ IR
n
.
With f(t) given, the attainability domain A[t] = A(t, t
0
, c(x
∗
, X
0
)) is deﬁned by the set
valued integral (1.3.1), which is now
A[t] = c(x
∗
, X
0
) +
t
_
t
0
c(p(t) + f(t), P(t))dt. (3.1.11)
With f(t) continuous, the setvalued function A[t] satisﬁes the evolution equation, based
on (1.3.3) or (1.14.6)
lim
σ→0
σ
−1
h(A[t +σ], A[t] + σc(p(t) + f(t), P(t)) = 0 (3.1.12)
with boundary condition
A[t
0
] = c(x
∗
, A
0
)
for the attainability tube A[]. On the other hand, with terminal set / being an ellipsoid,
/ = c(m, M), m ∈ IR
n
, M ∈ L(IR
n
, IR
n
), M > 0, (3.1.13)
we have an evolution equation
lim
σ→0
σ
−1
h(W[t −σ], W[t] −σc(p(t) + f(t), P(t)) = 0 (3.1.14)
W[t
1
] = c(m, M)
for the solvability tube .
166
Passing to an uncertain system with f(t) measurable, bounded by restriction (3.1.3), we
come to the equation for the solvability tube under uncertainty, which is
16
lim
σ→0
σ
−1
h(W[t −σ] + σc(q(t), Q(t)), W[t] − σc(p(t), P(t))) = 0,
W[t
1
] = c(m, M). (3.1.15)
After the introduction of an additional ellipsoidalvalued state constraint
Gx(t) ∈ c(y(t), K(t)) (3.1.16)
with K(t) ∈ L(IR
k
, IR
k
), K(t) > 0, y(t) ∈ IR
k
, y(t), K(t) continuous, the equation for the
solvability tube under both uncertainty and state constraints is as follows
lim
σ→0
σ
−1
h
+
(W[t −σ] + σc(q(t), Q(t)), W[t] ∩ c(y(t), K(t)) −σc(p(t), P(t))) = 0
W[t
1
] = c(m, M) ∩ c(y(t
1
), K(t
1
)). (3.1.17)
(The solvability tube is the maximal solution to (3.1.17)).
If the function u(t) is given, and the constraint (3.1.16) is due to a measurement equation
with observed values y(t) (assuming u(t), y(t) to be continuous), then the attainability
domain A[t] for system (3.1.1), (3.1.6), (3.1.7), (3.1.16) is the corresponding information
domain which satisﬁes the evolution equation
(3.1.18)
lim
σ→0
σ
−1
h
+
(A[t + σ], A[t] ∩ c(y(t), K(t)) + σc(q(t) + u(t), Q(t))) = 0,
A[t
0
] = c(x
∗
, A
0
).
(A[t] is the maximal solution to this equation). The set A[t] gives a guaranteed estimate of
the state space vector x(t) of system (3.1.11) (u(t) given) , under unknown but bounded
disturbances f(t) ∈ c(q(t), Q(t)), through the measurement of vector
y(t) ∈ G(t)x(t) + c(0, K(t)). (3.1.19)
As we have observed in Part II, the sets A[t], W[t] generated by the solutions to the
evolution equations of this Section, are not obliged to be ellipsoids. We shall therefore
introduce external and internal ellipsoidal approximations of these within a scheme that
would generalize the results of Part II, propagating them to continuoustime dynamic
processes. Our further subject is therefore the one of ellipsoidalvalued dynamics. Following
the sequence of topics of Part I, we start from the simplest attainability problem.
16
We recall that this equation was introduced under nondegeneracy Assumptions 1.7.1 or 1.7.2 which
imply that the tube W[t], t ∈ [t
0
, t
1
] contains an internal tube β(t)S, β(t) > 0 .
167
3.2 Attainability Sets and Attainability Tubes:
the External and Internal Approximations
Our ﬁrst subject is to consider the diﬀerential inclusion
˙ x ∈ c(p(t), P(t)) + f(t), t
0
≤ t ≤ t
1
, (3.2.1)
x(t
0
) = x
0
, x
0
∈ c(x
∗
, X
0
),
and to approximate its attainability domain A[t] = X(t, t
0
, c(x
∗
, X
0
)), where
A[t] = c(x
∗
, X
0
) +
t
_
t
0
c(p(s) + f(s), P(s))ds. (3.2.2)
The external ellipsoidal approximation for such a sum had been indicated in Section 2.7,
particularly, through relations (2.7.20), (2.7.24), (2.7.22). Applying these relations to the
present situation and changing the notations to those of (3.2.2), we have
A[t] ⊆ c(x
∗
(t), X
+
(t)), (3.2.3)
where
˙ x
∗
(t) = p(t) + f(t), (3.2.4)
˙
X
+
(t) = π(t)X
+
(t) + π
−1
(t)P(t), π(t) > 0, (3.2.5)
x
∗
(t
0
) = x
∗
, X
+
(t
0
) = X
0
. (3.2.6)
Here X
+
(t) actually depends on π(), so that if necessary, we shall also use the notation
X
+
(t) = X
+
(t[π()).
It follows from Theorem 2.7.2 and the substitution (2.7.23) that the inclusion
168
A[t] ⊆ c(x
∗
(t), X
+
(t[π()) (3.2.7)
is true, whatever is the function π(t) > 0 that allows representation (2.7.20), (2.7.23) with
p(t) > 0. Moreover, the equality
A[t] = ∩¦c(x
∗
(t), X
+
(t[π())[π()¦ (3.2.8)
is true if the intersection is taken over all the functions π() of the type indicated above.
We leave to the reader to observe that (3.2.7) remains true if the intersection is taken over
all piecewise continuous or even continuous functions π(t) > 0. This ﬁnally leads to the
proof of the following assertion
Theorem 3.2.1 The external ellipsoidal approximation to the attainability domain A[t] =
c(x
∗
(t), X
+
(t)) of the diﬀerential inclusion (3.2.1) is given by the inclusion (3.2.7) with
exact representation (3.2.8), where the intersection may be taken over all piecewise–
continuous (or even continuous) functions π(t) > 0.
Let us now return to the last theorem, approaching it through another scheme  the tech
nique of funnel equations. Following Sections 1.4, 3.1, we observe that the tube A[t]
satisﬁes the funnel equation (3.1.12). This allows to write
A[t + σ] ⊆ A[t] + σc(p(t), P(t)) + o(σ)S,
where σ
−1
o(σ) → 0 if σ → 0, and S is a unit ball in IR
n
, as before.
With A[t] being an ellipsoid of type A[t] = c(x
∗
(t), X
+
(t)), we may apply the expansion
(2.5.6), so that the external approximation to A[t + σ] would be
A[t + σ] ⊆ c(x
∗
(t + σ), X
+
(t + σ)) , (3.2.9)
where
X
+
(t + σ) = X
+
(t) + σπ
−1
(t)X
+
(t) + σπ(t)P(t) + σ
2
P(t), (3.2.10)
with π(t) > 0 continuous. Relations (3.2.9) , (3.2.10) are true for any σ > 0 and any π(t)
of the indicated type.
169
Dividing the interval [t
0
, t] into subintervals with subdivision
= ¦σ
1
, ..., σ
s
¦,
τ
0
= t
0
, τ
s
= t
0
+
s
i=0
σ
i
, t = τ
s
,
where
σ > 0,
s
i=0
σ
i
= t − t
0
,
we have :
A(τ
1
) = c(x
∗
, X
+
(t
0
)) + σ
1
c(p(t
0
) + f(t
0
), P(t)) ⊆
⊆ c(x
∗
(τ
1
), X
+
(τ
1
)) = c
+
[τ
1
],
where
x
∗
(τ
1
) = x
∗
+ σ
1
(p(t
0
) + f(t
0
)) (3.2.11)
X
+
(τ
1
) = (1 + σ
1
π
−1
(t
0
))X
+
(t
0
) + σ
1
π(t
0
)P(t
0
) + σ
2
1
P(t
0
). (3.2.12)
We further have:
A
+
(τ
k
) ⊆ c(x
∗
(τ
k−1
), X
+
(τ
k−1
)) + σ
k
c(p(τ
k−1
+ f(τ
k−1
), P(τ
k−1
)) ⊆
⊆ c(x
∗
(τ
k
), X
+
(τ
k
)), (k = 1, ..., s),
where
x
∗
(τ
k
) = x
∗
(τ
k−1
) + σ
k
(p(τ
k−1
) + f(τ
k−1
)), (3.2.13)
X
+
(τ
k
) = (1 + σ
k
π
−1
(τ
k−1
))X
+
(τ
k−1
) + σ
k
π(τ
k−1
)P(τ
k−1
). (3.2.14)
Dividing relations (3.2.13), (3.2.14) by σ
k
and passing to the limit, with
170
max¦σ
k
[k = 1, ..., s¦ → 0 , s → ∞,
and t being ﬁxed as the endpoint of the interval [t
0
, t), whatever is the subdivision
and the integer s, we again come to equations (3.2.4), (3.2.5) with initial condition (3.2.6).
This gives an alternative proof for the relation (3.2.7) of Theorem 3.2.1.
Let us now assume A(t) ,≡ 0. Then Theorem 3.2.1 transforms into
Corollary 3.2.1 For every t ∈ [t
0
, t
1
] the following equality is fulﬁlled
A(t) = ∩¦c(x(t), X
+
(t[π())[π())¦,
where X
+
(t) = X
+
(t[π()), are the solutions of the following diﬀerential equations
˙ x = A(t)x + p(t); x(t
0
) = x
∗
,
˙
X
+
= A(t)X
+
+ X
+
A
(t) + π
−1
(t)X
+
(t) + π(t)P(t); X
+
(t
0
) = X
0
.
We shall now indicate that with a certain modiﬁcation this result remains true for the
special case when c(p(t), P(t)) is a product of ellipsoids and therefore does not fall under
the requirements of Section 3.1. Let us start from a generalization of Lemma 2.2.1.
Lemma 3.2.1 Suppose E
1
= c(q
1
; Q
1
), E
2
= c(q
2
; Q
2
) where
Q
1
=
_
A
1
0
0 0
_
, Q
2
=
_
0 0
0 A
2
_
,
A
1
(A
2
) is a symmetric positively deﬁned mm (respectively, k k−) matrix, m+k = n.
Then
E
1
+E
2
= ∩¦c(q
1
+ q
2
, Q(p))[p > 0¦,
Q(p) =
_
(1 + p
−1
)A
1
0
0 (1 + p)A
2
_
.
171
Proof. The upper estimate
E
1
+ E
2
⊆ c(q
1
+ q
2
; Q(p)), p > 0,
can be obtained along the lines of the proof of Lemma 2.3.1.
Consider now an arbitrary vector v = ¦l, b¦ ∈ IR
n
, l ∈ IR
m
, b ∈ IR
k
such that l ,= 0, b ,= 0.
It is not diﬃcult to demonstrate that
ρ(v[E
1
+ E
2
) = v
(q
1
+ q
2
) + (l
A
1
l)
1
2
+ (b
A
2
b)
1
2
=
= v
(q
1
+ q
2
) + (v
Q(p)v)
1
2
for p = (l
A
1
l)
1
2
/(b
A
2
b)
1
2
.
This yields
ρ(v[E
1
+ E
2
) = ρ(v[c(q
1
+ q
2
, Q(p))),
for every direction v = ¦l, b¦ with l ,= 0, b ,= 0. From the continuity of the support
functions of the convex compact sets E
1
+E
2
and of the set ∩¦c(q
1
+q
2
, Q(p))[p > 0¦ we
conclude that equality
ρ(v[E
1
+ E
2
) = ρ(v[ ∩ ¦c(q
1
+ q
2
; Q(p))[p > 0¦)
is true for all v ∈ IR
n
. The last relation implies the assertion of this Lemma Q.E.D.
Denote symbol Π
+
[t
0
, t
1
] to stand for the set of all continuous positive functions from [t
0
, t
1
]
to IR
1
.
Combining the last Lemma with Corollary 3.2.1, we come to the conclusion
Corollary 3.2.2 Consider the diﬀerential inclusion
˙ x ∈ A(t)x +T(t),
x(t
0
) ∈ c(x
∗
, X
0
), t
0
≤ t ≤ t
1
,
with
172
T(t) = c
k
(s(t), S(t)) c
m
(q(t), Q(t)),
where c
k
(s(t), S(t)) ⊂ IR
k
, c
m
(q(t), Q(t)) ⊂ IR
m
, k + m = n.
For every t ∈ [t
0
, t
1
] the following equality is true
X(t) = ∩¦c(z(t), Z(t[π(), χ())[π(), χ()¦
where ¦π(), χ()¦ ∈ Π
+
[t
0
, t
1
], and
z : [t
0
, t
1
] → IR
n
, Z : [t
0
, t
1
] → L(IR
n
, IR
n
)
are the solutions to diﬀerential equations
˙ z = Az + v(t), v(t) = ¦s(t), q(t)¦, z(t
0
) = x
∗
.
˙
Z = A(t)Z + ZA
(t) + χ
−1
(t)Z + χ(t)Q[t],
Q[t] = Q(t, π(t)) =
_
(1 + π
−1
(t))S(t) 0
0 (1 + π(t))Q(t)
_
and Z(t
0
) = X
0
.
In order to deal with internal approximations, we will follow the last scheme, dealing now
with funnel equation (3.1.12). This time we have
A[t] + σc(p(t), P(t)) ⊆ A[t +σ] + o
1
(σ)S, (3.2.15)
where
σ
−1
o
1
(σ) → 0 , σ → 0.
With A[t] being an ellipsoid of type
173
A[t] = c(x
∗
(t), X
−
(t)) ,
we may apply formula (2.3.3) with (2.3.1). Changing the respective notations, namely,
taking
Q
1
= X
−
(t) , Q
2
= σ
2
P(t) , S = H(t) ,
we have, in view of (3.2.15)
A[t + σ] ⊇ c(x
∗
(t + σ), X
−
(t + σ)),
where
X
−
(t + σ) = H
−1
(t)[H(t)X
−
(t)H
(t) + σ(H(t)X
−
(t)H
(t))
1
2
(H(t)P(t)H
(t))
1
2
+
+ σ(H(t)P(t)H
(t))
1
2
(H(t)X
−
(t)H
(t))
1
2
+ σ
2
H(t)P(t)H
(t)]H
−1
(t),
and
x
∗
(t +σ) = x
∗
(t) + σ(p(t) + f(t)).
After a discretization and a limit transition in the last equations, similar to the one for the
external approximations of the above , we come to ordinary diﬀerential equations which
are equation (3.2.4) and
dX
−
(t)/dt = H
−1
((H(t)X
−
(t)H
(t))
1
2
(H(t)P(t)H
(t))
1
2
+ (3.2.16)
+ (H(t)P(t)H
(t))
1
2
(H(t)X
−
(t)H(t))
1
2
)H
−1
,
with initial conditions
x
∗
(t
0
) = x
0
, X
−
(t
0
) = X
0
.
174
What follows from here is the inclusion
A[t] ⊇ c(x
∗
(t), X
−
(t)), (3.2.17)
where x
∗
(t), X(t) satisfy (3.2.4), (3.2.16) and H(t) is a continuous function of t with val
ues in Σ  the variety of symmetric matrices. A detailed proof of the same inclusion
follows from Theorem 2.9.1, where one just has to change notations S(t), q
0
(t), Q
−
(t) to
H(t), x
∗
(t), X
−
(t) respectively. The given reasoning allows to formulate
Theorem 3.2.2 The internal approximation of the attainability domain A[t] =
X(t, t
0
, c(x
∗
, X
0
)) of the diﬀerential inclusion (3.2.1) is given by the inclusion (3.2.17)
,where x
∗
(t), X
−
(t) satisfy the equations (3.2.4), (3.2.16). Moreover, the following repre
sentation is true
A[t] = ∪¦c(x
∗
(t), X
−
(t))[H()¦, (3.2.18)
where the union is taken over all measurable matrixvalued functions with values in Σ.
Relation (3.2.18) is a direct consequence of Corollary 2.8.1 .
One may remark, of course, that all the earlier conclusions of this Section were made under
the assumptions that ellipsoids c(p(t), P(t)), c(x
∗
, X
0
) are nondegenerate. However, the
given relations still remain true under relaxed assumptions that allow degeneracy in the
following sense.
Consider system
˙ x ∈ A(t)x + B(t)c(p(t), P(t)), (3.2.19)
x(t
0
) = x
0
, x
0
∈ c(x
∗
, X
0
),
where B(t) is continuous, p(t) ∈ IR
m
, P(t) ∈ L(IR
m
, IR
m
), m < n.
The parameters of this sytem allow to generate the setvalued integral
A
∗
[t] =
t
_
t
0
S(τ, t)B(τ)c(0, P(τ))dτ, (3.2.20)
where matrix S(τ, t) is deﬁned in Section 1.1, see (1.1.6).
175
Assumption 3.2.1 There exists a continuous scalar function β(t) > 0, t > t
0
such that
the support function
ρ(l[A
∗
[t]) ≥ β(t)(l, l)
1/2
,
for all t > t
0
.
This assumption implies that the attainability domain A[t] of system (3.2.19) has a
nonempty interior ( intA[t] ,= ∅). It is actually equivalent to the requirement that system
( 3.2.19) with unrestricted control u(t) would be completely controllable, [147], [212], on
every ﬁnite time interval [t
0
, t].
Under the last Assumption the analogies of Theorems 3.1, 3.2 for system (3.2.19) still
remain true. Namely, taking equations
(3.2.21)
˙
X
+
(t) = A(t)X
+
(t) + X
+
(t)A
(t) + π(t)X
+
(t) + π
−1
(t)B(t)P(t),
˙ x
∗
(t) = A(t)x
∗
+ B(t)p(t) + f(t), (3.2.22)
we have the following assertion.
Lemma 3.2.2 Suppose Assumption 3.2.1 for system (3.2.19) is true. Then the results
of Theorems 3.2.1 and Corollary 3.2.1 (for the attainability domain A[t] of this system)
remain true with equations (3.2.5),(3.2.4) substituted by ( 3.2.2), (3.2.21) .
The details of the proof follow the lines of Section 2.7 and the reasoning of the present
Section.
Remark 3.2.1 An assertion similar to Lemma 3.2.2 is also true for internal ellipsoidal
approximations ( under the same Assumption 3.2.1). System (3.2.16) is then substituted
by
dX
−
(t)/dt = (3.2.23)
= A(t)X
−
(t) + X
−
(t)A(t) + H
−1
((H(t)X
−
(t)H
(t))
1
2
(H(t)P(t)H
(t))
1
2
+
+ (H(t)P(t)H
(t))
1
2
(H(t)X
−
(t)H(t))
1
2
)H
−1
,
176
Exercise 3.2.1,
Prove the statement of Remark 3.2.1.
Remark 3.2.2 It is now possible to single out individual ellipsoidal tubes that approximate
A[t] externally or internally. This, particularly, may be done as described in Lemma 2.7.4
(due to a ”local” optimality condition) or due to nonlocal criteria, of types (2.7.33), for
example ( see Exercise 2.7.3).
We emphasize once again that that functions π(t), H(t) in equations (1.2.5), (1.2.16) may
be interpreted as controls which, for example, may be selected on an interval [t
0
, τ] so as
to optimalize the terminal ellipsoid c(x
∗
(τ), X
+
(τ)) or c(x
∗
(τ), X
−
(τ)) in the classes of
ellipsoids determined by equations (3.2.21)(3.2.23).
The next natural step would be to introduce ellipsoidal approximations for solvability tubes.
Prior to that, however, we shall introduce some evolution equations for ellipsoidalvalued
mappings.
3.3 Evolution Equations
with Ellipsoidal  Valued Solutions
Having found the external and internal ellipsoidal approximations for the attainability
domains A[t] and recalling that A[t] satisﬁes an evolution ” funnel” equation, we come to
what seems to be a natural question : do the ellipsoidal mappings that approximate A[t]
satisfy, in their turn, some evolution equations with ellipsoidalvalued solutions ? Let us
investigate this issue.
Writing down the evolution equation (3.1.12) for X[t] with the ellipsoidal data of Section
3.1, we have
lim
σ→0
σ
−1
h(A[t + σ], A[t] + σc(p(t), P(t) + f(t))) = 0, (3.3.1)
A[t
0
] = c(x
0
, X
0
).
As indicated in the above, it should be clear that in general the solution to (3.3.1) is not
ellipsoidalvalued.
Let us now introduce another equation, namely,
177
lim
σ→0
h
−
(c[t + σ], c[t] + σc(p(t), P(t) + f(t)l)) = 0, (3.3.2)
c(t
0
) = c(x
0
, X
0
).
Deﬁniton 3.3.1 A function c
+
[t] is said to be a solution to the evolution equation (3.3.2)
if
(i) c
+
[t] satisﬁes (3.3.2) almost everywhere,
(ii) c
+
[t] is ellipsoidalvalued,
(iii) c
+
[t] is the minimal solution to (3.3.2) with respect to inclusion.
From the deﬁnition of the semidistance h
−
and of the solution c
+
[t] (points it (i),(ii)), it
follows that always
c
+
[t] ⊇ A[t],
so that c
+
[t] is an external approximation of A[t].
Lemma 3.3.1 The external approximation
c
+
[t] = c(x
∗
(t), X
+
(t[π()))
is a solution to the equation (3.3.2) in the sense of Deﬁnition 3.3.1, provided π(τ) > 0 is
selected as
π(τ) = (l
∗
(τ), P(τ)l
∗
(τ))
1
2
, (l
∗
(τ), l
∗
(τ)) = 1, t
0
≤ τ ≤ t, (3.3.3)
where l
∗
(τ) is a measurable function of t.
This follows, due to (3.3.1), (3.2.7), from the inclusion
c
+
(x
∗
(t + σ), X
+
(t + σ[π()) + o(σ)S ⊇ c(x
∗
(t), X
+
(t[π())) + σc(p(t), P(t)),
that ensures the ellipsoidalvalued function c[t] to satisfy (3.3.2) and from the equalities
ρ(l[c
+
[t]) = ρ(l[I
0
[t
0
, t
1
]),
I
0
[t
0
, t
1
] = c(x
0
, X
0
)) +
_
t
1
t
0
c(p(t), P(t))dt,
(taken for π(t) selected due to (3.3.3)) that ensure the minimality property (iii) of Deﬁnition
3.3.1.
The last lemma indicates that the solution to equation (3.3.2) is not unique. This is all
the more true due to
178
Corollary 3.3.1 The ellipsoidal function c
+
[t] = c(x
∗
(t), X
+
(t[π())) is a solution to
(3.3.2) , whatever is the measurable function π(τ) > 0, t
0
≤ τ ≤ t
1
selected due to the
inequalities
min¦(l, P(τ)l)[(l, l) = 1¦ ≤ π(τ) ≤ max¦(l, P(τ)l)[(l, l) = 1¦.
The proof of this corollary follows from the results of Sections 2.3, 2.7.
For a given function π(t) and given initial pair x
∗
, X
∗
, we shall also denote
E
+
(t, τ, c(x
∗
, X
∗
)) = c(x
∗
(t, τ, x
∗
), X
+
(t, τ, X
∗
)),
where x
∗
(t, τ, x
∗
), X
+
(t, τ, X
∗
) satify (3.2.4),(3.2.16) with
x
∗
(τ, τ, x
∗
) = x
∗
, X
+
(τ, τ, X
∗
) = X
∗
.
Then, obviously
E
+
(t, τ, c(x
∗
, X
∗
)) ⊇ c(x
∗
, X
∗
) +
_
t
τ
c(p(s), P(s))ds
and a direct substitution leads to
Lemma 3.3.2 The following inclusions are true
E
+
(t, τ, E
+
(τ, t
0
, c(x
0
, X
0
))) ⊇ E
+
(t, t
0
, c(x
0
, X
0
)), (3.3.4)
t
0
≤ τ ≤ t.
Relations (3.3.4) describe the dynamics of the external ellipsoidal estimates E
+
(t, τ, c(x
∗
, X
∗
))
. They thus deﬁne an ”upper” semigroup property of the respective mappings. The sets
E
+
(t, t
0
, c(x
0
, X
0
)) are sometimes referred to as supperattainability domains .
Together with (3.3.2) consider equation
lim
σ→0
σ
−1
h
+
(c[t + σ], c[t] + σc(p(t), P(t))) = 0, (3.3.5)
c[t
0
] = c(x
0
, X
0
).
Deﬁniton 3.3.2 A function c
−
[t] is said to be a solution to the evolution equation (3.3.5)
if
(i) c
−
[t] satisﬁes (3.3.5) almost everywhere,
(ii) c
−
[t] is ellipsoidalvalued,
(iii) c
−
[t] is the maximal solution to (3.3.5) with respect to inclusion.
179
From the deﬁnition of the semidistance h
+
and of the solution c
−
[t] (points (i),(ii)) it
follows that always
c
−
[t] ⊆ A[t].
Thus, we have
Lemma 3.3.3 Any solution c
−
[t] to (3.2.5) that satisﬁes points (i),(ii) of Deﬁnition 3.3.2
is an internal approximation for A[t].
Moreover, representation (3.2.18) yields the fulﬁllment of the requirement of point (iii)
of Deﬁnition 3.3.3 for any function c
−
[t] = c(x
∗
(t), X
−
(t)) generated by the solutions
x
∗
(), X
−
() to equations (3.2.4),(3.2.16).This leads to
Theorem 3.3.1 The internal approximation
c
−
[t] = c(x
∗
(t), X
−
(t))
is a solution to the evolution equation (3.3.5), whenever x
∗
(t), X
−
(t)) are the solutions to
diﬀerential equations (3.2.4), (3.2.16).
For a given function H(t) and a given initial pair x
∗
, X
∗
denote
E
−
(t, τ, c(x
∗
, X
∗
)) = c(x
∗
(t), X
−
(t)),
where x
∗
(t), X
−
(t) are the solutions to (3.2.4), (3.2.16) with initial conditions
x
∗
(τ) = x
∗
, X
−
(τ) = X
∗
.
Then, clearly,
E
−
(t, τ, c(x
∗
, X
∗
)) ⊆ c(x
∗
, X
∗
)) +
_
t
τ
c(p(s), P(s))ds,
and a direct substitution leads to
Lemma 3.3.4 The following relation is true
E
−
(t, τ, E
−
(τ, t
0
, c(x
0
, X
0
))) ⊆ E
−
(t, t
0
, c(x
0
, X
0
)), t
0
≤ τ ≤ t. (3.3.6)
The last relation describes the dynamics of the internal estimates E
−
(t, τ, c(x
∗
, X
∗
)) for
the attainability domains A[t], deﬁning thus a ”lower” semigroup property for the respec
tive mappings. The sets E
−
(t, t
0
, c(x
0
, X
0
)) are sometimes referred to as subattainability
domains. A similar type of descriptions may now be introduced for solvability tubes.
180
3.4 Solvability in Absence of Uncertainty
We shall now pass to the treatment of solvability tubes for the simplest case of systems
without uncertainty and state constraints. Our aim is to approximate these tubes by
ellipsoidalvalued functions.
Returning to relation (1.4.4), we recall that in our case
W[t] = c(m, M) −
_
t
1
t
c(p(t) + f(t), P(t))dt. (3.4.1)
Then, following the approximation schemes of Sections 2.8, 2.9, 3.2, 3.3, with obvious
changes of signs, we come to the diﬀerential equations
˙ x = p(t) + f(t), (3.4.2)
˙
X
+
(t) = −π(t)X
+
(t) −π
−1
(t)P(t), (3.4.3)
˙
X
−
(t) = −H
−1
(t)[(H(t)X
−
(t)H
(t))
1
2
(H(t)P(t)H
(t))
1
2
+ (3.4.4)
+ (H(t)P(t)H
(t))
1
2
(H(t)X
−
(t)H(t))
1
2
]H
−1
(t),
with boundary conditions
x(t
1
) = m , X
+
(t
1
) = M, X
−
(t
1
) = M. (3.4.5)
Denote the solutions to (3.4.2)(3.4.4) with boundary conditions (3.3.5) as x(t), X
+
(t), X
−
(t)
respectively. Similarly to (3.1.3), ( 3.1.17), we then come to
Theorem 3.4.1 The following inclusions are true
c
−
(x(t), X
−
(t)) ⊆ W[t] ⊆ c
+
(x(t), X
+
(t)), (3.4.6)
whatever are the solutions to diﬀerential equations (3.1.2)(3.1.5) with π(t) > 0, H(t) ∈ Σ.
As in the previous Sections 3.2, 3.3 , the last assertion develops into exact representations
181
Theorem 3.4.2 (i) The following ”external” representation is true
W[t] = ∩¦c
+
(x(t), X
+
(t))[π()¦, (3.4.7)
where the intersection is taken over all measurable functions π(t) that satisfy the inequalities
min¦(l, P(t)l)[(l, l) = 1¦ ≤ π(t) ≤ max¦(l, P(t)l)[(l, l) = 1¦, (3.4.8)
(ii) The following ”internal ” representation is true
W[t] = ∪¦c
−
(x(t), X
−
(t))[H()¦, (3.4.9)
where the union is taken over all measurable functions H(t) with values in Σ.
The next issue is to write down an evolution equation with ellipsoidalvalued solutions for
each of the approximating functions c
−
(t), c
+
(t).
Consider equations
lim
σ→0
σ
−1
h
−
(W[t −σ], W[t] −σc(p(t) + f(t), P(t)) = 0, (3.4.10)
lim
σ→0
σ
−1
h
+
(W[t −σ], W[t] −σc(p(t) + f(t), P(t)) = 0 (3.4.11)
with boundary condition
W[t
1
] = c(m, M). (3.4.12)
The solution to these equations are not obliged to be ellipsoidalvalued. Therefore, in
analogy with (3.1.2), we introduce another pair of equations, namely,
lim
σ→0
h
−
(c[t −σ], c[t] −σc(p(t) + f(t), P(t)) = 0, (3.4.13)
lim
σ→0
h
+
(c[t −σ], c[t] −σc(p(t) + f(t), P(t)) = 0, (3.4.14)
with the same boundary condition
c[t
1
] = c(m, M). (3.4.15)
182
Deﬁniton 3.4.1 A function c
+
[t] ( respectively, c
−
[t]), is said to be a solution to the
evolution equation (3.4.13) (resp. (3.4.14) ) ,if
(i)c
−
[t] (resp.c
+
[t]) satisﬁes (3.4.13) (resp. (3.4.14) ) almost everywhere,
(ii) c
+
[t] (resp.c
−
[t]) is ellipsoidalvalued,
(iii) c
+
[t] (resp.c
−
[t]) is the minimal ( resp. , maximal ) solution to (3.4.13) (resp.
(3.4.14)) with respect to inclusion.
From the deﬁnitions of the semidistances h
−
, h
+
and of the solutions c
+
[t], c
−
[t], (properties
(i),(ii)) it follows that always
c
−
[t] ⊆ W[t] ⊆ c
+
[t]. (3.4.16)
It also follows that the last relations are true for any functions c
+
[t], c
−
[t], that satisfy
properties (i),(ii) of Deﬁnition 3.4.1.
The minimality and maximality properties of the respective solutions are described simi
larly to Sections 3.2, 3.3.This gives
Theorem 3.4.3 (i) The ellipsoidalvalued function c
+
[t] = c(x(t), X
+
(t)) generated by
the solutions x(t), X
+
(t) to the diﬀerential equations (3.4.2), (3.4.3), (3.4.5) is a solution
to the evolution equation (3.4.13), whatever is the measurable function π() selected due to
the inequalities (3.4.8),
(ii) The ellipsoidalvalued function c
−
[t] = c(x(t), X
−
(t)) generated by the solutions to
the diﬀerential equations (3.4.2), (3.4.4), (3.4.5) is a solution to the evolution equation
(3.4.14), whatever is the measurable function H(t) with values in Σ.
For a given pair of functions π(), H() and a given pair of boundary values m
∗
, M
∗
we
shall denote
E
+
(t, τ, c(m
∗
, M
∗
)) = c(x(t, τ, m
∗
), X
+
(t, τ, M
∗
)), (3.4.17)
E
−
(t, τ, c(m
∗
, M
∗
)) = c(x(t, τ, m
∗
), X
−
(t, τ, M
∗
)), (3.4.18)
where x(t, τ, m), X
+
(t, τ, M
∗
), X
−
(t, τ, M
∗
) satisfy (3.4.2), (3.4.3), (3.4.4), with boundary
condition
x(τ, τ, m
∗
) = m
∗
, X
+
(τ, τ, M
∗
) = X
−
(τ, τ, M
∗
) = M
∗
.
183
Then, obviously,
E
−
(t, τ, c(m
∗
, M
∗
)) ⊆ c(x
∗
, M
∗
)) −
_
τ
t
c(p(s), P(s))ds ⊆ E
+
(t, τ, c(m
∗
, M
∗
)),
and a direct substitution leads to
Lemma 3.4.1 The following inclusions are true with t ≤ τ ≤ t
1
:
E
−
(t, τ, E(τ, t
1
, c(m, M))) ⊆ E
−
(t, t
1
, c(m, M)), (3.4.19)
E
+
(t, t
1
, c(m, M)) ⊆ E
+
(t, τ, E
+
(τ, t
1
, c(m, M))). (3.4.20)
Relations (3.4.20), (3.4.21) describe the dynamics of the ellipsoidal estimates for W[t], re
spectively deﬁning, (now , in backward time, however) the ”lower” and ”upper” semigroup
properties of the corresponding mappings.
Exercise 3.4.1
Assume that the original system (3.1.1) is given in the form (1.1.1), that is with A(t) ,= 0.
By direct calculations prove that equations (3.4.2)  (3.4.4) are then substituted by
˙ x = A(t)x + p(t) + f(t), (3.4.21)
˙
X
+
(t) = A(t)X
+
(t) + X
+
(t)A(t) −π(t)X
+
(t) −π
−1
(t)P(t), (3.4.22)
˙
X
−
(t) = A(t)X
−
(t) + X
−
(t)A(t) − (3.4.23)
H
−1
(t)[(H(t)X
−
(t)H
(t))
1
2
(H(t)P(t)H
(t))
1
2
−
+ (H(t)P(t)H
(t))
1
2
(H(t)X
−
(t)H(t))
1
2
]H
−1
(t),
with same boundary condition (3.4.5) as before.
The next step is to proceed with the approximations of solvability tubes for systems with
uncertainty.
184
3.5 Solvability Under Uncertainty
In this section we discuss solvability tubes for uncertain systems with unknown input
disturbances. Taking equation (3.1.15), (3.1.16) for the solvability tube of such a system,
we again observe that in general its setvalued solution W[t] is not ellipsoidalvalued. How
should we construct the ellipsoidal approximations for W[t] now, that f(t) is unknown but
bounded ?
Since we do not expect ellipsoidalvalued functions to be produced by solving (3.1.15),
(3.1.16), we will try, as in the previous Section, to introduce other evolution equations than
(3.1.15), constructing them such, that their solutions, on one hand, would be ellipsoidal
valued and on the other, would form an appropriate variety of external and internal ellip
soidal approximations to the solution W[t] of (3.1.15). Being interested in the solvability
set (under uncertainty) we shall further presume that W[t] is inclusionmaximal namely,
as shown in Section 1.6 1.7, the one that gives precisely the solvability set.
Indeed, relation ( 3.1.15), t ∈ [t
0
, t
1
], yields
17
W[t −σ] + σc(q(t), Q(t)) ⊆ W[t] −σc(p(t), P(t)) + o(σ)S. (3.5.1)
We shall now look for an ellipsoidalvalued function c(x(t), X(t)) = c[t] that would ensure
an internal approximation for the lefthand side of (3.5.1) and an external approximation
for the righthand side. Due to (2.4.1),(2.4.2), this would give
W[t −σ] + σc(q(t), Q(t)) ⊇ (3.5.2)
⊇ c(x(t −σ), X(t −σ)) + σc(q(t), Q(t)),
c(x(t −σ), X(t −σ)) + σc(q(t), Q(t)) ⊇ (3.5.3)
⊇ c(x(t −σ) + σq(t), H
−1
[(H(t)X(t −σ)H
(t))
1
2
+ σ(H(t)Q(t)H
(t))
1
2
]
2
H
−1
),
and
W[t] −σc(p(t), P(t)) ⊆ c(x(t), X(t)) −σc(p(t), P(t)), (3.5.4)
c(x(t), X(t)) −σc(p(t), P(t))) ⊆ (3.5.5)
17
This equation is treated under nondegeneracy Assumptions 1.7.1, 1.7.2, see footnote for formula
(3.1.15)
185
⊆ c(x(t) −σp(t), (1 + σπ(t))X(t) + σ
2
(1 + (σπ(t))
−1
)P(t)).
Combining (3.5.3), (3.5.5) and requiring that the righthand parts of these inclusions are
equal (within the terms of ﬁrst order in σ), we require the equality
c(x(t −σ) + σq(t), X[t −σ] + σH
−1
(t)[(H(t)X(t −σ)H
(t))
1
2
(H(t)Q(t)H
(t))
1
2
+
+(H(t)Q(t)H
(t))
1
2
(H(t)X(t −σ)H
(t))
1
2
)]H
−1
(t) =
= c(x(t) −σp(t), (1 + σπ(t))X(t) + σ
2
(1 + (σπ(t))
−1
)P(t)), (3.5.6)
which is ensured if x(t), X(t) satisfy the following equalities
x(t −σ) + σq(t) = x(t) − σp(t), (3.5.7)
and
X(t −σ) + σH
−1
(t)[(H(t)X(t −σ)H
(t))
1
2
(H(t)Q(t)H
(t))
1
2
+ (3.5.8)
+(H(t)Q(t)H
(t))
1
2
(H(t)X(t −σ)(H
(t))
1
2
]H
−1
(t) =
= (1 + σπ(t))X(t) −σ
2
(1 + (σπ(t))
−1
)P(t).
Dividing both parts by σ and passing to the limit (σ → 0), we come to the diﬀerential
equations (with further notation X = X
+
)
˙ x = p(t) + q(t), (3.5.9)
and
˙
X
+
(t) = −π(t)X
+
(t) −π
−1
(t)P(t) + (3.5.10)
+H
−1
(t)[(H(t)Q(t)H
(t))
1
2
(H(t)X
+
(t)H
(t))
1
2
+
+(H(t)X
+
(t)H
(t))
1
2
(H(t)Q(t)H
(t))
1
2
]H
−1
(t),
which have to be taken with boundary conditions
186
x(t
1
) = m, X
+
(t
1
) = M. (3.5.11)
Let us introduce an evolution equation
(3.5.12)
lim
σ→0
σ
−1
h
−
(c[t −σ] + σc(q(t), Q(t)), c[t] −σc(p(t), P(t))) = 0,
with boundary condition
c[t
1
] = c(m, M) (3.5.13)
Deﬁniton 3.5.1 A solution to (3.5.12), (3.5.13) will be deﬁned as an ellipsoidalvalued
function c[t] that satisﬁes (3.5.12) almost everywhere together with boundary condition
(3.5.13).
A solution to the evolution equation (3.5.12) obviously satisﬁes the inclusion
c[t −σ] + σc(q(t), Q(t))l ⊇ (3.5.14)
⊇ c[t] −σc(p(t), P(t)) + o(σ)S
Lemma 3.5.1 The ellipsoid c(x(t), X
+
(t[π(), H())) = c
+
[t] given by equations (3.5.9)
(3.5.11) satisﬁes the inclusion (3.5.14).
Introducing the support function ρ(l[c
+
[t]) and calculating its derivative in t , we have
∂ρ(l[c
+
[t])/∂t =
1
2
(
˙
X
+
(t)l, l)(X
+
(t)l, l))
−
1
2
+ (l, p +q).
By equation (3.5.10) this implies
∂ρ(l[c
+
[t])/∂t = (l, p + q) − (3.5.15)
−
1
2
(P(t)l, l)
1
2
((X
+
(t)l, l)
1
2
π(t)(P(t)l, l)
−
1
2
+ π
−1
(t)(P(t)l, l)
1
2
(X
+
(t)l, l)
−
1
2
)+
+(X
+
(t)l, l)
−
1
2
((H(t)Q(t)H(t))
1
2
H
−1
l, (H(t)X
+
(t)H(t))
1
2
H
−1
l).
187
From inequality a + a
−1
≥ 2 and the inequality of CauchyBuniakowski it then follows (
for all l ∈ IR
n
)
∂ρ(l[c
+
[t])/∂t ≤ (l, p + q) − (l, P(t)l)
1
2
+ (l, Q(t)l)
1
2
. (3.5.16)
Integrating this inequality within the interval [t − σ, t] and having in view the continuity
of P(t), Q(t), we come to (3.5.14) and therefore, to the proof of the Lemma.
Lemma 3.5.2 The ellipsoid c
+
[t] is an external estimate for W[t]  the solvability set
under uncertainty, whatever are the parametrizing functions π(t) > 0, H(t) ∈ Σ.
Following the scheme of Section 1.6 and incorporating relation (3.5.1), we have
−∂ρ(l[W[t])/∂t ≤ ρ(−l[c(p(t), P(t))) −ρ(l[c(q(t), Q(t))), (3.5.17)
Together with (3.5.17), this gives
−∂(ρ(l[W[t]) −ρ(l[c
+
[t]))/∂t ≤ 0, (3.5.18)
and since the boundary conditions are W[t
1
] = c(m, M) = c
+
[t
1
], this yields relation
ρ(l[W[t]) ≤ ρ(l[c
+
[t]) = ρ(l[c(x(t), X
+
(t[π(), H())), ∀l, (3.5.19)
or , in other terms,
W[t] ⊆ c
+
[t] = c(x(t), X
+
((t[π(), H()),
for t
1
− σ ≤ t ≤ t
1
and consequently, the same inclusion for all t ∈ [t
0
, t
1
], whatever are
the respective functions π(), H().
We shall now indicate the inclusionminimal solutions to (3.5.12). Indeed, for a given
vector l , take
π(t) = (X
+
(t)l, l)
−
1
2
(P(t)l, l)
1
2
, (3.5.20)
and take H(t) due to a relation similar to (2.4.4) , so as to ensure
(X
+
(t)l, l)
−
1
2
((H(t)Q(t)H(t))
1
2
H
−1
l, (H(t)X
+
(t)H(t))
1
2
H
−1
l) = (l, Q(t)l)
1
2
. (3.5.21)
For a given vector l, with π, H selected due to (3.5.20), (3.5.21), the value of the respective
derivative
∂ρ(l[c
+
[t])/∂t = ρ(l[c(q(t), Q(t))) −ρ(−l[c(p(t), P(t)))
188
is clearly the largest among all feasible π, H :
∂ρ(l[c
+
[t])/∂t ≥ ∂ρ(l[c
∗
+
[t])/∂t,
where c
∗
is any other external estimate. Integrating the last inequality from t to t
1
and
having in view that c
+
[t
1
] = c
∗
[t
1
] = c(m, M) , we come to
ρ(l[c
+
[t]) ≤ ρ(l[c
∗
[t]).
This means that along the direction l there is no other external ellipsoid governed by
equations (3.5.9),(3.5.10) that could be ”squeezed” between W[t] and c
+
[t] if the last one
is chosen due to (3.5.20),(3.5.21). This implies
Lemma 3.5.3 With π[t], H[t] selected due to (3.5.20),(3.5.21),l ∈ IR
n
, the ellipsoid
c
+
[t] = c(x(t), X
+
(t[π(), H()) is inclusionminimal in the class of all external ellipsoids
governed by equations (3.5.9),(3.5.10).
The selected functions π(t), H(t) may be treated as feedback controls selected, as we shall
see, so as to ensure a ”tightest” external bound for W[t].
18
We now note that the maximal solution W[t] to (3.1.15) ensures an equality in (3.1.19), if
for each l the ellipsoid c
+
[t] = c(x(t), X
+
(t[π(), H()) is selected to be inclusionminimal,
due to (3.5.20),(3.5.21). Indeed, we observe this after integrating (3.5.18) and arriving at
(3.5.19), where for each l there is its own pair of functions π, H.
Combining this fact with the previous assertions, we now observe that the inclusion
minimal external ellipsoids c
+
[t] ensure that the approximation of W[t] is as ”tight” as
possible.
Namely, in view of the indicated relations, we come to the conclusion that for every vector
l there exists a pair π(), H(), such that
ρ(l[W[t]) ≤ ρ(l[c(x(t), X
+
(t[π(), H())), (3.5.22)
and also
ρ(l[c(x(t), X
+
(t[π(), H())) ≤ ρ(l[W[t]) + o(σ, l),
18
The explicit relations for (3.5.20), (3.5.21) indicate that π(), H() are continuous or at least measurable
in t.
With l = l(t) ,= 0 in (3.5.20),(3.5.21), the respective functions π(), H() are still measurable , continuous
or piecewise continuous depending on the properties of l(t).
189
for
t
1
−σ ≤ t ≤ t
1
,
Therefore, particularly, for the indicated values of t one has
ρ(l[W[t]) = inf¦ρ(l, [c(x(t), X
+
(t))[π(), H()¦ +o(σ),
and moreover, for every vector l we have
ρ(l[W[t]) = ρ(l[c(x(t), X
+
(t[π(), H())) (3.5.23)
for some π(), H(). This yields
W[t] = ∩¦c(x(t), X
+
(t[π(), H())[π(), H()¦ + o
∗
(σ)S (3.5.24)
We may consequently repeat this procedure indicating (see (1.7.6)) for
W(t
1
−σ
1
−σ
2
, t
1
−σ
1
, W[t
1
−σ
1
]) =
=
_
W[t
1
−σ
1
] +
_
t
1
−σ
1
t
1
−σ
1
−σ
2
c(p(t), P(t))dt
_
˙
−
_
t
1
−σ
1
t
1
−σ
1
−σ
2
c(q(t), Q(t))dt
and for a given vector l the existence of a pair π(), H() that again yield (3.5.24), now for
t ∈ [t
1
−σ
1
−σ
2
, t
1
−σ
1
].
Continuing the procedure yet further , now for the sets deﬁned by (1.7.7) and passing,
under Assumption 1.7.1, to the respective limit transition of Lemma 1.7.2 and (1.7.8), we
may observe, that with σ → 0 the equalities (3.5.23), (3.5.24),(o
∗
(σ) = 0), are true for all
t ∈ [t
0
, t
1
]. Relations (3.5.9)(3.5.11) thus deﬁne an array of external ellipsoidal estimates
for the solvability set W[t]. Summarizing the results of the above, we have
Theorem 3.5.1 Under Assumption 1.7.1 there exists, for every vector l ∈ IR
n
, a pair
π(), H() of measurable functions that ensure the following :
(i) The support function
ρ(l[W[t]) = ρ(l[c(x(t), X
+
(t[π(), H())), t ∈ [t
0
, t
1
]. (3.5.25)
(ii) The relation (3.5.24) (o
∗
(σ) ≡ 0) is true for t ∈ [t
0
, t
1
].
(iii) The external estimates c
+
[t] = c(x(t), X
+
(t[π(), H()) for the solutions W[t] to the
evolution equation (3.1.15) that are generated by solutions x(t), X
+
(t) to the diﬀerential
equations (3.5.9), (3.5.10) satisfy the evolution equation (3.5.12), (3.5.13) and are minimal
with respect to inclusion among all solutions to (3.5.12), (3.5.13).
190
The next stage is to arrive at a similar theorem for internal estimates. Returning to
(3.1.15), and considering the relation opposite to (3.5.1), we shall look for an ellipsoidal
function c(x(t), X(t)) = c[t] that ensures an external approximation for its lefthand side
and an internal approximation for its righthand side.
Due to (2.4.1),(2.4.2) this would give
W[t −σ] + σc(q(t), Q(t)) ⊆ (3.5.26)
⊆ c(x(t −σ), X(t −σ)) + σc(q(t), Q(t)),
and
c(x(t −σ), X(t −σ)) + σc(q(t), Q(t))) ⊆ (3.5.27)
⊆ c(x(t −σ) + σq(t), (1 + σπ(t))X(t −σ) + σ
2
(1 + (σπ(t))
−1
)Q(t))),
together with
c(x(t), X(t)) −σc(p(t), P(t)) ⊆ W[t] −σc(p(t), P(t)), (3.5.28)
c(x(t) −σp(t), H
−1
(t)[(H(t)X(t)H
(t))
1
2
+ σ(H(t)P(t)H
(t))
1
2
]
2
H
−1
(t)) ⊆
⊆ c(x(t), X(t)) −σc(p(t), P(t)). (3.5.29)
Equalizing the righthand side of (3.5.28) with the lefthand side of (3.5.29) , dividing both
parts by σ and passing to the limit with σ → 0 (similarly to (3.5.6)  ( 3.5.8)), we come
(with further notation X = X
−
) to equations (3.5.9) and
˙
X
−
(t) = π(t)X
−
(t) + π
−1
(t)Q(t) − (3.5.30)
−H
−1
(t)[(H(t)P(t)H
(t))
1
2
(H(t)X
−
(t)H
(t))
1
2
+
+(H(t)X
−
(t)H
(t))
1
2
(H(t)P(t)H
(t))
1
2
]H
−1
(t),
which have to be taken with same boundary conditions as before, namely,
x(t
1
) = m, X
−
(t
1
) = M. (3.5.31)
Let us introduce an evolution equation
(3.5.32)
lim
σ→0
σ
−1
h
+
(c[t −σ] + σc(q(t), Q(t)), c[t] −σc(p(t), P(t))) = 0
191
with boundary condition
c[t
1
] = c(m, M). (3.5.33)
A solution to the last equation is to be considered due to Deﬁnition 3.5.1, where (3.5.32),
(3.5.33) are to be taken instead of (3.5.12), (3.5.13). It obviously satisﬁes the inclusion
c[t −σ] + σc(q(t), Q(t))+ ⊆ (3.5.34)
⊆ c[t] −σc(p(t), P(t)) + o(σ)S.
Similarly to Lemma 3.5.1 one may prove the following:
Lemma 3.5.4 The ellipsoid c(x(t), X
−
(t[π(), H())) = c
−
[t] given by equations (3.5.9),
(3.5.32), (3.5.33) is a solution to the evolution equation (3.5.34)(3.5.35).
A further reasoning is similar to the one that preceded Theorem 3.5.1, except that the
ellipsoid c
−
[t] is now an internal estimate for the maximal solution W[t] to the evolution
equation (3.1.15) and that the respective representations are true relative to closures of the
corresponding sets (as in Theorem 2.5.1). We leave the details ( which are not too trivial,
though) to the reader, conﬁning ourselves to the formulation of
Theorem 3.5.2 Under Assumption 1.7.1, for every vector l ∈ IR
n
, the following relations
are true:
(i)
ρ(l[W[t]) = sup¦ρ(l[c(x(t), X
−
(t[π(), H()))[π(), H()¦, (3.5.35)
(ii)
W[t] = ∪¦c(x(t), X
−
(t[π(), H())[π(), H()¦, (3.5.36)
where π(t) > 0, H(t) are measurable functions.
The internal estimates c
−
[t] = c(x(t), X
−
(t, [π(), H()) for the maximal solutions W[t]
of the evolution equation (3.1.15) that are generated by the solutions x(t), X
−
(t) to the
diﬀerential equations (3.5.9), ( 3.5.30) , are also the maximal solutions to the evolution
equation (3.3.32), (3.5.33).
192
Maximality is treated here with repect to inclusion , as in the above.
Exercise 3.5.1. Under a nondegeneracy Assumption 1.7.2 ( or 1.7.1) prove that equations
(3.5.9), (3.5.10), (3.5.30) may also be derived through the results of Part II, (Theorems
2.4.1, 2.4.2), when the evolution equation (3.5.1) is by deﬁnition substituted by
W[t −σ] ⊆ (W[t] −σc(p(t), P(t)))
˙
−σc(q(t), Q(t)) + o(σ)S (3.5.37)
and the ellipsoidal estimates (external and internal) are taken accordingly. The nondegen
eracy assumption implies in this case that there exist numbers > 0, δ > 0 such that
(W[t] −σc(p(t), P(t)))
˙
−σc(q(t), Q(t)) ⊇ S, σ ≤ δ, t ∈ [t
0
, t
1
].
Exercise 3.5.2
By direct calculation prove that with A(t) ,= 0 , that is with system (3.1.1) given in the
form (1.1.1), equations (3.5.9), (3.5.10), (3.5.30) will be substituted by
˙ x = A(t)x + p(t) + q(t), (3.5.38)
and
˙
X
+
(t) = A(t)X
+
(t) + X
+
(t)A(t) −π(t)X
+
(t) −π
−1
(t)P(t) + (3.5.39)
H
−1
(t)[(H(t)Q(t)H
(t))
1
2
(H(t)X
+
(t)H
(t))
1
2
+
+(H(t)X
+
(t)H
(t))
1
2
(H(t)Q(t)H
(t))
1
2
]H
−1
(t),
˙
X
−
(t) = A(t)X
−
(t) + X
−
(t)A(t) + π(t)X
−
(t) + π
−1
(t)Q(t) − (3.5.40)
−H
−1
(t)[(H(t)P(t)H
(t))
1
2
(H(t)X
−
(t)H
(t))
1
2
+
+(H(t)X
−
(t)H
(t))
1
2
(H(t)P(t)H
(t))
1
2
]H
−1
(t),
with same boundary conditions (3.5.11), (3.5.31) as in the above.
Remark 3.5.1 The external and internal ellipsoidal approximations c
+
[t], c
−
[t] of the
solvability tubes W[t] (” under uncertainty ”) may be interpreted as approximations of
Krasovski’s ”bridges” or Pontryagin’s ”alternated” integrals.
193
Remark 3.5.2 Relations (3.5.39),(3.5.40) will be simpler and reduce to those of type
(3.4.20),(3.4.21) if c(p(t), P(t)), c(q(t), Q(t)) satisfy the matching condition of Remark
1.6.2 which means
c(p(t), P(t))
˙
−c(q(t), Q(t)) = c(p(t) −q(t)), γP(t)), 0 ≤ γ < 1
We ﬁnally mention an important property of the estimates c
−
[t], c
+
[t].
Using the notations of (3.4.17), (3.4.18), where now x(t, τ, m
∗
), X
+
(t, τ, M
∗
), X
−
(t, τ, M
∗
)
satisfy (3.5.9)(3.5.10),(3.5.30) with boundary conditions (3.4.19), one may verify that the
external and internal ellipsoidal approximation mappings E
+
(), E
−
() satisfy the external
and internal semigroup properties respectively.
Lemma 3.5.5 The external and internal approximation mappings (for the solvability
tube under counteraction W[t]), are deﬁned through relations (3.4.17)(3.4.19), (3.5.9),
(3.5.10), (3.5.30) and satisfy the upper and lower semigroup properties (3.4.20),(3.4.21).
Remark 3.5.3 The ellipsoids c
+
[t], c
−
[t] are nondominated ( inclusionminimal and max
imal respectively). Due to this and to the semigroup property of Lemma 3.5.5, the sets c
−
[t]
turn to be ellipsoidalvalued ”bridges” as will be indicated in Section 3.8
We are now prepared to deal with problems of control synthesis with the aim of using the
described relations as a basis for constructive techniques in analytical controller design.
3.6 Control Synthesis
Through Ellipsoidal Techniques
In this Section we shall apply the results of the previous paragraphs to the analytical design
of synthesizing control strategies through ellipsoidal techniques developed in the previous
Sections.
Let us return to the Control Synthesis problem 1.1.4 of Section I.4. There the idea of
constructing the synthesizing strategy (t, x) for this problem was that (t, x) should
ensure that all the solutions x[t] = x(t, τ, x
τ
) to equation
˙ x(t) ∈ (t, x(t)) + f(t), τ ≤ t ≤ t
1
,
with initial state x[τ] = x
τ
∈ J[τ], ( J[t] is the respective solvability set described on the
same Section ) would satisfy the inclusion
x[t] ∈ J[t], τ ≤ t ≤ t
1
194
and would therefore ensure x[t
1
] ∈ /. This exact solution requires, as we have seen, to
calculate the tube J[t] and then, for each instant of time t, to solve an extremal problem of
type (1.4.9) whose solution ﬁnally yields the desired strategy (1.4.12) (t, x) .This strategy
is thus actually deﬁned as an algorithm.
In order to obtain a simpler scheme, we will now substitute J[t] by one of its internal
approximations c
−
[t] = c(x
∗
, X(t)). The conjecture is that once J[t] is substituted by
c
−
[t], we should just copy the scheme of Section 1.4, constructing a strategy 
−
(t, x) such
that for every solution x[t] = x(t, τ, x
τ
) that satisﬁes equation
˙ x[t] = 
−
(t, x[t]) + f(t), τ ≤ t ≤ t
1
, x[τ] = x
τ
, x
τ
∈ c
−
[τ], (3.6.1)
the following inclusion would be true
x[t] ∈ c
−
[t], τ ≤ t ≤ t
1
, (3.6.2)
and therefore
x[t
1
] ∈ c(m, M) = / = c(x
∗
(t
1
), X
−
(t
1
)).
It will be proven that once the approximation c
−
[t] is selected “appropriately”, the desired
strategy 
−
(t, x) may be constructed again according to the scheme of Section I.4, except
that J[t] will now be substituted by c
−
[t], namely

−
(t, x) =
_
c(p(t), P(t)) if x ∈ c
−
[t]
p(t) −P(t)l
0
(l
0
, P(t)l
0
)
−1/2
if x ,∈ c
−
[t],
(3.6.3)
where l
0
= l
0
(t, x) is the unit vector that solves the problem
d[t, x] = (l
0
, x) −ρ(l
0
[c
−
[t]) = max¦(l, x) −ρ(l[c
−
[t])[l ≤ 1¦ (3.6.4)
a nd
d[t, x] = h
+
(x, c
−
[t]) = min¦x −s[s ∈ c
−
[t]¦ (3.6.5)
One may readily observe that relations (3.6.4),(3.6.3) coincide with (1.4.9),(1.4.12) if set
J[t] is substituted for for c
−
[t] and T(t) for c(p(t), P(t).
Indeed, let us start with the maximization problem of (3.6.4). It may be solved in more
detail than its analogue (1.4.9) in section 1.4 (since c
−
[t] is an ellipsoid).
195
If s
0
is the solution to the minimization problem
s
0
= arg min¦(x −s)[s ∈ c
−
[t], x = x(t)¦, (3.6.6)
then we can take
l
0
= k(x(t) −s
0
), k > 0, (3.6.7)
in (3.6.4), so that l
0
will be the gradient of the distance d(x, c
−
[t]) with t ﬁxed. (This can
be veriﬁed by diﬀerentiating either (3.6.4) or (3.6.5) in x).
Lemma 3.6.1 Consider a nondegenerate ellipsoid c = c(a, Q) and a vector x ,∈ c(a, Q).
Then the gradient
l
0
= ∂d(x, c(a, Q))/∂x
may be expressed through l
0
= (x −s
0
)/x −s
0
,
s
0
= (I + λQ
−1
)
−1
(x −a) + a, (3.6.8)
where λ > 0 is the unique root of the equation h(λ) = 0, with
h(λ) =
_
(I + λQ
−1
)
−1
(x −a), Q
−1
(I + λQ
−1
)
−1
(x −a)
_
−1
Proof. Assume a = 0. Then the necessary conditions of optimality for the minization
problem
x −s = min, (s, Q
−1
s) ≤ 1
are reduced to the equation
−x +s + λQ
−1
s = 0
where λ is to be calculated as the root of the equation h(λ) = 0, (a = 0). Since it is
assumed that x ,∈ c(0, Q), we have h(0) > 0. With λ → ∞ we also have
_
(I +λQ
−1
)
−1
x, Q
−1
(I +λQ
−1
)
−1
x
_
→ 0
196
This yields h(λ) < 0, λ ≥ λ
∗
for some λ
∗
> 0. The equation h(λ) = 0 therefore has a root
λ
0
> 0. The root λ
0
is unique since direct calculation gives h
(λ) < 0 with λ > 0. The
case a ,= 0 can now be given through a direct shift x → x −a. Q.E.D.
Corollary 3.6.1 With parameters a, Q given and x varying, the multiplier λ may be
uniquely expressed as a function
λ = λ(x)
Let us now look at relation (1.4.12). In the present case we have T(t) = c(p(t), P(t)) and
problem (1.4.12) therefore reduces to
arg max¦(−l
0
, u)[u ∈ c(p(t), P(t))¦ = 
−
(t, x), (3.6.9)
Relation (3.6.3) now follows from the following assertion:
Lemma 3.6.2 Given ellipsoid c(p, P) , the maximizer u
∗
for the problem
max¦(l, u)[u ∈ c(p, P)¦ = (l, u
∗
) , l ,= 0,
is the vector
u
∗
= p + Pl(l, Pl)
−
1
2
This Lemma is an obvious consequence of the formula for the support function of an
ellipsoid, namely
ρ(l[c(p, P)) = (l, p) + (l, Pl)
−
1
2
We will now prove that the ellipsoidal valued strategy 
−
(t, x) of (3.6.3) does solve the
problem of control synthesis, provided we start from a point x
τ
= x(τ) ∈ c
−
[τ]. Indeed,
assume x
τ
∈ c
−
[τ] and x[t] = x(t, τ, x
τ
) , τ ≤ t ≤ t
1
to be the respective trajectory.
We will demonstrate that once x[t] is a solution to equation (3.6.1), we will always have
the inclusion (3.6.2). (With isolated trajectory x[t] given, it is clearly driven by a unique
control u[t] = ˙ x(t) −f(t) a.e. such that u[t] ∈ c(p(t), P(t)).
Calculating
d[t] = d[t, x[t]] = max¦(l, x(t)) −ρ(l [ c
−
[t])[l ≤ 1¦,
for d[t] ≥ 0, we observe
197
d
dt
d[t] =
d
dt
[(l
0
, x[t]) −ρ(l
0
[c
−
[t])]
and since l
0
,= 0 is a unique maximizer,
d
dt
d[t] = (l
0
, ˙ x[t]) −
∂
∂t
ρ(l
0
[c
−
[t]) = (3.6.10)
= l(l
0
, u[t]) −
d
dt
[(l
0
, x(t)) + (l
0
, X(t)l
0
)
1/2
]
where c
−
[t] = c(x(t), X
−
(t)). For a ﬁxed function H() we have c
−
[t] = c(x(t), X
−
(t)) ,
where x(t), X
−
(t) satisfy the system (3.4.2),(3.4.4).
Substituting these relations into (3.6.10) and remembering the rule for diﬀerentiating a
maximum over a variety of functions, we have
d
dt
d[t] = (l
0
, u[t]) −(l
0
, p(t)) +
1
2
(l
0
, X
−
(t)l
0
)
−1/2
(l
0
, H
−1
(t)([H(t)X
−
(t)H(t)]
1/2
[H(t)P(t)H(t)]
1/2
+
+[H(t)P(t)H(t)]
1/2
[H(t)X
−
(t)H(t)]
1/2
)H
−1
(t)l
0
)
or, due to the BunyakovskySchwartz inequality,
d
dt
d[t] ≤ −(l
0
, p(t)) + (l
0
, P(t)l
0
)
1/2
+ (l
0
, u[t]), (3.6.11)
where
u[t] ∈ 
−
(t, x) ⊆ c(p(t), P(t)),
with equality
d
dt
d[t] = 0
attained if u[t] ∈ 
−
(t, x). Integrating dd
2
[t]/dt = dd
2
[t, x[t]]/dt from τ to t
1
, ( see notations
of Section 1.4 ), we come to the equality
d
2
[τ, x(τ)] = d
2
[t
1
, x(t
1
)] = h
2
+
(x(t
1
), /) = 0
which means x(t
1
) ∈ /, provided x(τ) ∈ X
−
(τ).
What follows is the assertion
198
Theorem 3.6.1 Deﬁne an internal approximation c
−
[t] = c
−
(x(t), X
−
(t)) with given
parametrization H(t) of (3.4.4). Once x[τ] ∈ c
−
[τ] and the synthesizing strategy is 
−
(t, x)
of (3.6.3) , the following inclusion is true:
x[t] ∈ c
−
[t], τ ≤ t ≤ t
1
,
and therefore
x[t
1
] ∈ c(m, M)
.
The ”ellipsoidal” synthesis described in this section thus gives a solution strategy 
−
(t, x)
for any internal approximation c
−
[t] = c
−
(x(t), X
−
(t)). With x ,∈ c
−
[t], the function

−
(t, x) is singlevalued, whilst with x ∈ c
−
[t] it is multivalued (
−
(t, x) = c
−
[t]), being
uppersemicontinuous in x, measurable in t and ensuring the existence of a solution to the
diﬀerential inclusion (3.6.1).
Remark 3.6.1 (i) Due to Theorem 3.4.2 (see (1.4.8)), each element x ∈ intW[t] belongs
to a certain ellipsoid c
−
[t] and may therefore be steered to the terminal set / by means of
a certain ”ellipsoidalbased” strategy 
−
(t, x). Due to the assumptions of the Section 3.1
intW[t] ,= ∅.
(ii) Relations (3.6.3),(3.6.7),(3.6.8) indicate that strategy 
−
(t, x) is given explicitly, with
the only unknown being the multiplier λ of Lemma 3.6.1 which can be calculated as the only
root of equation h(λ) = 0. But in view of Corollary 2.6.1 the function λ = λ(x) may be
calculated in advance, depending on the parameters of the internal approximation c
−
[t]
( which may also be calculated in advance). With this speciﬁcity, the suggested strategy

−
(t, x) may be considered as an analytical design.
(iii) The internal ellipsoids c
−
[t] satisfy the evolution equation (3.5.32) and therefore the
equation (1.7.9) which implies Theorem 1.8.1 and its ”ellipsoidal” version, Theorem 2.6.1.
The given facts are particularly due to the lower semigroup property of the respective map
pings (see Lemma 2.5.3). and the ”inclusionmaximal” property of the ellipsoids c
−
[t].
We shall now proceed with numerical examples that demonstrate the constructive nature
of the solutions obtained above.
3.7 Control Synthesis: Numerical Examples
Let us take system (1.1.1), (3.1.2) to be 4dimensional, and study it throughout the time
interval [t
s
, t
e
], t
s
= 0, t
e
= 5. We will seek for graphical representations of the solutions.
199
And as the ellipsoids appearing in this problem are four dimensional, we we shall present
them through their two dimensional projections. The ﬁgures below are therefore divided
into four windows, and each shows projections of the original ellipsoids onto the planes
spanned by the ﬁrst and second ( x
1
, x
2
), third and fourth(x
3
, x
4
), ﬁrst and third (x
1
, x
3
),
and second and fourth (x
2
, x
4
) coordinate axes, in a clockwise order, starting from bottom
left. The drawn segments of coordinate axes corresponding to state variables range from
−10 to 10 according to the above scheme. In some of the ﬁgures, where we show the graph
of solutions and of solvability set, the third, skew axis corresponds to time and ranges from
0 to 5. Let the initial position ¦0, x
0
¦ be given by
x
0
=
_
_
_
_
_
4
1
0
0
_
_
_
_
_
,
the target set / = c(m, M) by
m =
_
_
_
_
_
0
5
5
0
_
_
_
_
_
and
M =
_
_
_
_
_
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
_
_
_
_
_
at the ﬁnal instant t
1
= 5. We consider a case when the right hand side is constant:
A(t) ≡
_
_
_
_
_
0 1 0 0
−1 0 0 0
0 0 0 1
0 0 −4 0
_
_
_
_
_
,
describing the position and velocity of two independent oscillators. The restriction u(t) ∈
c(p(t), P(t)) on the control u, is also deﬁned by time independent constraints:
p(t) ≡
_
_
_
_
_
0
0
0
0
_
_
_
_
_
,
200
P(t) ≡
_
_
_
_
_
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
_
_
_
_
_
,
so that the controls do couple the system. Therefore, the class of feasible strategies is such
that
U
c
P
= ¦(t, x)¦, (t, x) ⊆ T = c(p(t), P(t)).
The results to be presented here we obtain by way of discretization. We divide the interval
[0, 5] into 100 subintervals of equal lengths, and use the discretized version of (3.4.22) and
(3.4.23) implemented through a standard ﬁrstorder scheme ( see, for example [63], [272],
[273] for technical details). Instead of the set valued control strategy (3.6.3) we apply a
single valued selection:
u(t, x) =
_
p(t) if x ∈ c
−
[t]
p(t) −P(t)l
0
(l
0
, P(t)l
0
)
−1/2
if x ,∈ c
−
[t],
(3.7.1)
again in its discrete version. The use of a singlevalued strategy in the discrete version
does not aﬀect the existence of solutions to the respective recurrence equations.
We shall specify the internal ellipsoid c
−
[t] = c(x(t), X
−
(t)) of (3.4.22),(3.4.23) to be used
here by selecting
H(t) = P
−1/2
(t), 0 ≤ t ≤ 5
in (3.4.18). The calculations give the following internal ellipsoidal estimate c
−
[0] =
c(x(0), X
−
(0)) of the solvability set J[0] = J(0, 5, /):
x(0) =
_
_
_
_
_
4.2371
1.2342
−2.6043
−3.1370
_
_
_
_
_
,
and
X
−
(0) =
_
_
_
_
_
31.1385 0 0 0
0 31.1385 0 0
0 0 12.1845 2.3611
0 0 2.3611 44.1236
_
_
_
_
_
Now, as it is easy to check, x
0
∈ c
−
[0] and therefore we may apply Theorem 3.4.1 with an
implication that the control strategy 
−
(t, x) of (3.6.3) should steer every solution of
201
˙ x = A(t)x +
−
(t, x) + f(t), (3.7.2)
x
0
= x(0), into /. For the discrete version this produces
x[5] =
_
_
_
_
_
0.0264
4.9512
4.0457
−0.0830
_
_
_
_
_
as a ﬁnal state. Figure 3.7.1 shows the graph of the ellipsoidal valued map c
−
[t], t ∈ [0, 5]
and of the solution of
x(t
k+1
) −x(t
k
) = σ
−1
(A(t
k
)x(t
k
) + u(t
k
, x(t
k
))), (3.7.3)
t
s
= t
0
= 0 ≤ t ≤ 5 = t
100
= t
e
; x[0] = x
0
: σ = t
k+1
−t
k
> 0, k = 0, ..., 100,
where we use u(t, x) of (3.7.1).
Equation (3.7.3) serves as a discretetime version of the diﬀerential equation
˙ x[t] = A(t)x[t] + u(t, x[t]) (3.7.4)
However, the last equation has a singlevalued but discontinuous righthand side which
leads to additional questions on the existence of solutions to this equation. There is
actually no such problem, however, for the discretetime system (3.7.3).We will therefore
avoid the singlevalued equation (3.7.4), but will interpret the limit (σ → ∞) of solutions
to (3.7.3) as a solution to the diﬀerential inclusion (3.7.2).
Figure 3.7.2 shows the target set / = c(m, M), (projections appearing as circles), the
solvability set c
−
[0] = c(x(0), X
−
(0)) at the initial instant t = 0, and the trajectory of the
solution of (3.7.2), which, within the accuracy of the computation, may be treated as a
solution of (3.7.2) constructed for the same tube c
−
[t] as in (3.7.3),(3.7.1).
———————!!! insertFigures 3.71, 3.7.2 !!! ————
202
In the next example we show by way of numerical evidence, what can happens if the initial
state x
0
,∈ c
−
[0]. Leaving the rest of the data to be the same, we change the initial state
x
0
in such a way that the inclusion
x
0
∈ c
−
[0]
is hurt, but “not very much”, taking
x
0
=
_
_
_
_
_
4
1
0
2
_
_
_
_
_
.
——————!!! insert Fig. 3.7.3 !!!—————–
Though Theorem 3.6.1 cannot be used, let us still apply formulae (3.7.1) and (3.7.3). Anal
ogously to Figure 3.7.2, Figure 3.7.3 shows the phase portrait of the result. The trajectory
of the solution to (3.7.3) is drawn with a thick line, as long as it is outside of the respective
ellipsoidal solvability set, and with a thin line if it is inside. The drawn projections of the
initial state are inside, except one, (upper left window). As the illustration shows, at one
point in time the trajectory enters the tube c
−
[t], the line changing into thin. After this
happens, Theorem 3.6.1 does take eﬀect, and the trajectory remains inside for the rest of
the time interval. In this way we obtain
x[5] =
_
_
_
_
_
0.0255
4.9528
4.0215
−0.1658
_
_
_
_
_
as a ﬁnal state. The above phenomenon indicates that
• the initial state must be inside the solvability set J[0] = J(0, 5, /), that is actually
x
0
∈ J(0, 5, /) ¸ c
−
[0],
as it was possible to steer the solution to (3.7.3),(3.7.1) into the target set /,
• in this particular numerical example the control rule works beyond the tube c
−
[t].
203
In the third example, we move the initial state x
0
further away, so that the control rule
does not work any more, (Figure 3.7.4):
x
0
=
_
_
_
_
_
4
1
0
3
_
_
_
_
_
,
and obtain as ﬁnal state
x[5] =
_
_
_
_
_
0.0460
4.9150
3.3668
−0.5540
_
_
_
_
_
.
—————!!! insert Fig. 3.7.4 !!!————————
Figures 3.7.5 and 3.7.6 show the eﬀect of changing the target set. We take the data of
the ﬁrst example except for the matrix M in the target set / = c(m, M) by setting the
radius to be 2:
M =
_
_
_
_
_
4 0 0 0
0 4 0 0
0 0 4 0
0 0 0 4
_
_
_
_
_
,
resulting in a ﬁnal state
x[5] =
_
_
_
_
_
0.5875
4.8914
3.0158
−0.0536
_
_
_
_
_
.
——————!!! insert ﬁgures 3.7.5, 3.7.6 !!!—————
204
The switching of the control, due to the speciﬁc form of (3.7.1), is clearly seen in Figure
3.7.6. and later in Figure 3.7.8. Taking again the data of the ﬁrst example, we allow more
freedom for the controls, changing the matrix P(t) in the bounding set T = c(p(t), P(t))
again by setting the radius to be 2:
P(t) ≡
_
_
_
_
_
4 0 0 0
0 4 0 0
0 0 4 0
0 0 0 4
_
_
_
_
_
with a ﬁnal state
x[5] =
_
_
_
_
_
0.0235
4.9565
4.0536
−0.1308
_
_
_
_
_
.
Numerical simulations were made on a SUN SparcStation.
————————!!! insert ﬁg.3.7.7, 3.7.8 !!!————–
Finally we shall consider two coupled oscillators, represented by a system with parameters
x
0
=
_
_
_
_
_
−5
0
−10
10
_
_
_
_
_
,
with target set / = c(m, M) deﬁned by
m =
_
_
_
_
_
10
0
0
10
_
_
_
_
_
and
M =
_
_
_
_
_
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
_
_
_
_
_
205
at ﬁnal instant t
1
= 3.
The system matrix A is constant:
A(t) ≡
_
_
_
_
_
0 1 0 0
−1 0.25 0 0
0 0 0 1
16 0 −16 0
_
_
_
_
_
,
and the constraint on the controls is deﬁned by
p(t) ≡
_
_
_
_
_
0
0
0
0
_
_
_
_
_
,
P(t) ≡
_
_
_
_
_
9 0 0 0
0 0.1 0 0
0 0 9 0
0 0 0 0.1
_
_
_
_
_
,
The target control problem is solved as before, in 100steps, with synthesizing strategy
calculated due to (3.7.1) through a diﬀerence scheme similar to the above. The four
dimensional ellipsoidal tubes and the synthesized control trajectory in phase space are
shown in an appropriate scale in Figures 3.7.9, 3.7.10, ( here note the relatively ”small
size” of the target set) .
———————!!! insert ﬁgures 3.7.9, 3.7.10 !!!—
3.8 ”Ellipsoidal” Control Synthesis
for Uncertain Systems
In this Section we shall further apply the results of the previous paragraphs to the analyt
ical ”ellipsoidal” design of synthesizing control strategies, this time constructing them for
uncertain systems.
206
Let us consider the Problem of Control Synthesis Under Uncertainty of Section 1.8, (Deﬁ
nition 1.8.1). There the idea was that the respective synthesizing control strategy (t, x)
should ensure that all the solutions x[t] = x(t, τ, x
τ
) to the diﬀerential inclusion
˙ x(t) ∈ (t, x(t)) +c(q(t), Q(t)), τ ≤ t ≤ t
1
,
with initial state x[τ] = x
τ
∈ J
∗
[τ], would satisfy the inclusion
x[t] ∈ J
∗
[t], τ ≤ t ≤ t
1
and would therefore ensure the desired terminal condition x[t
1
] ∈ /.
Here J
∗
[t] = J
∗
[t] is the solvability set of Deﬁnition 1.8.2 which, under Assumptions 1.7.1
or 1.7.2 presumed here, could be speciﬁed through the Alternated Integral (1.7.8), so that
the setvalued function J
∗
[t] would satisfy the evolution equation (1.7.9),(1.7.10).
The exact solution scheme requires, as we have seen, to calculate the tube J
∗
[t] and then,
for each instant of time t, to solve an extremal problem of type (1.8.10) whose solution
ﬁnally allows to specify the desired strategy (t, x) = 
0
(t, x) according to (1.8.9). The
strategy 
0
(t, x) is again actually deﬁned as an algorithm which, due to the presence of
uncertain items, is more complicated, of course, than in the absence of these.
To obtain a simpler scheme, we shall substitute J
∗
[t] by one of its internal ellipsoidal
approximations c
−
[t] = c(x
∗
, X
−
(t)). The conjecture is that once J
∗
[t] is substituted by
c
−
[t], we should just copy the scheme of Section 1.8. Namely, we should construct the new
approximate strategy 
0
−
(t, x) such that for every solution x[t] = x(t, τ, x
τ
) to the system
˙ x[t] = 
0
−
(t, x[t]) + f(t), τ ≤ t ≤ t
1
, x[τ] = x
τ
, x
τ
∈ c
−
[τ], (3.8.1)
the inclusion
x[t] ∈ c
−
[t], τ ≤ t ≤ t
1
(3.8.2)
would be true, whatever is the function f(t) ∈ c(q(t), Q(t). This would ensure the terminal
condition
x[t
1
] ∈ c(m, M) = / = c[t
1
].
It will be proven again that once the approximation c
−
[t] is selected “appropriately’,
namely, due to relations (3.5.9), (3.5.30), (3.5.31), the desired strategy 
0
−
(t, x) may be
constructed as in Section 1.8, except that J
∗
[t] should be substituted by c
−
[t]. Namely,
207

0
−
(t, x) =
_
c(p(t), P(t)) if x ∈ c
−
[t]
p(t) −P(t)l
0
(l
0
, P(t)l
0
)
−1/2
if x ,∈ c
−
[t],
(3.8.3)
where l
0
= l
0
(t, x) = ∂d(x, c
−
[t])/∂x is the unit vector that solves the problem
d[t, x] = (l
0
, x) −ρ(l
0
[ c
−
[t]) = max¦(l, x) −ρ(l[c
−
[t])[l ≤ 1¦. (3.8.4)
and as before
d[t, x] = d(x, c
−
[t]) = h
+
(x, c
−
[t]) = min¦x −s[s ∈ c
−
[t]¦ (3.8.5)
Remark 3.8.1 We emphasize again that the given scheme follows the lines of Section 3.6,
but the tube c
−
[t] = c
−
(t, x) taken here is deﬁned by relations (3.5.9), (3.5.30)
rather than by (3.4.2), (3.4.4) as in Section 3.6 . This reﬂects the uncertainty (3.1.6)
in the inputs f of the system.
The further reasoning is analogous to that of Section 3.6. Without repeating the similar ele
ments in the scheme, we have to underline that the main new point here is the calculation of
the derivative dd[t, x]/dt due to the diﬀerential inclusion (3.8.1) with c
−
[t] = c
−
(x
∗
, X
−
(t))
deﬁned by (3.5.9),(3.5.30).
The desired solution strategy 
0
−
(t, x) must satisfy a relation of type (1.8.9) which depends
on vector l
0
= l
0
(t, x) the maximizer for problem (3.8.4) a direct analogue of problem
(1.8.10) of Section 1.8. The respective relations may now be obtained in more detail than
in the general case of Section 1.8, since c
−
[t] is an ellipsoid.
The properties of l
0
are similar to those described in (3.6.7), (3.6.8) and in Lemma 3.6.1.
Further on, we notice that again T(t) = c(p(t), P(t)) is an ellipsoid, so that problem (1.8.9)
reduces to
arg max¦(−l
0
, u)[u ∈ c(p(t), P(t))¦ = 
0
−
(t, x), (3.8.6)
and therefore relation (3.8.3) follows from Lemma 3.6.2.
We will now prove that the ”ellipsoidal”  based strategy 
0
−
(t, x) of (3.8.3) does solve the
problem of control synthesis of Deﬁnition I.8.1, provided we start from a point x
τ
= x(τ) ∈
c
−
[τ]. Indeed, assume x
τ
∈ c
−
[τ] and x[t] = x(t, τ, x
τ
) , τ ≤ t ≤ t
1
to be the respective
trajectory. We will demonstrate that once x[t] is a solution to (3.8.1), (t, x) = 
0
−
(t, x),
it will always satisfy (3.8.2).
Calculating
208
d[t] = d[t, x[t]] = max¦(l, x(t)) −ρ(l [ c
−
[t])[l ≤ 1¦,
we observe
d
dt
d[t] =
d
dt
[(l
0
, x[t]) −ρ(l
0
[c
−
[t])]
and since l
0
is a unique maximizer,
d
dt
d[t] = (l
0
, ˙ x[t]) −
∂
∂t
ρ(l
0
[c
−
[t]) = (3.8.7)
= (l
0
, u[t] + f(t)l) −
d
dt
[(l
0
, x(t)) + (l
0
, X
−
(t)l
0
)
1/2
],
where c
−
[t] = c(x(t), X
−
(t)). For ﬁxed functions π(), H() we have c
−
[t] = c(x(t), X
−
(t))
, where x(t), X
−
(t) satisfy the system (3.8.9),(3.8.30),
Substituting this into (3.8.7) and diﬀerentiating the respective function of the ”maximum”
type due to equation
d
dt
x[t] = u[t] + f(t),
where u[t] ∈ (t, x[t]) is a realization of the feedback control strategy  and f(t) is an
input disturbance , we have
d
dt
d[t] = (l
0
, u[t] + f(t)) −(l
0
, p(t) + q(t)) −
1
2
(l
0
, X
−
(t)l
0
)
−1/2
¦(l
0
, (π(t)X
−
(t) + π
−1
(t)Q(t))l
0
) −(l
0
, H
−1
(t)([H(t)X
−
(t)H(t)]
1/2
[H(t)P(t)H(t)]
1/2
+
+[H(t)P(t)S(t)]
1/2
[H(t)X
−
(t)H(t)]
1/2
)H
−1
(t)l
0
)¦.
Applying inequality a
2
+ b
2
≥ 2ab and the the BunyakovskySchwartz inequality to the
righthand part of the previous formula, we come to
d
dt
d[t] ≤ (3.8.8)
≤ (l
0
, u[t] + f(t)) −(l
0
, p(t) + q(t)) + (l
0
, P(t)l
0
)
1/2
−(l
0
, Q(t)l
0
)
1/2
,
where
u[t] ∈ c(p(t), P(t)), f(t) ∈ c(q(t), Q(t),
209
In other terms we have
d
dt
d[t] ≤ (l
0
, u[t] + f(t)) + ρ(−l
0
[c(p(t), P(t))) − ρ(l
0
[c(q(t), Q(t)))
With u[t] ∈ 
0
−
(t, x) and any feasible f(t) this yields (almost everywhere)
d
dt
d[t] ≤ 0, x ,∈ c
−
[t],
due to (3.8.6).
This also gives
d
dt
V (t, x) ≤ 0 V (t, x) = d
2
[t].
Integrating dd
2
[t]/dt from τ to t
1
, we come to the inequality (d[t] = d
2
W
∗[t, x[t]], see notation
of (1.8.19) )
h
2
+
(x(t
1
), /) = d
2
W
∗[t
1
, x(t
1
)] ≤ d
2
J
∗
[τ, x(τ)] = h
2
+
(x(τ, X
−
(τ)),
so that x(t
1
) ∈ / if x(τ) ∈ X
−
(τ). What follows is the assertion
Theorem 3.8.1 Deﬁne an internal ellipsoidal approximation c
−
[t] = c
−
(x(t), X
−
(t)) to
the solvability set J
∗
[t], with given parametrization H(t), π(t) in (3.8.30). Once x[τ] ∈
c
−
[τ] and the synthesizing strategy is selected as 
0
−
(t, x) of (3.8.6), (3.8.3) , the following
inclusion is true:
x[t] ∈ c
−
[t], τ ≤ t ≤ t
1
,
whatever is the solution x[t] to the diﬀerential inclusion
d
dt
x ⊆ 
0
−
(t, x) + c(q(t), Q(t))
,
(3.8.9)
and therefore
x[t
1
] ∈ c(m, M),
whatever is the disturbance f(t) ⊆ c(q(t), Q(t)) in the synthesized system
d
dt
x = 
0
−
(t, x) + f(t).
210
The ”ellipsoidal” synthesis thus gives a solution strategy 
0
−
(t, x) for any internal approxi
mation c
−
[t] = c
−
(x(t), X
−
(t)) of the solvability tube J
∗
[t] . With x ,∈ c
−
[t], the function

0
−
(t, x) is singlevalued, whilst with x ∈ c
−
[t] it is multivalued (
0
−
(t, x) = c
−
[t]), being
therefore uppersemicontinuous in x, measurable in t and ensuring the existence of a so
lution to the diﬀerential inclusion (3.8.9). Theorem 3.8.1 indicates that each of the tubes
c[t] is an ellipsoidalvalued bridge ( see Remark 3.5.3).
Remark 3.8.2 (i)Due to Theorem 3.5.2 (see (3.5.36)),each element x ∈ intW
∗
[t] belongs
to a certain ellipsoid c
−
[t] and may therefore be steered to the terminal set / by means of
a certain ”ellipsoidal strategy” 
−
(t, x). Such a strategy may be speciﬁed in explicit form
except for a scalar multiplier λ = λ(x) , which may be calculated in advance, as indicated
in Remark 3.6.2. With this reservation, the suggested strategy 
0
−
may be interpreted as
an analytical design.
(ii) We emphasize once more that the constructions given in Sections 3.5,3.8 are derived
here under Assumption 1.7.1 , which implies that there exists an internal curve ˜ x(t) such
that ˜ x(t) +(t)S
1
(0) ⊆ J
∗
[t] for all t with continuous (t) > 0. Then clearly intJ
∗
[t] ,≡ ∅.
We shall now proceed with further numerical examples ( this time for uncertain systems)
that demonstrate the constructive nature of the suggested solution schemes.
3.9 Control Synthesis for Uncertain Systems: Nu
merical Examples
In this Section our particular intention ﬁrst is to illustrate through simulation the eﬀect
of introducing an unknown but bounded disturbance f(t) into the system. We shall do
this by considering a sequence of three problems where only the size of the bounding sets
for the disturbances f(t) increases from case to case, starting from no disturbance at
all (that is where the sets Q(t) = c(q(t), Q(t)), t ∈ [t
0
, t
1
] are singletons ) to ”more
disturbance” allowed, so that the problem still remains solvable. The result is that in
the ﬁrst case we obtain a “large” internal ellipsoidal estimate c
−
[t] of the solvability set
J
∗
[t
0
] = J
∗
(t
0
, t
1
, /) , while in the last it shrinks to be “small”. We also indicate
the behaviour of isolated trajectories of system (3.8.1) in the presence of various speciﬁc
feasible disturbances f(t) ∈ c(q(t), Q(t)).
For the calculations we use a standard ﬁrst  order discrete scheme for equations (3.8.9),
(3.8.30) by dividing the time interval — chosen to be [0, 5] — into 100 subintervals of
211
equal length ( the details of such schemes may found in [63], [272], [273]. Instead of the
set valued control strategy (3.8.3) we apply a single valued selection:
u(t, x) =
_
p(t) if x ∈ c
−
[t]
p(t) −P(t)l
0
(l
0
, P(t)l
0
)
−1/2
if x ,∈ c
−
[t].
(3.9.1)
again in its discrete version. (The discrete version obviously does not require any additional
justiﬁcation for using the singlevalued selection). We calculate the parameters of the
ellipsoid c
−
[t] = c(x(t), X
−
(t)) by chosing a speciﬁc parametrization which is
H(t) = P
−1/2
(t)
and
π(t) =
Tr
1/2
(X
−
(t))
Tr
1/2
(Q(t))
in equation (3.5.30). We consider a 4dimensional system of type (1.1.1), (3.1.2)(3.1.4)
with the initial position ¦0, x
0
¦ given by
x
0
=
_
_
_
_
_
2
−10
1
−6
_
_
_
_
_
,
at the initial moment t
0
= 0 and target set / = c(m, M) deﬁned by
m =
_
_
_
_
_
10
0
0
10
_
_
_
_
_
and
M ≡
_
_
_
_
_
100 0 0 0
0 100 0 0
0 0 100 0
0 0 0 100
_
_
_
_
_
at the ﬁnal moment t
1
= 5. We suppose the right hand side to be constant:
A(t) ≡
_
_
_
_
_
0 1 0 0
−1 0 0 0
0 0 0 1
0 0 −4 0
_
_
_
_
_
,
describing the position and velocity of two independent oscillators. (Through the con
straints on the control and disturbance, however, the system becomes coupled.) The
212
restriction u(t) ∈ c(p(t), P(t)) on the control and v(t) ∈ c(q(t), Q(t)) on the disturbance
is also deﬁned by time independent constraints:
p(t) ≡
_
_
_
_
_
0
0
0
0
_
_
_
_
_
,
P(t) ≡
_
_
_
_
_
9 0 0 0
0 1 0 0
0 0 9 0
0 0 0 1
_
_
_
_
_
The center of the disturbance is the same in all cases:
q(t) ≡
_
_
_
_
_
0
0
0
_
_
_
_
_
The diﬀerence between the three cases i = 1, 2, 3 appear in the matrices:
Q
(1)
(t) ≡
_
_
_
_
_
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
_
_
_
_
_
,
Q
(2)
(t) ≡
_
_
_
_
_
1 0 0 0
0 9 0 0
0 0 1 0
0 0 0 9
_
_
_
_
_
,
Q
(3)
(t) ≡
_
_
_
_
_
1 0 0 0
0 13.1 0 0
0 0 1 0
0 0 0 13.1
_
_
_
_
_
Clearly, case i = 1 is the one treated in Section 3.7, but note that in the cases i = 2, 3 the
data are chosen in such a way that neither the controls, nor the disturbances dominate the
other, that is, both T
˙
− Q = ∅ and Q
˙
− T = ∅.
Obviously, in these cases the problem can not be reduced to simpler situations without
disturbances. More precisely, in these cases Assumption I.6.1 that allows such a reduction
is not fulﬁlled. At the same time, the solvability set W
∗
[t] contains an internal trajectory
so that intW
∗
[t] ,= ∅ ( see Remark 3.8.2(ii)). Its internal ellipsoidal approximations c
−
[t]
exist and may be calculated due to schemes of Section 3.5.
213
The calculations give the following internal ellipsoidal estimate c
(i)
−
[0] = c(x(0), X
(i)
−
(0))
of the solvability set J
(i)
(0, /), i = 1, 2, 3:
x(0) =
_
_
_
_
_
2.4685
−8.4742
1.5685
−5.2087
_
_
_
_
_
,
and
X
(1)
−
(0) =
_
_
_
_
_
323.9377 30.2735 0 0
30.2735 341.4382 0 0
0 0 147.0094 61.1077
0 0 61.1077 469.5488
_
_
_
_
_
,
X
(2)
−
(0) =
_
_
_
_
_
46.3661 25.5502 0 0
25.5502 66.4791 0 0
0 0 45.3047 28.3397
0 0 28.3397 132.7509
_
_
_
_
_
,
X
(3)
−
(0) =
_
_
_
_
_
12.2863 21.2197 0 0
21.2197 37.8930 0 0
0 0 33.6241 22.3911
0 0 22.3911 98.7732
_
_
_
_
_
.
Now, as is easy to check, x
0
∈ c(x(0), X
(i)
−
(0)) for i = 1, 2, 3 and therefore Theorem 3.8.1
is applicable, implying that the control strategy of (3.8.3) steers the solution of (3.8.1) into
/ under any admissible disturbance f(t) ∈ c(q(t), Q
(i)
(t)) in all three cases. Also, as it
can be proved on the basis of their construction, we have the inclusions
c(x(0), X
(3)
−
(0)) ⊂ c(x(0), X
(2)
−
(0)) ⊂ c(x(0), X
(1)
−
(0))
holding, analogously to the corresponding inclusions between the original (nonellipsoidal)
solvability sets J
(i)
(0, 5, /).
Since the ellipsoids appearing in this problem are four dimensional, and since the objective
is to describe the solutions also through graphical representations, we present their two
dimensional projections. The ﬁgures are therefore divided into four windows, showing pro
jections of the original ellipsoids onto the planes spanned by the ﬁrst and second (¦x
1
, x
2
¦),
third and fourth (¦x
3
, x
4
¦), ﬁrst and third(¦x
1
, x
3
¦), and second and fourth (¦x
2
, x
4
¦) co
ordinate axes, in a clockwise order starting from bottom left. The drawn segments of
coordinate axes corresponding to the state variables range from −30 to 30. The skew axis
in
Figures 3.9.1  3.9.3 is time, ranging from 0 to 5. Figures 3.9.1  3.9.3 show the graph of
the ellipsoidal valued maps c
(i)
−
[t], t ∈ [0, 5], i = 1, 2, 3, respectively, and of the solutions
to equation
x(t
k+1
) −x(t
k
) = σ(A(t
k
)x(t
k
) + u(t
k
, x(t
k
)) + f(t
k
)) (3.9.2)
214
x[0] = x
0
, 0 = t
0
≤ t ≤ t
100
= 5, σ = t
k+1
−t
k
> 0, k = 1, ..., 100,
which is a discrete version of the equation
˙ x[t] = A(t)x[t] + u(t, x[t]) + f(t),
(There may be problems with deﬁning the existence of solutions to the last equation ,
however, since function u(t, x) may turn to be discontinuous in x . We will therefore avoid
this last equation and refer only to (3.9.2) and (3.7.1), see analogous situation in Section
3.7).
Here u(t, x) is deﬁned by (3.9.1) and we consider three diﬀerent choices of the disturbance
f(t), one being f(t) ≡ 0 and two other — so called extremal bangbang type — feasible
disturbances. The construction of these disturbances is the following. The time interval
[0, 5] is divided into subintervals of constant lengths. A value f is chosen randomly at the
boundary of c(q(t), Q
(i)
(t)) and the disturbance is then deﬁned by
f(t) = f
over all the ﬁrst interval and
f(t) = −f
over the second. Then a new value for f is selected and the above procedure is repeated
for the next pair of intervals, etc. The controlled trajectory, that is the solution to (3.9.1),
(3.9.2), is drawn in a thin line if it is inside the current ellipsoidal solvability set, and by
a thick line if it is outside. So the statement of Theorem 3.8.1 is that the control ensures
that a thin line cannot change into thick.
——————–!!! insert ﬁgures 3.9.1  3.9.3 !!!————
Figures 3.9.4, 3.9.5, 3.9.6 show the target set / = c(m, M), (projections appearing as
circles of radius 10), the solvability set c
(i)
−
[0] = c(w(0), W
(i)
−
(0)) at t = 0, and trajectories
of the same solutions of (3.9.1), (3.9.2) in phase space.
215
—————!!! insert ﬁgures 3.9.4  3.9.6 !!! —————–
The ellipsoids c
−
[0] are only subsets of the respective solvability sets J
∗
(0, 5, /), therefore
from Theorem 3.8.1 there does not follow a negative statement, like if the initial state is not
contained in c
−
[t
0
], then it is not true that the trajectory can be steered into the target
set / under any disturbance f(t) ∈ Q(t). However, if the ellipsoidal approximation
c
−
[0] ⊂ J
∗
(0, 5, /) is ”appropriate”, then it may occur that such a behaviour can be
illustrated on the ellipsoidal approximations. To show this, we return to the parameter
values of the previous examples and change the initial state only, by moving it in such a
way that
x
0
∈ c
(1)
−
[0] ¸ c
(2)
−
[0] (3.9.3)
holds, taking
x
0
=
_
_
_
_
_
−12
0
3
0
_
_
_
_
_
.
In Figures 3.9.7 and 3.9.8 it can be seen that relation (3.9.3) holds indeed. The trajectory
in Figure 3.9.7 successfully hits the target set / at t = 5. (This is case i = 1, so there is
no disturbance.)
Figure 3.9.8 shows two trajectories under two simulated feasible disturbances f(t) ∈
c(q(t), Q(t)). In one case the control rule deﬁned using the ellipsoidal tube c
(2)
−
[t] steers
the trajectory into the target /, while under the other disturbance, it does not succeed.
(One thick trajectory changing into thin is clearly seen in the right hand side windows, and
the projection of the endpoint of the other is outside in the lower left window. Compare
these examples with those of Section 3.7). There may of course be other control rules,
like the one based on the exact (nonellipsoidal) solvability sets J
∗
[0] = J
∗
(t, t
1
, /), that
could be successful, once x(0) ∈ J
∗
[0].
——————!!! insert ﬁgures 3.9.7, 3.9.8 !!!—————–
216
Finally we again consider a system that describes two coupled oscillators with matrix
A(t) ≡
_
_
_
_
_
0 1 0 0
−1 0 0 0
0 0 0 1
−1 0 −9 0
_
_
_
_
_
,
and with the other parameters ( x
0
, P, p, M, m, q) same as in the previous ﬁgures. Taking
the disturbances to be restricted by Q
(1)
, Q
(2)
, Q
(3)
of the above and simulating the respec
tive target control synthesis problem, we come to results shown in Figures 3.9.9, 3.9.10,
3.9.11 accordingly.
—————!!! insert ﬁgures 3.9.9, 3.9.10, 3.11 !!! ——
3.10 Target Control Synthesis within
Free Time Interval
Considering again the Problem of Control Synthesis Under Uncertainty of Deﬁnition 1.8.1,
we shall modify this deﬁnition by deleting the requirement that the terminal instant t
1
is
ﬁxed. Thus, we shall require that the terminal inclusion x(t) ∈ / could be reached at any
instant t ∈ (t
0
, t
1
] (namely, not later than at t
1
rather than at ﬁxed t
1
, as before). We
shall look for an ”ellipsoidal” control synthesis solution to this problem within a scheme
similar in nature to the one of Section 3.8. We have in view that the constraints on u, f
and the target set / are all ellipsoidalvalued, as in Sections 3.1, 3.8.
We shall now brieﬂy describe this problem without going into speciﬁc details with the main
aim to demonstrate a numerical example of a nonconvex solvability set.
Recall the solvability set of Section 1.8. For time interval [τ, t] it should be denoted,
according to the respective notations, as J
∗
(τ, t, /). Our new problem with free terminal
time t will then be solvable for a given position ¦τ, x¦ if and only if x ∈ W
f
(τ, /), where
W
f
(τ, /) =
_
¦J
∗
(τ, t, /) : t ∈ [τ, t
1
]¦.
217
Clearly, set W
f
(τ, /) is not bound to be convex. The results of the previous Sections
allow to formulate the following assertion, see [305]. ( Here the earlier symbols Π
+
, Σ of
Sections 3.2  3.5 for the classes of functions π(), H() are complemented by [t
0
, t
1
] that
symbolizes the interval where these functions are deﬁned).
Theorem 3.10.1 Fix continuous functions π(t) ∈ Π
+
[t
0
, t
1
] and H(t) ∈ Σ[t
0
, t
1
]
and deﬁne an internal approximation c
−
[τ, t, /] = c(x(τ), X
−
(τ[π(), H()); t, /)) of
J(τ, t, /) for τ ∈ [t
0
, t].
19
Once
x
τ
= x(τ) ∈ c
−
[τ, t, /] (3.10.1)
for some τ ∈ [t
0
, t] and x[t
] = x(t
, τ, x
τ
) is a solution to ( 3.8.1),(3.8.3), where c
−
[τ] is
substituted by c[τ, t, /] , the inclusion
x[t
] ∈ c(t
, τ, /)
shall be true for all t
∈ [τ, t] and in particular
x[t] ∈ c(m, M).
Hence, the strategy (3.8.3) taken for c
−
[τ] = c[τ, t, /] solves the terminal control problem
by time t, whatever is the disturbance f(t) ∈ c(q(t), Q(t).
Proof. Follows from the fact that
_
¦c
−
[τ, t, /][t ∈ [τ, t
1
]¦ ∈ W
f
(τ, /),
where c
−
[τ, t, /] = c(x(τ), X
−
(τ[π(), H()); t, /)) and the pair π(), H() is ﬁxed.
Q.E.D.
Denote
_
¦c
−
[τ, t, /][t ∈ [τ, t
1
]¦ = E
f
(τ, /[π(), H()).
Then , if
x
τ
∈ E
f
(τ, /[π(), H()),
there exists a minimal value t = t
∗
among those t that ensure x
τ
∈ c[τ, t, /]. This is due
to the continuity of the distance function
19
Here symbol c
−
[τ, t, /] = c(x(τ), X
−
(τ[π(), H()); t, /)) stands for the internal ellipsoid described
by equations (3.5.30) or (3.5.41), but with boundary condition (3.5.31) taken at instant t instead of t
1
.
218
d(x
τ
, c[τ, t, /]) = h
+
(x
τ
, c[τ, t, /]) = d[x
τ
, t]
in t, ( check this assertion), so that t
∗
is the minimal root of the equation
d[x
τ
, t] = 0. (3.10.2)
Denote d
e
(x
τ
, W
∗
(τ, t, /)) = d
e
[x
τ
, t] and t
∗
e
to be the minimal root of equation
d
e
[x
τ
, t] = 0, (3.10.3)
(the latter function d
e
is also continuous in t). Time t
∗
shall then be the exact ”optimal
time ”. But since 0 ≤ d
e
[x
τ
, t] ≤ d[x
τ
, t] , we further come to the following fact
Lemma 3.10.1 The ”optimal time” t
∗
e
≤ t
∗
, whatever is the internal tube c[τ, t, /] that
generates the value t
∗
.
Remark 3.10.1 One should be aware that in general the functions d[x
τ
, t], d
e
[x
τ
, t] are
not monotonous in t, so that the practical calculation of the roots of equations (3.10.1),
(3.10.2) may lead to unstable numerical procedures that require additional regularization.
Exercise 3.10.1.
Check the the following assertion.
Fix continuous functions π() ∈ Π
+
[t
0
, t] and H() ∈ Σ[t
0
, t] for all t ∈ [t
0
, t
1
] and deﬁne
an external approximation c
+
[τ, t, /] = c
+
(x(τ), X
+
(t[π(), H()); t, /) of J(τ, t, /) for
τ ∈ [t
0
, t].
Once
x
t
0
= x[t
0
] / ∈ c
+
[t
0
, t, /] (3.10.4)
for all t ∈ [t
0
, t
1
], then the problem of target control synthesis of this Section (under
uncertainty, with free target time), cannot be solved.
We shall now proceed with numerical examples. For the calculations we use the same
discrete scheme as in Section 3.9, ( dividing the time interval — chosen to be [0, 5] —
into 100 subintervals of equal lengths) and the control strategy oftype (3.9.1) is found here
through the same parametrization.
The parameters A, M, p, P, q are the same as in the examples of ﬁgure 3.9.13.9.3, except
that the initial position is given by
x
0
=
_
_
_
_
_
0
−20
0
5
_
_
_
_
_
,
219
at the initial instant t
0
= 0 and for the target set / = c(m, M) we have:
m =
_
_
_
_
_
20
0
0
−20
_
_
_
_
_
For the constraint c(q, Q) on the adversary f here the matrix Q = Q
(2)
of Section 3.9.
Note that the data are chosen in such a way that neither the controls, nor the distur
bances dominate the other, that is, neither c(p, P) = T ⊃ Q = c(q, Q) nor Q ⊃ T
holds. Obviously, in this case the problem cannot be reduced to simpler situations without
disturbances.
The numerical calculation on the basis of Theorem 3.10.1. is carried out in the follow
ing way: after creating the internal estimate E
f
(t
0
, /), we check whether x(t
0
) = x
0
∈
c
−
(t
0
, t, /), taking increasing values of t ∈ [t
0
, t
1
]. In such a way we obtain that this
relation holds for
t = t
∗
= 4.6
i. e.
x
0
∈ c
−
[0, 4.6, /] = c[0, t
∗
, /].
So t
∗
= 4.6 is an upper estimate of t
∗
e
 the closest time instant by which the set / can
be hit for any disturbance f.
According to Theorem 3.10.1 then we keep the trajectory in the ellipsoidal valued map
starting from the above ellipsoid c(0, t
∗
, /).
Figure 3.10.1 shows the internal estimate of the set W
f
(τ, /) at τ = 0 in the form of
_
¦c
−
[0, t, /][t ∈ [0, t
1
]¦.
In Figure 3.10.2 we see again the above set, the ellipsoidal valued map c
−
[t, t
∗
, /], t ∈
[0, t
∗
], as well as the controlled trajectories under two simulated disturbances f resulting
in that the trajectories arrive to the target set / at time t = t
∗
= 4.6.
The layout of the two last Figures is the same as before, with the drawn segments of
coordinate axes corresponding to the state variables ranging from −40 to 40.
—————!!! insert ﬁgures 3.10.1, 3.10.2 !!!————
220
Part IV. ELLIPSOIDAL DYNAMICS:
STATE ESTIMATION and VIABILITY
PROBLEMS
. .
Introduction
This last Part IV of the present book is concentrated around state estimation and viability
problems, emphasizing constructive techniques for their solution worked out in the spirit
of the earlier parts.
We emphasize that here the uncertain items  the initial states, system inputs and mea
surement ”noise” are assumed to be unkown in advance, with no statistical information on
them being available. The problem may then be further treated in two possible settings.
The ﬁrst one is when the bounds on the unknowns are speciﬁed in advance. This leads to the
problem of ”guaranteed state estimation” introduced in Section 1.12. A natural move in
this setting is to use the setmembership (”bounding”) approach. A key element here is the
notion of information set of states consistent with the system equations, the realization of
the measurement and the constraints on the uncertain items. The information set always
includes the unknown actual state of the system and thus gives a setvalued guaranteed
estimate of this state. It may also be useful to ﬁnd a single vectorvalued state estimator,
which may be selected, for example, as the center of the smallest ball that includes the
information set ( which is the socalled the Chebyshev center of this set). One of the main
problems here is to give an appropriate description of the evolution of the information sets
in time and of the dynamics of the vectorvalued estimators. A detailed description of the
bounding approach could be found in monographs [276], [181], [225], and reviews [226],
[187], [186].
The calculation of information sets, even for the ”linearconvex” problems of this book, is
not a simple problem, though. Indeed, it requires to describe more or less arbitrary types
of convex compact sets, which actually are inﬁnitedimensional elements. One may try to
approximate them by ﬁnitedimensional elements however, particularly, by ellipsoids, as
in the present book.
The approximation of information sets by only one or few ellipsoids was described in
[277], [73]. This approximation may turn to be useful in applied problems where compu
tational simplicity stands above accuracy of solution. On the other hand, in sophisticated
applications ( to some types of pursuitevasion games, for example), this rather rough ap
proximation may be misleading. As mentioned above, among the objectives of this book
221
is to produce an ellipsoidal approximation by a parametrized variety of ellipsoids, which,
in the limit, gives an exact representation of the information sets.
20
The parameters of the approximating external ellipsoids are described here as solutions
to systems of ordinary diﬀerential equations. Two types of such equations are given in
Sections 4.3.and 4.5. The latter is derived through the relations of Section 2.6, while the
former follows from Dynamic Programming (DP) considerations. The DP techniques allow
to link the bounding approach with another deterministic approach to state estimation.
This second approach to state estimation presumes that no bounds on the uncertain items
are known. Given is a ”measure of uncertainty” for the uncertain items and the vector
valued estimator is generated through a system which realizes the minimal norm of a certain
inputoutput map or a saddle point of an appropriate dynamic game. The estimators
are then calculated through the knowledge of the information state  the value function
of a certain problem in dynamic optimization calculated as a ” forward” solution of an
appropriate HJB equation. This second scheme is often referred to as the socalled H
∞
approach.
21
The important connection between the two approaches is that the information sets are
the level sets for the information states  the solutions to the HJB equation of the H
∞
approach, ( Section 4.3, see also [32]). Since systems with magnitude constraints on the
inputs generate HJB equations with no classical solutions, the latter equations could
be analyzed within the notions of generalized solutions ( of the ”viscosity” or ”minmax”
types, for example, [82], [289]). In this book these generalized solutions are not calculated
explicitly, but are rather approximated by classical solutions to systems of HJB equations
constructed for adequate classes of linearquadratic extremal problems. In terms of level
sets the last construction is again an ellipsoidal approximation. It is thus observed that
the connection between the two approaches to the deterministic treatment of uncertainty
in dynamics lies, basically, in the incorporation of the same DP equations to both settings.
The DP approach may as well be applied to the calculation of attainability domains. Par
ticularly, if one deals with magnitude constraints on the inputs, then the ellipsoidal approx
imations to these domains may again be achieved through the construction of level sets
for value functions of appropriate linear quadratic extremal problems. However, Section
4.4 indicates that the respective ellipsoids could be transformed to be the same as those
obtained through the purely geometrical considerations of Parts II and III, as described
in Sections 2.7 and 3.2. Similar assertions are also proved for the calculation of viability
kernels , [15].
Among the problems of viability and state estimation are those, where the viability re
striction or the state constraint induced by the measurement equation are not continuous
20
The idea of such representations was indicated in [181], ¸¸ 12.2, 15.1.
21
The H
∞
approach to estimation and feedback control was studied in many papers. Here we mention
[94], [231] and especially those of J.Baras and M.James who introduced the notion of information state,
[30].
222
in time. ( This particularly happens, when the noise in the observations is modelled by
discontinuous functions, that may turn to be only Lebesguemeasurable, for example). A
possible scheme for handling such situations lies in imbedding the original problem into
one with singular perturbations, ( Section 4.6). The new problem is constructed such that
it is free of the inadequacies of the original problem on one hand, and allows an approx
imation of the original one, on the other. A detailed description of this scheme for state
estimation and viability problems of general type is given in references [191], [192]. Section
4.6 presents an ”ellipsoidal ” version of the technique.
The ﬁrst three Sections ( 4.14.3)
223
4.1 Guaranteed State Estimation:
a Dynamic Programming Perspective
We shall begin this Section by discussing the two basic approaches to the deterministic
treatment of uncertainty in the dynamics of controlled processes, as mentioned in the
previous Introduction, treating them in the context of the problem of state estimation
with a further aim on using ellipsoidal techniques.
The ﬁrst of these , as we have seen in Section 1.12, is the bounding approach based on set
membership techniques. Here the uncertain items are taken to be unknown but bounded
with given bounds on the performance range. The estimate is then sought for in the form
of a set  ” the informational domain”, which was described by funnel equations (1.12.10)
or (1.12.11).
The second one is the socalled H
∞
approach based in its linear version on the calculation
of the minimalnorm inputoutput map for the investigated system and the error bound
for the system performance expressed through this norm.
Although formally somewhat diﬀerent, these two approaches appear to have close con
nections. These may be demonstrated particularly through the techniques of Dynamic
Programming that are the topic of this Section. Namely, it will be indicated that both
approaches may be handled through one and the same equation of the HJB type.
For the case of problems with ellipsoidal magnitude constraints on the uncertain items
that are treated in the next Section and are among the the main points of emphasis in
the present book, we shall indicate an approximation technique for solving the respective
HJB equation. The technique is based on an approximation of the original problem
with magnitude constraints by a parametrized variety of problems with quadratic inte
gral constraints. Such a scheme shall then allow a turn to ellipsoidal approximations of
attainability domains.
Let us start with a slightly more general problem than in Section 1.12. Consider again the
system (1.12.1),(1.12.5) with u(t) ≡ 0, rewriting it as
˙ x(t) = A(t)x(t) + f(t), x(t
0
) = x
0
, (4.1.1)
y(t) = G(t)x(t) + v(t), t
0
≤ t ≤ τ. (4.1.2)
We shall assume that the unknown items ζ() = ¦x
0
, f(t), v(t), t
0
≤ t ≤ τ¦ are now bounded
by the inequality
Ψ(τ, ζ()) =
_
τ
t
0
ψ(t, f(t), v(t))dt + φ(x
0
) ≤ µ
2
, (4.1.3)
224
where Ψ(τ, ζ()) reﬂects the accepted uncertainty index for the unknown items.
Particularly, the bounds may be of the quadratic integral type, namely , such that
φ(x
0
) = (x
0
−a, L(x
0
−a)), (4.1.4)
ψ(t, f(t), v(t)) = (f(t) −f
∗
(t), M(t)(f(t) −f
∗
(t))) + (4.1.5)
+(v(t) −v
∗
(t), N(t)(v(t) −v
∗
(t))),
where (p, q) (p, q ∈ IR
k
), stands for the scalar product in the respective space IR
k
; a ∈ IR
n
is a given vector; f
∗
(t), v
∗
(t) are given vector functions of respective dimensions, square
integrable in t ∈ [t
0
, τ]; M(t), N(t) are positive deﬁnite, continuous, and L > 0.
Another common type of restriction is given by magnitude bounds, a particular case of
which is described by ellipsoidal–valued constraints – the inequalities
22
I
0
(x
0
) = (x
0
−a, L(x
0
−a)) ≤ µ
2
(4.1.6)
I
1
(τ, f()) = esssup
t
(f(t) −f
∗
(t), M(t)(f(t) −f
∗
(t)) ≤ µ
2
, (4.1.7)
I
2
(τ, v()) = esssup
t
(v(t) −v
∗
(t), N(t)(v(t) −v
∗
(t)) ≤ µ
2
, (4.1.8)
t ∈ [t
0
, τ].
In this case the functional
Ψ(τ, ζ()) = max¦I
0
, I
1
, I
2
¦. (4.1.9)
As we shall observe in the sequel, the number µ in the restriction (4.1.1) may be or may not
be given in advance and the corresponding solution will of course depend on this speciﬁcity
of the problem. Despite of the latter fact, the aim of the state estimation (”ﬁltering”)
problem could be described as follows:
(a) determine an estimate x
0
(τ) for the unknown state x(τ) on the basis of the available
information: the system parameters, the measurement y(t), t ∈ [t
0
, τ],and the restrictions
on the uncertain items ζ() (if these are speciﬁed in advance).
22
In the coming Sections 4.14.3 the notations for the bounds on the unknowns are independent of those
introduced earlier, emphasizing that the treatment of the state estimation (ﬁltering) problem , as given
here, is independent of the earlier material. In Section 4.4 we shall synchronise these notations with the
earlier ones.
225
(b) calculate the error bounds for the estimate x
0
(τ) on the basis of the same information,
(c) describe the evolution of the estimate x
0
(τ) and the error bound in τ , preferably
through a dynamic recurrencetype relation, an ordinary diﬀerential equation, for example,
if possible.
Let us discuss the problem in some more detail. Suppose that the constraints (4.1.1) with
speciﬁed µ are given together with the available measurement y = y(t), t ∈ [t
0
, τ]. The
bounding approach then requires that the solution would be given through the information
domain A(τ) of Deﬁnition 1.12.1. With A(τ) calculated, one may be certain that for the
unknown actual value x(τ) we have : x(τ) ∈ A(τ) , and may therefore ﬁnd a certain point
ˆ x(τ) ∈ A(τ) that would serve as the required estimate x
0
(τ). As mentioned above, at the
end of the previous Section, this point ˆ x(τ) may be particularly selected as the ”Chebyshev
center” for A(τ) , deﬁned through the relation
min
x
max
z
(x −z, x −z) = max
z
(ˆ x −z, ˆ x −z), z ∈ A(τ), (4.1.10)
and is obviously the center of the smallest ball that includes the set A(τ) .The inclusion
ˆ x(τ) ∈ A(τ)
will be secured as A(τ) is convex.(This may not be the case for the general nonlinear
problem, however, when the conﬁguration of A(τ) may be quite complicated). The set
A(τ) gives an unimprovable estimate of the statespace variable x(τ) , provided the bound
on the uncertain items ( the number µ ) is given in advance .
On the other hand, in the second or H
∞
approach, the value µ for the bound on the
uncertain items is not presumed to be known, while the value of the estimation error
e
2
(τ) = (x(τ) − ˆ x(τ), x(τ) − ˆ x(τ))
is then estimated, in its turn, merely through the smallest number σ
2
that ensures the
inequality
e
2
(τ) ≤ σ
2
Ψ(τ, ζ()) (4.1.11)
under restrictions (4.1.2),(1.12.5).
Since we deal with the linear case, the smallest number σ
2
is clearly the square of the
minimal norm of the inputoutput mapping T , where
e(τ) = T(ζ())
with y = y(t) given. It obviously depends on the type of norm ( the type of functional
Ψ(ζ() ) selected to evaluate η() ). The latter ”worstcase” estimate is less precise than in
226
the ﬁrst approach ( since, as one may observe, it actually indicates a larger error bound).
However, this may sometimes suﬃce for the speciﬁc problem under discussion.
We shall use the upcoming discussion in Section 4.3 to emphasize the connections between
the two approaches and to indicate , through a Dynamic Programming (DP) technique, a
general framework that incorporates both of these, producing either of them, depending
on the ”a priori” information, as well as on the required accuracy of the solutions.
Let us start by introducing a scheme for describing the information domains A(τ), presum
ing y() to be given and restriction (4.1.3) to be of the quadratic integral type (4.1.3)(4.1.5),
to start with.
Denote
η() = ¦x
0
, f(t), t ∈ [t
0
, τ]¦, x(t, t
0
, x
0
, f()) = x(t, t
0
, η())
and
Φ(τ, η()) = (x
0
−a, P
0
(x
0
−a)) +
_
τ
t
0
((f(t) −f
∗
(t), M(t)(f(t) −f
∗
(t))+
+(y(t) −G(t)x(t, x(t, t
0
, η())) −v
∗
(t), N(t)(y(t) −G(t)x(t, x(t, t
0
, η())) −v
∗
(t)))dt
Clearly,
Φ(τ, η()) = ¦Ψ(τ, ζ())[v(t) ≡ y(t) −G(t)x(t, t
0
, η()) −v
∗
(t)¦ (4.1.12)
Deﬁne
V (τ, x) = inf
η(·)
¦Φ(τ, η()) [x(τ, t
0
, η()) = x¦ (4.1.13)
With L, N(t) > 0 the operation ” inf” in the line above may be substituted for ”min”.
Deﬁniton 4.1.1 Given measurement y(t), t ∈ [t
0
, τ] and functional Φ(τ, η()) of (4.1.11),
the respective function V (τ, x) will be referred to as the information state of system
(4.1.1),(4.1.2), relative to measurement y() and criterion Φ.
An obvious assertion is given by
Lemma 4.1.1 The informational domain A(τ) is the level set
A(τ) = ¦x : V (τ, x) ≤ µ
2
¦ (4.1.14)
for the information state V (τ, x).
227
It should be emphasized here that both V (τ, x) ≥ 0 and A(τ) depend on the given mea
surement y(t) as well as on the type of functionals Ψ, Φ and that A(τ) ,= ∅, provided
V
0
(τ) = inf¦V (τ, x) [ x ∈ R
n
¦ ≤ µ
2
(4.1.15)
Since Lemma 4.1.1 indicates that the A(τ) is a level set for V (τ, x), the knowledge of
V (τ, x) will thus allow to calculate the sets A(τ).
We emphasize once more the main conclusions:
(i)the information domain A(τ)is the level set for the informational state V (τ, x) that
corresponds to the given number µ.
(ii)the information state depends both on y() and on the type of functional Φ .
The crucial diﬃculty here is the calculation of the sets A(τ), the function V (τ, x) and
further on ,of the estimate x
∗
(τ) for the unknown state x(τ). The calculations are relatively
simple for an exceptional situation  the linearquadratic case.
As already emphasized above, apart from their separate signiﬁcance, the linearquadratic
solutions will be important in organizing ellipsoidal approximations for systems with mag
nitude constraints.
Let us therefore introduce a DP  type of equation, taking V (τ, x) to be the value function
for the linearquadratic problem (4.1.13) when Ψ(τ, ζ()) is given by (4.1.3) (4.1.5). The
respective function Φ(τ, η()) then obviously satisﬁes the Optimality Principle of Dynamic
Programming [109].
Applying standard techniques , [53], [109], we may observe
is
∂V
∂τ
+ max
v
¦(
∂V
∂x
, (A(t)x + f(t) −(f −f
∗
(t), M(t)(f −f
∗
(t)))−
−(y(t) −G(t)x, N(t)(y(t) −G(t)x))¦ = 0
so that, after an the elimination of f, that the respective HJB equation is as follows
∂V
∂t
+ (
∂V
∂x
, Ax + f
∗
) +
1
4
(
∂V
∂x
, M
−1
(t)
∂V
∂x
)− (4.1.16)
(y(t) −G(t)x, N(t)(y(t) −G(t)x)) = 0
with boundary condition
V (t
0
, x) = (x −a, L(x −a)) (4.1.17)
228
Its solution is a quadratic form
V (τ, x) = (x −z(τ)), T(τ)(x −z(τ)) + k
2
(τ) (4.1.18)
where T(t), z(t), k
2
(t) are the solutions to the following wellknown equations [149], [57],
[276], [181].
˙
T = −TA(t) −A
(t)T −TM
−1
(t)T + G
(t)N(t)G(t), (4.1.19)
T(t
0
) = L,
˙ z = A(t)z +T
−1
G
(t)N(t)(y(t) −G(t)z −v
∗
(t)) + f
∗
(t), (4.1.20)
z(t
0
) = a,
˙
k
2
= (y(t) −G(t)z −v
∗
(t), N(t)(y(t) −G(t)z −v
∗
(t))), (4.1.21)
k
2
(t
0
) = 0
Equations (4.1.19)(4.1.21) are derived by direct substitution of V (t, x) into equation
(4.1.16).
23
An obvious consequence of the given reasoning is the following assertion
Lemma 4.1.2 Under restrictions (4.1.3)  (4.1.5) on the uncertain inputs ζ() =
¦η(), v()¦ the informational domain A(τ) for the system (4.1.1),(4.1.2) is the level set
(4.1.14) for the informational state V (τ, x) ,being an ellipsoid E(z(τ), T(τ)) given by the
relation
A(τ) = E(z(τ), T
−1
(τ)) = (4.1.22)
= ¦x : (x −z(τ), T(τ)(x −z(τ))) ≤ µ
2
−k
2
(τ)¦,
where z(τ), T(τ) > 0, k
2
(τ) are deﬁned through equations (4.1.19)(4.1.21) .
Remark 4.1.1 Note that the matrixvalued function T(t) does not depend on the mea
surement y(), while the scalar function k
2
(t) depends on the measurement. The estimation
error is given by an errror set 1(τ) = A(τ)−‡(τ) which therefore depends only on k
2
(τ).
23
One may easily observe that the ﬁrst two equations (4.1.9), (4.1.20) are the same as in stochastic
”Kalman” ﬁltering theory. However, the third one, (4.1.21), is not present in stochastic theory. It is
speciﬁc for the setmembership approach and reﬂects the dynamics of the ”size” of the information set.
229
Formula (4.1.21) immediately indicates the worstcase realization y
∗
(t) of the measurement
y(t) which yields the ”largest” set A(t) ( with respect to inclusion). Namely, if it possible
to obtain the speciﬁc measurement y
∗
(t) through the triplet
η() = η
∗
() = ¦a, v
∗
()¦, v() = v
∗
()
, (among other possible triplets), then
y
∗
(t) = G(t)x(t, t
0
, η
∗
()) + v
∗
(t)
is the worstcase realization of the measurement and the respective value
V
∗
(τ) = V
0
(τ)[
y(·)=y
∗
(·)
= 0
In order to check the last assertions, let us introduce an equation for the function h(t) =
x(t)−z(t), where x(t) is the realization of the actual trajectory , generated due to equation
˙ x = A(t)x + v, x(t
0
) = x
0
, (4.1.23)
Subtracting (4.1.23) from (4.1.20), we come to
˙
h = /(t)h(t) + K(t)(v(t) −v
∗
(t)) + f(t) −f
∗
(t), h(t
0
) = 0, (4.1.24)
where
/(t) = A(t) +T
−1
(t)G
(t)N(t)G(t), K(t) = T
−1
(t)G
(t)N(t),
If the actual realization x(t) is generated by x
0
= a, f(t) ≡ f
∗
(t), so that x(t) =
x(t, t
0
, η
∗
()), and the realization of the measurement ”noise” is v(t) ≡ v
∗
(t), then (4.1.24),
(4.1.21) yield h(t) ≡ 0, k
2
(t) ≡ 0. We therefore come to
Lemma 4.1.3 The worstcase realization y(t) = y
∗
(t) of the measurement is a function
that ( among other possible triplets ζ()) ) may be generated by the triplet ¦x
0
= a, f(t) ≡
f
∗
(t), v(t) ≡ v
∗
(t)¦ which yields k
2
(τ) = 0 .
The worstcase error set is the ellipsoid
1(τ) = A(τ) −z(τ) = E(0, µ
2
T(τ)) = ¦e : (e, T(τ)e) ≤ µ
2
¦
The other extreme situation is when the measurement is the best possible.
Lemma 4.1.4 There exists a function ( measurement ”noise”) v = ¯ v(t), such that the
triplet ζ() = ¦a, f
∗
(t), ¯ v(t)¦ generates , due to system (4.1.1),(4.1.2), (u(t) ≡ 0), a mea
surement ¯ y() that ensures k
2
(τ) = µ
2
, so that in this case the informational set A(τ)) is
a singleton and
A(τ) = ¦x(τ, t
0
, η
∗
())¦
230
Returning to equation (4.1.24) and rewriting (4.1.21) in view of measurement equation
y(t) = G(t)x(t) + v
∗
(t),
we come to
(4.1.25)
˙
k
2
(t) = (G(t)q(t) + ¯ v(t) −v
∗
(t), N(t)(G(t)q(t) + ¯ v(t) −v
∗
(t))),
and (with f(t) ≡ f
∗
(t) )
˙
h(t) = /(t)h(t) + K(t)(v(t) − ¯ v(t)) (4.1.26)
we shall require that k(t), h(t) satisfy the following boundaryvalue problem
k
2
(t
0
) = 0, k
2
(τ) = µ
2
; h(t
0
) = 0, h(τ) = 0, (4.1.27)
The solution ¯ v(t) to this problem obviously satisﬁes the requirements of the last Lemma,
ensuring particularly, at instant τ, the equalities
x(τ) = x(τ, t
0
, η
∗
()) = z(τ), ¦x(τ)¦ = A(τ).
We leave to the reader to verify that such a solution ¯ v(t) does exist .
Finally , let us assume that there is no measurement equation, so that we simply have the
standard system
˙ x = A(t)x + f(t), x(t
0
) = x
0
, (4.1.28)
with quadratic constraint (4.1.3)(4.1.5),N(t) ≡ 0 Then the set A(τ) is merely the attain
ability domain for system (4.1.28) under the constraint (4.1.3)(4.1.5),N(t) ≡ 0. We may
therefore follow the calculations of the above, setting N(t) ≡ 0.
The procedure then ”automatically” gives the following result
Lemma 4.1.5 Under restrictions (4.1.3)  (4.1.5),N(t) ≡ 0 on the inputs η() = ¦x
0
, f()¦
the attainability A(τ) for system (4.1.21) is the level set (4.1.25) for the function
V (τ, x) = (x −z(τ), T(τ)(x −z(τ))),
where T(t), z(t) are the solutions to the equations
231
˙ z = A(t)z + f
∗
(t), (4.1.29)
z(t
0
) = a,
˙
T +TA(t) + A
(t)T +TM
−1
(t)T = 0, (4.1.30)
T(t
0
) = L,
being an ellipsoid E(z(τ), T(τ)) given by relation
A(τ) = E(z(τ), T(τ)) = (4.1.31)
= ¦x : (x −z(τ), T
−1
(τ)(x −z(τ))) ≤ µ
2
¦.
We are now prepared to extend the results of this Section to problems with magnitude
constraints.
4.2 From Dynamic Programming to
Ellipsoidal State Estimates
Let us now specify the information state V (τ, x) of (4.1.13) for the case of magnitude
constraints, presuming Φ is deﬁned through relations (4.1l.4)  (4.1.7). One may observe
that Φ(τ, η) again satisﬁes the Optimality Principle (and is thus a quasipositional functional
in terms of [170]). One may therefore again calculate V (τ, x) through the HJB equation
or, if necessary, through its generalized versions that deal with nondiﬀerentiable functionals
(see [ooo],[ooo]). We shall not pursue the last direction, but shall rather apply yet another
scheme which will be of direct further use in this book.
Denote
Λ(τ, η(), α, β(), γ()) = α(x
0
−a, L(x
0
−a))+
+
_
τ
t
0
(β(t)(f(t) −f
∗
(t), M(t)(f(t) −f
∗
(t))) + γ(t)(v(t) −v
∗
(t), N(t)(v(t) −v
∗
(t))))dt
Lemma 4.2.1 Assuming M(t), N(t), t ∈ [t
0
, τ] continuous, the restrictions (4.1.6) 
(4.1.8) are equivalent to the system of inequalities
Λ(τ, η(), α, β(), γ()) ≤ 1 (4.2.1)
whatever are the parameters
α ≥ 0, β(t) ≥ 0, γ(t) ≥ 0 (4.2.2)
232
where
α +
_
τ
t
0
(β(t) + γ(t))dt = 1 (4.2.3)
The functions β(), γ() are taken to be measurable, with inequalities (4.2.1) , (4.2.2) being
true for almost all t.
We further denote the triplet ¦α, β(), γ()¦ = ω() and the variety of triplets ω that satisfy
(4.2.1) (4.2.2), as Ω = ¦ω()¦.
Proof. With (4.2.1)  (4.2.3) given , take any triplet ω() , then multiply (4.1.6) by α ,
(4.1.7) by β(t) and (4.1.8) by γ(t), then integrate the last two relations over t ∈ [t
0
, τ].
Adding the results , we obtain (4.2.1) due to (4.2.3).
Conversely, assume (4.2.1) to be true for any ω() ∈ Ω. Taking α = 1, β(t) ≡ 0, γ(t) ≡ 0,
one comes to (4.1.4). Further on , assume for example, that (4.1.7) is false and therefore
that I
1
(t, f()) ≥ > 0 on a set e of measure mes(e) > 0.Then taking α = 0; β(t) ≡
(mes(e)
−1
, t ∈ e; β(t) ≡ 0, t ,∈ e ; γ(t) ≡ 0, one comes to a contradiction with (4.2.4)
and thus (4.2.7) turns to be true. Similarly, the third condition (4.2.8) also follows from
(4.2.1).
24
. Q.E.D.
Using a similar reasoning the reader may now verify the following assertion
Lemma 4.2.2 The function Φ(τ, η()) of (4.1.12),(4.1.9) may be expressed as
Φ(τ, η()) = sup¦Λ(τ, η()), ω()) [ ω() ∈ Ω¦ (4.2.4)
The proof of an analogous fact may be also found in [181].
For further calculations we emphasize the following obvious property
Lemma 4.2.3 The functional Λ(τ, η(), ω()) is convex in η() = ¦x
0
, f()¦ on the set of
elements η() restricted by the equality ¦x : x(τ, η()) = x¦.
24
With slight modiﬁcations the present Lemma 4.2.1 may as well be proved if β(t), γ(t) are taken to be
continuous
233
Due to Lemma 4.2.2 we have
V (τ, x) = inf
η
¦Φ(τ, η()) [ x(τ, η()) = x¦ = (4.2.5)
= inf
η
sup
ω(·)
¦Λ(τ, η(), ω())¦
under restriction ω() ∈ Ω,
x(τ, η()) = x (4.2.6)
The functional Λ(τ, η(), ω()) is linear ( therefore, concave) in ω and convex in η , according
to Lemma 4.1.3. In view of minmaxtype theorems, (see [101], [86]), the order of operations
inf, sup may be interchanged. We therefore come to the relation
V (τ, x) = sup
ω
min
η(·)
Λ(τ, η(), ω()) (4.2.7)
under restrictions ω() ∈ Ω and (4.2.6).
The internal problem of ﬁnding
V (τ, x, ω()) = min¦Λ(τ, η(), ω()) [ η(), x(τ, η()) = x¦ (4.2.8)
may be solved through equation (4.1.16) ( see Remark 4.2.2 ) with V (τ, x) substituted for
V (τ, x, ω()) and M(t), N(t) for β(t)M(t), γ(t)N(t) respectively with boundary condition
being
V (t
0
, x, ω()) = α(x −a, L(x −a)) (4.2.9)
This leads to
Lemma 4.2.4 The information state (4.2.7) is given by
V (τ, x) = sup¦V (τ, x, ω()) [ ω() ∈ Ω¦ (4.2.10)
where V (τ, x, ω) is the solution to equation (4.1.16), under boundary condition (4.2.9) ,
with M(t), N(t) substituted for β(t)M(t), γ(t)N(t),.
Solving problem (4.2.10), we observe
V (τ, x, ω()) = (4.2.11)
= (x −z(τ, γ()), T(τ, ω())(x −z(τ, γ()) + k
2
(τ, γ()),
where T = T(t, ω()), z = z(t, γ()), k = k(t, γ()) satisfy the equations
234
˙
T = −TA −A
T −β
−1
(t)TM
−1
(t)T +γ(t)G
(t)N(t)G(t), (4.2.12)
(4.2.13)
˙ z = A(t)z + γ(t)T
−1
(t)G
(t)N(t)(y(t) −G(t)z −v
∗
(t)) + f
∗
(t),
(4.2.14)
dotk
2
(t) = γ(t)(y(t) −G(t)z −v
∗
(t), N(t)(y(t) −G(t)z −v
∗
(t)))
P
t
0
= αL, z(t
0
) = x
0
, k(t
0
) = 0 (4.2.15)
Finally this develops into the assertion
Theorem 4.2.1 For the system (4.1.1),(4.1.2) the information state V (τ, x) relative to
measurement y() and nonquadratic ( ”magnitude”) criterion (4.1.9), is the upper envelope
V (τ, x) = sup¦V (τ, x, ω()) [ ω() ∈ Ω¦ (4.2.16)
of a parametrized family of quadratic forms V (τ, x, ω()) of type (4.2.11) over the functional
parameter ω() = ¦α, β(), γ()¦ , where ω() ∈ Ω.
As we have observed in the previous sections, the informational domain A(τ) =
E(z(τ), T
−1
(τ)) is deﬁned by V (t, x) through inequality (4.2.13) with µ given. Moreover,
for each of the ellipsoidal level sets
A(τ, ω()) = (4.2.17)
= E(z(τ, ω()), T
−1
(τ, ω()) = ¦x : V (τ, x, ω()) ≤ µ
2
¦
where V (τ, x, ω()) is a nondegenerate quadratic form(!), we obviously have
A(τ) ⊆ X(τ, ω()) = E(z(τ, ω()), (µ
2
−k
2
(τ))T
−1
(τ, ω())), ∀ω() ∈ Ω,
so that (4.2.16) yields the following fact
Theorem 4.2.2 For the system (4.1.1),(4.1.2), with criterion (4.1.9), the informational
set A(τ) is the intersection of ellipsoids
A(τ, ω()) = E(z(τ, ω()), (µ
2
−k
2
(τ))T
−1
(τ, ω()))
namely,
235
A(τ) = ¦∩E(z(τ, ω()), (µ
2
−k
2
(τ))T(
−1
τ, ω())) [ ω() ∈ Ω¦ (4.2.18)
where
z(t) = z(t, γ()), T(t) = T(t, ω()), k
2
(t) = k
2
(t, γ())
are deﬁned through equations (4.2.12)(4.2.15).
The worst case measurement y(t) = y
∗
(t) is a function that may be generated ( among
other possible triplets) by triplet ζ
∗
() , where x
0
= a, f(t) = f
∗
(t), v(t) = v
∗
(t) .This
yields k
2
(τ) = 0 and
V
∗
(τ) = V
0
(τ)[
y(·)=y
∗
(·)
where
V
0
= inf¦V (τ, x)[x ∈ IR
n
¦ = 0
The last part of the theorem that deals with the worstcase measurement y
∗
() may be
checked by substituting ζ() into (4.2.13), (4.2.14) and following the reasoning of the pre
vious Section.
Remark 4.2.1 Observe that again function k
2
(t) depends upon the measurement y(),
while T(t) does not depend upon y().
Remark 4.2.2 The fact that functions β(t), γ(t) are taken measurable does not forbid to
use equation (4.1.16) and the further schemes of Section 4.1 for the function V (t, x, ω()).
This particularly is due to the unicity of solution to the extremal problem (4.2.8). Besides
that, α(t), β(t) may be assumed continuous ( see footnote after Lemma 4.2.1).
In the absence of state constraints induced by the measurement (N(t) ≡ 0), one should
simply delete the restriction (4.1.8) and set γ() = 0 in the previous Theorem. This also
gives k(t) ≡ 0.
Corollary 4.2.1 In the absence of the state constraint (4.1.8) relations (4.2.17),(4.2.18)
generated by equations (4.2.12)  (4.2.15) remain true , provided γ(t) ≡ 0. The set A(t) is
then the attainability domain of Section 1.2 for system (4.1.1) under ellipsoidal magnitude
constraints (4.1.6),(4.1.7).
Further, in Section 4.4, we shall rearrange the results obtained here in terms of earlier
notations and compare them with those obtained Parts IIIII. But prior to that we shall
discuss the calculation of error bounds for the estimation problems.
236
4.3 The State Estimates, Error Bounds
and Error Sets
Let us now pass to the discussion of the estimates and the error bounds. Consider the
informational domain A(τ) to be speciﬁed. Under the assumptions of Sections 4.1, 4.2 set
A(τ) will be closed and bounded. Let us calculate the Chebyshev center of A(). Following
formula (4.1.10), we have to minimaximize the function
min
z
max
x
(x −z, x −z) = max
x
(x − ˆ x, x − ˆ x)
under the restriction
V (τ, x) ≤ µ
2
.
Applying the conventional generalized Lagrangian technique [69], [260], [265], we have
min
z
max
x
¦(x −z, x −z) − λ
2
µ
V (τ, x)¦ (4.3.1)
Since ˆ x(τ) is the center of the smallest ball that includes A(τ)  a convex and compact set,
the inclusion ˆ x(τ) ∈ A(τ) is always true.
Here the number λ
2
µ
is the Lagrange multiplier which generally depends on µ as also does
ˆ x(τ) = ˆ x
µ
(τ). With V (τ, x) being a quadratic form of type (4.1.6), the solution to (4.3.1)
is the center of the ellipsoid (4.1.22), namely, ˆ x(τ) = z(τ), whatever is the value of µ.
Summarizing the results, we have
Lemma 4.3.1 The minmax estimate ˆ x(t) (the Chebyshev center) for the informational
domain A(τ) of Section 1.12, satisﬁes the property
ˆ x(τ) ∈ A(τ)
and in general depends on µ : ˆ x(τ) = ˆ x
µ
(τ).
In the linearquadratic case (4.1.3)(4.1.5) the vector
ˆ x(τ) = z(τ)
is the center z(τ) of the ellipsoid E(τ, T
−1
(τ)) described by the (4.1.22) and does not depend
on the number µ .
237
In order to compare the setmembership (bounding) and the H
∞
approaches, let us ﬁnd
the estimate ¯ x(τ) for the H
∞
approach to state estimation. Then we have to solve the
following problem:
Find the smallest number σ
2
that ensures
min
z
max
ζ(·)
¦(x −z, x −z) −σ
2
Ψ(τ, ζ())¦ ≤ 0
under the conditions
x(τ, η()) = x; G(t)x(t, t
0
, η()) + v(t) ≡ y(t); t
0
≤ t ≤ τ.
This, however, is equivalent to the problem of ﬁnding the smallest number σ
2
= σ
2
0
that
ensures
min
z
max
x
¦(x −z, x −z)− (4.3.2)
σ
2
¦inf
η(·)
Φ(τ, η())[x(τ, t
0
, x
0
, η()) = x¦ ≤ 0
or, equivalently,
min
z
max
x
¦(x −z, x −z) −σ
2
V (τ, x)¦ ≤ 0 (4.3.3)
It is not diﬃcult to observe the following;
Lemma 4.3.2 In the quadratic case (4.1.3)(4.1.5) the Lagrange multiplier λ
µ
of (4.3.1)
satisﬁes the equality
λ
2
µ
= σ
2
0
, ∀µ
and the solution ¯ x(τ) to (4.3.2), (4.3.3)) satisﬁes
¯ x(τ) = ˆ x(τ) = z(τ) = z
µ
(τ), ∀µ.
This conclusion follows from standard calculations in quadratic programming and linear
quadratic control theory.
As an exercise in optimization we oﬀer the reader to prove the next proposition:
Lemma 4.3.3 In the case (4.1.3), (4.1.9) of magnitude constraints the Lagrange multiplier
λ
µ
of (4.3.1) and the number σ = σ
0
of (4.3.2), (4.3.3) are related as follows
λ
2
µ
→ σ
2
0
, (µ → ∞)
with estimate
ˆ x
µ
(τ) → ¯ x(τ), (µ → ∞)
238
Remark 4.3.1 Among the recommended estimates for deterministic state estimation prob
lems of the above one may encounter the following one,([129]):
z
∗
(τ) = argmin¦V (τ, x) [ x ∈ R
n
¦.
Such a selection of the estimate is certainly justiﬁed for the linearquadratic problem since
then, as we have seen,
z
∗
(τ) = ˆ x(τ) = ¯ x(τ) = z(τ), (4.3.4)
and all the estimate types coincide (!).
However, as soon as we apply a nonquadratic functional Φ(τ, η()), like the one given by
(4.1.3), (4.1.9,), one may observe that all of the estimates (4.3.4) may turn to be diﬀerent,
despite the linearity of the system.
One of the basic elements of the solutions to the state estimation problem are the error
bounds for the estimates. For the setmembership ( bounding) approach, when the bounds
on the uncertain items ζ() are speciﬁed in advance, these are naturally given in the form
of error sets. Here the error set may be taken as
1(τ) = A(τ) − ˆ x(τ).
As indicated above, the set 1 will be the ”largest possible” (with respect to inclusion) if the
realizations of the uncertain items ζ() will generate the worstcase measurement y
∗
(t). As
we have seen, for the problems treated here these are ¦x
∗
0
= a; v(t) ≡ v
∗
(t); w(t) ≡ w
∗
(t)¦.
On the other hand, the set 1 is the ”smallest possible” if it a singleton 1 = 0, in which
case it is generated by the bestcase measurement ¯ y(τ) .
For the quadratic integral constraint the bestcase measurement is described in Lemma
4.1.4. The principles for identifying such measurements for magnitude constraints are
indicated in references [186], [187].
As for the H
∞
approach, the estimation error e
2
(τ) will depend upon the number σ
2
in
the inequality (4.3.2). The smallest possible value σ
2
0
of this of number depends in general
on the given measurement y(t) that determines the restriction (4.3.2).
Among all possible measurements, the largest possible value of σ
2
0
will be attained again
at the worst case measurement y
∗
(t) speciﬁed in Lemmas 4.1.3 and 4.2.2. On the other
hand, the bestcase measurement is the one that would ensure σ
2
0
= 0.
Exercise 4.3.1 In the H
∞
setting, for the case of quadratic integral index Ψ(τ, ζ()), check
whether the bestcase function ¯ y()) of Lemma 4.1.4 does yield the value σ
2
0
= 0.
239
Remark 4.3.2 Given measurement y(), suppose we have calculated number σ
2
0
for the H
∞
approach. If, moreover, we are also given the number µ in (4.1.3), then, in the quadratic
integral case (4.1.3)(4.1.5), the number µσ
0
≥ 0 will be the radius of the smallest sphere
that surrounds the error set 1(τ) :
1(τ) = A(τ) − ˆ x(τ) ⊆ ¦x : (x, x) ≤ µ
2
σ
2
0
¦.
The properties of the Chebyshev centers for the setmembership and the H
∞
solutions in
the nonlinear case yield yet more diversity in the estimates. This however leads us beyond
the scope of the present book.
We shall now compare the ellipsoidal relation derived in Sections 4.1, 4.2 with those intro
duced earlier, in Parts II, III.
4.4 Attainability Revisited.
Viability Through Ellipsoids
In this Section we again deal with the main object of this book, namely, with systems
restricted by magnitude constraints. We shall ﬁrst rearrange the relations of Section 4.2
using the notations of the earlier parts.
Let us start with system
˙ x = A(t)x + u + f(t), (4.4.1)
under constraints
u ∈ (c(p(t), P(t)), x(t
0
) ∈ X(a, X
0
),
taking f(t) to be ﬁxed. We also take A(t) ≡ 0 , which gives no loss of generality, as indicated
in Section 1.1. Then , due to Corollary 4.2.1, the attainability domain A(t) = A(τ, t
0
, X
0
)
at time τ, from set A
0
= A(t
0
) ( see also Deﬁnition 2.1 ) may be derived from Theorem
4.2.2 if one substitutes γ(t) ≡ 0, A(t) ≡ 0. Reformulating the Theorem for this particular
case, we have, with µ = 1,
Theorem 4.4.1 The attainability domain A[τ] = A(τ, t
0
, A
0
) for system
˙ x = u + f(t),
under restrictions (4.1.2), (4.1.4), is the intersection of ellipsoids
A[τ] = A(τ, χ()) = c(z(τ), T
−1
(τ, χ())),
240
namely,
A[τ] = ¦∩c(z(τ), T
−1
(τ, χ()))[χ()¦, (4.4.2)
where χ() = ¦α, β()¦,
α > 0, β(t) > 0, α +
_
τ
t
0
β(t)dt = 1,
and z(t) = z(t), T(t) = T(t, χ()) are deﬁned through equations
˙ z = p(t) + f(t), x(t
0
) = a, (4.4.3)
˙
T = −β
−1
(t)TM
1
(t)T, T
t
0
= αL, (4.4.4)
In order to compare the last result given in terms of equations (4.4.2),(4.4.3) to the one
given in Theorem 3.2.5 in terms of equation (3.2.5), we make the following substitutions
X(t) = T
−1
(t), X(t
0
) = X
0
= L
−1
, P(t) = M
−1
(t), a = x
∗
,
using also the relation
˙
T
−1
= −T
−1
˙
TT
−1
. This gives
˙
X(t) = β
−1
(t)P(t), X(t
0
) = (1 −
_
τ
t
0
β(s)ds)
−1
X
0
, (4.4.5)
or , after integration from t
0
to τ,
X(τ) = (1 −
_
τ
t
0
β(t)dt)
−1
X
0
+
_
τ
t
0
β
−1
(t)P(t)dt, (4.4.6)
Let us now take equation (3.2.5) of Section 3.2 and also integrate it from t
0
to τ with same
initial set X
0
. This gives
X
+
(τ) = κ(τ, t
0
)
_
τ
t
0
κ
−1
(t, t
0
)π
−1
(s)ds + κ(τ, t
0
)X
0
, (4.4.7)
where
κ(τ, t
0
) = exp
__
τ
t
0
π(t)dt
_
Comparing (4.4.6) and (4.4.7), one may observe, by direct calculation, that by setting
241
π(t) = β(t)(1 −
_
t
t
0
β(s)ds)
−1
, t
0
≤ t ≤ τ, (4.4.8)
relation (4.4.7) is transformed into (4.4.6).
Indeed, taking
exp −
__
t
t
0
π(t)dt
_
= 1 −
_
t
t
0
β(s)ds,
and diﬀerenting both parts in τ, we have
π(t) exp −
__
t
t
0
π(s)ds
_
= β(t), (4.4.9)
which gives (4.4.8) , on one hand, and transforms (4.4.7) into (4.4.6) on the other.
One may observe from (4.4.9) that for any function π(t) > 0 deﬁned on the interval
t
0
≤ t ≤ τ there always exists a function β(t) of the type
β(t) > 0, t
0
≤ t ≤ τ, (1 −
_
τ
t
0
β(t)dt) > 0,
Since for any β(t) tof the last type there exists a function π(t), due to (4.4.8), we are in
the position to formulate
Theorem 4.4.2 The attainability domain A[τ] of (4.4.2) allows an equivalent represen
tation
A[τ] = ¦∩c(z(τ), X(τ))[β()¦ = ¦∩c(x
∗
(τ), X
+
(τ))[π()¦ (4.4.10)
where z(τ) = x
∗
(τ) and the parametrized varieties of matrices ¦X(τ)¦, ¦X
+
(τ)¦ are de
scribed by equivalent relations (4.4.6), (4.4.7). The respective parameters β(), π() are
related due to (4.4.8), (4.4.9).
The respective diﬀerential equations for A[t], A
+
[t] t
0
≤ t τ are given by (4.2.5),(4.2.6) and
(4.4.5). These equations may be transformed one into another through the substitutions
given above, so that at instant τ they would yield the same solutions.
Let us now pass to the discussion of dynamic programming techniques for the viability
problem.
242
Consider system (4.1.1) under constraints (4.1.2), (4.1.4) on u, x
0
and viability constraint
x(t) ∈ c(q(t), Q(t)), t
0
≤ t ≤ τ, (4.4.11)
with function f(t) ≡ 0. Constraint (4.4.11) follows from (4.1.2), (4.1.8) if y(t) ≡ 0, G(t) ≡
I, N
−1
(t) ≡ Q(t), v
∗
(t) = −q(t) ( however, (4.4.11) may now be taken to be true every
where).
We shall look for the viability set W[τ] at given instant τ which is the set of all points
x = x(τ) for each of which there exists a control u = u(t) that ensures the viability
constraint:
x[t] = x(t, t
0
, x) ∈ c(q(t), Q(t)), τ ≤ t ≤ t
1
,
Set W[τ] coincides with the solvability set J[τ] = J(τ, t
1
, /) of Deﬁnition 1.9.4 if we
take G(t) ≡ I, ¸(t) ≡ c(q(t), Q(t)), / = c(q(t
1
, Q(t
1
)), .
We shall determine W[τ] as the level set
W[τ] = ¦x : V
v
(τ, x) ≤ 1¦
for the viability function V
v
(τ, x) , which we deﬁne as the solution to the following problem:
V
v
(τ, x) = min
u(·)
¦Φ(τ, u())[x[t] = x(t, τ, x[u())¦ = (4.4.12)
where
Φ(τ, u()) = max¦J
0
, J
1
, J
2
¦,
and
J
0
(x[t
1
]) = (x[t
1
] −q(t
1
), Q(t
1
)(x[t
1
] −q(t
1
))), (4.4.13)
J
1
(τ, u()) = esssup(u(t) −p(t), P(t)(u(t) −p(t)))), (4.4.14)
J
2
(τ, x[t]) = max(x[t], Q(t)x[t]), (4.4.15)
with t ∈ [τ, t
1
] and x[t] = x(t, τ, x[u()) being the trajectory of system (4.4.1) that starts
at position ¦τ, x¦ and is steered by control u(t).
243
The solution to this problem may be described by a certain ”forward” dynamic program
ming (HJB) equation, [109]. In order to avoid generalized solutions of this equation, we
shall follow the scheme of Section 4.4.2 by solving a linearquadratic control problem. It
is to minimize
Λ(τ, x, u(), ω()) =
_
t
1
τ
(γ(t)x[t]Q(t)x[t] + β(t)(u(t) −p(t), P(t)(u(t) −p(t)))dt+
+α(x[t
1
], Q(t
1
)x[t
1
])
over u(
˙
), with x[t] = x(t, τ, x[u()). Here ω() = ¦α, β(), γ())¦ and
α > 0, β(t) > 0, γ > 0, α +
_
t
1
τ
(β(t) + γ(t))dt = 1. (4.4.16)
The variety of such elements ω() is furthe denoted as Ω.
Then, in analogy with the previous Section 4.2, we have
V
v
(τ, x) = min
u(·)
¦Φ(τ, u())[x[t] = x(t, τ, x[u())¦ = (4.4.17)
= min
u(·)
sup
ω(·)
Λ(τ, x, u(), ω()).
The operations of min and sup may again be interchanged. Doing this, we denote
V
v
(τ, x) = sup ω()V
∗
(τ, x, ω), (4.4.18)
where
V
v
(τ, x, ω()) = min u()Λ(τ, x, u(), ω()).
We again look for this function as a quadratic form
V
v
(τ, x, ω) = (4.4.19)
= (x −z(τ, γ()), T(τ, ω())(x −z(τ, γ())),
where T[t] = T(t, ω()), z[t] = z(t, γ()), k = k(t, γ()) satisfy the equations
˙
T = −TA(t) −A
(t)T + β
−1
(t)TP(t)T −γ())Q
−1
(t), (4.4.20)
244
˙ z = A(t)z −γ(t)T
−1
Q
−1
(t)(z + v
∗
(t)) + p(t), (4.4.21)
k
2
(t) = −γ()(z + v
∗
(t), Q
−1
(t)(z + v
∗
(t))), (4.4.22)
T(t
1
) = αQ
−1
(t
1
), z(t
1
) = q(t
1
), k(t
1
) = 0. (4.4.23)
It may be more convenient to deal with matrix X
v
(t) = T
−1
[t] which satisﬁes equation
˙
X
v
= A(t)X
v
+ X
v
A
(t) + +γ())X
v
Q
−1
(t)X
v
−β
−1
(t)P(t), (4.4.24)
X
v
(t
1
) = α
−1
Q(t
1
), . (4.4.25)
Following the reasoning of Section 4.2, we formulate the following assertion
Lemma 4.4.1 The viability function V
v
(τ, x) is the upper envelope
V
v
(τ, x) = sup¦V
v
(τ, x, ω())[ω() ∈ Ω¦ (4.4.26)
of a parametrized variety of quadratic forms V
v
(τ, x, Ω()) of type (4.4.19) over the func
tional parameter ω() = ¦α, β()), γ())¦, where ω() ∈ Ω.
Since the level sets for V
v
(τ, x, ω()) are ellipsoids, namely
W[τ, ω()] = c(z[τ], (1 −k
2
[τ])
−1
X
v
[τ])
and since W[τ] is a level set for V
v
(τ, x), we are able, due to Lemma 4.4.1, to come to
Theorem 4.4.3 The viability set W[τ] is the intersection of ellipsoids , namely
W[τ] = ¦∩c(z[τ], (1 −k
2
[τ])
−1
X
v
[τ])[omega() ∈ Ω¦, (4.4.27)
where z, k, X
v
are deﬁned through equations (4.4.21)(4.4.25).
The setvalued function W[t], τ ≤ t ≤ t
1
is known as the viability tube which may therefore
also be approximated by ellipsoids along the schemes of this Section.
245
4.5 The Dynamics of Information Domains.
State Estimation as a Tracking Problem
As we have remarked before , the information domains of Sections 4.1.1 or 4.1.12 are
nothing else than attainability domains under state constraints when the last are given,
for example, by inequalities (4.1.8), (4.1.2). These domains A(τ) were therefore already
described through Dynamic Programming approaches in Sections (4.1.1)(4.1.3). However,
some other types of ellipsoidal estimates and their dynamics may be derived for A(τ)
directly, through the funnel equations of Sections 1.9, 1.12 and the ”elementary” formulae
of Part II.
In this Section we consider the attainability domain A[τ] for system
˙ x = u(t) + f(t), (4.5.1)
under the constraints and notations of Section 4.4 for u(t), x
0
and state constraint
x(t) ∈ c(y(t), K(t)), (4.5.2)
where the matrix valued function K(t) > 0, K(t) ∈ L(IR
n
, IR
n
) and the function y(t) ∈ IR
n
(the observed output in the state estimation problem) are assumed to be continuous .
25
To treat this case we shall ﬁrst follow the funnel equation of type (1.9.21). For the attain
ability domain A(t) under state constraint (4.5.2) this gives
A(t + σ) = (A(t) + σc(p(t), P(t) + f(t)) ∩ c(y(t + σ), /(t + σ)) + o(σ), , σ > 0.
Presuming A(t) = c(x(t), X(t)), we shall seek for the external ellipsoidal estimate c(x(t +
σ), X(t + σ)) of A(t + σ) . To do this we shall use relations ( 2.3.1) and (2.7.14).
Namely, using (2.3.1), we ﬁrst take the estimate
c(x(t), X(t)) + σc(p(t), P(t) + f(t)) ∈ c(˜ x(t),
˜
X(t)),
where ˜ x(t) = x(t) + σp(t), and
25
The case of measurable functions y(t) which allows more complicated discontinuities in y(t) and is of
special interest in applications is considered lower in Section 4.6.
246
˜
X(t) = (1 + q)X(t) + (1 + q
−1
)σ
2
P(t), q > 0, (4.5.3)
Further, using (2.7.4), we have
c(˜ x,
˜
X) ∩ c(y(t), K(t)) ∈ c(x(t + σ, X(t + σ)),
where
x(t + σ) = (I −M)(x(t) + σp(t)) + My(t + σ), (4.5.4)
and
X(t + σ) = (1 + π)(I −M)
˜
X(t)(I −M)
(4.5.5)
+(1 + π
−1
)MK(t + σ)M
, π > 0.
Making the substitutions
q = σ¯ q, π = σ¯ π, M = σ
¯
M,
collecting (4.4.3) (4.4.5) together and leaving the terms of order ≤ 1 in σ, we come to
x(t + σ) −x(t) = σp + σ
¯
M(y(t + σ) −x(t))
and also
X(t + σ) −X(t) = σ((¯ π + ¯ q)X(t) −
¯
MX −X
¯
M
+ ¯ q
−1
P + ¯ π
−1
¯
MK(t + σ)
¯
M
).
Dividing both parts of the previous equations by σ > 0 and passing to the limit σ → +0,
we further come, in view of the continuity of y(t), K(t), to diﬀerential equations ( deleting
the bars in the notations)
˙ x = p(t) + M(t)(y(t) −x), (4.5.6)
˙
X = (π(t) + q(t))X +q(t)
−1
P −M(t)X −XM
(t)+ (4.5.7)
+π
−1
M(t)K(t)M
(t),
247
where
x(t
0
) = x
0
, X(t
0
) = X
0
, (4.5.8)
and π(t) > 0, q(t) > 0, M(t) are continuous functions.
What further follows from Theorem 1.3.3 (formula (1.3.31)) and Lemma 2.7.3 (formula
2.7.11) is the assertion
Theorem 4.5.1 The attainability domain A(τ) for system (4.5.1) under restrictions
(4.1.2),(4.1.4) and state constraints (4.5.2) ( with y(t), K(t) continuous) satisﬁes the in
clusion A(τ) ∈ c(x(τ), X(τ)), where x(t), X(t) satisfy the diﬀerential equations (4.5.6)
(4.5.8) within the interval t
0
≤ t ≤ τ.
Moreover, the following relation is true
A(τ) = ∩¦c(x(t), X(t))[π(), q(), M()¦ (4.5.9)
where π(t) > 0, q(t) > 0, M(t) are continuous functions.
Exercise 4.6.1 Suppose the state constraint (4.5.2) is substituted by relation
G(t)x(t) ∈ c(y(t), K(t)) (4.5.10)
where y(t) ∈ IR
m
, K(t) ∈ L(IR
m
, IR
m
), G(t) ∈ L(IR
n
, IR
m
) and G(t) is continuous.
Prove that in this case the previous relations together with Theorem 4.5.1 still hold with
obvious changes. Namely, (4.1.6), (4.1.7) should be substituted by
˙ x = p(t) + M(t)(y(t) −G(t)x), (4.5.11)
˙
X = (π(t) + q(t))X + q(t)
−1
P(t)− (4.5.12)
−M(t)G(t)X −XG
M
(t) + π
−1
M(t)K(t)M
(t),
with same boundary conditions (4.5.8).
Remark 4.5.1 To obtain the equations for A(τ) through Dynamic Programming, we just
have to take relations (4.2.12)  (4.2.15) and set v
∗
(t) ≡ 0, N(t) ≡ K
−1
(t).
248
We thus have two sets of relations for A(τ), namely, the one given in Section 4.1.2 and the
one given in the present section.
Then each of the approaches leads to a variety of ellipsoidal sets that include A(τ) , on
one hand and allow exact representations of types (4.2.18) ( variety 1) and (4.5.9) (variety
2), on the other.
It is not diﬃcult to observe that variety 2 of ellipsoids given in (4.5.9) depends on more
parameters than in (4.2.18) and is therefore ”richer” than variety 1. This has the following
implication: if among the varieties 1 or 2 we are to select an ellipsoid optimal in some
conventional sense ( see, for example, Section 4.2), then we may expect that variety 2 (
the ”richer” one) will produce a ”tighter” optimal ellipsoid than variety 1. Elementary
examples of such kind are given above, in Section 2.6.
Remark 4.5.2 System(4.5.6)(4.5.8) was derived under the assumption that function y(t)
is continuous. However, we may as well assume that y(t) is allowed to be piecewise
continuous (”from the right”). Then the respective value in equations (4.5.6), (4.5.11)
should be y(t) = y(t + 0).
One could observe, that the funnel equation used earlier in the proof of Theorem 4.5.1,
is the one given in (1.12.10). A similar derivation is possible, however, if we use funnel
equation (1.12.11), which for constraint (4.5.2) is as follows:
lim
σ→+0
σ
−1
h
+
(A(t + σ), A(t) ∩ c(y(t), K(t)) + σc(p(t), P(t))) = 0.
Then for the maximal solution of this equation, with A[t
0
] = X
0
, we have
A[t +σ] = A[t] ∩ c(y(t), K(t)) + σc(p(t), P(t)) + o(σ),
and subsequently follow the operations: ﬁrst take the estimate
A[t] ∩ c(y(t), K(t)) ⊆ c(ˆ x(t),
ˆ
X(t)),
then the estimate
c(ˆ x(t),
ˆ
X(t)) + σc(p(t), P(t)) ⊆ c(x(t + σ), X(t + σ)).
With obvious modiﬁcations the futher reasoning is similar to proof of Theorem 4.5.1 (
see (4.5.3)(4.5.5) and following relations). The conclusion is that ﬁnally we again come to
249
equations (4.5.6) (4.5.8), except that this time we did not use the continuity of y(t) , having
implicitly used its piecewisecontinuity ”from the right.” (We have implicitly required that
at each point t we have
σ
−1
_
t+σ
t
y(s)ds → y(t), σ → +o,
This is ensured by the latter piecewise continuity).
Lemma 4.5.1 Theorem 4.5.1 remains true if y(t) is piecewisecontinuous from the right.
Exercise 4.6.1  a. Solve Exercise 4.6.1 under conditions of Lemma 4.6.1.
We shall now give a ”control” interpretation of the state estimation problem. Consider
again the state estimation ( attainability) problem for system (4.5.1),(4.5.10),(4.1.2),(4.1.4),
f(t) ≡ 0. Though the set A[τ] may be approximated both externally and internally by
ellipsoidalvalued functions, we shall again further deal only with the former case. (An
indication on a general scheme for internal ellipsoidal approximations of intersections of
ellipsoids is given at the end of Section 2.6 ).
As in Section 3.3, let us introduce an ”ellipsoidal” funnel equation, but for the present
problem now. Consider the evolution equation
lim
σ→+0
σ
−1
h
−
(c(t + σ), c(t) ∩ c(y(t), K(t)) + σc(p(t), P(t))) = 0,
(4.5.13)
t
0
≤ t ≤ t
1
, c[t
0
] = c(x
0
, X
0
).
A setvalued function c
+
[t] will be deﬁned as a solution to (4.5.13) if it satisﬁes (4.5.13)
for almost all t and is ellipsoidalvalued. Obviously the solution c
+
[t] is nonunique and
satisﬁes the inclusion
c
+
[t] ⊃ A[t], t
0
≤ t ≤ t
1
, c
+
[t
0
] = A[t
0
].
Moreover, as a consequence of Lemmas 2.2.1 and 2.6.3, one may come to
Theorem 4.5.2 For any t
0
≤ t ≤ t
1
the following equality is true
A[t] =
¦c
+
[t][c
+
[] is a solution to (4.5.13) ¦.
250
The ellipsoidal solutions c
+
[] = c(x
−
(), X
−
()) to (4.5.13) allow explicit representations
through appropriate systems of ODE’s for the centers x
−
() and the matrices X
−
() >
0 of these ellipsoids. One may check that among these are, particularly, the solutions
x(t, M), X(t, M, π, q) to system (4.5.6)(4.5.8). A more complicated problem is to ﬁnd the
”tightest” estimates , or, in other terms, the minimal ( with respect to inclusion ) ellipsoidal
solutions to (4.5.13).
Exercise 4.6.2. Check ,whether it is possible to select parameters M(t), π(t), q(t) in (4.5.6)
(4.5.7), so as to produce an inclusionminimal external ellipsoidal estimate c(x(τ), X(τ)) ⊇
A[τ] at time τ.
The indications on how to select such estimates for the static case are given in Sections
2.3,2.6.
As we have earlier observed in similar situations, parameters M, π, q may be interpreted
as controls and the problem of specifying the ”best” ellipsoids  as control problems. This
also leads to the following considerations.
Denote the external ellipsoid of system (4.5.11),(4.5.12) as c
ω
[t] = c(x(t, M), X(t, M, π, q)),
where ω = ¦M(), π(), q()¦. The center x(t, M) of the tube c
ω
[t], t
0
≤ t ≤ t
1
, satisﬁes
equation (4.5.11) with x(t
0
) = x
0
.
Let us denote the actual trajectory to be estimated, as x
∗
(). By construction, the inclusions
c
ω
[t] ⊇ A
∗
[t] ⊇ x
∗
(t), t
0
≤ t ≤ t
1
,
are true. Therefore the approximate estimation procedure is that the estimator x(t, M)
tracks the unknown trajectory x
∗
(t) and the ellipsoid c
ω
[t] around it plays the role of a
guaranteed conﬁdence region. The set c(0, X(t, M, π, q)) then estimates the error set
¦A
∗
[t] −x(t, M)¦ of the estimation process.
The trajectory of the estimator x(t, M) depends on the measurement output y(s), t
0
≤
s ≤ t, and therefore realizes a feedback procedure. (The parameter M may be also chosen
through feedback from y()).
The tracking procedure described here is similar in nature to a diﬀerential game of obser
vation.
26
.
Example 4.6.1
Given is a 4dimensional system
˙ x = A(t)x + u(t), G(t)x = y(t) + v(t), (4.5.14)
26
A feedback duality theory for diﬀerential games of observation and control was described in [179]
251
u(t) ∈ c(p(t), P(t)), v(t) ∈ y(t) +c(0, K(t)),
over the time interval [0, 5]. We ﬁrst describe the the attainability domain under state
constraints, ( or , interpreting y(t) as the observation,  the information domain).
The initial state is bounded by the ellipsoid A
0
= c(x
0
, X
0
) at the starting time t
0
= 0
with
x
0
=
_
_
_
_
_
1
0
1
0
_
_
_
_
_
and X
0
=
_
_
_
_
_
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
_
_
_
_
_
.
The matrix A is constant:
A(t) ≡
_
_
_
_
_
0 1 0 0
−8 0 0 0
0 0 0 1
0 0 −4 0
_
_
_
_
_
,
It describes the position and velocity of two independent oscillators. The unknown inputs
u(t) ∈ c(p(t), P(t)) are bounded by constant constraints
u(t) ∈ c(p(t), P(t)) where
p(t) ≡
_
_
_
_
_
0
0
0
0
_
_
_
_
_
, and P(t) ≡
_
_
_
_
_
1 0 0 0
0 0.01 0 0
0 0 1 0
0 0 0 0.01
_
_
_
_
_
,
(this form of the bounding sets makes the system coupled).
27
The state constraint is deﬁned by the data
G(t) ≡
_
_
_
0 1 0 0
0 0 0 0
0 0 0 1
_
_
_, k(t) ≡
_
0
0
_
, K(t) ≡
_
16 0
0 25
_
.
In Figure 4.5.1 we show the graph of external ellipsoidal estimates of the 4dimensional
state space variable x(t) — with and without state constraints — presenting them in four
windows, being conﬁned to projections onto the planes spanned by the ﬁrst and second,
third and fourth, ﬁrst and third, and second and fourth coordinate axes, in a clockwise
27
Following Section 1.1, we could transform this system to an equivalent form , where A = 0 and
P = P(t) is timedependent.
252
order starting from bottom left. The drawn segments of coordinate axes corresponding to
the output variables range from −30 to 30. The skew axis in Figure 4.5.1 is time, ranging
from 0 to 5.
Calculations are based on the discretized version of the system (4.6.14) and the schemes
of this Section. The parameters m, π, q in each step are selected as traceminimal along
the results of section 2.6. Figure 4.5.2 shows the trajectory of the centers, initial sets and
the ellipsoidal estimates of the state space variables x, projected on to the planes spanned
by two coordinate axes ( chosen with the same arrangement of the four windows as in Fig.
4.5.1), with drawn segments ranging from −10 to 10.
We turn now to the guaranteed state estimation problem interpreted as a tracking prob
lem, as described above. We keep the above parameter values of the time interval, A(t),
c(x
0
, X
0
), c(p(t), P(t)), and G(t) and the same calculation schemes.
We model the trajectory x
∗
(t) , — the one to be tracked — by using the following construc
tion for the triplet ζ
∗
() = ¦x
∗
0
, u
∗
(), v
∗
()¦. The initial value x
∗
0
is a (randomly selected)
element at the boundary of the initial set A
0
= c(x
0
, X
0
). The input u
∗
() is of the so
called extremal bangbang type: The time interval is divided into subintervals of constant
lengths. A value u is chosen randomly at the boundary of the respective bounding set,
that is in case of the input u
∗
(t), of set T(t) = c(p(t), P(t)) and its value is then deﬁned as
u
∗
(t) = u over all the ﬁrst interval and as u
∗
(t) = −u over the second. Then a new random
value for u is selected and the above procedure is repeated for the next pair of intervals,
etc. For modeling the measurement noise v
∗
() (generating together with x
∗
0
and u
∗
() the
actual measurement y
∗
()), we use a similar procedure.
As is well known, the size of the error set of the estimation depends on the nature of
v
∗
(). According to [181], if we chose it in such a way that it takes a constant value at
the boundary of c(0, K(t)) over all the time interval under study, then it corresponds to
the worst case. This means in large conﬁdence regions, while using e. g. the extremal
bangbang construction, ’good’ noises are created, reducing the conﬁdence regions’ size.
Figure 4.5.3 shows the process developing over time, — the drawn segments of coordinate
axes corresponding to the output variables range from −20 to 20. In Figure 4.5.4 the
initial sets of uncertainty (appearing as circles) are displayed in phase space, as well as
the conﬁdence region at the ﬁnal intant. Coordinate axes range here from −10 to 10. The
trajectory drawn with the thick line is the actual output x
∗
(t). The thin line represents
the trajectory of the centers x(t, M) of the projections of the tracking ellipsoids. Figures
4.5.5, 4.5.6 show how much the estimation can improve if the noise changes from worst
to better — although we obtain here only external ellipsoidal estimates of the true error
sets. Opposed to Figures 4.5.3,4.5.4, where the noise was constant, we chose its range to
be within [−0.5, 0.05]. The range of the coordinate axes is again [−20, 20].
253
4.6 Discontinuous Measurements and
the Singular Perturbation Technique
The idea of applying singular perturbation techniques to the state estimation problem of
the present book is motivated by the necessity to treat measurements y(t) that are of ”bad”
nature, possibly discontinuous. Indeed, in this Section we shall allow Lebesguemeasurable
realizations y(t) of the measurement output.
Consider system ( 1.12.19), (1.12.3)(1.12.5), where all the sets involved are ellipsoids:
˙ x ∈ A(t)x +c(p(t), P(t)), (4.6.1)
x(t
0
) ∈ c(x
0
, X
0
), (4.6.2)
G(t)x(t) ∈ y(t) +c(0, K(t)). t
0
≤ t ≤ τ (4.6.3)
Here
p : [t
0
, t
1
] → IR
n
, y : [t
0
, t
1
] → IR
m
,
P(t) ∈ L(IR
n
, IR
n
), K(t) ∈ L(IR
m
, IR
m
), x
0
∈ IR
n
,
the matrices X
0
, P(t), K(t) are symmetric and positively deﬁnite.
Our goal will be to ﬁnd the exact external ellipsoidal estimate for the attainable set X[τ]
for the system (12.6.1)(12.6.3).
After collecting the preliminary results of Sections 1.12 and 2.2, and using the notations
similar to those of Section 1.12, we are in a position to formulate the following result:
Theorem 4.6.1 Given instant τ ∈ [t
0
, t
1
], the following exact formula is true
X[τ] = X(τ, t
0
, X
0
) = (4.6.4)
= Π
x
(∩¦c(z(τ, L), Z(τ, L, π, χ)¦[L(), π(), χ()¦
and the continuous functions
L(t) ∈ L(IR
m
, IR
m
); π(t) > 0, χ(t) > 0.
Here
254
z(t, L) = ¦x(t), s(t)¦ is a solution to the system
˙ x = A(t)x + p(t), (4.6.5)
L(t) ˙ s = −G(t)x + y(t), (4.6.6)
x(t
0
) = x
0
, s(t
0
) = 0,
and Z
i
(t), i = 1, 2, 3, of
Z(t, L, π, χ) =
_
Z
1
(t) Z
2
(t)
Z
2
(t) Z
3
(t)
_
being the solutions to the matrix diﬀerential equations
˙
Z
1
= A(t)Z
1
+ Z
1
A
(t) + χ
−1
(t)Z
1
+ +χ(t)(1 + π
−1
(t))P(t),
L(t)
˙
Z
2
= −G(t)Z
1
+ L(t)Z
2
A
(t) + L(t)χ
−1
(t)Z
z
,
L(t)
˙
Z
3
= −G(t)Z
2
−L(t)Z
2
G
L
−1
(t)+
+χ
−1
(t)L(t)Z
3
+χ(t)(1 + π(t))K(t)L
−1
(t),
Z
1
(t
0
) = X
0
, Z
2
(t
0
) = 0, Z
3
(t
0
) = I,
where the identity matrix I ∈ L(IR
m
, IR
m
).
Proof. Introducing the perturbed system
˙ x ∈ A(t)x +c(p(t), P(t)), (4.6.7)
L(t) ˙ s ∈ −G(t)x +c(y(t), K(t)) (4.6.8)
¦x
0
, s
0
¦ ∈ c(¦x
0
, 0¦,
˜
X
0
),
˜
X
0
=
_
X
0
0
0 I
_
.
Applying consequently theorem 1.12.5 and Corollary 3.2.2 to systems (4.6.5), (4.6.6.) and
(4.6.1)(4.6.3), we come to the equality (4.6.4). Q.E.D.
255
The proposed scheme does not require the measurement y(t) to be continuous.
We shall further illustrate the procedures of this Section through a numerical example.
Example 4.6.1
28
To illustrate the Singular Perturbation Technique, we chose a system of two dimensions,
and a scalar measurement equation, taking righthand side constant:
A(t) ≡
_
0 1
−8 0
_
, (4.6.9)
The unknown inputs v(t) are bounded by time independent constraints v(t) ∈ c(p(t), P(t))
with
p(t) ≡
_
0
0
_
, and P(t) ≡
_
1 0
0 0.01
_
, (4.6.10)
and the initial state x
(
t
0
) by the ellipsoid c(x
0
, X
0
), there
x
0
=
_
1
0
_
and X
0
=
_
1 0
0 10
_
. (4.6.11)
Further we take the measurement equation to be 1dimensional:
G(t) ≡ ( 0 1 ) , y(t) ≡ 1, K(t) ≡ ( 1 ) .
Additionally we suppose the initial condition:
s(0) ∈ o
0
, o
0
= [−10
−5
, 10
−5
].
Therefore, we have
Z
0
= c(x
0
, X
0
) o
0
∈ IR
2
IR,
The time interval was divided into 100 subintervals of equal lengths and the calculations
were based on a discretized version of system (4.6.1)(4.6.3) with data (4.6.9)(4.6.11).
We further calculate the ellipsoidal estimate
A[τ] = Π
x
(c(z(τ, L
+
), Z(τ, L
+
, π, χ) ∩ c(z(τ, L
−
, π, χ),
for the following two choices for the function L:
28
The calculation of this Example belongs to K.Sugimoto.
256
L
+
(t) =
_
1 if t ∈ [0, 3.5]
0.3 if t ∈ (3.5, 5] ,
L
−
(t) =
_
1 if t ∈ [0, 3.5]
−0.3 if t ∈ (3.5, 5] .
with the range of coordinate axes being −30 to 30.
Parameters π, χ are chosen as
π(t) =
Tr
1/2
(P(t))
Tr
1/2
(K(t))
, χ(t) =
Tr
1/2
(Z(t))
Tr
1/2
(K(t))
.
It is useful to note that in general
Π
x
(c
1
∩ c
2
) ⊆ Π
x
(c
1
) ∩ Π
x
(c
2
). (4.6.12)
An illustration of that is given in Fig. 4.6.1, where the thin lines denote the projections
of two 3dimensional ellipsoids on the plane spanned by the ﬁrst two coordinates ( upper
left window), the ﬁrst and third coordinate (upper right ) and second and third ( lower
right). The thicker line denotes the projection of their intersection on the same planes.
Here (4.6.12) is a proper inclusion.
Returning to our numerical example, we illustrate it in Fig. 4.6.2 , where the upper left
window shows the projections onto the plane spanned by the two state variables. Here
they coincide as expected. In the upper right we see the projection of the two estimating
tubes ( corresponding to L
+
, L
−
) onto the plane of the measurement variable and the ﬁrst
state variable, while in the lower window  onto the plane of the measurement variable and
the second state variable. In Figure 4.6.3 we see the estimates (in the same arrangement
of the windows and in the same scale) at instant t = 4.25, drawn by thin lines, and the
projection of their intersection, drawn by a thicker line. It is to be noted here, that in the
space of the ﬁrst two variables, the projections of the two estimates coincide again, but
the projection of their intersection is a proper subset. We leave to the reader to try these
techniques with various types of discontinuous realizations y(t).
———!!! Insert ﬁgures 4.6.1  4.6.3————–
257
References
[1] AGRACHEV A.A., GAMKRELIDZE R.V., Quadratic Mappings and Smooth Vector
Functions: Euler Characteristics of Level Sets, in ser.: Modern Problems of Mathe
matics. Newest Achievements, v.35, R.V.Gamkrelidze ed., VINITI, Moscow, 1989,
pp.179239, ( in Russian).
[2] ALEFELD G., HERZBERGER J., Introduction to Interval Computation, Academic
Press, N.Y., 1983.
[3] AKIAN M., QUADRAT JP, VIOT M., Bellman Processes, 11th Intern. Conf. on
Systems Analysis and Optimization, SophiaAntipolis, France, 1994.
[4] ALEKSANDROV A.D. On the Theory of Mixed Volumes of Convex Bodies, pp.IIV,
Mat.Sbornik, v2., NN 5,6, 1937, v3., NN 1,2, 1938, (in Russian).
[5] ANANIEV B.I., Minimax MeanSquare Estimates in Statistically Uncertain Systems,
Diﬀerencialniye Uravneniya, N8, 1984, pp.12911297, ( in Russian ) (Translated as
”Diﬀerential Equations”).
[6] ANANIEV B.I., A Guaranteed Filtering Scheme for Hereditary Systems with no In
formation on the Initial State, Proc.ECC95, Rome, v2. pp.966971.
[7] ANANIEV B.I., PISCHULINA I.Ya., Minimax Quadratic Filtering in Systems with
TimeLag, in: Diﬀerential Control Systems, A.V.Kryazhimski ed., Ural Sci.Cent.,
1979, pp.312.
[8] ANTOSIEWICZ H.A., Linear Control Systems, Archive for Rat.Mechanics and Anal
ysis, v.12, N 4, 1963.
[9] ARNOLD V.I. The Theory of Catastrophes , 3d ed., Nauka , Moscow, 1990, ( in
Russian).
[10] ARTSTEIN Z., A Calculus for SetValued Maps and SetValued Evolution Equations,
SetValued Analysis, v.3, 1995, pp. 213261.
[11] ATTOUCH H., AUBIN JP., CLARKE F., EKELAND I.,Eds. Analyse Nonlin´eaire,
GauthiersVillars, C.R.M. Universit´e de Montr´eal, 1989.
[12] AUBIN JP., Motivated Mathematics, SIAM News, v.18, NN 1,2,3, 1985.
[13] AUBIN JP., Diﬀerential Games : a Viability Approach, SIAM J. on Control and
Optimization, v. 28,1990, pp. 12941320.
[14] AUBIN JP. , Fuzzy Diﬀerential Inclusions, Problems of Control and Information
Theory v. 19, 1990, pp. 5567.
258
[15] AUBIN J.P. , Viability Theory, Birkh¨auser, Boston, 1991.
[16] AUBIN J.P., Initiation `a l’Analyse Appliqu´ee. Masson, Paris, 1994.
[17] AUBIN J.P., Mutational and Morphological Analysis. Tools for Shape Regulation
and Optimization. Preprint CEREMADE, Univ. ParisDauphine, Sept.1995.
[18] AUBIN, J.P., CELLINA A., Diﬀerential Inclusions, Springer Verlag, Berlin, 1984.
[19] AUBIN J.P., DA PRATO G., Stochastic Viability and Invariance, Annali Scuola Nor
male di Pisa,27, 1990, pp.595694.
[20] AUBIN J.P., EKELAND I., Applied Nonlinear Analysis, WileyInterscience, 1984.
[21] AUBIN, J.P., FRANKOWSKA H., SetValued Analysis, Birkh¨auser, Boston, 1990.
[22] AUBIN, J.P., FRANKOWSKA H., Inclusions aux D´eriv´ees Partielles Gouvernant
des Contrˆ oles de R´etroaction. ComptesRendus de l’Acad´emie des Sciences, Paris,
311, 1990, pp. 851856.
[23] AUBIN, J.P., FRANKOWSKA H., Viability Kernels of Control Systems, in Nonlinear
Analysis, Ch.Byrnes, A.Kurzhanski, Eds., Ser. PSCT9, Birkh¨auser, Boston, 1990.
[24] AUBIN JP., SIGMUND K., Permanence and Viability, Journal of Comput. and Appl.
Mathematics v.22, 1988, pp. 203209.
[25] AUMANN R. J., Integrals of Setvalued Functions, Journal of Math. Anal. and Appl.,
v.12, 1965.
[26] BAIER R., LEMPIO F., Computing Aumann’s Integral, in: Modeling Techniques for
Uncertain Systems, Kurzhanski, A.B., Veliov V.M., eds.,ser. PSCT18, 1994, pp. 7192.
[27] BALAKRISHNAN A.V. Applied Functional Analysis Springer Verlag, 1976,
[28] BANKS H.T. , JACOBS M.Q., A Diﬀerential Calculus for Multifunctions. Journal of
Math.Anal. and Appl., v.29, 1970, pp. 246272.
[29] BARAS J.S., BENSOUSSAN A., JAMES M.R., Dynamic Observers as Asymptotic
Limits of Recursive Filters: Special Cases, SIAM J. on Appl.Math. v.48,N 5,1988, pp.
11471158.
[30] BARAS J., JAMES M., Nonlinear H
∞
Control, Proc. 33d IEEE, CDC, Lake Buena
Vista, Fl., USA, v.2, 1994, pp. 14351438.
[31] BARAS J.S., JAMES M.R., ELLIOT R.J., OutputFeedback, RiskSensitive Control
and Diﬀerential Games for ContinuousTime Nonlinear Systems ,Proc. 32nd IEEE,
CDC , San Antonio, Texas, USA, v.4.,1994, pp. 33573360.
259
[32] BARAS J.S., KURZHANSKI A.B., Nonlinear Filtering: the SetMembership (Bound
ing) and the H
∞
Approaches. Proc. of the IFAC NOLCOS Conference, Tahoe, CA
USA, Plenum Press, 1995.
[33] BARBASHIN E.A., On the Theory of Generalized Dynamic Systems. Scientiﬁc Notes
of the Moscow Univ., Mathematics, v. 2, N 135, 1949 ( in Russian).
[34] BARBASHIN E.A., Liapunov Functions, Nauka, Moscow, 1970 (in Russian).
[35] BARBASHIN E.A., ALIMOV Yu.I., On the Theory of Dynamical Systems with Multi
Valued and Discontinuous Characteristics, Dokl. AN SSSR, 140, 1961, pp. 911, (in
Russian).
[36] BARBU V., DA PRATO G., HamiltonJacobi Equations in Hilbert Spaces : Vari
ational and Semigroup Approach, Annali di Matematica Pura e Applicata, CXLII,
1985, pp. 303349.
[37] BASAR T., BERNHARD P., H
∞
Optimal Control and Related Minimax Design Prob
lems ser. SCFA, Birkh¨auser, Boston, 1991.
[38] BASAR T., KUMAR P.R., On WorstCase Design Strategies, Comput. and Math.
with Appl., v.13(13), 1978, pp. 239245.
[39] BASAR T., DIDINSKY G., PAN Z. A New Class of Identiﬁers for Robust Parameter
Identiﬁcation and Control in Uncertain Systems. Proc. of Workshop on ” Robust Con
trol via Variable Structure and Liapunov Techniques, Benvenetto, Italy, Sept. 1994.
[40] BARMISH B.R., LEITMANN G., On Ultimate Boundedness Control of Uncertain
Systems in Absence of Matching Conditions, IEEE Trans.Aut.Control, AC27, 1982,
p.1253.
[41] BARTOLINI G., ZOLEZZI T., Asymptotic Linearization of Uncertain Systems by
Variable Structure Control, Syst.Cont. Letters,v.10, 1988, pp.111117.
[42] BASAR T., MINTZ M., Minimax Terminal State Estimation for Linear Plants with
Unknown Forcing Functions, Intern. Journ. of Control v.16(1), 1972 , pp. 4970.
[43] BASILE G., MARRO G. Controlled and Conditioned Invariants in Linear Systems
Theory, Prentice Hall, Englewood Cliﬀs, 1991.
[44] BEHREND F., Uber die kleinste umbeschriebene und die gr¨oste einbeschriebene El
lipse eines konvexen Bereichs, Math.Annalen, v.115, 1938, pp.379411.
[45] BELL D.J., Singular Problems in Optimal Control  a Survey, Int. J. Control, 21
(1975), pp.319331.
[46] BELLMAN R. Dynamic Programming Princeton Univ. Press., Princeton NJ, 1957.
260
[47] BENSOUSSAN A., DA PRATO G., DELFOUR M.C., MITTER S.K., Representation
and Control of InﬁniteDimensional Systems, Ser. SCFA, Birkh¨auser, Boston,v.I, 1992,
v.2., 1993.
[48] BERGER M., G´eom´etrie3, CEDIC, Fernand Nathan, Paris, 1978.
[49] BERKOVITZ L.D. Characterization of the Values of Diﬀerential Games, Appl.Math.
and Optim. v.17, 1988, pp.177183.
[50] BERKOVITZ L.D. Diﬀerential Games of Survival, J. of Math.Anal.Appl. v.129,2,
1988, pp.493504.
[51] BERNHARD P. A DiscreteTime Certainty Equivalence Principle. Systems and Con
trol Letters , 1994.
[52] BERNHARD P. Expected Values, Feared Values and Partial Information Control,
Preprint, INRIA, SophiaAntipolis, Jan.1995.
[53] BERTSEKAS D.P., Dynamic Programming Deterministic and Stochastic Models,
Prentice Hall, Englewood Cliﬀs, NJ, 1987.
[54] BERTSEKAS D.P., RHODES I.B., On the Minimax Reachability of Target Sets and
Target Tubes, Automatica N 7, 1971, pp. 233247.
[55] BERTSEKAS D.P., RHODES I.B., Recursive State Estimation for a SetMembership
Description of Uncertainty. IEEE Trans. Aut. Control, AC16, 1971, pp.117128.
[56] BERTSEKAS D.P., RHODES I.B., Suﬃciently Informative Functions and the Min
max Feedback Control of Uncertain Dynamic Systems. IEEE Trans. Aut. Control,
AC18(2), 1973, pp.117123.
[57] BITTANTI S.,ed., The Riccati Equation in Control, Systems and Signals, Pitagora
Editrice, 1989.
[58] BIRGE J.R., WETS R.J.B., Sublinear Upper Bounds for Stochastic Programs with
Recourse, Math.Prog., v.43, 1989, pp. 131149.
[59] BLAGODATSKIH V.I., FILIPPOV A.F., Diﬀerential Inclusions and Optimal Control.
Trudy Mat. Inst. AN SSSR, 169, 1985, pp. 194252, (Translated as Proc. of the Steklov
Math.Inst.,4, North Holland, Amsterdam, 1986)
[60] BONNENSEN T., FENCHEL W., Theorie der Konvexen K¨orper, Springer, Berlin,
1934.
[61] BROCKETT R.W., FiniteDimensional Linear Systems, J.Wiley, NY, 1970.
261
[62] BROCKETT R.W., Pulse Driven Dynamical Systems. Systems, Models and Feedback:
Theory and Applications, A.Isidori, T.J.Tarn Eds., 1992. ser. PCST 12, Birkh¨auser,
Boston.
[63] BRYSON A.E., HO Y.C., Applied Optimal Control, Ginn, Waltham, MA, 1969.
[64] BURAGO Yu.D., ZALGALLER V.A., Geometrical Inequalities, Nauka, Leningrad,
1980, (in Russian).
[65] BUSEMANN H., A Theorem on Convex Bodies of the BrunnMinkowski Type,
Proc.Nat.Acad.Sci. USA, , v.35, 1949, pp.2731.
[66] BYRNES Ch.I., ISIDORI A., Feedback Design from the Nonzero Dynamics Point of
View Computation an Control ser. PSCT1, Birkh¨auser, Boston.
[67] BYRNES Ch.I., KURZHANSKI A.B., eds., Nonlinear Synthesis, ser. PCST9,
Birkh¨auser, Boston, 1989.
[68] CASTAING C., VALADIER M., Convex Analysis and Measurable Multifunctions,
Lecture Notes in Mathematics, Vol.580, Springer Verlag, 1977.
[69] CEA J., Optimisation: Th´eorie et Algorithmes, Dunod, Paris, 1971.
[70] CELLINA A., ORNELAS A. Convexity and the Closure of the Solution Set to Diﬀer
ential Inclusions. Preprint SISSA, Trieste, 1988.
[71] CELLINA A., ORNELAS A. Representation of the Solution Set to Lipschitzian Dif
ferential Inclusions. Preprint SISSA, Trieste, 1988.
[72] CHERNOUSKO F. L., Optimal Guaranteed Estimates of Indeterminacies with the
Aid of Ellipsoids, I., II., III., Izv. Acad. Nauk SSSR, Tekhn. Kibernetika 3,4,5 1980,
(In Russian). Translated as Engineering Cybernetics.
[73] CHERNOUSKO, F. L. State Estimation for Dynamic Systems, CRC Press, 1994.
[74] CHEUNG MF., YURKOVICH S., PASSINO K.M., An Optimal Volume Algorithm
for Parameter Set Estimation, IEEE, Trans.Aut.Cont. v.39, n6, 1994, pp. 12681272.
[75] CHISCI L., GARULLI A., ZAPPA G., Recursive State Bounding by Parallelotopes,
Preprint, Univ.Firenze, 1995.
[76] CHIKRII A.A. ConﬂictControlled Processes, Naukova Dumka, Kiev, 1992, ( in Rus
sian).
[77] CHIKRII A.A., ZHUKOVSKII V.I., LinearQuadratic Diﬀerential Games, Naukova
Dumka, Kiev, 1994, ( in Russian).
262
[78] CLARKE F.H., Optimization and Nonsmooth Analysis, Wiley  Interscience, New
York, 1983.
[79] COLOMBO G., Approximate and Relaxed Solutions of Diﬀerential Inclusions.
Preprint SISSA, Trieste, 1988.
[80] COMBA J.L.D., STOLFI J., Aﬃne Arithmetic and its Applications to Computer
Graphics, Anais do SIBGRAPI VI, 1993, pp. 918.
[81] CORLESS M., LEITMANN G., Adaptive Control for Uncertain Dynamical Systems,
Dyn.Syst. and Microphys., Contr. Theory and Mech., Acad. Press, 1984, pp.91158.
[82] CRANDALL M.G., LIONS P.L. Viscosity Solutions of HamiltonJacobi Equations,
Trans.Amer.Math.Soc. , 277, 1983, pp.142.
[83] CRANDALL M.G., ISHII H., LIONS P.L. User’s Guide to Viscosity Solutions of Sec
ond Order Partial Diﬀerential Equations. Bull.of the Amer.Math.Soc.,v.27,N1, 1992,
pp.167.
[84] DANZER, L., LAUGWITZ, D., LENZ, H.
¨
Uber das L¨ownersche Ellipsoid und sein
Analogon unter den Einem Eik¨orper Einbeschriebenen Ellipsoiden, Arch. Math. 8,
1957, pp. 214219.
[85] DEIMLING K. Multivalued Diﬀerential Equations on Closed Sets. Diﬀerential and
Integral Equations, 1, 1988, pp.2330.
[86] DEMIANOV V.F., Minimax: Directional Diﬀerentiation, Leningrad University Press,
1974.
[87] DEMIANOV V.F., LEMARECHAL C., ZOWE J., Approximation to a SetValued
Mapping, I: a Proposal, Appl. Math. Optim., 14 (1986), pp.203214.
[88] DEMIANOV V.F., RUBINOV A.M., Foundations of Nonsmooth Analysis and Qua
sidiﬀerential Calculus, Nauka, Moscow, 1990, ( in Russian).
[89] DEMIANOV, V. F., VASILIEV L. V. Nondiﬀerentiable Optimization, Nauka,
Moscow, 1981, (In Russian)
[90] DI MASI G., GOMBANI A., KURZHANSKI A.,eds., Modeling, Estimation and Con
trol of Systems with Uncertainty, ser. PCST10, Birkh´auser, Boston, 1990.
[91] DONTCHEV A.L., Perturbations, Approximations and Sensitivity Analysis of Op
timal Control Systems, Lect. Notes in Control and Inform. Sciences, 52, Springer 
Verlag, 1986.
[92] DONTCHEV A.L., LEMPIO F., Diﬀerence Methods for Diﬀerential Inclusions: a
Survey, SIAM Rev.,,v.34, 1992.
263
[93] DONTCHEV A.L., VELIOV V.M., Singular Perturbation in Mayer’s Problems for
Linear Systems, SIAM J. Control Optim., 21 (1983), pp.566581.
[94] DOYLE J.C., FRANCIS B.A., TANNENBAUM A.R., Feedback Control Theory, N.Y.,
McMillan, 1992.
[95] DUBROVIN B.A., NOVIKOV S.P., FOMENKO A.T., Modern Geometry. Methods
and Applications, URSS Pub., Moscow, 1994.
[96] DUNFORD, N., SCHWARTZ, J. T. Linear Operators, Part I. Interscience Publishers
Inc. New York, 1957.
[97] DWYER R.A., EDDY W.F., Maximal Empty Ellipsoids, in : Proc. of the Fifth Annual
ACMSIAM Symposium on Discrete Algorithms, Arlington, VA, USA, New York,
ACM, 1994, pp. 98102.
[98] EGGLESTON H., Convexity, Cambr.Univ.Press, 1963.
[99] EFIMOV N.V. Advanced Geometry, Nauka, Moscow, 1971, ( in Russian).
[100] EKELAND I., TEMAM R., Analyse Convexe at Probl`emes Variationnelles, Dunod,
Paris, 1974.
[101] FAN KY Minmax Theorems, Proc. Nat.Acad. of Sci. USA, v.39, N1, 1953, pp. 4247.
[102] FILIPPOV A.F. On Some Problems of Optimal Control Theory .Vestnik Mosk. Univ.,
Math.,N2, 1958, pp.2532 ( in Russian), English version in SIAM J. Control, N1, 1962,
pp.7684.
[103] FILIPPOV A.F. Classical solutions of Diﬀerential Equations with a Multivalued
RightHand Side, SIAM J. on Control, N5, 1967, pp.609621.
[104] FILIPPOV A.F. Diﬀerential Equations with Discontinuous Righthand Side, Kluwer
Acad. Pub., 1988.
[105] FILIPPOVA T. F., A Note on the Evolution Property of the Assembly of Viable
Solutions to a Diﬀerential Inclusion, Computers Math. Applic., 25 (1993), pp.115121.
[106] FILIPPOVA T.F., On the Modiﬁed Maximum Principle in the Estimation Problems
for Uncertain Systems, IIASA Working Paper, Laxenburg 1992.
[107] FILIPPOVA T.F., KURZHANSKI A. B., SUGIMOTO K., VALYI I., Ellipsoidal
Calculus, Singular Perturbations and the State Estimation Problem for Uncertain
Systems, in Bounding Approaches to System Identiﬁcation , M.Milanese, J.Norton,
H. PietLahanier, E.Walter,eds., Plenum Press, 1995, also in IIASA Working Paper,
Laxenburg,WP9251, 1992.
264
[108] FLEMING W.H., RISHEL R.W., Deterministic and Stochastic Optimal Control,
SpringerVerlag, 1975.
[109] FLEMING W.H., SONER H.M., Controlled Markov Processes and Viscosity Solu
tions, SpringerVerlag, 1993.
[110] FLEMING W.H., SOUGANIDIS P.E., On the Existence of Value Function for a
TwoPlayer ZeroSum Diﬀerential Game, Indiana Univ. Math. J., v.38, 1989, pp. 293
314.
[111] FOGEL E. System Identiﬁcation via Membership Set Constraints with Energy Con
strained Noise, IEEE, Trans.Aut.Contr. v.24, N5, 1979, pp. 752757.
[112] FORMALSKI A.M., On the Corner Points of the Boundaries of Attainability Re
gions, Prikl.Mat.Meh., v.47, 1983, ( Translated as Appl.Math. and Mech.)
[113] FRANK M., WOLFE P., An Algorithm for Quadratic Programming, Naval Res. Log.
Quart. , v3., 1956, pp.95110.
[114] FRANKOWSKA H. L’
´
Equation d’HamiltonJacobi Contingente, ComptesRendus
de l’Acad´emie des Sciences, Paris, 304, ser.1, 1987, pp. 295298.
[115] FRANKOWSKA H. Contingent Cones to Reachable Sets of Control Systems, Siam
J. on Control Optim., v.27, 1989, pp.170198.
[116] FRANKOWSKA H., LowerSemicontinuous Solutions of HamiltonJacobiBellman
Equations, SIAM J. Control Optim. v.31, N1, 1993, pp. 257272.
[117] FRANKOWSKA H., PLASCACZ M., RZEZUCHOWSKI T., Measurable Viabil
ity Theorems and HamiltonJacobiBellman Equations, J.Diﬀ.Equat., v.116,N2, 1995,
pp.265305.
[118] FRANKOWSKA H., QUINCAMPOIX M., Viability Kernels of Diﬀerential Inclu
sions with Constraints, Math. Syst., Est. and Control, v.1, N3, 1991, pp.371388.
[119] FRIEDMAN, A. Diﬀerential Games, Interscience, N.Y., 1971.
[120] GANTMACHER, F. R. Matrix Theory, III. Chelsea Publishing Co. New York, 1960.
[121] GRUBER P., Approximation of Convex Bodies, in Convexity and Applications,
P.Gruber, J.Wills eds., Birkhauser, 1993.
[122] GIARR
`
E L., MILANESE M., TARAGNA M., H
∞
 Identiﬁcation and Model Quality
Evaluation, Preprint, Politecnico Torino, 1995.
[123] GRAVES C.H., VERES S.M., Using MATLAB Toolbox ”GBT” in Identiﬁcation
and Control. in: IEE Colloquium on ’Identiﬁcation of Uncertain Systems’, Digest
No.1994/105.
265
[124] GR
¨
UNBAUM B., Convex Polytopes, Itersc.,London, 1967.
[125] GUSEINOV KH. G., USHAKOV V. N., Strongly and Weakly Invariant Sets with
Respect to a Diﬀerential Inclusion. Soviet Mathematical Doklady, Vol. 38, 3. 1989, (
in Russian).
[126] GUSEINOV KH.,G, SUBBOTIN A.I., USHAKOV V. N., Derivatives for Multivalued
Mappings with Applications to Game theoretical Problems of Control, Problems of
Control and Inform. Theory, v.14, 3 , 1985, pp.155167.
[127] HADWIGER H., Vorlesungen
¨
Uber Inhalt, Oberﬂ¨ache und Isoperimetrie, Springer
Verlag, Berlin, 1957.
[128] HALE J.K., LIN X.B., Symbolic Dynamics and Nonlinear Flows, Annali Mat. Pura
e Applicata, v.144, 1986, pp.229260.
[129] HIJAB O.B. MinimumEnergy Estimation, Ph.D. Thesis, UCBerkeley, 1980.
[130] HINRICHSEN D., MARTINSEN B.,eds., Control of Uncertain Systems, ser. PSCT6,
Birkh´auser, Boston, 1990.
[131] HOFBAUER J., SIGMUND K. The Theory of Evolution and Dynamical Systems
Cambridge Un. Press, 1988.
[132] HONIN V. A. Guaranteed Estimation of the State of Linear Systems by Means of
Ellipsoids, Evolutionary Systems in Estimation Problems, A. B. Kurzhanski, T. F.
Filippova Eds., Sverdlovsk, 1980, (In Russian).
[133] IOFFE A.D., TIKHOMIROV V.M., The Theory of Extremal Problems, Nauka,
Moscow, 1979.
[134] ISAACS R., Diﬀerential Games J.Wiley, NY, 1965.
[135] ISIDORI A. Nonlinear Control Systems, SpringerVerlag, 1989.
[136] ISIDORI A., Attenuation of Disturbances in Nonlinear Control Systems, in: Systems,
Models and Feedback: Theory and Applications, Isidori A., Tarn T.J.eds. ser. PCST
12, Birkh¨aser, Boston, 1992.
[137] JOHN F. Extremum Problems with Inequalities as Subsidiary Conditions, in: K.
Friedrichs, O. Neugebauer, J. J. Stoker Eds. Studies and Essays, Courant Anniversary
Volume John Wiley and Sons. Inc., New York, 1948.
[138] JOULIN L., WALTER E., Guaranteed Estimation of Stability Domain via Set
Inversion, IEEE Trans.Aut.Cont., v.39, N4, 1994, pp. 886889.
[139] JUNG H.E.W,
¨
Uber die Kleinste Kugel, die Eine R´aumliche Figur Einschliesst,
J.Reine Anew.Math. v.123, 1901, pp. 241257.
266
[140] KAC I.Ya., KURZHANSKI A.B., Minimax Estimation in Multistage Systems,
Sov.Math.Dokl., v.16, N2, 1975, pp.374379.
[141] KAHAN W., Circumscribing an Ellipsoid About the Intersection of Two Ellipsoids,
Can.Math.Bull., v.11, 1968.
[142] KAILATH T., Linear Systems Prentice Hall, Englewood Cliﬀs, NJ, 1980.
[143] KAILATH T. The Innovations Approach to Detection and Estimation Theory, . Proc.
of the IEEE, v.58, 1970, pp.680695.
[144] KALL P. Stochastic Programming with Recourse: Upper Bounds and Mo
ment Problems  a Review, in: Advances in Math.Optimization, ( Dedicated
to Prof. F.Noˇziˇcka), J.Guddat, B.Bank, H.Hollatz, P.Kall, D.Klatte, B.Kummer,
K.Lommatzch, K.Tammer, M.Vlach eds. AkademieVerlag, Berlin, 1988.
[145] KALL P., STOYAN D., Solving Stochastic Programming Problems with Recourse
Including Error Bounds, Math. Operationsforsch.Statist.,Se.Opt, v.13, 1982, pp.431
447.
[146] KALL P., WALLACE S.W., Stochastic Programming , J.Wiley, Chichester, 1994.
[147] KALMAN R.E., On the General Theory of Control Systems, Proc. 1st. IFAC
Congress, v.1, Butterworths, London, 1960.
[148] KALMAN R.E., Nine Lectures on Identiﬁcation Lect. Notes on Economics and Math.
Systems, SpringerVerlag, 1993.
[149] KALMAN R.E., BUCY R., New results in Linear Filtering and Prediction Theory,
Trans. AMSE, 83, D, 1961.
[150] KANTOROVICH L.V., AKILOV G.P., Functional Analysis, Pergamon Press, 1982.
[151] KARL W.C., VERGHESE G.C, WILLSKY A.S., Reconstructing Ellipsoids from
Projections, Graph.Mod.and Image Proc, v.56, N2, 1994, 124139.l
[152] KHACHIYAN
L.D., A Polynomial Algorithm in Linear Programming, Sov.Math.Dokl, v.20, 1979,
pp. 191194.
[153] KHACHIYAN L.D., An Inequality for the Volume of Inscribed Ellipsoids, Discr.and
Comput.Geom., v.5, N3, 1990, pp.219222.
[154] KHACHIYAN L.D., TODD M., On the Complexity of Approximating the Maximal
Inscribed Ellipsoid for a Polytope, Cornell Univ.Tech.Rep. N 893,Ithaca, NY,1990.
267
[155] KISELEV O.N., POLYAK B.T., Ellipsoidal Estimation with Respect to a General
ized Criterion, Avtomatika i Telemekhanika, v.52, N 9, 1991, pp.13345, Translation:
Automation and Remote Control, Sept. 1991, v.52, N 9, pp.128192.
[156] KLEE V.L., Extremal Structure of Convex Sets,p.I, Arch.Math., v.8, 1957, pp. 234
240, p.II, Math.Zeit., v.69, 1958, pp. 90104.
[157] KLIR G.J., FOLGER T.A., Fuzzy Sets, Uncertainty and Information, Prentice Hall,
Englewood Cliﬀs, NJ , 1988.
[158] KNOBLOCH H., ISIDORI,A., FLOCKERZI D., Topics in Control Theory,
Birkh¨auser, DMVseminar, Band 22, 1993.
[159] KOENIG H., PALLASCHKE D., On Khachian’s Algorithm and Minimal Ellipsoids,
Num.Math., v.36, 1981, pp.211223.
[160] KOKOTOVIC P., BENSOUSSAN A., BLANKENSHIP G., eds., Singular Perturba
tions and Asymptotic Analysis in Control Systems, Lect. Notes in Contr. and Inform.
Sci., 90, SpringerVerlag, 1986.
[161] KOLMOGOROV A. N., FOMIN S. V., Elements of Theory of Functions and Func
tional Analysis, Nauka, Moscow, 1968.
[162] KOMAROV V.A. The Equation for Attainability Sets of Diﬀerential Inclusions in
the Problem with State Constraints, Trudy Matem.Instit. Akad. Nauk SSSR, v.185,
1988, pp.116125.
[163] KOMAROV V.A., The Estimates of Attainability Sets of Diﬀerential Inclusions, Mat.
Zametki, 37, 1985, pp.916925.
[164] KONSTANTINOV G.N., SIDORENKO G.V., Outer Estimates of Reachable Sets of
Controlled Systems, Izv. Akad. Nauk SSSR, Teh. Kibernet, N3, 1986, ( Translated as
Engineering Cybernetics).
[165] KOSCHEEV A. S., KURZHANSKI A. B., On Adaptive Estimation of Multistage
Systems Under Conditions of Uncertainty, Izvestia AN SSSR, Teh. Kibernetika, N4,
1983, (in Russian), pp.7293, Translated as ” Engineering Cybernetics”, N2, 1984,
pp.5777.
[166] KRASOVSKII N. N., On the Theory of Controllability and Observability of Linear
Dynamic Systems, Priklad.Mat.i Meh., v.23,N 4, 1964, ( in Russian), Translated as
J.of Appl.Math. and Mech.
[167] KRASOVSKI N.N., The Theory of Control of Motion, Nauka, Moscow, 1968, (in
Russian).
268
[168] KRASOVSKII N. N., GameTheoretic Problems on the Encounter of Motions Nauka,
Moscow, 1970, (in Russian), English Translation : Rendezvous Game Problems Nat.
Tech.Inf.Serv.,Springﬁeld, VA, 1971.
[169] KRASOVSKII N.N. The Control of a Dynamic System, Nauka, Moscow, 1986, (in
Russian).
[170] KRASOVSKI N.N., KRASOVSKI A.N. Feedback Control Under Lack of Information,
Birkh¨auser, Boston, 1995.
[171] KRASOVSKI N.N., SUBBOTIN A.N., Positional Diﬀerential Games, Springer
Verlag, 1988.
[172] KRASTANOV M., KIROV N., Dynamic Interactive System for Analysis of Linear
Diﬀerential Inclusions, in: Modeling Techniques for Uncertain Systems, Kurzhanski,
A.B., Veliov V.M., eds.,ser. PSCT18, 1994, pp. 123130.
[173] KRENER A., Necessary and Suﬃcient Conditions for WorstCase H
∞
Control and
Estimation. Math.Syst.Estim. and Control, N4, 1994.
[174] KUMAR P.R., VARAIYA P., Stochastic Systems: Estimation, Identiﬁcation and
Adaptive Control, Prentice Hall, Englewood Cliﬀs, NJ, 1986.
[175] KUNTSEVICH V.M., LYCHAK M., Guaranteed Estimates, Adaptation and Robust
ness in Control Systems, Lect. Notes in Contr.Inf.Sci., v.169, SpringerVerlag, 1992.
[176] KURATOWSKI R., Topology vv.1,2, Academic Pres, NY, 1966.
[177] KURZHANSKI A.B. On the Duality of Problems of Optimal Control and Observa
tion, Prikl.Mat.,Meh.,v.34, N3, 1970, (in Russian), Translated as J.of Appl.Math. and
Mech.
[178] KURZHANSKI A. B., Diﬀerential Games of Approach Under Constrained Coor
dinates, Dok.Akad.Nauk SSSR ,v.192, N3,1970.( in Russian), Translated as Soviet
Math.Doklady, v.11, N3, 1970, pp.658672.
[179] KURZHANSKI A. B., Diﬀerential Games of Observation, Sov. Math. Doklady, v.13,
N6, 1972, pp.15561560.
[180] KURZHANSKI A. B., On Minmax Control and Estimation Strategies Under Incom
plete Information, Problems of Contr. Inform. Theory, 4, 1975, pp. 205218.
[181] KURZHANSKI A. B., Control and Observation under Conditions of Uncertainty,
Nauka, Moscow, 1977.
[182] KURZHANSKI A. B., Dynamic Decision Making Under Uncertainty, in: State of
the Art in Operations Research , N.N.Moiseev ed., Nauka, Moscow, 1979, pp.197235,
( in Russian).
269
[183] KURZHANSKI A. B., On The Estimation of Dynamic Control Systems Under Un
certainty Conditions, Problems of Control and Information Theory, part I, v.9, 1980,
pp.395406; part II, v.10, 1981, pp.3342.
[184] KURZHANSKI A.B., Evolution Equations for Problems of Control and Estimation of
Uncertain Systems, in Proceedings of the Int. Congress of Mathematicians, Warszawa,
1983, pp.13811402.
[185] KURZHANSKI A.B., On the Analytical Description of the Bundle of Viable Trajec
tories of a Diﬀerential System, Soviet Math.Doklady v.33, 1986, pp.475478.
[186] KURZHANSKI A.B., Identiﬁcation  a Theory of Guaranteed Estimates, in From
Data to Model, J. C. Willems, ed., Springer  Verlag, 1989.
[187] KURZHANSKI A.B., The Identiﬁcation Problem  a Theory of Guaranteed Es
timates, Automation and Remote Control, (translated from ”Avtomatika i Tele
mekhanika”), v.52, N4, pt.1, 1991, pp. 447465.
[188] KURZHANSKI A.B., Inverse Problems of Observation and Invertibility for Dis
tributed Systems, Vestnik Mosk. Univers, Vychislit. Mat., N1, 1995, ( in Russian).
[189] KURZHANSKI A. B., FILIPPOVA T. F., On the Description of the Set of Viable
Trajectories of a Diﬀerential Inclusion, Sov. Math.Doklady v.34, 1987.
[190] KURZHANSKI A. B., FILIPPOVA T. F., On the SetValued Calculus in Problems
of Viability and Control for Dynamic Processes: the Evolution Equation, L´es Annales
de l’Institut Henri Poincar´e, Analyse Nonlin´eaire, 1989, pp.339363.
[191] KURZHANSKI A. B., FILIPPOVA T. F., On the Method of Singular Perturbations
for Diﬀerential Inclusions, Sov. Math. Doklady, v.44, 1992, pp.705710.
[192] KURZHANSKI A. B., FILIPPOVA T. F., Diﬀerential Inclusions with State Con
straints. The Singular Perturbation Method, Trudy Matem. Inst. Ross. Akad. Nauk,
1995, (in Russian).
[193] KURZHANSKI A. B., FILIPPOVA T. F., On the Theory of Trajectory Tubes 
a Mathematical Formalism for Uncertain Dynamics, Viability and Control, in Ad
vances in Nonlinear Dynamics and Control: a Report from Russia, A.B.Kurzhanski
ed., ser.PSCT 17, Birkh¨auser, Boston, 1993, pp.122188.
[194] KURZHANSKI A.B., NIKONOV O.I. On the Problem of Synthesizing Control
Strategies. Evolution Equations and SetValued Integration, Doklady Akad. Nauk
SSSR, 311, 1990, pp.788793, Sov. Math. Doklady, v.41, 1990.
[195] KURZHANSKI A., NIKONOV O.I Evolution Equations for Tubes of Trajectories
of Synthesized Control Systems. Russ.Acad.of Sci. Math. Doklady , v.48, N3, 1994,
pp.606611.
270
[196] KURZHANSKI A., OSIPOV Yu.S., On Optimal Control Under Restricted Coor
dinates, Prikl.Mat.Meh., v.33, N4, 1969, (Translated as Applied. Math. and Mech)
.
[197] KURZHANSKI A., OSIPOV Yu.S., On One Problem of Control Under Bounded
Coordinates, Izv.Akad.Nauk SSSR. Teh. Kiber., N5, 1970, (Transl. as Engineering
Cybernetics).
[198] KURZHANSKI A.B., TANAKA M., On a Uniﬁed Framework for Deterministic and
Stochastic Treatment of Identiﬁcation Problems, IIASA Working Paper, WP89013,
Laxenburg, 1989.
[199] KURZHANSKI A.B., PSCHENICHNYI B. N., POKOTILO V. G., Optimal Inputs
for Guaranteed Identiﬁcation, Problems of Control and Information Theory v.20, N1,
1991, pp.1323.
[200] KURZHANSKI A. B., SIVERGINA I. F., On Noninvertible Evolutionary Systems:
Guaranteed Estimates and the Regularization Problem, Sov. Math. Doklady, v.42, N2,
1991, pp.451455.
[201] KURZHANSKI A. B., SUGIMOTO K., VALYI I., Guaranteed State Estimation for
Dynamic Systems: Ellipsoidal Techniques, Intern.Journ. of Adaptive Contr. and Sign.
Proc., v.8., 1994, pp. 85101.
[202] KURZHANSKI A. B., V
´
ALYI I. Set Valued Solutions to Control Problems and
Their Approximation, in: A. Bensoussan, J. L. Lions Eds. Analysis and Optimization
of Systems, Lecture Notes in Control and Information Systems, Vol. 111, Springer
Verlag, 1988. pp. 775785.
[203] KURZHANSKI A. B., V
´
ALYI I. Ellipsoidal Techniques for Dynamic Systems: the
Problems of Control Synthesis, Dynamics and Control, 1, 1991, pp. 357378.
[204] KURZHANSKI A. B., V
´
ALYI I. Ellipsoidal Techniques for Dynamic Systems: Con
trol Synthesis for Uncertain Systems, Dynamics and Control, 2, 1992, pp. 87111.
[205] KURZHANSKI A.B., VELIOV V.M., eds., SetValued Analysis and Diﬀerential In
clusions, Ser.PCST16, Birkh´auser, Boston, 1993.
[206] KURZHANSKI A.B., VELIOV V.M., eds., Modeling Techniques for Uncertain Sys
tems, ser. PSCT18, Birkh´auser, Boston, 1994.
[207] KWAKERNAAK H., SIVAN.R., Linear Optimal Control Systems, Wiley
Intersc.,NY, 1972.
[208] LAKSHMIKANTAM V., LEELA S., Nonlinear Diﬀerential Equations in Abstract
Spaces, Ser.in Nonlin. Math.: Theory, Methods, Appl., 2, Pergamon Press, 1981.
271
[209] LAURENT P. J., Approximation and Optimization, Hermann, Paris, 1972.
[210] LEDIAYEV Yu. S., Criteria for Viability of Trajectories of Nonautonomous Diﬀer
ential Inclusions and Their Applications, Preprint CRM , 1573, Univ. de Montr´eal,
1992.
[211] LEDIAYEV Yu. S., MISCHENKO E.F., Extremal Problems in the Theory of Dif
ferential Games, Trud. Mat. Inst. AN SSSR , v.185, 1988, pp. 147170, 1988, ( in
Russian).
[212] LEE E.B., MARCUS Foundations of Optimal Control Theory, Wiley, N.Y.,1961.
[213] LEICHTWEISS K., Konvexe Mengen, VEB Deutsch.Verl. der Wissensch., Berlin,
1980.
[214] LEITMANN G., Deterministic Control of Uncertain Systems via a Constructive Use
of Lyapunov Stability Theory, Proc. 14th IFIP Conference on Systems Modeling and
Optimization, Leipzig, July, 1989.
[215] LEITMANN G., One Approach to the Control of Uncertain Dynamical Systems,
Proc. 6th Workshop on Dynamics and Control, Vienna, 1993.
[216] LEMARECHAL C., ZOWE J., Approximation to a MultiValued Mapping: Exis
tence, Uniqueness, Characterization, Math.Institut, Universit¨at Bayreuth, 1987, Re
port N 5.
[217] LIAPUNOV A.M., Probl`eme Generale de Stabilit´e de Mouvement,
Ann.Fac.Toulouse, v.9, 1907, pp.207474.
[218] LIONS JL., Controllabilit´e Exacte, Perturbations et Stabilization des Syst`emes Dis
tribu´es, vv.1,2, Masson, Paris, 1990.
[219] LIONS PL., SOUGANIDIS P.E., Diﬀerential Games, Optimal Control and Direc
tional Derivatives of Viscosity Solutions of Bellman’s and Isaac’s Equations, SIAM
J.Cont.Opt., v.23, 1995, pp.566583.
[220] LOTOV A.V., Generalized Reachable Sets Method in Multiple Criteria Problems, in:
Methodology and Software for Interactive Decision Support, Lecture Notes in Econ.and
Math.Systems, v.337, SpringerVerlag, 1989.
[221] MARCHUK G.I., Methods of Numerical Mathematics, SpringerVerlag, 1975.
[222] MARKOV S.M., ANGELOV R., An Interval Method for Systems of ODE, Inter
val Mathematics 1985, Proc.Int.Symp.Freiburg, Lecture Notes in Comput.Sci., v.212,
Springer Verlag, 1986.
[223] MELKMAN A.A., MICCHELLI C.A., Optimal Estimation of Linear Operators in
Hilbert Spaces from Inaccurate Data, SIAM J. Num.Anal., v.16., N1, 1979, pp. 87105.
272
[224] MILANESE M., Properties of LeastSquares Estimates in SetMembership Identiﬁ
cation, Automatica, v.31, N2, 1995, pp.327332.
[225] MILANESE M., NORTON J., PIETLAHANIER H., WALTER E.,eds., Bounding
Approaches to System Identiﬁcation , Plenum Press, 1995,
[226] MILANESE M., VICINO A., Optimal Estimation for Dynamic Systems with Set
Membership Uncertainty: an Overview, Automatica, v.27, 1991, pp. 9971009.
[227] MILANESE M., VICINO A., InformationBased Complexity and Nonparametric
WorstCase System Identiﬁcation , Journ. of Complexity v.9, 1993, pp.427446.
[228] MOHLER R.R., Nonlinear Systems: v.II Application to Bilinear Control, Prentice
Hall, Englewood Cliﬀs, NJ, 1991.
[229] MOORE R.E., Methods and Applicationsn of Interval Analysis, SIAM, Phil.,1979.
[230] MOREAU JJ. Raﬂe par un Convexe Variable, Analyse Convexe, Montpellier, p.I
Ex.N15, 1971, p.II Ex.N3, 1972.
[231] NAGPAL K.M., KHARGONEKAR P.P., Filtering and Smoothing in an H
∞
Setting,
IEEE, Trans. on Aut. Contr., v.36, N2, 1991, pp.152166.
[232] NATANSON, I. P. A Theory of Functions of the Real Variable, Nauka, Moscow, 1972.
(In Russian)
[233] NEMYTSKI V.V., STEPANOV V.A., Qualitative Theory of Diﬀerential Equations,
Princeton Univ.Press, 1960.
[234] NESTEROV YU.N., NEMIROVSKII A.S., Selfadjustment Functions and Polynomial
Algorithms in Convex Programming, CEMI Akad.Nauk SSSRl, Moscow, 1989, ( in
Russian).
[235] NEUMAIER A., Interval Methods for Systems of Equations, Cambridge Univ. Press,
1990.
[236] NIKOLSKI M.S., On the Approximation of the Attainability Domain of Diﬀerential
Inclusions, Vestn. Mosk. Univ., Ser. Vitchisl. Mat. i Kibern., 4, 1987, pp.3134, ( in
Russian).
[237] NORTON J.P., Identiﬁcation and Application of BoundedParameter Models, Auto
matica, v.23, N4, 1987, pp.497507.
[238] NURMINSKI, E. A., URYASIEV, S. P. The Diﬀerence of Convex Sets, Doklady AN
Ukrainian SSR, Ser. A, 1. 1985, (In Russian).
[239] OLECH C., The Characterization of Weak Closures of Certain Sets of Integrable
Functions, . SIAM J.Contr.,Opt., v.12, 1974.
273
[240] OSIPOV Yu.S. The Alternative in a DiﬀerentialDiﬀerence Game, Sov.Math.Dokl.,
v.12, N2, 1971, pp. 619624.
[241] OSIPOV Yu.S., KRYAZHIMSKII A.V., Inverse Problems of Ordinary Diﬀerential
Equations: Dynamical Solutions, Gordon and Breach, 1995.
[242] OVSEEVICH, A. I. Extremal Properties of Ellipsoids Approximating Attainability
Sets, Problems of Control and Information Theory , Vol.12, N1, pp. 111, 1983.
[243] OVSEEVICH A. I., RESHETNYAK Yu.N., Approximation of the Intersection of
Ellipsoids in Problems of Guaranted Estimation, Sov.J.Comput.Syst.Sci., v.27, N1,
1989.
[244] PANASYUK A. I., PANASYUK V. I., On an Equation Resulting from a Diﬀerential
Inclusion, Mathem.Notes, v.27, N3, 1980, pp.213218.
[245] PANASYUK A. I., Dynamics of Attainable Sets of Control Systems, Diﬀ.Eqns., v.24,
N12, 1988.
[246] PANASYUK A. I., Equations of Attainable Set Dynamics, pp.1,2, JOTA , v.64, 1990,
p.349.
[247] PETROSIAN L.A., Diﬀerential Games of Pursuit, Leningrad Univ.Press, 1977.
[248] PETROVSKI I.G., Lectures on Ordinary Diﬀerential Equations, 6th ed., Nauka,
Moscow, 1970.
[249] PETTY C.M., Ellipsoids, in Convexity and Its Applications, P.M.Gruber, J.M.Wills
eds., Birkhauser, 1983.
[250] PISIER G., The Volume of Convex Bodies and Banach Space Geometry,
Cambr.Univ.Press, 1989.
[251] POLOVINKIN E. S., The Theory of Multivalued Mappings MFTI, Moscow, 1983.
(in Russian),
[252] POLOVINKIN E. S., SMIRNOV G.V., Diﬀerentiation of Multivalued Mappings and
Properties of Solutions of Diﬀerential Inclusions, Sov.Math.Dokl., v.33, 1986, pp.662
666.
[253] POLYAK B.T., SCHERBAKOV P., SCHMULYIAN S., Circular Arithmetic in Ro
bustness Analysis, in: Modeling Techniques for Uncertain Systems, Kurzhanski, A.B.,
Veliov V.M., eds.,ser. PSCT18, 1994, pp. 229243.
[254] POLYAK B.T., TSYPKIN Ya.Z., Robust Identiﬁcation, Automatica, v.16, N1, 1980.
[255] POLYAK B.T., TSYPKIN Ya.Z., Robust Stability Under Complex Parameter Per
turbations, Autom. and Remote Control, v.52, 1991, pp.10691077.
274
[256] PONTRYAGIN, L. S., On Linear Diﬀerential Games, I., II., Soviet Mathematical
Doklady, V.8, 1967, pp.769771 and pp.910912.
[257] PONTRYAGIN, L. S. Linear Diﬀerential Games of Pursuit, Mat. Sbornik, vol. 112
(154):3(7), 1980, (In Russian).
[258] PONTRYAGIN L.D., BOLTYANSKI V.G., GAMKRELIDZE R.V., MISCHENKO
E.F., Mathematical Theory of Optimal Processes, Interscience, NY, 1962.
[259] PSCHENICHNYI B.N., The Structure of Diﬀerential Games, Sov. Math. Dokl., v.10,
1969, pp. 7072.
[260] PSCHENICHNYI B.N., Necessary Conditions of Extremum, Marcel Decker, 1972.
[261] PSCHENICHNYI B.N., Convex Analysis and Extermum Problems, Nauka, Moscow,
1980, ( in Russian).
[262] PSCHENICHNYI B.N., OSTAPENKO V.V., Diﬀerential Games, Naukova Dumka,
Kiev, 1992, ( in Russian).
[263] PSCHENICHNYI B.N., POKOTILO V.G., KRIVONOS I.V., On Optimization of
the Observation Process, Prik.Mat.Meh, v.54, N3, 1990, ( Translated as Appl.Mat.
and Mech.).
[264] QUINCAMPOIX M., Diﬀerential Inclusions and Target Problems, SIAM
J.Cont.Opt., v.30, 1992.
[265] ROCKAFELLAR, R. T. Convex Analysis, , Princeton University Press, 1970.
[266] ROCKAFELLAR, R. T., State Constraints in Convex Problems of Bolza, SIAM J.
Control,v.10, 1972, pp.691715.
[267] ROCKAFELLAR R. T., The Theory of Subgradients and its Application to Problems
of Optimization of Convex and Nonconvex Functions, Helderman Verlag, W.Berlin,
1981.
[268] ROCKAFELLAR R. T., LinearQuadratic Programming and Optimal Control ,
SIAM J.Cont.Opt., v.25, 1987, pp.781814.
[269] ROCKAFELLAR R.T., WETS R.J.B., Generalized LinearQuadratic Problem of
Deterministic and Stochastic Optimal Control in Discrete Time, SIAM J. Cont.Opt.,
v.28, 1990, pp.810822.
[270] ROXIN E., On the Generalized Dynamical Systems Deﬁned by Contingent Equa
tions, J.Diﬀ. Eqns, 1, 1965, pp.188205.
[271] SAINTPIERRE P. Approximation of the Viability Kernel, Applied Math. and Op
tim., v.29, 1994, pp.187209.
275
[272] SAMARSKII A.A., The Theory of Diﬀerence Schemes, Nauka, Moscow, 1983, ( in
Russian).
[273] SAMARSKI, A.A., GULIN A.V., Numerical Methods, Nauka, Moscow, 1989, (in
Russian).
[274] SCHLAEPFER F., SCHWEPPE F., ContinuousTime State Estimation Under Dis
turbances Bounded by Convex Sets, IEEE Trans. Autom. Contr., AC17, 1972,
pp.197206.
[275] SCHNEIDER R., Steiner Points of Convex Bodies, Isr.J. of Math., v.9, 1971, pp.
241249.
[276] SCHWEPPE F. C., Recursive State Estimation: Unknown but Bounded Errors and
System Inputs, IEEE, Trans. Aut. Cont., AC13, 1968.
[277] SCHWEPPE F. C., Uncertain Dynamic Systems, Prentice Hall, Englewood Cliﬀs,
NJ, 1973.
[278] SEEGER A., Direct and Inverse Addition in Convex Analysis and Applications,
J.Math.Anal.Appl, v.148, N2, 1990, pp.317349.
[279] SERRA J., Image Analysis and Mathematical Morphology, Academic Press, 1982.
[280] SHARY P., On Controlled Solution Set to Interval Algebraic Systems, Interval Com
putations, v4., N6, 1992, pp. 6675, (in Russian).
[281] SHOR N.Z., BEREZOVSKI O.A., New Algorithms for Constructing Optimal Cir
cumscribed and Inscribed Ellipsoids , Optimization Methods and Software, v.1, N4,
1992, pp.283299.
[282]
ˇ
SILJAK D.D., Decentralized Control of Complex Systems, Acad. Press., 1991.
[283] SKOWRONSKI J. M. A Competitive Diﬀerential Game of Harvesting Uncertain
Resources, in: D. F. Batten, P. F. Lesse Eds. New Mathematical Advances in Economic
Dynamics, Croom Helm, London Sydney, 1985. pp. 105118.
[284] SNYDER J.M., Interval Analysis for Computer Graphics, ACM Computer Graphics,
v.26, N2, 1992, pp.121130.
[285] STOER J., WITZGALL C., Convexity and Optimization in Finite Dimensions I,
SpringerVerlag, 1970.
[286] STOER J., SONNEVEND G., The Theory of Analytical Centres, Appl.Math. and
Opt., 199?.
[287] SUBBOTIN A.I., A Generalization of the Basic Equation of Diﬀerential Games,
Sov.Math.Dokl., v.22, N2, 1980, pp. 358362.
276
[288] SUBBOTIN A.I., Existence and Uniqueness Results for HamiltonJacobi Equations,
Nonlinear Analysis, v.16, 1991, pp. 683689.
[289] SUBBOTIN A.I., Generalized Solutions of FirstOrder PDE’s. The Dynamic Opti
mization Perspective, Ser. SC , Birkh¨auser, Boston, 1995.
[290] SUBBOTIN A.I., CHENTSOV A.G., Optimization of Guarantees in Problems of
Control, Nauka, Moscow, 1981, (in Russian).
[291] SUBBOTINA N.N., The Maximum Principle and the Superdiﬀerential of the Value
Function, Problems of Contr. and Inf.Theory, v.18, N3, 1989, pp.151160.
[292] TANAKA M., OSADA A., TANINO T., Ellipsoidal Approximations in Set
Membership State Estimation of Linear DiscreteTime Systems, Transactions of the
Society of Instrument and Control Engineers, v.27, N.12 , 1991, pp.137481, ( in
Japanese).
[293] TARASYEV A.M., Approximation Schemes for Construction of the Generalized So
lution of the HamiltonJacobi (BellmanIsaacs) Equation, Prikl.Mat.Meh., v.58, 1994,
pp.2226, ( in Russian).
[294] TIKHONOV A.N., On the Dependence of the Solutions of Diﬀerential Equations on
a Small Parameter, Matem.Sbornik, v, 22, 1948, pp.198204.
[295] TIKHONOV A.N., Systems of Diﬀerential Equations Containing a Small Parameter
Multiplying the Derivative, Matem.Sbornik, v.31, 1952, pp.575586.
[296] TITTERINGTON D.M., Estimation of Correlation Coeﬃcients by Ellipsoidal Trim
ming, Appl.Stat., v.27, N3, 1978, pp.227234.
[297] TODD M., On MinimumVolume Ellipsoids Containing Part of Given Ellipsoid,
Math. of Operat.Res., v.7, N2, 1982, pp. 253261.
[298] TOLSTONOGOV A.A., Diﬀerential Inclusions in Banach Space, Nauka, Novosi
birsk, 1986.
[299] TRAUB J.F., WASILKOWSKI G.W., WOZNIAKOWSKI H., InformationBased
Complexity, Acad.Press, 1988.
[300] TSAI W.K., PARLOS A.G., VERGHESE G.C., Bounding the States of Systems with
Unknownbutbounded Disturbances. Int.J. of Contr. v.52, N 4, 1990, pp.881915.
[301] USORO P.B., SCHWEPPE F.C., WORMLEY D.N., GOULD L.A., Ellipsoidal Set
Theoretic Control Synthesis, J.Dyn.Syst.Measur.Cont., v.104, 1982, pp.331336.
[302] USTIUZHANIN A.M., On the Problem of Matrix Parameter Identiﬁcation, Problems
of Control and Inf. Theory, v.15, N4, 1986, pp. 265274.
277
[303] VALENTINE F., Convex Sets, MacGrawhill, N.Y., 1964.
[304] VALYI I., Ellipsoidal Approximations in Problems of Control, in: Modelling and
Adaptive Control Ch.Byrnes, A. B. Kurzhanski, eds, Lect.Notes in Contr. and In
form.Sci., v.105, SpringerVerlag, 1988.
[305] VALYI I., Ellipsoidal Methods in TimeOptimal Control, in: Modeling Techniques
for Uncertain Systems, A.B.Kurzhanski, V.M.Veliov eds., ser. PCST 18, Birkh¨auser,
Boston, 1994.
[306] VARAIYA P. On the Existence of Solutions to a Diﬀerential Game, SIAM J. Control,
v.5, N1, 1967, 153162.
[307] VARAIYA, P., LIN, J. Existence of Saddle Points in Diﬀerential Games, SIAM J.
Control, Vol. 7, 1, 1969, pp. 142157.
[308] VELIOV V. M., Second Order Discrete Approximations to Strongly Convex Diﬀer
ential Inclusions, Systems and Control Letters, v.13, 1989, pp.263269.
[309] VELIOV V.M., Approximations to Diﬀerential Inclusions by Discrete Inclusions,
IIASA Working Paper WP89017, Laxenburg, 1989.
[310] VICINO A., MILANESE M., Optimal Inner Bounds of Feasible Parameter Set in
Linear Estimation with Bounded Noise, IEEE, Trans. Aut.Contr., v.36, 1991, pp.
759763.
[311] VINTER R.B., A Characterization of the Reachable Set for Nonlinear Control Sys
tems, SIAM J. Contr. Optim., v.18 , 1980, pp.599610.
[312] VINTER R. B., WOLENSKI P., Hamilton  Jacobi Theory for Optimal Control
Problems with Data Measurable in Time, SIAM J.Contr. Optim., v.28 , 1990, pp.1404
1419.
[313] WAZEWSKI T., Syst`emes de Commande et
`
Equations au Contingent,
Bull.Acad.Pol.Sci., 9, 1961, pp.151155.
[314] WIENER N. Extrapolation, Interpolation and Smoothing of Stationary Time Series,
MIT Press, Cambridge, Ma, 1949.
[315] WILLEMS J.C., Dissipative Dynamical Systems, Part I: General Theory,
Arch.Rat.Mech.Anal., v.45, 1972, pp.321351.
[316] WILLEMS J.C. Feedback in a Behavorial Setting, in: Systems, Models and Feed
back: Theory and Applications, Isidori A., Tarn T.J., eds., Ser. PCST 12, Birkh¨auser,
Boston, 1992.
[317] WITSENHAUSEN H. S., Set of Possible States of Linear Systems Given Perturbed
Observations, IEEE Trans.Autom. Cont., AC13, 1968, pp.556558.
278
[318] WOLENSKI P. R. The Exponential Formula for the Reachable Set of a Lipschitz
Diﬀerential Inclusion, SIAM J.Contr. Optim., v.28, 1990, pp.11481161.
[319] WONHAM W.M., Linear Multivariable Control. A Geometric Approach, Springer
Verlag, 1985.
[320] YOUNG L.C., Lectures on the Calculus of Variations and Optimal Control Theory,
Saunders, Phil.,1969.
[321] ZADEH L.A., Fuzzy Sets,Inform. and Control, v.8, 1965, pp.338353.
[322] ZAGUSKIN V.L., On Circumscribed and Inscribed Ellipsoids of the Extremal Vol
ume, Uspekhi Mat.Nauk., v.13, 6(64), 1958, ( in Russian). Translated as Soviet
Math.Surveys.
279
Contents
Preface Part I. Evolution and Control. The Exact Theory Introduction 1.1 The System 1.2 Attainability and the Solution Tubes 1.3 The Evolution Equation 1.4 The Problem of Control Synthesis: a Solution Through SetValued Techniques 1.5 Control Synthesis Through Dynamic Programming Techniques 1.6 Uncertain Systems. Attainability Under Uncertainty 1.7 Uncertain Systems : the Solvability Tubes 1.8 Control Synthesis Under Uncertainty 1.9 State Constraints and Viability 1.10 Control Synthesis Under State Constraints i 1
6 6 9 12 15
22
31
40
47 53 62
68
1.11 State Constrained Uncertain Systems. Viability Under Counteraction. 1.12 Guaranteed State Estimation : the Bounding Approach 1.13 Synopsis 1.14 Why Ellipsoids ? Part II. THE ELLIPSOIDAL CALCULUS Introduction 2.1 Basic Notions : the Ellipsoids 2.2 External Approximations: the Sums. Internal Approximations: the Diﬀerences 2.3 Internal Approximations: the Sums. External Approximations: the Diﬀerences 2.4 Sums and Diﬀerences: the Exact Representation 2.5 The Selection of Optimal Ellipsoids 2.6 Intersections of Ellipsoids 2.7 Finite Sums and Integrals: External Approximations 2.8 Finite Sums and Integrals: Internal Approximations ii
73
76 83 90 93 93 94
105
119
124 127 136
149
158
Part III. ELLIPSOIDAL DYNAMICS: EVOLUTION and CONTROL SYNTHESIS Introduction 3.1 EllipsoidalValued Constraints 3.2 Attainability Sets and Attainability Tubes: the External and Internal Approximations 3.3 Evolution Equations with Ellipsoidal  Valued Solutions 3.4 Solvability in Absence of Uncertainty 3.5 Solvability Under Uncertainty 3.6 Control Synthesis Through Ellipsoidal Techniques 3.7 Control Synthesis: Numerical Examples 3.8 ”Ellipsoidal” Control Synthesis for Uncertain Systems 3.9 Control Synthesis for Uncertain Systems: Numerical Examples 3.10 Target Control Synthesis within Free Time Interval Part IV. ELLIPSOIDAL DYNAMICS: STATE ESTIMATION and VIABILITY PROBLEMS Introduction iii
164 164 165
168
177 181 185
194 199
206 211
217
221 221
4.1 Guaranteed State Estimation: a Dynamic Programming Perspective 4.2 From Dynamic Programming to Ellipsoidal State Estimates 4.3 The State Estimates, Error Bounds and Error Sets 4.4 Attainability Revisited. Viability Through Ellipsoids 4.5 The Dynamics of Information Domains. State Estimation as a Tracking Problem 4.6 Discontinuous Measurements and the Singular Perturbation Technique References
224
232
237
240
246
254 258
iv
population . should match the abilities of modern computer technology . for example. whether directed at traditional ( aerospace. Its aim is to present a uniﬁed framework for eﬀectively solving these problems and their generalizations to the end through modern computer tools. The setmembership model of uncertainty reﬂects many actual informational situations in applied problems. desirably. processes with limited numbers of 1 .to use the data to specify or reﬁne the mathematical model. among the possible approaches. regulation. then to analyse the model through available or new mathematical tools and further on . the best or the better control laws. These. Among the speciﬁc features of the controlled processes in the mentioned areas are usually their dynamic nature and the uncertainty in their description. It is on how to construct.Preface It is well known that the emphasis of mathematical modelling on the basis of available observations is at ﬁrst . a solution theory that allows analytical designs relatively simple for practical implementations or leads. what is this book about? The present book is devoted to an array of selected key problems in dynamic modelling.to use this analysis in order to predict or prescribe ( control ) the future course of the modelled process. But it is not any type of theory that is desired. So. Not the least objective is to develop. communication and transport). It is also to indicate how the level of uncertainty and the ” amount” of information used for designing the feedback control laws aﬀects the result of the controlled process. to eﬀective numerical algorithms. at least. ﬁnances and economics. The given principles are also among the objectives of modern control theory . the values of the cost functions or the aspired ”guaranteed” performance levels. Following this reasoning. An important component of the overall process is to verify the model and its performance over the actual course of events. Particularly. technology) or relatively new applications ( environment. with setmembership description of the uncertain items. This is particularly done by specifying feedback control strategies (policies) that realize the desired goals. biomedical issues. one may claim that control theory is a science of assigning feedback control or regulation laws for dynamic processes on the basis of available information on the system model and its performance goals (given both through online observations and through data known in advance). . it appears relevant in estimating nonrepetitive processes. mechanics. in some appropriate sense.state estimation . viability and feedback control under uncertainty. particularly . These are taken to be unknown but bounded with preassigned bounds and no statistical information whatever.allow parallel calculations and graphic animation. The model of uncertainty considered here is deterministic.
but to indicate ellipsoidal representations of the exact solutions. The book is divided into four parts designed along the following lines. incomplete knowledge of the problem data and no available statistics. unions or intersections). 1 The problems treated here are described through setvalued functions and are thus to be treated through setvalued analysis. feedback control and gametype dynamics not only allows to present the rather sophisticated mathematical solutions in ”visible” forms ( and literally. Namely. 1 The references and some historical comments are given in the introductions to each part and throughout the text. under unmodelled dynamics. with an emphasis on those directly related to the topics of this publication and those that would allow. The authors believe that it may also give new insights in to the mathematical structure of the solutions. it also reﬂects the research preferences. but a calculus that would allow eﬀective solutions of the selected problems and their generalizations with fairly simple control designs and possibility of graphic animation. It is necessary to indicate that the ellipsoidal techniques of this particular book are not conﬁned to approximation of convex sets by one or several ellipsoids only as done in other publications. as we hope. The scheme includes approximations by single ellipsoids as a particular element of the overall approach. which. It is a common approach in pursuitevasion diﬀerential games. Needless to say . A particular emphasis of this book is on the possibility of computergraphic representations. 2 . each convex set or convex setvalued function considered here is represented by a parametrized variety of ellipsoids or ellipsoidalvalued functions. some assertions of general nature proved in this book were ﬁrst ” noticed” during the animation experiments). a more and more accurate representation with exact one in the limit. representative. The attempt is based on introducing an ”ellipsoidal calculus” that allows to represent the exact setvalued solutions of the respective problems through ellipsoidalvalued functions. rather than exhaustive number of publications available to them. These are certainly not conﬁned only to control applications . but cover a broad variety of problems in systems modelling. interests and experiences of the authors. to pursue the further directions indicated here. However. This further allows to parallelize the calculations and to animate the solutions through computer graphic tools. The solutions are thus constructed of elements that involve only ellipsoidal sets or ellipsoidalvalued functions and operations over such sets or functions.attainability (reachability) and solvability. while their number increases. The authors apologize that among the enormous literature on the subject they were able to mention only a very limited. control and disturbance attenuation.observations . they also demonstrate an approach applicable to many other situations that spread quite beyond the topics addressed here. to peer into the multidimensional spaces through computer windows or more sophisticated computer tools). The ﬁrst part gives exact solutions to the problems of evolution . particularly. (Thus. The animation of the problems in estimation . the aim of this book is not to produce any type of setvalued technique. The authors also hope that though applying their techniques to a specially selected array of problems. in robust stabilization. jointly allow ( through their sums .
a direct interpretation in terms of the above mentioned ”bridges” . Particularly. to parallelizable computation schemes. experiment planning. through the solution of a simple algebraic equation. interval analysis and its generalizations. The third part also deals with feedback control. except for a scalar parameter whose dependence on the state space vector may be calculated in advance.the key elements of gametheoretic feedback control. In the latter case these are more complicated.N. image processing. particularly. Though written for applications to Problems of Part I. These analytical designs are possible due to the fact that the internal ellipsoidal tubes that approximate the solvability domains under uncertainty are precisely such. that they posses the property of being an ”ellipsoidalvalued bridge”. It also indicates the applicability of the suggested schemes to problems posed within the socalled H∞ approaches. being related to reachability or solvability under uncertainty or counteraction and allowing.Krasovski as a basis for further ”aiming rules”. problems in probability and statistics. The latter property arrives due to two basic features : the fact that the respective ellipsoidalvalued mappings posses a generalized semigroup property and the fact that the internal tubes are inclusionmaximal among all other internal ellipsoidal tubes. or through the evolution of support functions or through level sets of appropriate HJB ( HamiltonJacobiBellman) equations. The respective control designs are based on applying ”ellipsoidal” versions of the exact solutions. of course. The third and fourth parts indicate the applications of ellipsoidal calculus to problems of Part I. the text of this chapter may be also considered as a separate theory with motivations and applications coming from topics other than discussed here. from optimization under uncertainty and multiobjective optimization. the third part describes ( both in forward and backward time) the internal and external ellipsoidal representations of attainability (reachability) tubes for systems without and with uncertainty. The feedback control designs are based on using setvalued solvability tubes which may be interpreted as ”bridges” introduced by N. through ellipsoidalvalued representations . The principal schemes for this framework are speciﬁcally directed towards the desired transition. Thus. when the 3 . The second part of the book describes the ellipsoidal calculus itself.as well as of estimation . adaptive systems and robotics. It covers external and internal ellipsoidal representations for basic setvalued operationsgeometrical (”Minkowski”) sums and diﬀerences as well as intersections of ellipsoids and integrals of ellipsoidalvalued functions. as well as on Dynamic Programming considerations. viability and feedback terminal target control. This leads to a nonlinear control synthesis in the form of analytical designs . The exact theory ( both known and new ) is rewritten within a uniﬁed framework that involves trajectory tubes and their setvalued descriptions either through evolution equations of the ”funnel type”. The fourth part deals with stateestimation under unknown but bounded errors. with attainability under state constraints and viability problems. mathematical morphology and related areas of theoretical and applied research.
particularly. the Dynamic Programming approaches applied here open the routes to further penetration into generically nonlinear classes of systems. 2 Another important aspect hardly discussed here is the accuracy and computational complexity of the underlying algorithms. becoming an eﬀective tool in nonlinear control theory.6 deals with problems of state estimation under ”bad noise”. This book is not a collection of numerous facts or artifacts in setvalued analysis or control theory. The narrative stops just short of control under measurement feedback and adaptive control. as we believe. The respective challenges are beyond the material of the present book. this book indicates a uniﬁed concise framework for problems of state estimation. As already mentioned . and thus to solve the problem to the end. In this text we are conﬁned to ”linearconvex” systems and problems. These are areas which require separate serious consideration and explanation. particularly. [289]). the control synthesis given here is nonlinear and the synthesized systems are nonlinear systems. the application of ellipsoidal techniques would be especially useful in these areas . It introduces an ellipsoidal calculus . which are approached through the incorporation of singular perturbation techniques. For covering the needs of this book we use its simple versions that do not extend beyond the use of subdiﬀerential calculus. Other topics include new types of dynamic relations for the treatment of information sets and vectorvalued ”guaranteed estimators” as well as an interpretation of the state estimation problem as one of tracking an unknown motion under unspeciﬁed but bounded errors. our aim was also to stimulate and encourage further investigation in the spirit of the present approach as well as implementations in reallife modelling. The ﬁnal Section 4. where the ﬁrst of the authors had earlier worked for many years. This links the problem with those of viability under counteraction.Krasovski and his associates at in Ekaterinburg (Sverdlovsk). Moreover. However. 2 4 . Numerical examples complement the theoretical parts.value of the error bound is not speciﬁed. It is rather a book on basic problems and principles for calculating their solutions through setvalued models. citecranish. The conceptual approaches to controlled dynamics that served as a background for this work were inﬂuenced by the research of N. viability and feedback control under setmembership uncertainty for systems with linear dynamics. However. (The latter issue could be the topic of a separate monograph). In the The ”nondiﬀerentiable” version of Dynamic Programming has been substantially developed in the recent years (see references [82]. This is due to the involvement of Dynamic Programming techniques. Whether reached or not. This approach allows to treat discontinuous observations described by measurable functions and also to deal with viability problems under ”measurable constraints”. of one and the same HJB equation for the treatment of uncertain dynamics in both of the settings investigated here.N. The algorithmization and animation in these cases is certainly a worthy challenge.to develop the solutions from theory to algorithms and computer animation.
E.Nikonov.Beschler. A. A. R. K. W. M.Mitter.Tikhomirov. Our special thanks go to C. to E. M.Leitmann. the SDS secretaries E.Pﬂug. G.Schepunova for her help in arranging the ﬁnal version of the manuscript. O.SaintPierre.Robinson. A.Tanino.Willems. J. The principal parts of this book and the underlying ellipsoidal ”representation” approach for the setvalued functions were worked out throughout the authors participation at the SDS (Systems and Decision Sciences) Program of IIASA . H. P. R.Krener.Gusev. pleasant working conditions and possibility of regular contacts with a broad spectrum of researchers certainly stimulated our work at the Institute and the direction of our eﬀorts. A. A.Baras. P.T.Wierzbicki for their stimulating discussions and support. P. B. the late Vladimir a Sergeevich Michalevich. V.the International Institute of Applies Systems Analysis at Laxenburg. Yu.Artstein.Norton.Knobloch. G.Runggaldier. M. We thank T. J.Rusczynski. The authors are grateful to the Directors of IIASA . Throughout the last years we had the pleasant opportunity to discuss the topics treated in this book with Z.Filippova. P. the Publisher and to the Birkhauser staﬀ for their support. K.Veliov.Thoma. patience and understanding of the problems faced by the authors in preparing the manuscript.P.Milanese. Their valuable comments certainly helped to shape the contents.necessity of studying and applying setvalued analysis we believe to share the views of JP. M.Sugimoto who have coauthored some of our IIASA Working Papers used here and O. particularly of our own investigations. Aubin.A. We wish to thank our colleagues at IIASA and its Advisory Committee on Methodology — J. R. T.Thomas Lee.Varaiya. The serious but friendly atmosphere.Kenderov. S.Wets.Rockafellar.Aubin and his colleagues.Frankowska. 5 . C.Chernousko.Makowski for their help in computer graphics.Tanaka. K.Martin. S.Gaivoronski.Fedra and M. F.Kalman. H.Subbotin.Pschenichnyi.Mischenko.Veres.Sigmund. Robert Pry and Peter De J´nosi and to the Chairman of the IIASA Council in 19871992 .Gruber and C.the Editor of the Birkhauser Series on Systems and Control: Foundations and Applications. for their support of methodological research at IIASA.Osipov. V. J.Kall.Isidori.Byrnes . A.EnzlbergerVaughan for preparing papers and manuscripts used in this book. Austria. S.
[270]). [193]. However. x0 . ˙ with magnitude bounds on the controls u and the uncertain items f (t).see [34]. The class of systems treated here are linear timevariant systems x = A(t)x + u + f (t). It presents a uniﬁed approach to these topics based on descriptions involving the notions of trajectory tubes and evolution equations for these tubes. [17] 3 6 . The ﬁrst step is the description of attainability (reachability) domains X [t] for systems without uncertainty. these equations indicate setvalued discretetime schemes important for calculations. [298]. The evolution of these in time is naturally described by setvalued (”Aumann”) integrals with variable upper time limit. This idea is Among the recent investigations on evolution equations for trajectory tubes are papers [246].Part I. Namely. it is not unimportant to introduce some sort of evolution equation with setvalued statespace variable that would describe the dynamics of sets X [t] in time. This immediately leads to setvalued ”attainability tubes” whose crossections are the attainability domains. Evolution and Control. t1 ]. The objective is nevertheless reached through evolution equations of the funnel type that do not involve such derivatives. The target control processes and the estimation problems are considered within ﬁnite time intervals : t ∈ [t0 . The Exact Theory Introduction The present ﬁrst part of the book is a narrative on constructive techniques for modeling and analyzing an array of key problems in uncertain dynamics. then the synthesized control strategy should be designed to keep the trajectory within this tube (or ”bridge”) throughout the process. estimation and control. The solvability tubes are used here in synthesizing feedback control strategies for problems of terminal target control. 3 . The serious obstacle for deriving such an equation in a diﬀerential form is the diﬃculty in deﬁning an appropriate derivative for setvalued functions. x(t0 ) = x0 . if the solvability tube ends at the target set M.Roxin . The present topic thus diﬀers from problems where the objective lies only in the achievement of an appropriate asymptotic behaviour of the trajectories with perhaps some desired quality of the transient process.Barbashin and E. Though somewhat cumbersome at ﬁrst glance. (Among the ﬁrst investigations with this emphasis are the works of E. The tubes X [t] could then be interpreted as ”trajectories ” of some generalized dynamic systems. The attainability tubes may be also constructed in backward time in which case they are referred to as solvability tubes . This requires a rather detailed investigation of the system dynamics.
x} we come to the value function V (t. [181]. when calculated over closed loop controls is diﬀerent. for example. In this case the attainability set under counteraction (in forward time ) and solvability set under uncertainty ( in backward time) are in general far more complicated than in the absence of uncertainty (see Sections 1. It is the former value that may be described through a respective HJBI ( HamiltonJacobiBellmanIsaacs) equation. from endpoint x[t1 ] to the target set M . The function V (t. One may also apply Dynamic Programming to the mentioned uncertain systems. in general.Pschenichnyi [259]. Under some nondegeneracy conditions.7.6). [168]. for example (see Section 1.the openloop solvability tubes from the closedloop solvability tubes (Section 1.N. The backward procedure for solvability tubes is also similar in nature to the schemes introduced by P. for example. [219]. we note that now it should be minimaximized over the control u and the input disturbance f respectively. nevertheless. Selecting a starting position {t. x). Namely. The next stage is the treatment of systems with input uncertainty f (t) ( unknown but bounded. x = x(t). There is a particular case. Taking the cost function d2 (x(t1 ). x) then satisﬁes a corresponding generalized HJB (HamiltonJacobiBellman) equation.Pontryagin’s alternated integral.7).6). (see [109]. One should note that in the absence of uncertainty the value function V (t. the latter ones may be again interpreted as Krasovski’s ”bridges” ( now for uncertain systems) and may be used for designing feedback strategies through the extremal aiming rules (Sections 1.the essence of the ”extremal aiming rule” introduced by N. M). A key element that allows to use the solvability tubes for the control synthesis problem is that the respective multivalued maps satisfy a semigroup property and therefore generate a generalized dynamic system with setvalued trajectories. The calculation of the solvability tubes is then as simple as in the absence of uncertainty.8).) There is an exception again. But the value of this minmax. [307] and B. [289]. It is indicated here that they also satisfy some special evolution equations of the funnel type. A similar type of strategy may be derived through a Dynamic Programming technique with cost function being the square of the Euclid distance d2 (x[t1 ]. [167]. [257]. x) is the same both for openloop (programmed) control and for closedloop (positional) feedback control. Finding the optimal value of the cost function for any position {t. [169] and used in Section 1.4. x) may be therefore calculated through standard methods of convex analysis used traditionally for solving related problems of openloop control.5). then 7 . [171].8). if the matching conditions are satisﬁed . For the ”linearconvex” problems of these Sections the function V (t. x}.Varaiya et al. from its value calculated for open loop controls.Krasovski. [266].61. In the linearconvex case considered here the constructive description of solvability tubes may be given by a setvalued integral known as L. when the openloop and closedloop solvability tubes coincide. however. with magnitude bounds). M). 1. (Section 1. This is when the system satisﬁes the socalled matching condition which mean that the bounds on the controls u and the disturbances f are similar in some sense (Section 1. One should now distinguish. we may minimize the cost function by selecting an appropriate optimal control ( in the class of either openloop or closedloop controls).S.
[50]. in another interpretation). Having in mind the previous remarks. The last problem of the ﬁrst part is the one of state estimation under unknown but bounded errors and disturbances (Section 1. while also reaching the terminal set ( Section 1. [119] ). [178]. [317]. [179].11). one may observe that the value functions V (t. We now proceed with the main text. Namely.P. whether calculated over openloop or closedloop controls. x) used in this book play the role of Liapunov functions used in respective approaches to the design of feedback controllers for uncertain systems ( see [214]. The latter ones then coincide with viability kernels introduced by J. together with the result of the measurement. 4 The main objects of investigation here are the information sets consistent with the system dynamics. [215]). in other words.Baras and M.10) or viability under counteraction (or under persistent disturbances. ([15]). [94]. where one may also ﬁnd some connections with other approaches to deterministic ﬁltering (particularly. The further problems are similar to the previous ones but complicated by state constraints (”viability restrictions”). commencing with the basic notions. it is to ensure the system to be steered by control u to the target set M ( at given time t1 ) under persistent disturbances f rather than to ﬁgure out the saddle points of a positional (feedback) dynamic game between two ”equal” players u and f which is the emphasis of the theory of diﬀerential games (see [37]. whilst constructive techniques are introduced in parts III. We should also emphasize that the problem treated here is to reach the goal in ﬁnite time attenuating the unknown disturbances. 8 .9). in the interpretation of J. The control synthesis problem is now to ensure viability (Section 1. The evolution equation for the information set acts a ”guaranteed ﬁltering equation” and the guaranteed state estimate is then the ”Chebyshev center” of this set ( namely. the H∞ approach. The information sets are actually the attainability domains under a state constraint that is induced by the measurement equation and therefore arrives online. 4 The ﬁrst investigations of state estimation problems under unknown but bounded inputs date to papers [166].James [30]). A synopsis of the results and some suggestions on why ellipsoids were taken to be studied ﬁnalize this part. the available measurement and the constraints on the uncertain items. the value function V (t.Aubin. within the framework of viability theory (Section 1. This ﬁrst part of the book gives but a general introduction into the problem. [276].12). A systematic investigation of the setvalued approach in continuous time seems to have started with [54]. [181]. x) is the same. IV. Evolution ” funnel” equations are introduced for the dynamics of attainability sets under state constraints (in forward time) and respective solvability sets ( in backward time). the center of the smallest ball that includes the information set ). [171].the minmax of the cost function or.
measurable on T (a “measurable control”). The values u of the controls are assumed to be restricted for almost all t by a magnitude or ”geometrical ” constraint (1.1 The System In this book we consider dynamic models described in general by a linear timevariant system (1. x(t)) + f (t).4) where (1.3) x(t) = A(t)x(t) + u(t) + f (t) ˙ with u(t) ∈ P(t).1. being a function of the position {t.1. with u(t). The class of functions u(·) = u(t). t ∈ T . ˙ U(t. and — closed loop. x} of the system.2) u ∈ P(t) . x(t)) is a multivalued map.1) x(t) = A(t)x(t) + u + f (t) ˙ with ﬁnitedimensional state space vector x(t) ∈ IRn and inputs u (the control) and f (t) (the disturbance or external forcing term). In the second case we come to a nonlinear diﬀerential inclusion (1.5) x(t) ∈ A(t)x(t) + U(t.measurable in t ∈ T and restricted as in (1. . when u = u(t) is a function of time t. continuous in t.1. The n × n matrix function A(t) is taken to be continuous on a preassigned interval of time T = {t ∈ IR : t0 ≤ t ≤ t1 } within which we consider the forthcoming problems.1.1. 9 t ∈ T. f (t) assumed Lebesguemeasurable in t ∈ T . In the ﬁrst case we come to a linear diﬀerential equation (1.1. x) ⊆ P(t).2) is further denoted as UP . We shall further consider two types of controls which are: — open loop. when u = U(t. being an openloop control.1. namely U : T × IRn −→ convIRn measurable in t and upper semicontinuous in x. where P(t) is a multivalued function P : T → convIRn . t ∈ O T .
the solution to (1.3) and (1. being restricted by (1.1. upper semicontinuous in x. The initial values for these are related 10 .3) with initial value (1.1. t1 )x(t) and substituting x for z in (1.3) is a standard property of linear diﬀerential equations.5) and such that equation (1.7) is given by the formula x(t) = S(t0 . τ )A(t). ∂t ∂ S(t.6) ∂ S(t.1. Clearly. t1 )x0 .1. c The class UP = {U(t.9) respectively. Systems (1. x(t)) + f (t) ˙ for almost all t ∈ T .1.1. τ ).3) we come to the equation (1.1. t) = I. t ∈ T .1. z(t) = S(t.3).1.4) may be transformed into simpler relations. x)} of feasible control strategies consists of all convex compactvalued multifunctions that are measurable in t. As it is well known. (1.1. [248]).1. τ ) = −S(t. t)x0 + Taking the transformation (1. ∂τ S(τ. t1 )u(t) + S(t.10) z(t) = S(t. there is a onetoone correspondence of type (1. τ ) = I. τ ) stand for the matrix solution to the equation (1.1.1.1. The latter means that there exists an absolutely continuous function x(t). t1 )f (t) ˙ z 0 = z(t0 ) = S(t0 .1. Let S(t.is a feedback (closedloop) control strategy .8) t t0 x(t0 ) = x0 S(τ. The existence of solution for system (1. which also satisﬁes the equation S(t.9) (1. t)(u(τ ) + f (τ ))dτ. [167].6) between the solutions x(t) and z(t) to equations (1. [142]. ([61].4) does have a solution extensible to any ﬁnite time interval T for any x0 = x(t0 ) ∈ IRn . τ ) = A(τ )S(t. that yields the inclusion x(t) ∈ A(t)x(t) + U(t.
10).1. t1 )f (t) S(t. t)z) S(t. z(t) ∈ W(t. but the time range for which the respective substitution is true will be {t ≥ t0 }.1.through (1. t1 )u(t) S(t.1. Therefore. that the described substitution (1. z) P0 (t) = = = = S(t. z) ∈ UP0 . z) + g(t) ˙ w(t) ∈ P0 (t). Then the original system may be again .1.8) allows to consider the forthcoming problems for A(t) ≡ 0 within the time range {t ≤ t1 }.4) with A(t) ≡ 0.4). z) belong to the class deﬁned by the constraint P0 (t). In other terms.7). (1. however. constraint (1.1.1.14) z(t) = S(t. It is compulsory however that the constraint function P0 (t) would be timevariant. ˙ z(t0 ) = S(t0 . The “new” feedback strategies W(t.1. t1 )U(t.1. One should realize . without loss of generality.2)–(1.1.1. we come to (1.1. The reader will always be able to return to A(t) ≡ 0 as a healthy exercise. S −1 (t0 . without loss of generality we may further follow the notations of (1. Without loss of generality we may therefore further treat systems (1.2) and (1.3).1.1. The ﬁrst issue to discuss is the description of the set of states that can be reached in ﬁnite time due to systems (1.1.2) and initial condition (1.1. but otherwise the same as before: c W(t.1. 11 . (1.13) Here obviously w(t) g(t) W(t.1.4).1.2)–(1. taken with A(t) ≡ 0 . t1 )P(t) z(t) = w(t) + g(t). instead of the systems (1.3).5).11)–(1.A similar result may be also obtained by substitution (1.1.13) rather than (1.11) (1. t1 )x0 .4) under restriction (1.12) with constraint (1. t0 ) x(t). and the setvalued function P0 (t) remains continuous. We shall often make use of the indicated facts in the sequel in the hope that this will enable to demonstrate the basic techniques without overloading the text with unessential but cumbersome procedures.1.
t0 .) at time t. from set X 0 . Let x[t] = x(t.2. (1.1) x(t) = u(t) + f (t).2. ˙ 12 . ·) : compIRn → compIRn it is not diﬃcult to check that it satisﬁes the following semigroup property: X (t.3) (or (1. t0 . deﬁnes a solution tube to the diﬀerential inclusion (1.2. Therefore we denote X [t] = X (t.2 Attainability and the Solution Tubes Taking system (1. x0 ) : x0 ∈ X 0 .1. X 0 ) = {x(t. x0 ) denote an isolated trajectory of system (1. X (τ. being driven by a certain control u(t). u(t) ∈ P(t). τ.1). τ with t1 ≥ t ≥ τ ≥ t0 .1) that starts at instant t0 from state x0 . (1.2. Deﬁniton 1.2.1. t ∈ T }.2. X 0 ∈ comp IRn . t0 . with constraint (1.1. X 0 ) is referred to as the attainability domain for system (1. t0 . The setvalued map X [t] = X (t. t0 .2. X 0 ) = X (t. ˙ t ∈ T. X 0 ) . One of the ﬁrst questions that arise in control theory is to describe the variety of all states x = x(t) that can be reached by the system trajectories that start at a prescribed set X 0 . t0 .3) x(t) ∈ P(t) + f (t).1. whatever are the values t.1 The set X [t] = X (t.2) for A(t) ≡ 0.2) u(t) ∈ P(t). t0 .3). For the mapping X (t.2. t0 . We will be further interested in the union of all such isolated trajectories over all possible initial states x0 ∈ X 0 and measurable controls u(t) ∈ P(t). we have (1.1. t ∈ T.2) x0 ∈ X 0 . The attainability domain X [t] is often said to be the reachability domain . We also presume that the initial state x0 = x(t0 ) is restricted by the inclusion (1. X 0 )).
t ∈ T.1.5) x(t) ∈ U(t. x[t] = x∗ . [181]. t0 .1 The multifunction X [t] is convex compactvalued (X [t] ∈ convIRn ) and continuous on the interval T . ∞)}.1. t0 ≤ τ ≤ t. X [t0 ] = X 0 is also known as the solution tube to system (1.3) and set X [t0 ] = X . Investigate the problem of Remark 1. An aﬃrmative answer will indicate that any point in IRn may be reached in ﬁnite time O through a bounded control u(·) ∈ UP .2.that starts from set X 0 . from set X 0 . It is clear that the control u(t) = x[t] − f (t) ˙ is the one that corresponds to x[t].2.1.1 One of the popular problems studied on the subject of attainability is the following: given X 0 = {0}. ˙ The multivalued function X [t]. coincide with the whole space IRn ?.Otherwise one is to specify X as a subset of IRn . ˙ (1. the set X [t] = {x∗ } consists of all those vectors x∗ for each of which there exists an isolated trajectory x[τ ] = x(τ.2. Exercise 1. t0 ≤ τ ≤ t.2.6) 5 0 c U(·. Remark 1. t ∈ [t0 .2.2).4) X [t] = {x[t] : x[τ ] − f (τ ) ∈ P(τ ). Passing to the diﬀerential inclusion (1. x[t] ∈ X0 }.2. of (1. ·) ∈ UP . 13 . Other terminology says that X [t] is the trajectory assembly generated by system (1.2.2.2. for the interval t ∈ [t0 . x) + f (t).2. will the set X 0 = ∪{X [t]. so that we could also indicate (1.1). In other words. 5 As a preliminary exercise it is not diﬃcult to prove the following Lemma 1. under restriction (1. x0 ).3) that satisﬁes the boundary conditions x[t0 ] ∈ X 0 . t1 ] = T .
x0 ) be the crossection of the set of all isolated trajectories x[t] that satisfy the relation x[t] ∈ U(t. 0 Let XU [t] = XU (t. To show the opposite assume there exists a trajectory x∗ [t] = x∗ (t. namely x∗ [τ ] ∈ U ∗ (t. The next question is whether it would be possible to describe the evolution of sets X [t] in time t through some type of evolution equation with setvalued states X = X [t]. so that (1.1. x). x[t]) + f (t). (1.1. X 0 ) = and further X ∗ [t] = X ∗ (t. ˙ c c for a given multivalued map U(·. ·) ∈ UP .1). x). t ∈ T .1 With X ∗ [t0 ] = X [t0 ] = X 0 the following relation is true: X [t] ≡ X ∗ [t]. (1. t0 .4) this yields x∗ [t] ∈ X [t]. x0 ) : x0 ∈ X 0 }. x∗ [τ ]) + f (τ ) ⊆ P (τ ) + f (τ ) ˙ and due to (1.2.2. ·) = U ∗ (·.1. t0 . x[t0 ] = x0 . t0 . ·) ∈ UP }. x0 ∈ X . t0 .2. t0 . which satisﬁes the inclusion x∗ [t] ∈ X ∗ [t]. ˙ c for some U(·.2.5).2. Then one may want to know what is the relation between the tubes X ∗ [t] obtained for the closedloop system (1. x0 ).2. the solvability tube X [t] for system (1. X 0 ) : U(·. t ∈ T . Denote for a ﬁxed U of (1. x∗ [τ ]) + f (τ ). t ∈ T .3)? Theorem 1.6) and X [t] obtained for the openloop system (1. A particular element of UP is the setvalued map P(t) itself.1) with u = u(t.3) could be viewed upon as a particular case of equation (1. t0 . Therefore X [t] ⊆ X ∗ [t].2)). x) ≡ P(t).2) taken in the class UP of open loop controls u[t] is the same as the solvability tube c X ∗ [t] taken in the class UP of closedloop controls U(t.1.2.2. This conclusion is also true when the closed loop controls are selected among appropriate classes of singlevalued functions u = u(t.1. we come to the following questions.5). when U(t. X 0 ) = c {XU (t. t ∈ T.1. Then obviously x∗ [τ ] ∈ U ∗ (τ. To prove this assertion we observe that every singlevalued function u(t) can be treated as c an element of UP . {XU (t. 14 .1). x) ∈ P(t) that allow the existence and prolongation of solutions of (1.1 is such that with function f (t) given (there is no uncertainty in system (1.6): XU [t] = XU (t.2. The main conclusion given by Theorem 1. 0 (1. ·) ∈ UP .c where the class of feasible feedback strategies UP is as deﬁned in Section 1. t ∈ T .
1 The Hausdorﬀ semidistance h+ (X . Y) = 0 implies X ⊆ Y (and h− (X . Y) = h+ (Y.1.3 The Evolution Equation We shall now introduce an evolution equation with state space variable X ∈ convIRn . Y ∈ convIRn is introduced as h(X . Y) = 0 implies Y ⊆ X ). Y) = max{h+ (X .3. x y Similarly h− (X . Y). Y ∈ conv IRn is introduced as h+ (X . Deﬁniton 1. Y) = min{γ ≥ 0 : X ⊆ Y + γS} or equivalently h+ (X . (ii) h+ (X . Deﬁniton 1.3. Y) ≥ h+ (X . Y) between sets X . Y) = max min{(x − y. X ). Y). 15 1 . Y) between sets X . h− (X . Y)}.2 The Hausdorﬀ distance h(X . The question is therefore whether one could construct an evolution equation for describing X [t]. Z) + h+ (Z. x − y) 2 x ∈ X .1. The following properties are true for X . whose solution will be precisely the tube X [t] of Section 1. y ∈ Y}. where the second term in the right hand is the setvalued Lebesgue integral (“the Aumann integral”.1) X [t] = X 0 + t0 P(τ )dτ + t0 f (τ )dτ. [25]) for the function P.3. Y.2. Obviously t t (1. Z ∈ convIRn : (i) h+ (X .
x)  x ∈ X } which is a positively homogeneous convex function of l. Y) = 0 implies X = Y. for X . α1 + α2 = 1. ∀l ∈ IRn . X ∈ convIRn is equivalent to the inequality (l. where α1 > 0. Y) = max { ρ(lX ) − ρ(lY) : l ≤ 1} 16 . α2 ≥ 0. Direct calculation gives us the following formulae: h+ (X . h(X .3. a closed convex set X ∈ convIRn may be described by its support function ρ(lX ) = sup{(l. Y ∈ convIRn As it is well known. Y) = max { ρ(lX ) − ρ(lY) : l ≤ 1} (1.1 The inclusion x ∈ X . A well known property is given by Lemma 1. namely ρ(αlX ) = αρ(lX ) for α ≥ 0 and ρ(α.Obviously (iii).3. x) ≤ ρ(lX ).2) h(X . For X ∈ compIRn we have ρ(lX ) < ∞. ∀l ∈ IRn . l1 + α2 l2 X ) ≤ α1 ρ(l1 X ) + α2 ρ(l2 X ).
4 A multivalued function Z : T → convIRn is said to be a solution of (1.3. X [ti ]) ≤ ε. t) is absolutely continuous in t ∈ T uniformly in l ∈ S.5) ρ(lX [t]) = ρ(lX0 ) + t0 ρ(lP(τ ))dτ + t0 (l. together with (1.3.3).2 A function X : T → IRn is absolutely hcontinuous if the support function ρ(lX [t]) = f (l.3 A function X : T → convIRn is said to be absolutely hcontinuous on T if for any ε > 0 there exists a δ > 0 such that condition (ti − ti ) < δ i yields h(X [ti ]. Lemma 1.3) in the sense of the last deﬁnition.3) with “initial value” (1.3.3.3. Now we may consider the “equation” (1. X [t] + σP(t) + σf (t)) = 0 σ→0 Deﬁniton 1.3) for almost all t ∈ T . Let us see whether X [t] is a solution to (1.Deﬁniton 1.1) in terms of support functions. lim σ −1 h(X [t + σ].3.4) if it is absolutely hcontinuous and satisﬁes (1. l) ≤ 1}. 17 .3.4) X [t0 ] = X 0 . Rewriting (1. S = {l : (l.3.3.3. (1.3. f (τ ))dτ.3. i The deﬁnition of absolute h+ continuity is given by mere substitution of h by h+ in Deﬁnition 1.3. we come to t t (1.4).3.
t) = t [(ρ(lP(τ )) + (l. Taking into account the equality H(σ. t+σ R(l. σ → 0 for all t and t+σ (1. To calculate h(X [t + σ].7) σ −1 t ρ(lP(τ ))dτ → ρ(lP(t)). σ → 0 if f (t) is not continuous. [232].6) is still true. A similar remark is true for (1. t) = max{R(l. t) : and the relation 18 l = 1} .5) at ﬁrst we have R(l. due to (1. X [t] + σP(t) + σf (t)) = H(σ.3.7) with ρ(lP(t)) measurable in t. f (t)) In case of continuous f (t) and P(t) we further have t+σ (1. [19]. relation (1. t).3).3.6) σ −1 t f (τ )dτ → f (t).3.3.3. which are the “points of density” of f (t). σ.Here we made use of the fact that for a continuous map P : T → convIRn . t) = ρ(lX [t + σ]) − ρ(lX [t]) − σρ(lP(t)) − σ(l.3. being only measurable. σ. the following is true t t ρ(l t0 P(τ )dτ ) = t0 ρ(lP(τ ))dτ. (1. σ. f (t)). but now only for almost all of the values of t. f (τ ))]dτ − σρ(lP(t)) − σ(l.
1 The map X ∗ [t] of Section 1. 19 .1 The map X : T → convIRn .8) turn to be when X [t] = {x[t]} and P(t) = {p(t)} are singlevalued. h(x . Theorem 1.3.7). x ) = (x − x .2.3. x ) = d(x .3. (1.3.8) A solution X [t] to (1. (I + σA(t))X [t] + σP(t) + σf (t)) = 0. It is not uninteresting to write down a formal analogy of equation (1. (1. we observe lim σ −1 H(σ.3.3) substituted by (1.3. This proves the following assertion: Theorem 1. t) = 0. This is as follows: σ→0 lim σ −1 h(X [t + σ].9) x[t + σ] = (I + σA(t))x[t] + σp(t) + σf (t) + o(σ). X 0 ∈ convIRn is one that satisﬁes Deﬁnition 1.3.3. that follows from (1.7) with given initial state X [t0 ] = X 0 .1 implies the following Corollary 1.3. Let us now have a look at what would equation (1.3) when A(t) ≡ 0. Then.3).σ→0 lim σ −1 R(l.3.3.8). t) = 0 σ→0 for almost all t ∈ T . but with equation (1.3).1 is a solution to the evolution equation (1. being true for almost all t ∈ T . σ.3. x − x )1/2 and (1. is a solution to the evolution equation (1. clearly.4.2.3.6).3. and is uniform in l ∈ S.
3. that follows from (1.3. To conclude this paragraph we shall introduce another version of the evolution equation (1. This yields σ −1 (x[t + σ] − x[t]) = A(t)x[t] + p(t) + f (t) + σ −1 o(σ) for almost all t ∈ T or after a limit transition σ → 0: (1.3. Z[t] + σP(t) + σf (t)) = 0 . t ∈ T. where equation (1.3.3) may be integrated.5). From (1.3.3.3.3. l ∈ IRn x[t] = A(t)x[t] + p(t) + f (t). The integral form for its solution X [t] is given by (1.13) lim σ −1 h+ (Z[t + σ]. in presenting an ODE in the form (1.3).11) (1. Thus.3.9).4): Lemma 1.3. The derivative ∂ρ/∂t exists for almost all t due to the properties of the integrals in (1. (1.11).3. which is (1.8) is clearly a setvalued analogy of the ordinary diﬀerential equation (ODE) (1. There is no special point.3 The solution X [t] to equation (1.8). however.3).9).3. ∂t ρ(lX [t0 ]) = ρ(lX 0 ).10) for almost all t ∈ T .4) is unique.[25].3.12) for almost all t ∈ T and all l ∈ IRn . Equation (1.3.3. particularly because here we avoid the unpleasant operation of subtraction of sets or setvalued functions. This gives (1. (1. The support function for X [t] satisﬁes the partial diﬀerential equation ∂ρ (lX [t]) = ρ(lP(t)) + (l.3. f (t)).3.3.8) may be quite convenient.3. It is not so for setvalued maps. equation (1.3.10) that may also be presented in a form similar to (1.3. (1.3.1) and is described by a multivalued Lebesgue integral.3).where σ −1 o(σ) → 0 with σ → 0.12) its is not diﬃcult to observe that X [t] is the only solution to (1. namely by substituting the Hausdorﬀ distance h() for a semidistance h+ (). 20 σ→0 .5). x[t0 ] = x0 ˙ (1.
16).16) ρ(lZ[t0 ]) ≤ ρ(lX 0 ).3.12). (⊆ instead of =.4 The solutions Z[t].(1. It is not diﬃcult to observe that (1. This leads us to the assertion Lemma 1. We emphasize again that (1.3.3.3.13).3.3. in general.4) has a unique solution. A solution Z[t] to (1.3.3. (1.3.3).13). f (τ )))dτ = ρ(lX [t]).) which would be the case for (1. Particularly.3.3.13).4) and (1. rather than an equality.13). The initial condition also satisﬁes an inequality (1.12) satisﬁes an inclusion Z[t + σ] ⊆ Z[t] + σP(t) + σf (t) + o(σ)S. for all t ∈ T.3.15) and (1.3.4) respectively.3.13) has a nonunique solution. (1.3.14) Z[t0 ] ⊆ X 0 .3.15) for the solution Z[t].3). the solution to (1. This directly yields a partial diﬀerential inequality ∂ρ (lZ[t]) ≤ ρ(lP(t)) + (l.3. X [t] to the evolution equations (1.3.3) (1. is speciﬁed as in Deﬁnition 1. 21 .3. any single valued trajectory x(t) driven by a control u(t) ∈ P(t) with x0 ∈ X 0 will be one of these. Integrating (1. in view of (1. Here a solution Z[t] to (1. for all t ∈ T . (1.3) substituted by (1.3. f (t)) ∂t (1. while. satisfy the inclusion Z(t) ⊆ X [t]. l ∈ IRn .3. we come to t ρ(lZ[t]) ≤ ρ(lX [t0 ]) + t0 (ρ(lP(τ )) + (l.4) is nonunique.3.4 but with equation (1.
5 The maximal solution Z 0 [t] to (1. As an exercise the reader may prove the following Lemma 1. t1 ).13) is maximal if Z[t] ⊆ Z 0 [t]. for any solution Z[t] to (1.4).3.2) and a “terminal set” M ∈ convIRn .3.8). xτ ∈ W ∗ (τ.1) x(t) ∈ U(t.4 is not an isolated approach.4 The Problem of Control Synthesis: a Solution Through SetValued Techniques Consider system (1. but allows an equivalent representation in conventional terms terms of Dynamic Programming as applied in either a standard or a ”nondiﬀerentiable” version.1. We shall ﬁrst present a constructive technique for Control Synthesis based on setvalued calculus and further used here in the Sections devoted to ellipsoidalvalued dynamics. t1 .3).14) exists and coincides with the unique solution X [t] to (1. We will further use the evolution equations (1. ·) ∈ UP such that all the solutions to the diﬀerential inclusion (1. 1.3).3. (1. (1.3.4).3.4.1. (1.3.18) and their generalizations as an essential tool for describing the topics of this book. xτ = x[τ ]. ∀t ∈ T. U(·.Deﬁniton 1. Among the ﬁrst of these in the problem of Control Synthesis. A still further Section 1.3.3.1. 22 .1) .3.3. (1.1 The problem of control synthesis consists in specifying a solvability set c W ∗ (τ.4. x) + f (t) ˙ that start from any given position {τ.3. t1 . xτ }. M).5 A solution Z 0 [t] to (1. would reach the terminal set M at time t1 : x[t1 ] ∈ M.13). τ ∈ [t0 .3.5 is intended to indicate that the technique of Section 1.13) with the same initial condition (1. M) and a feedback control strategy u = U(t.1.(1. Deﬁniton 1. x).
Its solution is obviously t1 t1 (1. on the interval T . Taking W ∗ (t. Equation (1. but is treated in backward time.1 and the deﬁnition of W[t] is the following Theorem 1. M) = W(τ. τ ≤ t ≤ t1 .4. τ ≤ t ≤ t1 that steers the system from xτ to M due to a respective trajectory x[t]. The set W[τ ] = W(τ. and x[t1 ] ∈ M.1.3). M) = ∅. where the solvability set W ∗ (τ. namely. W(t1 ) = M is deﬁned as the “open loop” solvability tube for set M. M)) for all t0 ≤ τ ≤ t ≤ t1 . The respective map W[t]. so that x[τ ] = xτ . from t1 to τ .1).4.1. M) is the set of all states xτ ∈ IRn such that there exists a control u(t) ∈ P(t).4.1 The setvalued function W[t] satisﬁes the evolution equation (1. t1 . t1 . from set M. t1 . t1 .4. M) is nothing more than the attainability domain at instant τ for system (1. M) = W ∗ [τ ] is the ”largest” set of states from which the solution to the problem of control synthesis does exist at all. W[t] − σP(t) − σf (t)) = 0 σ→0 (1.3. To describe the tube W ∗ [t] we ﬁrst start from the following Deﬁniton 1. A direct consequence of Theorem 1. t ∈ T .2) is the same as (1.4) W[t] = M − t P(τ )dτ − t f (τ )dτ.2) lim σ −1 h(W[t − σ].4. t1 . also the same. t1 . The deﬁnition of the solution is naturally. t1 ].3) and the semigroup property W[t1 ] = M W(τ.2 The “openloop” solvability set W(τ. (the “solvability tube”) where W ∗ [t1 ] = M. t1 . we come to a setvalued map W ∗ [t] = W ∗ (t.The deﬁnition is nonredundant provided xτ ∈ W ∗ (τ. 23 .3. (1. t. M) for any instant t ∈ [t0 .2). t ∈ T . W(t.4. but calculated in backward time.(More precisely this will be speciﬁed below). t1 . M).
4. .4.3 The “closed loop” solvability set W ∗ (τ. Tube W ∗ [t] therefore satisﬁes the evolution equation (1. namely W[t] ≡ W ∗ [t].4.7) Z[t1 ] ⊆ M which is the same as (1. With A(t) ≡ 0 we have (1.Deﬁniton 1. t ∈ T . t1 . (1. From Theorem 1.4. or (1.3).3.7).5) σ→0 lim σ −1 h (W[t − σ].14) but taken in backward time.5). The deﬁnition of its solution and maximal solution are analogies of those given in the previous section for direct (”forward”) time. (1.2).4.4.6).4. x[τ ] = xτ . By analogy with Lemma 1.4. (1. 24 . M) is the set of all states c xτ ∈ IRn such that there exists a control strategy u = U(t.1 we come to Lemma 1.3) are unique and given by convex compactvalued maps.4.2 With W[t1 ] = M. U(·.4.4.4.13). W ∗ [t1 ] = M deﬁnes the “closedloop” solvability tube W[·] for set M.3. The respective map W ∗ [t] = W ∗ (t. t ∈ T .2). to end in set M : x[t1 ] ∈ M. the map W[t].3. M). (I − σA(t)) W[t] − σP(t) − σf (t)) ≡ 0. Z[t] − σP(t) − σf (t)) = 0 σ→0 (1. x). is the maximal solution to (1.4.5 we also come to Lemma 1. which are W[t] and W ∗ [t]. t ∈ T.1 With W[t1 ] = W ∗ [t1 ] = M the open loop and closed loop solvability tubes. (1.4.4.1) that starts at τ . do coincide. ·) ∈ UP that ensures every trajectory x[t] of the diﬀerential inclusion (1. Substituting Hausdorﬀ distance h() for semidistance h+ () we come to the equation (1.2. (1.3).6) lim σ −1 h+ (Z[t − σ]. t1 . The solutions to (1.
t1 . namely.2. x) dt along the trajectories of system (1. as we shall see. and V [τ. Clearly. W[τ ]). x) is the tube W[t]. M) c is necessary and suﬃcient for the existence of a synthesizing strategy U(·. Assume x ∈ IRn and set W[τ ] to be given. x) = d2 [τ.2.This is a consequence of the deﬁnition of the solvability sets and it is certainly important to emphasize that the inclusion xτ ∈ W(τ.1) . x]. x] = h+ (x. We may now investigate the derivative d V (t.1). An essential element in constructing the strategy U = U(t. x) = arg min{ d V (t. u ∈ P(t)}. where d[τ. V [τ. h+ (x. The control set U 0 (t. U 0 (t. x) ≤ 0} is the level set for V (t. ·) ∈ UP . (1. W[τ ]) = min{ x − w  w ∈ W[τ ]}. x) will then consist. x)). x) dt 25 implies x ∈ W[τ ]. of all the values u(t) ∈ P(t) that minimize this derivative. x] = 0 implies x ∈ W[τ ]. x] > 0. Let us introduce a ”synthesizing function” V (t. (One may observe that W[τ ] = {x : V (t.
1) where (1. l ≤ 1} = and where l0 = 0.4.2. it is not diﬃcult to prove the following property. l0 = 1. u(t)) + ρ(−l0 P(t)). dt 26 (1. (1.1) d d[t. x) − ρ(lW[τ ]) : = (l0 . We further need the derivative dd[t. x] dt ). A direct diﬀerentiation yields (1. x) − ρ(l0 W[t]). x(t) ˙ dt ∂t ∂x ∂ = (l0 . x) dt = 2d[t.2.2.11) . Since W[t] is absolutely continuous.4.4.4. x(t)] = (l0 . x] > 0.9) d[τ. W[t]) > 0. Then the function d∗ (t) is absolutely continuous on the same interval. which is x(t) = u(t) + f (t). f (t)). x(t)) − ˙ ρ(l0 W[t]) ∂t = (l0 . and therefore d d[t. x])/dt when d[t. x] = 0. is a unique maximizer for d[τ.10) d ∂ ∂ d[t.1). Lemma 1. ˙ For this we obtain: (1. u(t) + f (t)) + ρ(−l0 P(t)) − (l0 . due to the system (1. We will always choose the maximizer to be l0 = {0} if d[τ. x] = d[t.8) d V (t. x]( (1. x] > 0. x].Let us specify this in detail. x] + d[t. x] = max{(l.3 Let x∗ [t] be an absolutely continuous function on an interval e where d∗ (t) = h+ (x∗ [t].4.
9). l0 ) = + dt ∂t where l0 = arg max {H(t. x) will have to be speciﬁed both in the domain {x ∈ W[t]} (or V (t. x(t)) = max{H(t. x) > 0.3.2. x) > 0) and in {x ∈ W[t]} (or V (t.1) that indicates the system along whose solutions we calculate the derivative).4.x . Then U 0 (t. Remark 1.10) for d[t. x. The strategy U 0 (t. [261]. we observe. x). ˙ ∂x l = 1}. x). 27 . following [86].15) turns to be an equality. l0 ) . dt (We further omit the index (1. l) : with unique maximizer l0 we have dh(t. x.4.Here we have used the formula t1 t1 ρ(lW[t]) = ρ(lM) + t ρ(−lP(τ ))dτ − t (l. x) will be deﬁned as U 0 (t. we should avoid diﬀerentiation in l0 . x) = arg min d d[t.4) and the fact that in calculating the derivative (1. x] which is the maximum over l in (1. that for a diﬀerentiable function of type h(t. x(t). x) = 0).1 The direct calculation of ∂ρ(lW[t])/∂t introduced in (1. x) u(t) ∈ P(t) . We now proceed with specifying the feedback strategy U 0 (t. x] u(t) ∈ P(t) dt = = arg min d V (t.4. l) : l = 1}. Assume V (t. f (τ ))dτ that follows from (1.9) also indicates that with Z(t) = W[t] the inequality (1. Since l0 depends on t and x we further use the notation l0 = l0 (t. x) ∂H(t. ∂H(t.4. [265].4. Indeed. x.
11).4. x). x).2 The strategy u(t) = U 0 (t.11). what is the same.4.8). (1.5 For any position {t.1) .(1. x) = 0 we have l0 = 0 and therefore U 0 (t.4. x) = arg max{(−l0 (t.4. dt (1.10) yield the following assertion Lemma 1.4.4.14) where U 0 (t. u) u(t) ∈ P(t)} or. we come to Lemma 1. x).4. 28 d V (t. that with V (t. x) deﬁned by equation (1. x) dt ≤0 . x}the derivative (1.12). dt d d[t.4.12) U 0 (t.11). x) The latter relations allow to prove Theorem 1.(1. x) = arg min{(l0 (t. (1.1. Combining this with (1.4. x] ≥ 0 for any u ∈ P. x] = 0 for u(t) ∈ U 0 (t. this turns into U 0 (t. Relations (1. (One should observe. does solve the problem of control synthesis speciﬁed in Deﬁnition 1.4.4.9). (1. x) = P(t)). u) u(t) ∈ P(t)}.Due to (1.4.4.4 With d[t. x] > 0 the derivative d d[t.4.4.15) provided u ∈ U 0 (t.13) (1.11).2. x) is deﬁned by relation (1.4.
satisﬁes the relation X [tU 0 ] ⊆ W[t].4. t0 . x[t0 ]) ≤ 0 for any t ∈ (t0 . in general. We shall prove that x[t] ∈ W[t]. t1 ]. (U = U 0 ) . x[τ ]) dτ = V (t. this yields V (t. t0 . The proof of Theorem 1.4. and therefore X [t1 U 0 ] ⊆ M. with x0 ∈ W[t0 ] or equivalently. Assume x[t] = x(t. with V (t0 .1) generated by strategy U(t. that V (t. which. X 0 U 0 ) to the diﬀerential inclusion (1. U = U 0 . x[t0 ]) ≤ 0 dτ due to (1. therefore X [t1 U 0 ] ⊆ M. of isolated trajectories x0 [t] = x0 (t. t ∈ T . t ∈ T .2 is based on the following Lemma 1.Since x[t0 ] ∈ W[t0 ] .4. x0 ) is a trajectory of inclusion (1.6 from which Theorem 1. for any value of t ∈ (t0 .4. x0 ) to (1.6 The tube X [tU 0 ]. U 0 ] = X (t.1 follows directly. Corollary 1. x). t1 ] and thus proves Lemma 1. t0 . generates a tube X [t.4. t ∈ T. t0 .4. x). X [t0 U 0 ] = x0 . t ∈ T.1 With X 0 = X (t0 U 0 ) ⊆ W[t0 ] the respective tube of solutions X [tU 0 ] = X (t. or equivalently.Assume that x0 ∈ W[t0 ] and that the inclusion (1. The same property is true if x0 is substituted by a set X 0 ∈ W[t0 ]. x0 ∈ W[t0 ].1). x[t]) ≤ V (t0 .4.4. x) = U 0 (t. x0 U 0 ) = {x0 (t. x[t]) ≤ 0.15).4. Then . t0 .1). for all t ∈ T .1) is run by strategy u(t) = U 0 (t. Proof. x0 ) ≤ 0 and x[t] ∈ X [tU 0 ]. x[t]) − V (t0 .4. t0 dV (τ. we observe that the integral t t ∈ T. x0 }. satisﬁes the inclusion X [tU 0 ] ⊆ W[t]. 29 .
∀l ∈ IRn . ∀l ∈ IRn . t ∈ Tτ . (1. ρ(lP(t)) − (q.19) (l0 . Assume f (t. for the whole trajectory tube X [tU 0 ].4. U = U 0 . A comparison of (1. l) = ρ(lP(t)). It is now obvious that the largest strongly invariant set for time τ . U = U 0 that evolves from set Xτ remains within W[t].1). We recall that a subdiﬀerential ∂l f (t. t1 ] generated by system (1. l) convex in l. l − l0 ).4. q) ≥ 0. is the set of all vectors q such that (1. ∀l ∈ IRn .17) f (t. that (1. l − l0 )..18). a vector q ∈ ∂l f (t. whence q ∈ P(t). ρ(l0 P(t)) − (l0 . q) ≥ 0. with l = 0. According to the terminology of . The tube X [tU 0 ] therefore satisﬁes (1. at point l = l0 ) of a function f (t.4. for time τ . relative to W[t]. l0 ) (in the variable l.4. [265]. is W[τ ] itself.4. The feedback strategy U 0 (t. [17] Xτ is strongly invariant relative to tube W[t]. we come to (1.19) and a substitution of q for u yields 30 . l) ≥ (l0 P(t)) − (l0 . [261]. From here it follows (taking l = 2l0 ). q) ≥ ρ(l0 P(t)).16) as a state constraint. On the other hand. l0 ) ≥ (q. Tτ = [τ. [267]. x) may be rewritten in terms of the notion of subdiﬀerential.1).e.18) and therefore. t ∈ Tτ .We thus observe that if for an instant τ ∈ T the inclusion Xτ = X [τ U 0 ] ⊆ W[τ ] is true. t ∈ Tτ .4. l0 ) if and only if ρ(lP(t)) − ρ(l0 P(t)) ≥ (q. The latter means that every trajectory of the inclusion (1. Then.i. then (1. l) − f (t.4.4. due to the deﬁnition.16) X [tU 0 ] ⊆ W[t] will hold for all t ≥ τ .4.
x) = ∂l f (t.1. one may always take A(t) ≡ 0 as shown in Section 1. as introduced in the above and further used in the sequel. l) = ρ(lP(t)) and l0 (t. the respective subdiﬀerential ∂l f (t. l Summarizing the reasoning of the above. −l0 (t. a ”conventional” way of looking at the problems under discussion. and therefore U 0 (t. 1. that the strategy U 0 (t. x) is the maximizer for (1. x) = ∂l f (t. l) is given by: ∂l f (t. at ﬁrst glance .7 With f (t.2. l0 ) = P(t). as an exercise. in a very concise form.1) to be given together with a terminal set M ∈ convIRn .4. Let us indicate a cost criterion I(τ.3 The feedback strategy U 0 (t. c One may check. x) that solves the problem of control synthesis may be deﬁned as (1. presuming that our objective will be to ﬁnd 31 . −l0 (t. It may be demonstrated .Lemma 1. Clearly. x)). to be somewhat unusual. x) for the problem of control synthesis. Assume a position {τ.4. We therefore feel obliged to indicate. for l0 = 0 we have ∂l f (t. u) = ρ(l0 P(t))}.4. l0 ) = {u ∈ IRn  (l0 . where l0 (t. x)). may seem.9).4. where f (t. we conclude the following. l) = ρ(lP(t)). however. x) ∈ UP and belongs to the class of feasible strategies introduced in Section 1. that they are quite in line with the wellknown fundamentals of control theory. x) is the maximizer for problem (1. Theorem 1.2).9).5 Control Synthesis Through Dynamic Programming Techniques For a controltheorist experienced in the conventional methods of this theory the geometrical techniques of setvalued calculus.20) U 0 (t.4. x} due to system (1.(Although matrix A(t) ≡ 0 is present in the ﬁrst part of this Section.
1. (The respective theory may be found in [109]. x) ∈ UP that minimizes this criterion.5.5. x} will be further referred to as the value function V(τ. x) = I 0 (τ. It is obvious that V(τ. x∗ ) The optimal value c I 0 (τ. (1. [53]). This is (1.1) along the c trajectories of system (1. M]). x) ≤ 0}.1) I(τ. + where x[t1 ] = x(t1 . x) ∈ UP .5. x) = min {I(τ.3).5. x) u(·. x) by formally writing down the H − J − B (HamiltonJacobiBellman) equation for the problem of minimizing cost criterion (1. Let us now calculate function V(τ. let us assume (1.4) ∂V(t. x∗ ) = h2 (x[t1 ].2) ∂V(t.3) or. A(t)x + u + f (t) u ∈ P(t) ∂x = 0 with same boundary condition (1. x) > 0 if x[t1 ] ∈ M Therefore. More speciﬁcally. x) = h2 (x. the solvability domain W [τ ] of Section 1. x) .5.c an ”optimal” control strategy u = u(t. ·) ∈ UP } when taken for any position {τ. 32 . more precisely. x)) P(t) + f (t) ∂x = 0 V(t1 .1) with u = u(t. τ. x). if x[t1 ] ∈ M and V(τ.4 is actually the level set W [τ ] = {x : V(τ. x) + min ∂t with boundary condition (1. M) + ∂V(t. x ∂x − ρ ∂(−V(t. x) .5. x) = 0 . x) + ∂t A (t) ∂V(t.
5.5) Our problem is to ﬁnd I 0 (τ. l). the function φ(x) = h2 (x.We have min φ(x(t1 )) = min max{(l.We have to check. with φ(l) → ∞ as l → ∞. 4 (1. whether these formal operations are justiﬁed. Obviously. This indicates that the operations of min and max are interchangeable ( [101] ). x − q)x ∈ IRn }q ∈ M} = max{(l. l)q ∈ M} 4 which is φ∗ (l) = ρ(lM) + 1 (l. x(t1 )) − φ∗ (l)} = u(·) u(·) l = max{min{(l.1. t1 . (1. x(t1 )) − φ∗ (l)}} l u(·) The function in the brackets in the righthand side is linear in u(·) and concave in l . x − q)q ∈ M} + has a conjugate φ∗ (l) = max{(l. s(t1 ) = l. through the technique of convex analysis. τ ≤ t ≤ t1 } over the trajectories of system (1. We shall do that by calculating the value I(τ. x) − (x − q. x}.1).2) with given initial position {τ. ˙ 33 . x) directly. x) − φ(x)x ∈ IRn } = 1 = max{max{(l. however. t ≤ t1 . l) to be the solution of the adjoint equation s = −sA(t) . Denoting s(t.2. q) + (l. x) = min{φ(x(t1 ))u(t) ∈ P(t). M) = min{(x − q.
t1 . Φ(τk . l0 (τk . x. x(k) ) = l∗ Taking into account the obvious inequalities.6) is therefore attained at a single vector l0 = l0 (τ. we may rewrite t1 (l. Lemma 1. x(k) )) 34 . x(k) } → {τ. The maximum in (1. due to the properties of Φ(τ.7) . k → ∞. x(k) )} will be equibounded. Then. x.5. x. whatever is the position {τ. x}.5.6). Φ(τ. x) of (1. l)x + τ s(t. S(t. l) in l. l0 (τ. due to the quadratic term). the respective sequence of maximisers {l0 (τk .(1.1 The maximizer l0 (τ. l) = s(τ. t1 . t1 )u(t) + f (t)) dt = t1 = s(τ.1.5. l0 (τk . t1 . The function Φ(τ. x. x). so that there exists a limit k→∞ lim l0 (τk . x(k) )) ≥ Φ(τk .6) where I 0 (τ. t1 . x)) Φ(τ. x.( s is a vectorrow) and using the notations of (1. S(τ. even strictly concave. l0 (τ.5. x}. Hence (1. l)l ∈ IRn }.1. l)P(t) + f (t)) dt − φ∗ (l). x. Indeed. x)) ≥ Φ(τ. l)x − t1 − τ ρ(−s(t. x. suppose the sequence {τk . x(t1 )) = (l.6) is continuous in {τ. t1 )x) + τ (l. l) is concave in l (moreover. x(k) . l)(u(t) + f (t)) dt. x}. x) = max{Φ(τ.
and passing to the limits (k → ∞). x) in x now gives (1.5. x)) ≥ Φ(τ. x.5. l)l ∈ IRn } = {0}.5. t1 .(1. x) this yields l∗ = l0 (τ. x) = max{(l. l0 (t.3) . so that Φ(τ. l)l} = Φ(τ. the respective formula is as follows ∂ ∂ max{Φ(τ. l0 (τ. M). t1 . t1 .2 For any x ∈ X 0 [τ ] we have l0 (τ. l) ). x)). (1. l)P(t) + f (t)) − s(τ.5.5. In order to check the boundary condition (1. l0 (τ. x. Φ(τ. x) − ρ(lM) − (l. x. 4 1 ψ(l) = ρ(lM) + (l. t1 . x))P(t)) − s(t. we have. x) = arg max{Φ(τ. x) = s(τ. x.5. x) and the Lemma is thus proved. x. Due to the unicity of l0 (τ. ∂x ∂x l=l0 (τ.). x) = ρ(−s(τ. l∗ ) ≥ Φ(τ.6) (1. in view of (1. ∂τ Taking V(t. l). x) is unique. x) and substituting into (1.9) ρ(−s(t.8).8) ∂I 0 (τ. l0 (t.5. x. l) = φ∗ (l) 4 ∗ ∗ ∗ due to (1. x. we observe from formula (1.5. x)).10) where 1 I 0 (t1 . Φ(τ.5. l∗ ) = Φ(τ. l) in all of its variables) . x. x) = I 0 (t. x. l∗ ). t1 . x) ≤ 0} Lemma 1.5.2). l0 (τ. Hence ψ (x) = (φ ) (x) = φ(x) = h2 (x. x))(A(t)x + f (t))+ min{s(t.5. l)l ∈ IRn } = ψ ∗ (x).x) Similarly (1. A direct diﬀerentiation of I 0 (τ. x)). x. + Therefore.7) ∂I 0 (τ. x. Denote X 0 [τ ] = {x : I 0 (τ. This follows from the explicit expression for Φ(τ.7). ∂x (Recall that since l0 (τ. l0 (t. the following assertion turns to be true 35 . x))(A(t)x + u + f (t))u ∈ P(t)} = 0. we observe (due to the continuity of Φ(τ. l)A(τ )x. l0 (τ.5. t1 .
(1. Theorem 1. x. l) = (l. x] 36 . x) given by formula (1. x) ≡ P(t).6) satisﬁes the Dynamic Programming (HJB) equation (1. M)u(t) ∈ P (t). x] = max{Ψ(τ. x] and therefore leads us to Lemma 1.12)..11) u0 (t.5. .4. with V(t.5. x) is the value function I 0 (t. x] = d[τ.5.6). Following the scheme for calculating (1. where V(t. x) − τ t1 ρ(−lP(t) + f (t))dt − ρ(lM) = = (l. x) is given by by relations (1.12) u0 (t. x) deﬁned in Section 1.5. x) = arg max ∂(−V(t. x) in (1. where Ψ(τ. x) − ρ(lW[t]) and W[t] is deﬁned by (1. l) l ≤ 1}. ∂x Particularly. x) = V(τ.5. u u ∈ P(t) . x) = d2 [τ. we have r[τ. (1.5.(1. x)/∂x) plays the role of vector l0 (t. τ ≤ t ≤ t1 } assuming A(t) ≡ 0. x.1 The value function V(t.5.5. The control u0 (t.5. This yields r[τ.5. x) = 0.5. while (−∂V(t. (This reﬂects that ∂V(t.Theorem 1. this gives (1.3). x)/∂t = 0 if V(t. x) is thus similar to U 0 (t. x) = I 0 (t.4.4. x) is then formally determined from (1.2).4).12).2) ((1.5.9) as (1.4.20).4)) with boundary condition (1.2.5. Let us now calculate the value r[τ.5.11). x) = 0 .in view of Lemma 1. We therefore come to an equivalent of Theorem 1.2).2 The solution strategy u0 (t. x)) .3 With A(t) ≡ 0 the value function I 0 (τ.5. x). x] = min{h+ (x[t1 ]. The respective control u = u(t.4.
while the antigradient (−∂V/∂x) is collinear with l0 (t. of course. In the more general case of nonlinear systems and an arbitrary terminal cost φ(x) the main inconvenience is that there may be no diﬀerentiable function V(t. aimed particularly at the applicability of ellipsoidal calculus .4. are those formulated in Section 1. Let us write down equation (1.(1. This integration will be an essential tool for the treatment of those nonlinear problems for which the techniques of this book cease to be eﬀective. then the partial derivatives of V may not be continuous or may not even exist at all.10).12).5. particularly.11). [82]. (W 0 [t] ≡ W[t](!)). as we have seen).1.5. [289]. x) that solves the DP ( HJB ) equation (a nonlinear analogy of equation (1. with boundary condition (1.5.5.5. M ). Nevertheless. perhaps) . it may be interpreted as a viscosity solution . of course.2). the value functions will turn to be convex in x. or its equivalent . calculated by integrating equation (1.One just has to recognise that d2 [τ. under condition A(t) ≡ 0 ( which does not imply any loss of generality.4) or its analogies ( in a generalized sense. x) in (1. x) through DP techniques. x) > 0}. P(t) are nondegenerate ellipsoids.4) but through direct constructive techniques which .5. whereas if we look for nondiﬀerentiable functions. x)/∂x. in the problems of this book.5.and M. shall be somewhat more complicated in the case of uncertain systems and state constraints.4 through setvalued techniques is precisely the one derived in this Section through Dynamic Programming. Particularly. one may observe that for deﬁning u0 (t. M − 1(x − m)) ≤ 1} = E(m. the elements V(t. one needs to know the following elements: • the level sets W 0 [t] = {x : V(t.namely. [109]. Looking at the solution (1. Example 1. For the problems treated in this book these elements may be determined without integrating equation (1. ∂V(t. the solution given in Section 1. x) ≤ 0} • the partial derivatives ∂V/∂x in the domain {x : V(t.3) for the particular case when the system is autonomous . may be.4. 37 .Thus.5. We shall further continue to indicate the Dynamic Programming interpretations of the outcoming relations which. x). x) and therefore that the level set W 0 [t] is the solvability set W[t] . M = {x : (x − m.the minmax solution .2)). The solution to the DP equation should then be interpreted in some generalized sense. A ≡ 0. Needless to say. x] = V(τ. They will be directionally diﬀerentiable and therefore allow a more or less clear propagation of the notions of Dynamic Programming (DP).
5.5. x) + . Relation (1. t1 ]. P − 1(u − p)) ≤ 1} = E(p.14) = 0 V(t. x) −1 ∂V(t.10) by direct calculation. Exercise 1. x) = h2 (x. 1 38 . M (x − m))(1 − (x − m. x) = (x − m. x. M. x) + ∂t ∂V(t. We have ∂V(t. M ). Taking system (1.15) where IT = (x(t1 ) − m. I1 }. x}. P ). M (x(t1 ) − m)).1: with A(t) ≡ 0 indicate the cost criterion I ∗ for which the value function V ∗ would be V ∗ (t. x ∈ E(m. M (x − m))− 2 )2 .5. x ∈ E(m.5.1) and position {τ.5. f (t) + p ∂x 1 2 (1. t ∈ [τ. t1 ]. u(·)) = max{IT .13) + ∂V(t.5.P ∂x ∂x with (1. P > 0. Ψ(τ. x = x(τ ). M ). I1 = esssupt (u(t) − p(t). x) = 0 . solve the following problem: minimize the functional (1.14) follows from (1. W[t]).2. + where W[t] is the solvability set of Section 1. P (t)(u(t) − p(t))). x) . τ ∈ [t0 .P(t) = {u : (u − p. Let us now indicate another relation for the solvability set W [τ ].4. V(t.
39 τ t1 t1 τ (l. First – consider set Wµ [τ ] = m + µE(0.5. M ) − t1 τ (p(t) + f (t) + µE(0.17) V (τ. P (t)). We have x ∈ Wµ [τ ] iﬀ (l. x). or otherwise (l. (l. M ). Second – for a given position {τ.Introduce value function V (τ. clearly (1. x) ≤ 1}. x) ≤ ρ(lWµ [τ ]). We shall now indicate an explicit relation for V (τ. M l)1/2 + This immediately yields Lemma 1. x∗ (t1 ) = m.4) with M = m+µE(0. x.5.16) W [τ ] = {x : V (τ. M l)1/2 + where x∗ (t) is the solution to system ˙ x∗ = p(t) + f (t).4. x − x∗ (τ ))H −1 (t. This set is similar to set W[τ ] of (1. P (t)))dt. x∗ (τ )) + µ (l.4 The value function (1. Its support function ρ(lWµ [τ ]) = (l. l) = (l. x) = min{Ψ(τ. . ∀l. u(·))u(·)} Then. where H(t. ∀l ∈ IRn . P(t) = p(t)+µE(0. x − x∗ (τ ))H −1 (t. P (t)l)1/2 dt . l)l ∈ IRn }.5. P (t)l)1/2 dt. x) = max{(l. x} ﬁnd smallest µ for which x ∈ Wµ [τ ]. l) ≤ µ.
rather than solving it explicitly. Check whether this equation does have a classical solution.2) and (1.4).6.(We oﬀer the reader to specify the formulation of such a problem). ˙ 0 that reﬂects the availability of only openloop controls u(·) ∈ UP and also has an unknown disturbance f (t) subject to a given constraint (1. (1.1) f (t) ∈ Q(t). Our next subject will be the issue of uncertainty in the knowledge of the system inputs.6. such that X [t] would be the level set for the respective value function. it is possible to formulate an optimization problem . Exercise 1. x) be considered a solution to this equation? Would it be a viscosity solution. x. we shall discuss this issue in conjunction with ellipsoidal techniques. the maximum is attained). u(·)).1). since W[t] is similar to the attainability domain X [t]. Indeed.2. in Part IV.6 Uncertain Systems.2.4.4. x(t0 ) ∈ X 0 . Try to write a formal HJB equation for cost criterion Ψ(τ. t ∈ T.1. Attainability Under Uncertainty We are returning to systems (1.1 . 40 .1.5.6.( check that here. In what sense could function V (τ. we shall indicate an approach for approximating the solution of the HJB equation of Exercise 1. but now the disturbance (or “forcing term”) f (t) will be taken to be unknown but bounded.14. if the latter would be calculated in backward time. namely. the information on f (t) will be restricted to the inclusion (1.1).1. Naturally. We therefore come to the following systems: (i) the linear diﬀerential equation (1.5. in Sections 4. [82]. continuous in t. the description of attainability domains also allows an application of DP. where Q(t) is a given multivalued map Q : T → convIRn .3.2) x = u(t) + f (t). 1. with M > 0. Later. [109] ? Later. sections 4.
What would be the notion of attainability now. that the input f (t) is unknown? It is quite obvious that the respective deﬁnitions for both openloop and closed loop controls could be presented in several ways. u∗ (·) ∈ UP . ˙ X (t0 ) = X 0 . t0 .6. such that for any f ∗ (·) ∈ Q(·) there exists a pair {x0∗ . f ).6. u∗ (·)}. X 0 ) of Section 1. x∗ [t1 ] = x∗ . f ) is the crosssection at instant t of the solution tube to the linearconvex diﬀerential inclusion x ∈ P(t) + f (t).6. Deﬁniton 1. In other terms X [t] = X (t. x) + f (t).1 may then be presented as (1. t ∈ T.4) X [t1 ] = X (t1 . to the notation of the attainability domain X [t] = X (t.(ii) the nonlinear diﬀerential inclusion (1. ∗ that generates an isolated trajectory x [t] of system x = u∗ + f ∗ . t0 .6. t0 . ·) ∈ UP and f (t) is unknown. ˙ c where U(·.1 An ”openloop” domain of attainability under counteraction for system (1. Let us further add the symbol f (·).1). t0 . (1. but bounded by constraint (1.6.2 emphasizing its dependence on a given input f (·). x0 ∈ X 0 .3) x ∈ U(t. X 0 .6. X 0 .1) from set X 0 = X [t0 ] at time t1 is deﬁned as the set X (t1 . X 0 ) = 41 . We shall start with the following open loop construction that will be used in the sequel.6. The set X[t1 ] of Deﬁnition 1. namely X [t] = X (t. t0 . This reﬂects the availability of closedloop (feedback) controls. ˙ that satisﬁes the boundary conditions x∗ [t0 ] ∈ X 0 . x∗ [t0 ] = x0∗ . X 0 ) = X [τ ] 0 of all states x∗ .2).
4). We invite the reader to investigate this issue.1. may be deﬁned by introducing operations of either intersection ∩ or union ∪ over x0 . f (τ ))dτ : f (·) ∈ Q(·) (1. Let us see. The function h(t. 42 . namely h(t. whether it is possible to derive an evolution equation for the tube X [t].4) and taking X [t]. for any t ∈ T . t0 .= f x0 X (t1 . in an order other than (1. {x0 }. f )) ∀l ∈ IRn .6. f ) : f (·) ∈ Q(·) . t0 . f (·) ∈ Q(·) or X [t1 ] = X (t1 . αl) = αh(t.1 Other types of attainability domains. Obviously x∗ ∈ X [t] if and only if (l. or (l. x∗ ) ≤ ρ(lX (t. x∗ ) ≤ h(t.6. X 0 .6. l).6. Here direct calculation gives t t ∀f (·) ∈ Q(·) h(t. f ) f (·) ∈ Q(·) Remark 1.6.5) h(t. we come to the open loop solution tube (under counteraction). t0 . Returning to (1. l) = inf ρ(lX (t. X 0 . f ) x0 ∈ X 0 . than in Deﬁnition 1. l) = ρ(lX ) + t0 0 ρ(lP(τ ))dτ + inf t0 t (l. X 0 . l). t0 . for all α > 0. l) = ρ(lX ) + t0 0 (ρ(lP(τ )) − ρ(−lQ(τ )))dτ. l) is positively homogeneous in l. f (·). with h(t.
l) > −∞.6. A standard result in convex analysis consists in the following.6.4. ˙ ρ(lR(τ )) = ρ(lP(τ )−(−Q(τ ))) = ρ(lP(τ )) − ρ(lQ(τ )) = ρ(lP(τ )) − ρ(lQ(τ )) and R(τ ) = ∅.6.1 the function g(τ.1 Under Assumption 1.1 The function g(τ. 43 X [t0 ] = X 0 .6) lim σ −1 h(X [t + σ].1 Under Assumption 1. Under Assumption 1.6.6. ˙ Here P −Q stands for the geometrical (“Minkowski”) diﬀerence between P and Q: ˙ P −Q = {x : x + Q ⊆ P}.1 the function g(τ.1 we come to Theorem 1. namely.6.1 the tube X [t] satisﬁes the following evolution equation (1. The setvalued map R(τ ) is continuous. ∀l.Assumption 1. [265]. l) is convex in l. l) = ρ(lP(τ )) − ρ(−lQ(τ )) is convex in l (and ﬁnite valued: g(τ. X [t] + σR(t)) = 0. It is also positively homogeneous in l and therefore turns to be the support function for a certain set R(τ ) that evolves continuously in time. σ→0 . Following the proof of Theorem 1. l) = ρ(lP(τ )) − ρ(−lQ(τ )) is the support function for set ˙ R(τ ) = P(τ )−(−Q(τ )). ∀τ ∈ T ).6. Lemma 1.
X (τ. t0 . ∀l ∈ IRn .2 The geometrical diﬀerence of the following two integrals is not empty: t t ˙ P(τ )dτ − t0 t0 (−Q(τ ))dτ = ∅. l) − (l. This is given by Lemma 1.2 The set X [t] is nonvoid (X [t] = ∅) if and only if there exists a vector c ∈ IRn such that the function h(t. 44 . This case is known as the matching condition for the constraints on u. t0 . Assumption 1.1).6.2 A typical example for Assumption 1. X 0 )) = X (t.6. X 0 ) to equation (1. Once X 0 = {x0 } is a singleton.A consequence of Theorem 1. for t0 ≤ τ ≤ t ≤ t1 . f in equation (1.6. c ∈ IRn . so that ˙ P −(−Q) = (1 − α)P − c. X 0 ). c) ≥ 0.6.2. the function h(t. Remark 1.6. t0 .1 is Corollary 1.6.1 is when 0 ∈ P.1 The solution X [t] = X (t.6. It is not diﬃcult to formulate the necessary and suﬃcient condition for X [t] to be nonvoid. 0 < α < 1. l) satisﬁes the condition of Lemma 1. τ.6) satisﬁes the semigroup property X (t. The proof of this assertion is a standard exercise in convex analysis. −Q = αP + c.6.2 if the following assumption is fulﬁlled.
[265]. l where h∗∗ (t. (1. Then x∗ ∈ X [t] if and only if (l. This is given by the relation (col h)(t. since t t t ˙ P(τ )dτ ) − t0 t0 Q(τ )dτ ⊇ t0 ˙ (P(τ ) − Q(τ )) dτ (prove this inclusion). l) in the variable l.6.3 The results of convex analysis imply a formal calculation for determining the closed convex hull (col h)(t. where (col h) (t. l) = h∗∗ (t.6. p) − k(l) ∈ IRn }. Recall that l k ∗∗ (l) = (k ∗ )∗ (l). where k ∗ (p) = sup{(l.Thus the set X [t] = ∅ if and only if (col h)(t. [100].6. l).6.8) X [t] = (X [t0 ] + t0 ˙ P(τ )dτ ) − t0 Q(τ )dτ It follows under Assumption 1. l) is the closed convex hull of h(t. 45 .1. Remark 1. l) in the variable l. x∗ ) ≤ (col h)(t. that the set X [t] = ∅ for any convex compact set X [t0 ] and any t ∈ T . for all l. l).7) Therefore t t ∀l ∈ IRn (col h)(t. l) = −∞. l) = ρ(lX [t]). l). or (1. l) is the second conjugate of h(t.
In other words. f ) = c X (t. t0 . t0 . c Given a strategy u = U ∗ (t.6. such that for any f (·) ∈ Q(·) there exists a vector x0∗ ∈ X 0 . t0 . X 0 ) can be described as (1. f U ∗ ) : x0 ∈ X 0 Deﬁniton 1.8) may not satisfy the semigroup property. U ∗ }. x) ∈ UP .3).6. x0 . X(t1 . t0 .9) x ∈ U ∗ (t. X 0 U) = 46 . f U ∗ ). x0 .6.4 It is not diﬃcult to check that the tube X [t] of (1. and f (·) ∈ Q(t). t0 . X 0 . so that X (t. f U) : U ∈ UP X (t. generates a solution tube X (t1 . f U ∗ ). X 0 . Let us now deﬁne an attainability domain under counteraction in the class of feedback (closedloop) control strategies.2) that satisﬁes the boundary condition x∗ ∈ X (t1 . f U ∗ ) = The union of such tubes will be X (t. x0∗ . t0 . t0 .6. such that the pair {x0∗ . f U ∗ ) to system (1.6. X 0 ) of all states x∗ .Remark 1. t0 .10) X(t1 .6. we shall deﬁne the respective solution tube to system (1. and a stratc egy U ∗ ∈ UP . x) + f (t).1) from set X 0 = X∗ [t0 ] at time t1 is deﬁned as the set X[t1 ] = X(t1 . x) .6. U ∗ (t. (1. t0 .6.2 A closedloop domain of attainability under counteraction for system (1. x0∗ . ˙ x(t0 ) = x0 as X (t. t0 . Exercise:Construct an example for this remark. X 0 .
10).6. X 0 . . t1 . of all those states x∗ such that for any f (·) ∈ Q(·) there exists a solution x[τ ]. U ∈ UP . A direct calculation similar to that of Section 1. u(·) ∈ UP that steers system (1. x0 . l). t ≤ τ ≤ t1 . t] to (1. such that for every O function f (τ ) ∈ Q(τ ). for terminal set M is the set W [t] = W (t.7. x∗ ) ≤ k(t. t0 . 47 ∀l ∈ IRn .6. t < t1 . there exists an openloop control u = u(τ ). t0 . so that x(t1 ) ∈ M.6. f U ∗ ) : f (·) ∈ Q(·) This also means that set X[t] = X(t.7 Uncertain Systems : the Solvability Tubes The notion further important for control synthesis is that of the solvability set.5)) gives: x∗ ∈ W[t] if and only if (l.1) x = u(τ ) + f (τ ) ˙ from x∗ = x(t) to set M. We shall start with a respective deﬁnition for the case of openloop controls.3) c generated by some x[t0 ] ∈ X 0 . M) of all states x∗ ∈ IRn . Other types of attainability domains under feedback and counteraction could be deﬁned by introducing operations of intersection or union over x0 . τ ∈ [t0 . 1. f.6 (see (1. f (·) ∈ Q(·) = = X (t1 . for any ﬁxed t ∈ T .1 The ”openloop” solvability set under counteraction at time t. f U) : x0 ∈ X 0 . U ∈ UP . Deﬁniton 1. X 0 U ∗ ) consists. U in an order other than in (1.= f U x0 c X∗ (t1 . t0 . such that x[t] = x∗ .7. This is left to the reader.
starting at t1 and ”moving” towards given instant t < t1 . The function W[t] = W(t. W[t] − σ(P(t)−(−Q(t))) = 0 σ→0 W[t1 ] = M.6.7. M) generates a multivalued map with convex compact values (provided M ∈ conv IRn ). however. This solution will require some preparatory work. This lemma follows directly from Theorem 1. t1 . 0 ∈ P(t). Lemma 1. however. but when the latter is taken in backward time. Let us formally construct a set W ∗ (t. but its solution is evolving in backward time. α ∈ (0.1 Under Assumption 1.1 the map W[t] satisﬁes the evolution equation (1.7.3) in this case is precisely the equation (1. is to devise a feedback control strategy for an uncertain system that operates under unknown but bounded input disturbances. that W[t] would immediately satisfy a semigroup property and therefore.7.2 (namely. since an additional assumption is required here.1. an evolution “funnel equation”.7. In terms of setvalued maps this allows the relation t1 t1 (1. 48 .6. t1 . A precise deﬁnition of the problem as well as its solution will be given in the next section. t1 . under the conditions of Remark 1.1.6.6.3) ˙ lim σ −1 h(W[t − σ]. 1)). Our aim. M) which shall be a certain superposition of the “openloop” sets W(t. it is clear that the inverse problem of ﬁnding W [t] is precisely the one of constructing the attainability domain of Deﬁnition 1.4) ˙ P(t)−(−Q(t)) = (1 − α)P(t) + c = R(t). M) deﬁned above. −Q(t) = αP(t) + c. Particularly.where t1 k(t. This does not mean. we have (1. l) = ρ(lM) + t (ρ(l − P(τ )) − ρ(lQ(τ )))dτ.2) W[t] = M + t (−P(τ ))dτ ˙ − t Q(τ )dτ.6. For the openloop case. considered here. The evolution equation (1.7.4).
. Following the procedure. t1 − σ1 . . . introduce a subdivision Σ = {σ1 . k t = t1 − i=1 σi .7. .7. t1 . . . Due to (1. W[t1 − σ1 ]) = t1 −σ1 t1 −σ1 W[t1 − σ1 ] + t1 −σ1 −σ2 ˙ (−P(τ ))dτ − t1 −σ1 −σ2 Q(τ )dτ and may ﬁnally calculate the value W t.2) this gives t1 t1 (1. t + σk + σk−1 . t1 − i=1 σi . As a ﬁrst step. . ﬁnd the openloop solvability set W[t1 − σ1 ] = W(t1 − σ1 .1 There exists a continuous function β(t) > 0. t1 .7. . W(t1 − σ1 . Σ). .7. σk }. . 49 . starting at instant t1 .7. . t ∈ T . we come to (1.5)–(1. M).) (1.Taking the interval t ≤ τ ≤ t1 .7) = J (t. such that all the sets j j−1 j−1 j−2 W(t1 − i=1 σi . t1 − σ1 . .7.7.7) involved in the construction to be nonvoid. M. W(t1 − i=1 σ1 . . W(t + σk . M) . t1 − i=1 σi . t1 . i=1 σi = t1 − t.6) = W(t − σ1 − σ2 . t + σk . t1 . Assumption 1. . .5) W[t1 − σ1 ] = M + t1 −σ1 ˙ (−P(τ ))dτ − t1 −σ1 Q(τ )dτ. = The formal procedure described here presumes all the sets W() of type (1. . . . where k σi > 0.
. σi > 0. . t1 . . . Σ) = ∅ for any subdivision Σ. t1 . .7. J (t.. t1 . whatever is the subdivision k Σ = {σ1 .7.7. ((M + t1 −σ1 ˙ (−P(τ )dτ )− t1 −σ1 ˙ Q(τ )dτ ) .. M. We shall refer to (1.5)–(1. − t Q(τ )dτ )). . . . . . k. M. t1 . M) = W ∗ (t.7.1 there exists a Hausdorﬀ limit lim h J (t. .2 Under Assumption 1.1 clearly ensures J (t.8) J (t. i=1 σi = t1 − t. σk }.7. . k} → 0.7). . . . we come to the analytical expression J (t. Lemma 1. t1 . .7. t1 . We may consider the limit of these sets with max {σi : i = 1. M) with k = 0 max{σi : i = 1.j . .2). (1. . k → ∞. i=1 σi = t1 − t. M.. Σ) . . M)) . Σ) is convex and compact for any subdivision Σ. Following (1. The set J (t. k} → 0. M. . W(t1 − σ1 . Σ) = t+σk t1 t1 = (. . . t1 .7.) − β(t1 − i=1 σi )S are nonvoid with j = 1. M) = W ∗ [t] as the “alternated” solvability domain and denote 50 . Assumption 1. t1 .
t. [257]. such that (1. t1 . W ∗ [t]) = t t = (W [t] + t−σ ∗ ˙ (−P(τ ))dτ )− t−σ Q(τ )dτ. Deﬁniton 1. (see [257]) .9) (1.7.3 Once W ∗ [t] = ∅. W[t] − σP(t)) = 0.7. σ→0 Obviously W ∗ [t1 ] = M. M) = ˙ ((−P(τ ))dτ −Q(τ )dτ ).8) will be referred to as the “alternated solvability tube”.11) yields the following: 51 . t. the setvalued function W ∗ [t] of (1.7. and following the deﬁnitions of these sets we observe that there exists a function γ(σ).7. the setvalued function W ∗ [t] satisﬁes for all t ∈ T the evolution equation (1. t ∈ T . M) is actually the value of a certain type of setvalued integral that is known as the “Alternated Integral of L. (1.7.12) W(t − σ. Due to σ > 0. W ∗ [t − σ] ⊆ W(t − σ.7. t1 .t J (t.M The set J (t. Pontryagin. and to the deﬁnition of geometric diﬀerence. Lemma 1.10) lim σ −1 h+ (W[t − σ] + σQ(t). σ −1 γ(σ) → 0 with σ → 0.7. W ∗ [t]) + γ(σ)S. Taking W ∗ at an arbitrary instant of time t and also W ∗ [t − σ].” The integral was introduced and described in detail in papers [256].11) where γ(σ) > 0. W[t1 ] = M.2 With t varying.S. relation (1. t1 .7.
together with (1. It also follows from Lemma 1.7.8. t1 . The prolongability of the solution W ∗ [t] towards time t0 follows from the condition W ∗ [t] = ∅. σP(τ ) = 0. M) = W ∗ (t.7.7. σQ(t) = 0.3 of the next Section.7.7. The continuity of P(τ ).7.1) give the inclusion (1. Lemma 1. The latter relations.4 The setvalued function W ∗ [t] is a maximal solution to equation (1.7. W ∗ (τ.14) is equivalent to the existence of a solution to (1. particularly. 52 W ∗ (t. M) satisﬁes the semigroup property (in backward time). τ. M)). where σ −1 α(σ) → 0 with σ → 0.7. Relation (1. that W ∗ [t] satisﬁes the following properties. .15) with t ≤ τ ≤ t1 .5 The setvalued map W ∗ (t. Namely (1. Studying equation (1. σ→0 lim σ h t−σ −1 Q(τ )dτ.7. t1 . and Q(τ ) implies t σ→0 lim σ h t−σ t −1 P(τ )dτ. Lemma 1. t1 .13) W [t − σ] + t−σ t ∗ Q(τ )dτ ⊆ ⊆ W ∗ [t] + t−σ (−P(τ ))dτ + γ(σ)S.9) it is possible to observe that its solution in nonunique (devise an example) and moreover.t (1. The proof of this assertion is left to the reader.14) W ∗ [t − σ] + σQ(t) ⊆ W ∗ [t] − σP(t) + α(σ)S.9).7.9) at any given instant t. t ∈ T and from the boundedness of the tube W ∗ [t]. This justiﬁes the assertion of Lemma 1. at t = t1 .7.3.
U(·.7. Hence.7.7. by means of a regularization procedure that would allow to keep up with the basic solution scheme. x) + Q(t) ˙ that start from any given position {τ.1) and terminal set M.1 or 1. such that x(t) + β(t)S ⊂ W ∗ [t]. all the propositions that follow in the sequel and involve the tube W ∗ [t] are true only under this assumption.1. t1 ). the tube W ∗ [t] will coincide with the solvability tube for the problem of control synthesis under uncertainty.2 The alternated solvability tube W ∗ [t] is nondegenerate. 53 . Deﬁniton 1. For future operation it may be sometimes more convenient to use an assumption of equivalent type Assumption 1. where S is a unit ball in IRn . x). xτ }. t0 ≤ t ≤ t1 .1 The assertions of this Section concerning the Alternated Solvability Tube W ∗ [t] have been all derived under Assumption 1.The proof of the relation (1. there exists an absolutely continuous function x(t) and a function β(t) > 0. Such situations are not discussed in this book and are left for additional treatment. M).7.1) x ∈ U (t.8. Remark 1.[257].7. xτ = x[τ ] ∈ W∗ (τ.1 The problem of control synthesis under uncertainty consists in specifying a solvability set W∗ (τ. would reach the terminal set M at time t1 : x(t1 ) ∈ M. t0 ≤ t ≤ t1 . t1 .8 Control Synthesis Under Uncertainty Consider system (1.14) follows from the additivity properties of the alternated integral J (t. 6 As we shall see in the next section. 1. there is a degenerate situation which has to be approached separately. t1 . 6 Once Assumptions 1.2 are not fulﬁlled. t1 .8. M).7. ·) ∈ UP such that all the solutions to the diﬀerential inclusion (1. Namely. M) and a setvalued feedback control strategy c u = U(t.7. τ ∈ [t0 .
4. 54 .11). x) is the maximizer of the expression for calculating dZ [τ. Z[τ ]).19).8. x) constructed similar to the one in Section 1.4.7. where. M) = W∗ [t]. We shall prove that the “alternated solvability tube” W ∗ [t] of Section 1. M) is the solvability set. x) = ∂l f (t.Deﬁniton 1. 1. which is dZ [t.15)). Assume Z(t) to be a solution to the evolution equation (1. Thus (1. provided xτ ∈ W∗ [t].1.4. Deﬁnition 1. 0 where f (t. M) = ∅.4 (see (1. such that there exists a control strategy U(t. M) for the problem of control synthesis under uncertainty is the set of all states xτ ∈ IRn . x) − ρ(lZ[t]) : l ≤ 1} = = (l0 .8. Taking W∗ (t. (l0 = {0} for dZ [t.8.6 does coincide with W∗ [t].8. x] = h+ (x. we come to a setvalued map (the “solvability tube”).3) is formally similar to the deﬁnition of the “extremal strategy” (1. x] = max{(l. as we have seen.4.2 The solvability set W∗ [t] = W∗ (t. −lZ (t. t1 . t1 . Relation (1.1 is nonredundant if W∗ (τ.2) 0 UZ (t. but would not necessarily be the largest solvability set and solvability tube as required by Deﬁnitions 1.8. t1 . W∗ (τ. x) − ρ(l0 Z[t]). t1 . x) that solves the problem of control synthesis of Deﬁnition 1. Let us ﬁrst try to ﬁnd a set and a tube that would provide an appropriate solution for the problem of control synthesis.14) and (1. x] = 0).1. which is the ”largest” set of states from which the solution to the problem does exist at all. For every solution Z(t) let us assign a feedback strategy UZ (t.2.8. l) = ρ(lP(t)) and l0 = lZ (t.9) with boundary condition Z(t1 ) ⊆ M and therefore an absolutely continuous setvalued map with convex compact values. x)).7. (1.
At a point {t.8. x] dZ [t.3)) is that with dZ [t. A consequence of Lemma 1. and further on.Consider the derivative d 2 d dZ [t.8.7. a direct calculation yields d ∂ρ(l0 Z(t)) dZ [t.1 (see (1. (1.8.1). x] > 0.7. after a limit transition σ → 0. l ∈ IRn .1 The following inequality is true ∂ρ(lZ(t)) ≥ ρ(lQ(t)) − ρ(−l − P(t)).4) The evolution equation (1. l)1/2 which gives. x) − ˙ .2).9) leads to the inclusion Z(t − σ) + σQ(t) ⊆ Z(t) − σP(t) + o(σ)S with σ −1 o(σ) → 0. the inequality σ −1 (ρ(lZ(t) − ρ(lZ(t − σ)) ≥ ρ(lQ(t)) − ρ(−lP(t)) + σ −1 o(σ)(l. x] = d[t. x] dt dt due to system (1. dt ∂t (1.3) Lemma 1.4) of the Lemma.8.8. ρ(lZ(t − σ)) + σρ(lQ(t)) ≤ ρ(lZ(t)) + σρ(−lP(t)) + o(σ)(l. x] = (l0 .8. x} that gives dZ [t.8. the result (1. ∂t (1. l)1/2 or otherwise. x] > 0 for the derivative the following inequality holds: σ → 0. 55 .
τ. Lemma 1.8.6) This leads us to ∀f (t) ∈ Q(t). τ < t1 . dt (1. x[τ ] = xτ . x] dt Z calculated due to the system x ∈ UZ (t.8) XZ [t] ⊆ Z[t]. dt u(t) ∈ P(t).8. where f (t) ∈ Q(t) −(l0 .8. the solution tube XZ [t] = XZ (t. the boundary condition Z(t1 ) ⊆ M. with u = u0 . and therefore.8. we will have d dZ [t. Therefore. Some further reasoning yields the next assertion Lemma 1. dt Z ∀f (t) ∈ Q(t).8. xτ f (·)) of system (1. τ ≤ t ≤ t1 .7) satisﬁes the inequality d 2 d [t. x] ≤ 0. u0 ) = ρ(−l0 P(t)).2 The derivative (1. satisﬁes the inclusion (1.8.5) where d dZ [t. f (t) ∈ Q(t).3 With xτ ∈ Z[τ ].8. x] ≤ (l0 .(1. τ ≤ t ≤ t1 . 56 . x] ≤ 0. u(t) + f (t)) − ρ(l0 Q(t)) + ρ(−l0 P(t)). x) + f (t) ˙ d 2 d [t.6).
8. x] = (l0 .1 with Z[τ ] being the solvability domain (but not necessarily the largest one). we observe that tube W ∗ [t] is the maximal solution to equation (1. The results of the above may be summarized into Theorem 1.7.8. x) is a synthesizing strategy that solves the problem of control synthesis of Deﬁnition 1.3. Lemma 1. τ ≤ t ≤ t1 . xτ ) ⊆ W ∗ [t].3 is similar to that of Lemma 1. 1. x) is the maximizer for the problem dW ∗ [τ. to solve the problem of control synthesis through any solution Z[t] of equation (1.8. xτ ) is the solution tube for system (1.4 Strategy U 0 (t. −lW ∗ (t.3. . The results of Lemmas 1.8. is to ﬁnd the maximal solvability domain W∗ [t] for the problem of Deﬁnition 1.9) with boundary condition Z[t1 ] ⊆ M.8. x) − ρ(lW ∗ [τ ] l ≤ 1} (1. x) − ρ(l0 W ∗ [τ ]). τ.8. x) = ∂l f (t.8.8.8.10) dW ∗ [τ.8. Referring to Lemmas 1.1 The synthesizing strategy U 0 (t. however. x).1 and the respective strategy U(t.7.3 imply Lemma 1. 0 where l0 = lW ∗ (t. x) ensures the inclusion (1.9) resolves the problem of control synthesis under uncertainty of Deﬁnition 1. It is therefore possible. with U(t. Here XW ∗ (t. The tube W ∗ [t] generates a strategy (1. 57 XW ∗ (t. x) = U 0 (t. x)).8.4. x[τ ] = xτ .4). The set of states for which the problem of Deﬁnition 1.11) provided xτ ∈ W ∗ [τ ].8.1).9) 0 U 0 (t. 1.8. Our problem.8.7.8.1. in principle.The proof of Lemma 1.8.3 as the main relation used in the proof is the inequality (1. τ. x] = max{(l.1 is solvable will then be restricted to Z[t].3 thus indicates that UZ (t.2. x).7. x) of (1.9) with an equality in the boundary condition (W ∗ [t1 ] = M).
the answer to the question is aﬃrmative. This gives d 0 0 ˙ dW ∗ (t. Let us calculate this derivative solving the extremal problem dW ∗ (t. x∗ ) − ρ(l∗ W ∗ [t]). dt 0 The calculation of the derivative ∂ρ(l∗ W ∗ [t])/∂t can be done using the representations (1. once x∗ ∈ W ∗ [τ ].8.1).10) exists and that intW∗ ≡ ∅. we have to answer the following question: is the maximal solution W ∗ [t] to equation (1.4). a strategy (1.1? As we shall see. This may be proved due to the inequality (1. in view of a relation of type (1.12)(1.(1. x∗ ) = max{(l.7.7.8.7.13) this further gives t+σ ρ(lW [t + σ]) − ρ(lW [t]) ≥ − ρ(l t ∗ ∗ (−P(τ )dτ − Q(τ )dτ )).10) also the maximal solution tube for the problem of control synthesis of Deﬁnition 1. x∗ )/dt due to system (1.1 type.9). x∗ ) − ∂ρ(l∗ W ∗ [t])/∂t.7.8. and 0 l∗ = 0 for dW ∗ (t. (1. v) = ρ(l0 Q(t))}.8. The result presumes however that the solution W∗ [t] to the evolution equation (1. Finally. to make the ends meet.7. x∗ ) = 0.6. This strategy will aﬀect the sign of the derivative ddW ∗ (t. Thus. x∗ ) − ρ(lW ∗ [t]) : l ≤ 1} 0 0 dW ∗ (t.7.12) V 0 (t.7. it is possible to select in the domain d[x∗ . x∗ ) = (l∗ .7.14). x) = {v  (l0 .1.9).1 It is necessary to emphasize that the last theorem is true in the absence of matching conditions of the Assumption 1. x∗ ) = (l∗ . The latter is ensured by Assumption 1. 58 .Remark 1.7. W ∗ [τ ]] > 0. Namely.
using (1. Diﬀerentiating dW ∗ (t. x∗ ) ≥ 0.1. x∗ ) = (l∗ . x).∂ρ(lW ∗ [t]) ≥ ρ(lQ(τ )) − ρ(l − P(τ )). the inequality (1.8.6. x∗ )/dt due to system (1.5 Under Assumption 1.7.1 the derivative dW ∗ (t. x) + f . we come to the following relation.8. x∗ ) = 0.8.15). ·) ∈ UP . W∗ [t]) ≥ rτ . dt f ∈V 0 ∀u ∈ P(t). that starts at a point x∗ ∈ W ∗ [t] or in other words. W∗ [τ ]) = rτ > 0. does satisfy c the inequality d(x[t].8. to the following relations d dW ∗ (t. (1. x). u + f ) + ρ(−l∗ P(t)) − ρ(l∗ Q(t)) dt Selecting f ∈ V 0 (t. Under Assumption 1. This implies that any solution x[t] to the diﬀerential inclusion dx ∈ U(t. u) ≤ ρ(l∗ P(t)). ∀u ∈ U 0 (t.6.8. x). t1 ] whatever is the strategy U(·.8.12) and observing that 0 0 u ∈ P(t) implies − (l∗ . with d(x∗ . we arrive.13) and also the respective rule indicated in Remark 1.1) is given by the relation (1. ∂t or.8.14) ∂ρ(lW ∗ [t]) = ρ(lQ(τ )) − ρ(l − P(τ )) = ∂t ˙ = ρ(l − P(τ )−Q(τ )) We shall further ﬁrst continue under this assumption so that.15) d 0 0 0 dW ∗ (t.6.4. f ∈ V 0 (t. dt f ∈V 0 d dW ∗ (t.5) turns to an equality. ∀u ∈ P(t). due to (1.1. with Z(t) = W ∗ (t). dt x[τ ] = x∗ . Lemma 1. under Assumption 1.8.1 we have therefore proved 59 .13) (1. (1. x∗ ) for x∗ ∈ W ∗ [t]. t ∈ [τ.
x).Theorem 1. the value v may be required to run around a variety v (i) of some extremal points of the set Q. τ ∈ [t0 .8. ∀u ∈ P(t). addressing the reader to monographs [169].1) and target set M.8. one has to substitute strategy V 0 (t. x) over all c the strategies U(·. our aim is to minimaxize the cost I(t.8.1). x) is the same as in (1. throughout any minor interval of time. Consider equation (1.(ii) in the general case. ·) ∈ UP . Let us now pass to the DP interpretation for this Section. [170]. f (·) ∈ Q(·)} u f where I(t. u + f (t)) + ρ(l0 (−P(t))−Q(t)) ≥ 0 dt f ∈ V ast (t. x)U(·. ·) ∈ UP and disturbances f (·) ∈ Q(·). however. where V ∗ has to be speciﬁed as a probabilistic measure concentrated on Q.5. f2 (l) = ρ(lQ).Introduce the value function c V∗ (t. where these topics are discussed in detail. (ii) The set W ∗ [τ ].6. Loosely speaking. x) looks as follows ∂V ∂V + min max{ .u + f } = 0 u f ∂t ∂x (1. namely W∗ [t] = W ∗ [t]. x) of (1.12) by another one V ∗ that would in some sense ensure a relation similar to the following d ˙ dW ∗ [t. f1 (l) = ρ(l − P).7. It exists. [171].16) ˙ ρ(l(−P)−Q) = co(f1 − f2 )(l). We shall not specify the rigorous deﬁnition and precise construction of such strategies V ∗ as this would require to discuss notions that are quite beyond the scope of this book.17) 60 . x] = (l0 (t).8. t 0 ≤ t ≤ t1 . To prove (i).2 (i) The alternated solvability tube W ∗ [t] coincides with the solvability tube W∗ [t] of the problem of control synthesis under uncertainty. on diﬀerential games.8. It should be emphasized that this theorem remains true without the Assumption 1.12). t1 ) is the largest solvability domain for this problem. Since in general we have (1. the desired strategy V ∗ may not exist in the explicit form of (1.Here the formal HJB equation for the value V∗ (t. in the class of mixed strategies (also known as relaxed controls). x) = min max{I(t.1.
Obviously W (1.18). x) ≤ V (t. x) ≤ 0.8. x) ∂V (t.22). x) in its variables and therefore. x].21).8. x) = h2 (x. provided 0 u ∈ UW ∗ (t. Presuming W ∗ [t] = ∅ . M) + Then in view of Lemma 1. x] > 0 ( lW ∗ (t.8. = ∂l f t. t. Indeed.(1.21). x) = arg min{ ∂V (t. x) = ∂d2 ∗ [t.18) V (t1 . x) and the strategy U ∗ (t.20) ensures that once V (t. consider the function V (t.1.8. x) satisﬁes the inequality (1. x) = U ∗ (t.8. then V (τ. inequality (1.8. x) . x]/∂x with dW ∗ [t. x) = d2 ∗ [t. (1. Then the value function V∗ (t.20) ∂V (t.8.8.8. x) + f (τ ). − Relations (1. dτ whatever is the disturbance f (τ ) that satisﬁes (1. x) = h2 (x.17) and (1. −lW ∗ (t. solves the problem of Control Synthesis of Deﬁnition 1.8. t1 ]. u u ∈ P(t)} = ∂x ∂V (t.20).21) U ∗ (t. x)) 0 0 where f (t.8. M) + (Having had a minmax operation involved.8. x) = 0 W with dW ∗ [t.19) V (t.1). τ ∈ [t. u + f f ∂t ∂x ≤ 0 and boundary condition (1. x] = 0). Denoting UW ∗ (t.(1. l) = ρ(lP(t)) and lW ∗ (t. x). x). of the diﬀerential inclusion (1.8.Isaacs and his contribution to diﬀerential games). x[τ ]) ≤ 0 for any trajectory x[τ ] = x(τ.6 Suppose W ∗ [t] = ∅. 61 .22) dx(τ ) = U ∗ (τ. the latter HJB equation is often referred to as the the HJBI equation with letter I being a reference to R.with boundary condition (1.8. x). x[t] = x. x) = ∂l f (t. x) + max .8.8.2 one may observe that V (t.6. we may rewrite (1. x) ∂x .The continuity of W ∗ [τ ] in τ implies the upper semicontinuity of U ∗ (τ.9) then imply Lemma 1. the existence of solutions to the diﬀerential inclusion (1.
where G is a given matrix of dimensions m × n (m ≤ n) and Y(t) is a multivalued function continuous in t with convex compact values (Y(t) ∈ comp Rn . x).8. x0 ) of system (1.8. We shall now introduce an additional state constraint (1.9.9. with a ﬁxed disturbance f (t). [290] ).8. taking it here to be continuous. 7 A detailed theory of viable trajectory tubes for diﬀerential inclusions may be found in [17].1).9. x) will be the value of a respective diﬀerential game ( see [171]. (1.Under Assumption 1.9.3) Gx(t) ∈ Y(t).9. We shall start from Deﬁniton 1.20) turns into an equality and V ∗ (t.3). ∀t). Then V (t. Our next issue is to deal with state constraints.1 relation (1. Our interest is in describing the tube of such trajectories. 1. ˙ t 0 ≤ t ≤ t1 . t 0 ≤ t ≤ t1 . x) = V (t. namely.1) x(t) ∈ P(t) + f (t).[193].8. In order to achieve this equality without such an assumption one has to allow the disturbance f to be selected as indicated in the comments after Theorem 1. in the class of functions generated by a mixed strategy which may result in sliding modes or socalled chattering functions f . (1.2. This section gives a very concise description of the subject.9. 7 62 .1 A trajectory x[t] = x(t.3) if it satisﬁes the state constraint (1.6. t0 .9 State Constraints and Viability Let us return to system (1.2) is said to be viable relative to constraint (1. being only an introduction to other parts of the book.2) x0 ∈ X 0 .
For a certain instant t = ϑ these relations yield (1.9. t t0 Gx(t) = Gx0 + t0 Gu(τ )dτ + Gf (τ )dτ.9.6) x0 ∈ X 0 .4) is equivalent to the equality x(θ) = x0 + ϑ t0 ϑ t0 (1.4) (1.9.9. The set X [t] is generated through relations (1.9. t0 ≤ t ≤ ϑ.3). with restrictions (1.5) x(ϑ) = x0 + ϑ t0 u(τ )dτ + t ϑ t0 f (τ )dτ.9.9.5) is equivalent to the equality (1.9.Deﬁniton 1. (1.2) relative to constraint (1.2) under a state constraint (1.8) ϑ t0 ϑ t0 λ (t)Gx(t)dt = ϑ τ ϑ t0 λ (t)Gx0 dt + ϑ t0 τ ϑ λ (t)Gdt u(τ )dτ + 63 λ (t)Gdt f (τ )dτ .9.1).9. It is obvious that X [t0 ] ∈ Y(t0 ).9.2 A viability tube X [t] = X (t. while (1.9.1)(1. u(t) ∈ P(t). x0 ) : x0 ∈ X 0 } over X 0 of all viable trajectories of system (1.7) u(τ )dτ + f (τ )dτ that should be true for any vector ∈ IRn . This is also the attainability domain of system (1.1).9. t0 . X 0 ) is the union X [t] = {x(t.9. Let us ﬁrst calculate the support function ρ( X [t]) of the crossection X [t] of the tube X [·] at time t. (1. t0 .3) and (1.3). It is not diﬃcult to observe that (1.9.9.9.3). t0 ≤ t ≤ ϑ.
it was proved in [181].9.11) 0≤− ϑ t0 λ (t)Gx(t)dt + ϑ t0 ρ(λ(t)Y(t))dt. ϑ].10) u(t) ≤ ρ( P(t)). the inclusions (1.9) and (1. On the other hand.11). where Φϑ ( .9. ϑ].9. while (1. ∀ ∈ IRn .7)–(1.9.9.11). t0 ≤ t ≤ ϑ.that should be true for any continuous vector function λ(t) ∈ C m [t0 .9.9) x0 ≤ ρ( X 0 ). ∀λ(·) ∈ C n [t0 .9. Following the theory of convex analysis. 64 . ϑ]. The set X [ϑ] will now consist of all those vectors x[θ] that satisfy (1. whatever are the elements ∈ IRn . (1. λ(·) ∈ C m [t0 . λ(·) ∈ C n [t0 .9.9)–(1.3) is equivalent to (1.9. [100]. that the latter requirement will be fulﬁlled if and only if x(θ) ≤ Φϑ ( .9.8) under restrictions (1.12) x(θ) ∈ X [ϑ] if and only if there exists a vector x0 and a function u(t) that respectively satisfy (1. λ(·)) for any ∈ IRn . collecting relations (1.9.7).6) are equivalent to the following inequalities (1. (1. In other terms.9. we observe that (1. λ(·)) = ρ( − ϑ t0 ϑ t0 ϑ t0 λ (t)G dtX 0 ) + ϑ τ ρ( − λ (t)G dtP(τ ) + f (τ ))dτ + ρ(λ(τ )Y(τ ))dτ.10) and also the inequality x(θ) ≤ ( − ϑ t0 ϑ t0 λ (t)Gdt)x0 + ϑ τ ϑ t0 − ϑ τ ϑ t0 λ (t)Gdt u(τ )dτ + ( − λ (t)G dt)f (τ )dτ + ρ(λ(τ )Y(τ ))dτ.9. ϑ].9. ∀ ∈ IRn .
9.This in its turn will be true if and only if (1. The inclusion (1. A solution to (1.9.9. Function Φϑ [ ] happens to be convex and positively homogenous (this may be veriﬁed as an exercise) and therefore. ϑ]} = Φϑ [ ].17) may be rewritten due to (1.9. Z[t] ∩ Y(t) + σP(t) + σf (t)) = 0 Z[t0 ] ⊆ X0 . due to Lemma 1.12) we come to the equality Xϑ = X [ϑ].9.3.14) ρ( X [ϑ]) = Φϑ [ ]. t 0 ≤ t ≤ t1 .9. A more detailed version of these calculations could be also found in paper [181].3.9.3.16) and X [t0 ] = X 0 ∩ Y[t0 ].1). P(τ )dτ + t+σ t f (τ )dτ ⊆ X [t] + σP(t) + σf (t) + o(σ)S that follows from the deﬁnition of the Hausdorﬀ semidistance h+ as well as from relations (1. (1.1 (b). t1 ] . λ(·))  λ(·) ∈ Cm [t0 .9.17) X [t + σ] ⊆ X [t] + t+σ t σ→0 X [t] ⊂ Y[t].2) and therefore. (1.3.9.3).13). obviously (1.13) x(θ) ≤ inf {Φϑ ( .9.5.15) almost everywhere and is also h+ – absolutely continuous in the sense of Deﬁnition 1.16) as (1. σ < t1 − t0 this yields the inclusion (1.9.6).3. For any t ∈ [t0 .9. Let us now introduce an evolution equation which will prove to be an appropriate description for X [t].9.9.3.13) is necessary and suﬃcient for (1.2) viable relative to constraint (1. At the same time X [t] = {x[t]} is a collection of all solutions to (1.15) lim σ −1 h+ (Z[t + σ].18) X [t + σ] ⊆ X [t] ∩ Y[t] + σP(t) + σf (t) + o(σ)S. is a solution to the evolution equation (1. is a support function of some set Xϑ .1 The support function for X [ϑ] is given by (1. Since (1. (1. due to Lemma 1.1).9. This will be (1. σ > 0.7). having proved Theorem 1.9.2. 65 .9. Then.15) is a multifunction Z[t] that satisﬁes (1.9. Let X [t] = {x[t]} be the union of all trajectories of (1.
the following assertion is true. Lemma 1.15).15). It is not diﬃcult to observe that an isolated trajectory x[t] = x(t. X 0 )). for any ∈ S. There is another . From the relations of the above it also follows that X [t] is absolutely h+ continuous in the sense of Deﬁnition 1. ) is absolutely continuous on T uniformly in We may now formulate 66 .15).9.9.2 The mapping X [t] = X (t.15).15). ) is absolutely continuous on the interval T .9.9. This deﬁnition ensures that f (t. t0 .1 The setvalued function X [t] is a maximal solution to (1. t0 .3 A convex compact multivalued function Y(t) ∈ compIRn is said to be absolutely continuous on an interval T = [t0 .3.9. We will precede this with a deﬁnition Deﬁniton 1.15) is true for t ∈ [t0 .9. Given X [t] and any other solution Z[t] to (1.9. X 0 ) satisﬁes the semigroup property (1. τ. ∈ S. X 0 ) = X (t.9.9. We thus come to the proposition Theorem 1.3) is also a solution to (1.2. X (τ.9. t0 . t0 . x0 ) that is viable relative to constraint (1.1 as well as the following assertion Lemma 1.2 The setvalued function X [t] is a solution to the evolution equation (1. namely Z[t] ⊆ X [t] for any other solution Z[t] to (1. stronger form of an evolution equation which should be mentioned in this context.9.19) X (t. t1 ]. t1 ] if its support function ρ( Y(t)) = f (t.9.9.The latter relation indicates that (1. We leave to the reader to verify both Lemma 1.
2) such that x(t) ∈ intY(t).9.9. This gives (1.9. Also assume that there exists a trajectory x(t) of system (1.9.3 Assume the multifunction Y(t) to be absolutely continuous on the interval T . (W[t] − σP(t) − σf (t)) ∩ Y(t − σ)) = 0. is the set of all states {xτ } = W0 [τ ] such that there exists a measurable function u(t) that generates a trajectory x(t.9.9.9. xτ ) = x[t].19) can also be treated in backward time.Theorem 1.19) is somewhat diﬀerent from (1. W[t1 ] = M. (X [t] + σP(t) + σf (t)) ∩ Y(t + σ)) = 0 X [t0 ] = X 0 .3 With Y(t) absolutely continuous.19) with A(t) ≡ 0.22) σ→0 lim σ −1 h(W[t − σ].8. [193].21) σ→0 lim σ −1 h(X [t + σ]. Let us now formally write the equation (1. Set W[τ ] allows the following interpretation. τ ≤ t ≤ t1 of system (1.8. ((I + σA(t))X [t] + σP(t) + σf (t)) ∩ Y(t + σ)) = 0. equation (1. 67 . particularly in having involved the Hausdorﬀ distance h rather than the semidistance h+ . Equation (1.9.1).3 is given in papers [190].9.20) σ→0 lim σ −1 h(X [t + σ].3).3). namely in the following form (1. The proof of Theorem 1.4 A solvability set W0 [τ ] under state constraints (1. Theorem (1.9.15).1) that satisﬁes the inclusion x[t1 ] ∈ M together with restriction (1.9.(1.9. Equation (1.21) has a unique solution deﬁned on the interval T .9.9. τ.3) obviously yields Lemma 1.9. Deﬁniton 1.9. Then the multifunction X [t] is the unique solution to the evolution equation (1.
9.3. t0 ≤ t ≤ t1 generates a solvability tube that is a crucial element for solving the problem of control synthesis under state constraints.).1 With M = IRn . Function W0 [t]. An alternative version of (1. [17].9.21) is given by the equation (1.8.valued trajectory of (1. single .4 The multifunction W[t] is the maximal solution to equation (1. the viability set W[t] is the collection of all 0 positions {t.A set W[t] with such a property is referred to as weakly invariant relative to constraint (1.23) and the proof of the respective existence theorem is similar to that of Theorem 1.9. so that W0 [τ ] ≡ W[τ ].9. the set W0 [τ ] is also known as the viability kernel (relative to constraint (1. With M = IRn . however.8.(see [17]).Here W0 [τ ] is obviously the same as the attainability set.9.2. We ﬁnally emphasize the following property that could be proved through standard procedures. A solution to (1.1).9. from each of which there exists a control u(·) ∈ UP that keeps the respective trajectory x(t) within the state constraint (1.3)).9. satisﬁes (1. but is taken in backward time and is clearly a solution to equation (1.9.9. Thus any viable . Lemma 1.(1. x) that solves the problem of control synthesis under state constraints. X 0 = IRn . Z[t] ∩ Y[t] − σP(t) − σf (t)) = 0 Z[t1 ] ⊆ M. x} .22) (Y[t] may be assumed to be merely continuous and even upper semicontinuous). 1.22).9.8.23) σ→0 lim σ −1 h+ (Z[t − σ]. 68 .9.23) exists under weaker assumptions than those for (1.2).3).9.10 Control Synthesis Under State Constraints Given the solvability tube W[t] of the previous paragraph we may construct a multivalued synthesizing strategy U(t. Remark 1. if taken in backward time.21). Its solution is nonunique.
x) + f (t) ˙ that start from any given position {τ. M) = ∅. l) = ρ(lP(t)) and l = lW (t. W 0 [τ ] is the ”largest” set of states xτ . (1. so that W 0 [τ ] ≡ W0 [τ ] ≡ W[τ ]. xτ }. as before. M) = W 0 [τ ] c and a setvalued feedback control strategy u = U(t. τ ∈ T .4) 0 U(t.10.10. Here 69 .5) or dW [t.2) x(t) ∈ Y(t). x) is the maximizer for the expression (1.9). τ ≤ t ≤ t1 . x] > 0 (otherwise l0 = 0).10. f (t.1 does exist at all.Deﬁniton 1. −lW (t. (1.10. and deﬁne a feedback strategy (1.1. where. x) = ∂l f (t.20). t1 .8. x] = max{(l.9. Following the same reasoning as in the absence of state constraints (see Sections 1.1) x(t) ∈ U(t. the problem of control synthesis under state constraints consists in specifying a solvability set W(τ.1 Given a terminal set M ∈ compIRn . Here.9. from which the solution to the problem of Deﬁnition 1. such that all the solutions to the diﬀerential inclusion (1.3.4. τ ∈ [t0 .4 ) .9. xτ = x[τ ]. xτ ∈ W(τ.10. x) − ρ(l0 W[t]) if dW [t. Deﬁnition 1. x). t1 ] would satisfy the restrictions (1. x)) 0 similar to that of (1.1 is nonredundant provided W 0 [τ ] = W(τ. t1 . x] = (l0 . U(·. τ ≤ t ≤ t1 .10.We shall further use notation W[τ ] for this set and for the respective tube ( τ ∈ T ).1. ·) ∈ UP . x) − ρ(lW[t]) l ≤ 1} dW [t. it may be shown that the set W 0 [τ ] will coincide with set W[τ ] of Section 1.3) x(t1 ) ∈ M. Let us now consider the tube W[τ ]. M). t1 .
Since h(W .7) d d d+ V (t. x) + f (t). W ) = max{ρ(lW ) − ρ(lW ) we observe that the increments ∆1 (σ) = σ −1 (ρ(lW[τ − σ] − ρ(lW[τ ])) and ∆2 (σ) = σ −1 (ρ(l(W[τ ] − σP(τ ) − σf (τ )) ∩ Y(τ − σ)) − ρ(lW[τ ])) are such that σ→0 l = 1} lim ∆1 (σ) − ∆2 (σ) = 0. x] dW [t. x − z)z ∈ W[t]} = V (t.10.10. x] = dW [t. In order to do that. namely ∂− ρ(lW [τ ]) ∂τ τ =t+0 where ∂− ρ(lW [τ ]) = lim σ −1 (ρ(lW[τ − σ]) − (ρ(lW[τ ]) σ→0 ∂τ n for a given direction l ∈ IR . 0) → 0 with σ → 0. let us ﬁrst calculate the left partial derivative in t of the support function ρ(lW [t]).6) d2 [t. 70 .8) We assume in this section that the support function ρ(lY(t)) of the multifunction Y(t) is absolutely continuous.(1. ˙ due to the inclusion (1. x] dW [t. x] = min{(x − z. particularly to calculate the increment ρ(lW[τ − σ]) − ρ(lW[τ ]) through the relation W[τ − σ] = (W[τ ] − σP(τ ) − f (τ )) ∩ Y(τ − σ) + r(σ) where σ −1 h(r(σ). x] dt dt dt x ∈ U(t. x) is a solution to our problem we have to calculate the derivative (1.10.9. We will further use the relation (1. x) = dW [t. x) W To prove that U(t.21).
(. − 71 ∂ (ρ(lY(t))} ∂t . This ﬁnally yields Lemma 1. Formula (1.11) = τ =t = τ =t min{ρ(−pP(t)) − (p. t = τ. Let us elaborate on this result.9) where (1.9) is taken over all p ∈ IRn that satisfy the equality ρ(pW[t]) + ρ(l − pY(t)) = ρ(lW[t]). Since the properties of W[t] imply W[t] ⊆ Y(t).0.10. f (t)) − ∂ (ρ(l − pY(t)) p ∈ F(t. ∂t F(t. l) = {p ∈ IRn : k(t. in [89].10.10. we have ρ(lW[t] ∩ Y(t)) = min{ρ(lW[t]).10.9).10. l)}. l) = ρ(lW[t]).10.9) actually yields ∂− ρ(lW[τ ]) ∂τ (1. l) − k(t.10. ρ(lY(t))} and therefore the minimum of g(0) over p is attained at either p = 0 (which is when ρ(lW[t]) ≤ ρ(lY(t)) or p = l (which is when ρ(lY(t) = ρ(lW[t])). is exact. p) = ρ(l − pY(t))} min{ρ(−lP(t)) − (l.Therefore it suﬃces to calculate the left derivative dg(σ) dσ for the function g(σ) = ρ(l(W[τ ] − σP(τ ) − σf (τ )) ∩ Y(τ − σ)) = min{ρ(p(W[τ ] − σP(τ ) − σf (τ ))) + ρ(l − pY(τ − σ))p ∈ IRn } since dg(σ) dσ = σ=0 σ=0 ∂− ρ(lW[τ ]) ∂τ The calculation then follows the techniques of directional diﬀerentiation given.1 The following relation holds ∂− ρ(lW[τ ]) ∂τ (1. Here the relations (1. for example.10) and k(t.10) reﬂect the fact that the “inﬁmal convolution” that deﬁnes g(0). f (t)) . namely the minimum in (1.
Relation (1.10.9) allows to calculate the right directional derivative d+ dW [t, x]/dt due to system (1.9.1) through formula (1.9.5). In view of the equality ∂− ρ(lW[τ ]) ∂τ ∂ρ(lW[τ ]) ∂τ
= −
τ =t
we come to d+ dW [t, x] ∂ρ(l0 W[t]) = (l0 , u + f (t)) + = dt ∂t = (l0 , u + f (t)) + ∂ ρ(l0 Y(t))} ∂t ≤ (l0 , u + f (t)) + ρ(−l0 P(t)) + (−l0 , f (t)), min{ρ(−l0 P(t)) − (l0 , f (t)), − which is true for almost all t. The last relation turns into an equality d+ dW [t, x]/dt = 0, if u ∈ U(t, x), where U(t, x) is given by (1.10.4). This conclusion produces Lemma 1.10.2 Once u ∈ U (t, x), where U(t, x) is deﬁned by to (1.10.4), then almost everywhere the derivative d+ dW [t, x]/dt ≤ 0, and therefore (1.10.12) d+ V (t, x) dt ≤ 0 a.e.
u∈U (t,x)
Similar to how it was before, in Sections 1.1.4 and 1.1.8, inequality (1.10.12) suﬃces to prove that strategy U(t, x) of (1.10.4) does solve the problem 1.10.1 of control synthesis under state constraints. We thus come to the proposition Theorem 1.10.1 The problem of control synthesis under state constraints of Deﬁnition 1.10.1 is solved by strategy U(t, x) of (1.10.4). The problem is obviously solvable if the starting position {t, x} is such that x ∈ W[t], where W[t] is the solvability set given by the unique solution to equation (1.9.22) or by the unique maximal solution to equation (1.9.23). It is not diﬃcult to prove though that W[t] is the “largest” set from which the solution does exist at all. The respective proof is similar to the one given in the last part of Section 1.8, so that the last theorem may be complemented by 72
Lemma 1.10.3 In order that the problem of Deﬁnition 1.10.1 would be solvable it is necessary and suﬃcient that xτ ∈ W[τ ]. The results of this section may be again explained through DP techniques . Exercise 1.10.1. (a)Introducing the value function
t1
V0 (t, x) = min
t
h2 (x(τ, Y(τ ))dτ + h2 (x(t1 ), M) u(τ ) ∈ P(τ ) , + +
check, whether it satisﬁes the corresponding HJB equation for system (1.10.8) and what would be the relations between the solutions to the problem of Deﬁnition 1.10.1 achieved through V0 (t, x) and through function V (t, x) of (1.10.6). (b) Taking Example 1.5.1, complement it by a state constraint (x(t) − n(t), N (t)(x(t) − n(t)) ≤ 1, N (t) > 0,
and ﬁnd the solvability set W 0 [τ ] of Deﬁnition 1.10.1 by following the schemes of (1.5.15)(1.5.17). Calculate the analogy of formula (1.5.17) for the given state constraint. We ﬁnally come to the next topic which incorporates all the diﬃculties speciﬁc for the previous Sections.This is the problem of control synthesis under both uncertainty and state constraints.
1.11
State Constrained Uncertain Systems. Viability Under Counteraction.
Consider system (1.7.1) with terminal set M, state constraint (1.10.2) and constraints (1.1.2), (1.6.1) on the control u and the uncertain input f . Deﬁniton 1.11.1 The problem of control synthesis under uncertainty and state constraint consists in specifying a solvability set W[τ ] = W∗ (τ, t1 , M) and a setvalued control strategy U(t, x) such that all the solutions x[t] = x(t, τ, xτ ) to the diﬀerential inclusion (1.8.1) that start at a given position {τ, xτ }, xτ = x[τ ] ∈ W∗ (τ, t1 , M), τ ∈ [t0 , t1 ), would reach the terminal set M at time t1 , so that x(t1 ) ∈ M, and would also satisfy the state constraint (1.10.2), namely x[t] ∈ Y(t), 73 τ ≤ t ≤ t1 .
Here the multivalued function Y(t) with values in compIRn is again taken to be absolutely continuous. It is clearly the strategy U(t, x) that is responsible for the solution x[t] to satisfy the state constraint (1.10.2), no matter what is the disturbance f (t). In this section we will have to combine the schemes of sections 1.1.7, 1.1.8 and sections 1.1.9, 1.1.10. The technicalities of this combination require a more or less sophisticated mathematical treatment, the details of which are not directly relevant to the topics of this book. They are the subject of other publications ( see [194], [195]). We will however give a concise presentation of the solution to this problem emphasizing the substantial interrelations important for the results. The solution strategy U(t, x) will again be determined by a relation of type (1.10.4), (1.10.5), where W[t] has to be substituted by W∗ [t] — the solvability set of Deﬁnition 1.11.1. The basic evolution equation for W∗ [t] now has the form
(1.11.1)
σ→0
lim σ −1 h+ (Z[t − σ] + σQ(t), Z[t] ∩ Y(t) − σP(t)) = 0, Z[t1 ] ⊆ M,
so that the following assertion holds. Lemma 1.11.1 The solvability set W∗ [t] for the problem of control synthesis under both uncertainty and state constraints as formulated in Deﬁnition 1.11.1 is the maximal solution to equation (1.11.1) with boundary condition Z[t1 ] = M. The proof of this assertion follows the lines of sections 1.1.7 – 1.1.10 . As we have seen above, the property important for Control Synthesis is the behavior of the directional derivative d V (t, x) dt V (t, x) = d2 ∗ [t, x] = h+ (x, W∗ [t]) W along the solutions to the diﬀerential inclusion (1.10.1) with f (t) unknown, but bounded: 74
(1.11.2)
f (t) ∈ Q(t),
t ∈ [t0 , t1 ].
Combining the calculations of sections 1.1.8 and 1.1.10 under Assumption 1.6.1 we come to Lemma 1.11.2 The derivative dd(x, W∗ [t])/dt is given by d ∂ ˙ d(x, W∗ [t]) = (l0 , u + f ) + min{ρ(l0 (−P(t)) − Q(t)) , − (ρ(l0 Y(t))} ≤ dt ∂t
(1.11.3)
≤ (l0 , u + f ) + ρ(l0  − P(t)) − ρ(l0 Q(t)).
0 The synthesizing strategy U∗ (t, x) may now be deﬁned in the same way as U 0 (t, x) of Section 1.1.8, that is according to (1.8.9), but with W ∗ [t] substituted by W∗ [t] of Deﬁnition 1.11.1, 0 (the notation U 0 (t, x) is also substituted by U∗ (t, x) ).
Similarly to Section 1.1.8 (Theorem 1.8.1), the previous lemma implies
0 Theorem 1.11.1 The strategy U∗ (t, x) deﬁned by (1.8.9) (with W ∗ [t] substituted by W∗ [t]) resolves the problem of control synthesis under uncertainty and state constraints of Deﬁnition 1.11.1.
Therefore every solution x[t] to the system (1.11.4)
0 x ∈ U∗ (t, x) + f (t) ˙
(1.11.5) satisﬁes the constraint (1.11.6) and therefore the inclusion
x(t0 ) ∈ W∗ [t0 ]
x[t] ∈ W∗ [t],
t0 ≤ t ≤ t1
x[t1 ] ∈ M. In this case we will say that system (1.11.3) is viable relative to constraint (1.10.2) under counteraction (1.11.2), provided x(t0 ) satisﬁes (1.11.6). 75
In other terms we may say that W∗ [t], t0 ≤ t ≤ t1 , is a tube of strongly invariant sets for system (1.11.3) under counteraction (1.11.2).( The latter term indicates that all the solutions to the diﬀerential inclusion (1.11.3) that start in W[t0 ], do satisfy the state constraint (1.10.2)). The assertions of this section ﬁnalize the concise description of the solutions to the problems of evolution and control synthesis in the presence of uncertainty and state constraints. A topic for further discussion is the application of set valued calculus to the problem of stateestimation.
1.12
Guaranteed State Estimation : the Bounding Approach
One of the basic problems of modelling and control is to estimate the state of an uncertain or incompletely deﬁned dynamic system on the basis of on or oﬀline observations corrupted by noise. Leaving aside the welldeveloped stochastic approach to these problems, and following the emphasis of the present book, we shall again assume the ”setvalued” interpretations of the respective problems. Namely, as in Section 1.6, an uncertain system is understood to be one of the following type (1.12.1) x(t) ∈ A(t)x(t) + u(t) + f (t), ˙ t0 ≤ t ≤ t1 , x(t0 ) = x0 ,
where A(t) ∈ L(IRn × IRn ), u(t) is a given function ( a preselected control) and f (t) ∈ IRn is the unknown but bounded input (disturbance). It is presumed that the initial state x0 ∈ IRn is also unknown but bounded, so that (1.12.2) (1.12.3) f (t) ∈ Q(t), t0 ≤ t ≤ t1 ,
x0 ∈ X 0 ,
where the set X 0 ⊂ convRn and the continuous setvalued function Q(t) ∈ compIRn , are given in advance. Equation (1.12.1) may be complemented, as we have seen earlier, in Section 1.9, by a state constraint (1.12.4) G(t)x(t) ∈ Y(t), t 0 ≤ t ≤ t1 76
with G(t) ∈ L(IRn × IRm ) and Y(t) ∈ conv IRm , m ≤ n. The constraint (1.12.3) may be particularly generated by a measurement equation (1.12.5) y(t) = G(t)x(t) + v(t), t 0 ≤ t ≤ t1 ,
with an unknown but bounded error (1.12.6) v(t) ∈ K(t), t 0 ≤ t ≤ t1
where K(t) ∈ conv IRm , t0 ≤ t1 is an absolutely continuous setvalued map, ( recall Section 1.9). With the realization y(·) being known, restriction (1.12.4), (1.12.5) turns into (1.12.7) so that Y(t) = {x : G(t)x ∈ y(t) − K(t)}, ( however, the whole function y(·) may not be known in advance, arriving online). Our objective will be to estimate the system output (1.12.8) z(t) = Hx(t), z ∈ IRr , r ≤ n, t 0 ≤ t ≤ t1 G(t)x(t) ∈ y(t) − K(t), t 0 ≤ t ≤ t1 ,
at any prescribed instant of time t. More precisely, the problem is to specify the range of the output z(t) that is consistent with relations (1.12.1)(1.12.4) (the Attainability Problem under State Constraints), or the set fo all outputs z(t) consistent with system (1.12.1)(1.12.3), and measurement equation (1.12.5), (1.12.6) ,with realization y(t) of the measurement being given (the Guaranteed State Estimation Problem). The solution to both problems is therefore given in the form of a set representing thus the bounding approach to state estimation. Our aim here is not to repeat the wellknown information ( [276], [181], [225]). on these issues, but to rewrite some theoretical results focusing them on the main objective, which is further to devise in Part IV some constructive algorithmic procedures based on ellipsoidal techniques that would allow a computer simulation with graphical representations. 8
8 The ﬁrst descriptions of state estimation ( observation ) problems under unknown but bounded errors may be traced to papers [166], [317], [276], [177]. The setvalued approach to such problems in continuous time appears to have started from publications [54], [178], [277], [181].
77
Let us specify the problems considered here. starting with the Attainability Problem. As indicated in Section 1.9, the attainability domain X (·, t0 , x0 ) for (1.12.1),(1.12.2) under state constraint (1.12.4) at time t ∈ [t0 , t1 ] from point x0 ∈ IRn is the crosssection at t ∈ [t0 , t1 ] of the tube of all trajectories x[·] = x(·, t0 , x0 ) that satisfy (1.12.1),(1.2.2),(1.12.4), namely, (1.12.9) X (·, t0 , x0 ) = {x(·, t0 , x0 )  x0 ∈ X 0 }
Deﬁne the map X [t] = X (t, t0 , X 0 ) as X (t, t0 , X 0 ) = {X (t, t0 , x0 )x0 ∈ X 0 }.
The multivalued map X [·] generates a generalized dynamic system. Namely the mapping X : [t0 , t1 ] × [t0 , t1 ] × convIRn → convIRn possesses a semigroup property, that is, whatever are the values t0 ≤ t ≤ τ ≤ θ ≤ t1 , we have X (θ, t, X [t]) = X (θ, τ, X (τ, t, X [t])). Also, the setvalued map X , or in other words, the tube X [t], (t0 ≤ t ≤ t1 ) satisﬁes an evolution equation — the ”funnel” equation of type (1.9.20), ([190], [193]) – which is (1.12.10)
σ→+0
lim σ −1 h(X [t + σ], ((I + A(t)σ)X [t] + σP(t)) ∩ Y(t + σ)) = 0, t0 ≤ t ≤ t1 , X [t0 ] = X0 ,
Equation (1.12.10) is correctly posed and has a unique solution that deﬁnes the tube X [·] = X (·, t0 , X0 ) for system (1.12.1)(1.12.4) if the map Y(·) is such that the support function ρ(  Y(·)) = max{( , p)  p ∈ K(t)} is absolutely continuous in t,[193]. Using only one of the Hausdorﬀ semidistances in (1.12.10) leads to the loss of uniqueness of the solutions, but allows to relax the requirements on the multivalued function Y(t). Consider the evolution equation of type (1.9.23) which, in our case transforms into (1.12.11) lim σ −1 h+ (Z[t + σ], ((I + A(t)σ)Z[t] ∩ Y(t)) + σP(t)) = 0, 78
σ→+0
10) one may also observe.(1. with given y(τ ). are consistent with the constraints (1.12.2).(1.12. t0 . Then the tube X t = X ∗ [·] of domains X ∗ [t] = X [t] = X (t.12.12.12. of states x = x(t) of system (1.1) on the basis of the available measurement y ∗ (τ ).12) ζ ∗ (t) = {x∗ .12.(1. if Y(t) is only upper semicontinuous in t). Equation (1.14) ∗ x∗ [t] = A(t)x∗ [t] + u(t) + f ∗ (t) ˙ x∗ ∈ X 0 . Each set X ∗ [t] therefore gives a guaranteed estimate of the state x∗ [t] of system (1. that X [·] = X+ [·].12.12.11) if it satisﬁes (1.12.12.12. t0 .12.11) has a unique maximal solution under relatively mild conditions ( for example.12. f ∗ (t). t0 ≤ τ ≤ t under the constraints (1. X 0 ) generated by (1.12.1).13) (1.12.12.3).1)(1.9. 0 t0 ≤ t ≤ t1 . X0 ).12.12. due to system (1. (1. (1. we obtain alternative descriptions for the multivalued map X [·]. t0 ≤ τ ≤ t .4) is given.(1. Particularly. that complies with the constraints (1.12.3).12.3). Deﬁniton 1.1) that.(1. such that X+ [t] ⊂ Z[t] for all t ∈ [t0 .12.3). and is generated by an unknown triplet (1.12.7). ( [194]). and Z[t0 ] = X0 .(1. A setvalued map X+ [·] will be deﬁned as a maximal solution to (1. it allows to treat a reasonably large class of discontinuous setvalued functions Y(t).12.11) for almost all t ∈ [t0 . suppose that the measurement y(·) = y ∗ (·).5). 0 y ∗ (t) = G(t)x∗ [t] + v ∗ (t). t1 ] and X+ [·] = W[·]. Namely. t1 ] and if there exists no other solution Z[·] .6).7) is referred to as the information domain relative to measurement y(·)). (1.6) and calculated due to the knowledge of the measurement y[·] = y ∗ [·]. As mentioned in Section 1. equation (1.with t0 ≤ t ≤ t1 . (1. t 0 ≤ t ≤ t1 . Under the conditions required for the existence and uniqueness of the solutions to (1.12. As we have observed earlier.5).10).12. the solution to this equation is nonunique. that is: (1. does always contain the unknown actual trajectory x∗ [·]. 79 . v ∗ (t)}.12.2).2). generated by ζ ∗ (·). The Guaranteed State Estimation Problem may now be formulated more precisely.1 The set X [t] = X (t.12. By complementing it with an extremality condition.12.11) is an alternative version relative to (1.
The information domain X [t] .(1.3). The solution of the guaranteed estimation problem is to specify the tube X [t] = X ∗ [t].or the ”feasibility domain”. [181].It is also the unique maximal solution to (1. if we use equation (1.7).12.11). We shall therefore further allow it to be only Lebesguemeasurable.1).2.12. ( [56]. or absolutely continuous. As in the earlier Sections by X [t] = X (t. z0 ∈ conv(IRn × IRm )).15) x ∈ A(t)x + P(t). the estimate of the output z(t) is the set Z(t) = HX ∗ [t]. it is the attainability domain X [t] for system (1.12.4) for all t ∈ [t0 . one ought to presume. ˙ (1. that y(t) is piecewisecontinuous (”from the right”).12.16) that consists of all those trajectories that start at X [t0 ] = X0 and satisfy the state constraint (1.1 (i) With Y(t) upper semicontinuous.2). we shall apply the idea of singular perturbation technique.12.4) and (1. 80 .12.12. the tube X ∗ [t] is the unique maximal solution to the evolution equation (1. But one must of course realize that this time the object of application is a diﬀerential inclusion and that the propagation of wellknown results [294]. (iii) Once X ∗ [t] is known. t0 ≤ t ≤ t1 . t0 .12.12.12.16) w ∈ −G(t)x + Y(t).12.The results of Section 1.12. for example. [295] in singular perturbation theory to trajectory tubes would require speciﬁc treatment. w(t0 )} ∈ Z0 t0 ≤ t ≤ τ. τ ] X [t] is obviously the tube of all viable trajectories relative to constraints (1. we denote the trajectory tube of system (1. (ii) With Y(t) absolutely continuous. In order to imbed this situation in the given schemes.17) {x(t0 ). the tube (setvalued function)X ∗ [t] satisﬁes the evolution equation (1. [225]). To conform with the assumptions on Y(t) of the above.10). [277]. Theorem 1.11).(1.10). deﬁned for a given measurement y(t) = y ∗ (t). As mentioned above.2.9 allow the following assertion.1. Here w ∈ IRm . Consider the system of diﬀerential inclusions ( > 0): (1. ˙ (1. if we use (1.12. X0 ). is also referred to as the ”domain of consistency”.11). It is important to emphasize that in many applied problems the observed measurement output y(t) is not obliged to be continuous.
4) directly.12. t0 ≤ t ≤ τ. L).6). Z0 . t0 .12.(1.18).19)(1.5).21) x ∈ A(t)x + P(t). 1. ) witll denote the tube of solutions z(t) = {x(t). ˙ L(t)y ∈ −G(t)x + Y(t) ˙ z0 = {x(t0 ).21) will be denoted as Z[t.10 and papers [192].12.12. τ ]. as in (1. ) for all > 0. X0 .12.17) on the interval [t0 .12. t1 ] z(τ ) = {x(τ ). The following analogy of Theorem 12.7) or (1.12. Then X [τ ] ⊆ Πx (∩{Z(τ. Here the constraint (1. but shall rather deal with the perturbed system (1.(1.12. t ∈ [t0 .12. Our aim is still to describe the tube X [t]. The following assertion is true.5). The latter system may then be fully treated within the ”standard” framework of Sections 1. [193].20) (1.12.12. ) > 0}) Let us now introduce another system of diﬀerential inclusions of type (1.15)(1. Z0 .16) .12. The ”bad” properties of y(t) are then clearly due to the ”bad” measurement ”noise” v(t) in (1. w0 } ∈ Z0 and for every τ ∈ [t0 .12.12.16).12.the realization of the observations . However. y(t)} to the system (1. The class of all continuous invertible matrix functions L(t) ∈ L(IRn . t0 . We will also use the notation Πx W for the projection of set W ⊂ IRn × IRm onto the space IRn of variables x.12. 81 .4) there exists a vector w0 ∈ IRm such that {x(t0 ). (1.12.12. Z0 .1 Assume (1. w(t0 )} ∈ Z0 .12.12.(1.18) X0 ⊆ Πx Z0 . we shall not study system (1.3). Theorem 1.9. IRn ).12. t0 .12.is allowed to be Lebesguemeasurable. L] = Z(t.(1.2 Assume that (1. t1 ] will be denoted as L and the solution tube to system (1. where function y(t) . in order to achieve this. w( τ )} ∈ Z(τ.15).19) (1.12.(1. Corollary 1.15). t0 . but with a timedependent matrix L(t) instead of the scalar > 0 : (1.18) to be true.16).12.4) may particularly be generated by a measurement equation.12. Then for every trajectory x(·) ∈ X [·] of (1.Following this notation symbol Z[t] = Z(τ.2 is true.
t1 ] z(τ ) = {x(τ ). 82 X [τ ] = Πx (∩{Z[τ.12. Then (1. t1 ]. Its details may be found in [193]. L].2 Assume relation (1.Theorem 1.23) yield an exact description of the set X[τ ] through the solution of the perturbed diﬀerential inclusions (1.12. t1 ] the following inclusion is true (1.19)(1. (1.12.18) to be true.4 Let us assume Πx Z0 ⊆ X0 . Then for every x(·) ∈ X [·] there exists a vector w0 ∈ IRm such that {x(t0 ).3 Assume relation (1.12.12. This result may be proved within the techniques of Sections 1. L]L(·) ∈ L}) ⊆ X [τ ].12. Relations (1.5 Under the assumption Πx Z0 = X0 the following formula is true (1. y(τ )} ∈ Z[τ.9. and for every τ ∈ [t0 . Corollary 1.12. 1. L]L(·) ∈ L}).10. The principal result of the singular perturbations method applied to the guaranteed estimation problem discussed here is formulated as follows Theorem 1.24) for any τ ∈ [t0 .22) X [τ ] ⊆ Πx (∩{Z[τ.12.12.23) Πx (∩{Z[τ.21) that are without any state constraints: Theorem 1.12. Then for every τ ∈ [t0 .18) to be true.12.12. L]L(·) ∈ L}) .22). whatever is the function L(·) ∈ L. w0 } ∈ Z0 .12.
in the ﬁrst Sections of Part 1V. which is usually taken to be the ”Chebyshev center” of set X (t).1) x(t) = u + f (t) .12. whether the results of the last few Sections could again be interpreted in some conventional way. = min max{ x − z z The Chebyshev center of a set X is the center of the smallest Euclidean ball that includes X . who has managed to reach these lines. The approximate calculation of Chebyshev centers is generating an increasing literature.25) x max{ x0 (t) − z z z ∈ X (t)} = x ∈ X (t). as the solution to the problem (1. [225]. where it will be further modiﬁed to suit the related ellipsoidal techniques. z ∈ X (t)}.[225]. . These questions are discussed further.13. [86]. [209]. recall that we have considered the system (1. in terms of DP . may require to ﬁnd the worstcase estimate of x(t) as a vector x0 (t). for example. [139].6.2) the unknown inputs (1. A less investigated problem is to ﬁnd the Steiner center.13. [69].13. [88]. Its calculation leads to mathematical programming problems of special type. t0 ≤ t ≤ t1 . Namely. The conventional theory of guaranteed state estimation as introduced in [181]. may be curious to know .13 Synopsis We shall now summarize the results of the previous Sections. as in Section 5 .The application of this theorem to the calculation of information sets will be illustrated in Section 4. namely . The interested reader. [275] of set X [t]. u ∈ P(t) .3) the initial state 83 f (t) ∈ Q(t) . ˙ with constraints on the controls (1. 1.
II System with input uncertainty and no state constraints (Sections 1.3 ): f (t) . Among the problems of control and estimation for this system we have singled out ﬁve for detailed treatment.3).unknown but bounded. 1. G(t)x ∈ Y(t) . G(t) ∈ L(IRn .2 . Y(t) ≡ IRm .13.11) : f (t) . but bounded. Y(t) ≡ IRm . otherwise.unknown. but bounded. equivalently.3). initial states and measurement noise : control u(t) . Q(t) ≡ f (t). input f (t) .unknown. IV’ System with measurement output (Section 1.same as in III.7 ): f (t) .14.given . IRm ). Y(t) ∈ convIRm .4) and the state space variables (1. due to (1. x(t0 ) = x0 . Y(t) ∈ convIRm and matrixvalued function G(t) taken to be continuous in t. Q(t) ∈ convIRn . 84 .9): f (t) .given. to demonstrate the suggested approach. state constraint given in the form y(t) ∈ G(t)x + K(t) or.absolutely continuous in t.Q(t) = f (t) . f (t) is allowed to be measurable in t). whenever Q(t) = f (t) is reduced to a singleton f (t).13. and with continuous in time set–valued functions P(t) ∈ convIRn .(1. Y(t) ∈ convIRm .1.singlevalued.6.3).given. due to (1. III System with state constraint but no uncertainty (Section 1.13.13. IV System with uncertainty and with state constraints (Section 1.12). due to (1. ( Recall that the presumed property of Q(t) being continuous is translated into the presumption that f (t) is continuous. These are the following I System with no input uncertainty and no state constraints ( Sections 1. with uncertainty in the inputs.5) x0 ∈ X 0 .
K(t) is the bound on the measurement error. (X [t] ∩ Y(t) + σP(t) + f (t)) + 0. namely. or equivalently.y(t) ∈ Y(t) . σ→0 (1. (X [t] + σP(t) + f (t)) ∩ Y(t + σ).13. σ→0 under condition (1.13. 85 σ→0 .6) for case II X [t0 ] = X 0 . X [t] + σ(P(t)−(−Q(t)))) = 0. X [t] + σP(t)) = 0. These were given through the solutions of the following evolution ”funnel” equations with setvalued solutions. X [t] + σP(t) + σf (t)) = 0. X [t] ∩ Y(t) + P(t) + f (t)) = 0.13. for case IV lim σ −1 h+ (X [t + σ] − σQ(t). Y(t) = y(t) − K(t).6) . σ→0 under (1. for case I lim σ −1 h(X [t + σ]. for case III σ→0 lim σ −1 h(X [t + σ]. The ﬁrst issue discussed was the calculation of the attainability domains and the attainability tubes . where y(t) is the available measurement.6) . ˙ lim σ −1 h+ (X [t + σ]. or lim σ −1 h+ (X [t + σ]. σ→0 lim σ −1 h+ (X [t + σ] − σQ(t).
The respective attainability domains are given through the respective unique solutions to the evolution equations when these are written in terms of the Hausdorﬀ distance h(·. [169] and also [171]. Both equations are considered under conditions (1.·) and through the maximal solutions ( with respect to inclusion) for the equations written in terms of the Hausdorﬀ semidistance h+ (·.Krasovski under the name of extremal aiming strategy with solution tubes W [t] being referred to as ”bridges”. σ→0 (if the function Y(t) = y(t) − K(t) is absolutely continuous in t ) or lim σ −1 h+ (X [t] + σ.under condition (1. 86 . dt dt u=U0 u=U (1. Here the objective is to reach a preassigned terminal target set M at given time c t = t1 by selecting a feedback control strategy U(t. particularly. x) = min : U ∈ UP .12. x) ∈ UP which in general turns to be nonlinear. The second group of issues consists of problems of goal . if K(t) is continuous and y(t) is piecewise continuous from the right). The overall synthesized system is then described by a nonlinear diﬀerential inclusion.6) .N. (X [t] + σ(u(t) + y(t)) + σQ(t)) ∩ K(t)) = 0.13. x) = d2 (x. see [168].oriented nonlinear control synthesis. W[t]) and W [t] is the crossection of the respective solvability tube. ·).IV this strategy U = U0 (t. x) V (t. for case IV’ lim σ −1 h(X [t + σ]. x) is selected in a standard way by minimizing the derivative d d c V (t. X [t] ∩ ((y(t) − K(t)) + σ(u(t) + Q(t))) = 0.The attainability domain for case IV’ is the informational domain for the guaranteed state estimation problem of Section 1. 9 9 This type of solution was introduced by N. as the controls are bounded here by magnitude bounds. σ→0 (if the function Y(t) is upper semicontinuous in t.6) . For each of the system types I .13.7) where V (t.13.
·) ∈ UP } + for case I. x) or V = V0 (t. c V ∗ (t. The control strategies are then determined from relation (1. x) may also be calculated directly.8.14. M)U(·. namely. f (·) ∈ Q(·) + + for case IV and with further application of (1. II and W[t] = {x : V0 (t. x) = min max{h2 (x(t1 . by solving for the respective problems the respective HJB equations with value functions c V ∗ (t. t0 . x) u(·) ∈ P(·) + + for case III. x) ≤ 0} for cases III. x) U(·. x) ≤ 0} for cases I. x) = min{h2 (x(t1 . x) = min max U f t c h2 (x(τ.1. t0 .1. ·) ∈ UP .1.13. Y(τ ))dτ + h2 (x(t1 . W[t] = {x : V ∗ (t.10. x). IV. x). f (·) ∈ Q(·)} + U f for case II.6) taken for the corresponding value functions.14. M)U(·. but. t0 . 87 . x). t1 V0 (t. For systems of type (1.1) the respective solvability sets W[t] will turn to be level sets for the corresponding value functions. t1 V0 (t.The strategy U(t. x) = min t h2 (x(τ. t0 . as indicated in Sections 1.11.5. Y(τ ))dτ + h2 (x(t1 . ·) ∈ UP . without introducing the tube W[t].6) to V = V ∗ (t.
7).9) or (1.13.13. σ→0 σ→0 σ→0 σ→0 σ→0 lim σ −1 h+ (W [t − σ]. (1. (1.13.The ability to calculate the solvability tubes W(·) eliminates the necessity to solve the HJB equation .10) for case III (1. ˙ lim σ −1 h+ (W [t − σ].13.12) for case IV (1. 88 .13) and the maximal solutions to equations (1.11) or (1. All of these equations have to be solved with boundary condition (1.11).10) with boundary condition (1.13.13.12) with same boundary condition give us the respective solvability tubes W[·] that produce the the crucial elements W[t] for calculating the required control strategies U(t.13.8).13.13) lim σ −1 h(W [t − σ]. σ→0 lim σ −1 h+ (W [t − σ] + σQ(t). W [t] − σP(t) − σf (t)) = 0. (W [t] − σP(t) − σf (t)) ∩ Y(t − σ)) = 0. x). but should be now taken in backward time. W [t] − σP(t)) = 0. The unique solutions to equations ( 1. Namely.8) for case II (1. lim σ −1 h(W [t − σ].13. W [t] ∩ Y(t) − σP(t) − σf (t)) = 0.13.13. The speciﬁc emphasis is that these tubes may be calculated through evolution equations which are precisely the ones introduced for the attainability domains . W [t] − σ(P(t)−(−Q(t))) = 0.13. W [t] ∩ Y(t) − σP(t)) = 0. (1.14) W [t1 ] = M.13. lim σ −1 h+ (W [t − σ] + σQ(t).13.9). we have introduced the following equations: for case I (1.(1.
This book indicates ellipsoidal approximations as an appropriate technique.18) for case IV (1.7)(1.13. where distance h is substituted by semidistance h+ ).7) is also the maximal solution to a modiﬁcation of this equation.19) W [t − σ] ˙ W [t] ∩ Y(t) − σ(P(t)−(−Q(t))) − σf (t).17) or (1. where σ −1 γ(σ) → 0. The last relations indicate that the basic setvalued operations for the topics of this book ˙ are the geometrical ( ”Minkowski”) sums (+.( The solution to (1.13.Since an arbitrary convex compact set is an inﬁnitedimensional element ( that may be identiﬁed with its support function.Needless to say.(1.13.13. equations (1. 10 .13.13. The indicated relations describe ﬁrst order approximations to the exact set .13.12) may serve to be the motivation and the basis for introducing discretized schemes with setvalued elements. W [t − σ] ˙ W [t] − σ(P(t)−(−Q(t))).8). W [t − σ] (W [t] − σP(t) − σf (t)) ∩ Y(t − σ) W [t − σ] W [t] ∩ Y(t) − σP(t) − σf (t). that equations (1.13.14) .13. we may loosely assume: for case I (1.18) are true relative to an error of order γ(σ). for example). The theory of secondorder approximations to solution tubes for diﬀerential inclusions was discussed in paper [308] 10 89 .11) are particular cases of equation (1.) and diﬀerences(−) of convex compact sets as well as their intersections (∩).16) for case III (1.13.13. W [t − σ] W [t − σ] − σP(t) − σf (t). σ → 0.7) is a particular case of (1.13.9).13.dimensional approximations. Now it should be probably clear.12) and equation (1. In other words.valued solutions of the above.10). The ”equalities” (1.(1.13.13. the respective numerical calculations require ﬁnite. (1.15) for case II (1.13.
among others. W[t] = ∪E− (t). W[t] are exact and are true for the solutions of each of the evolution equations indicated in Section 1. Each of these . W[·] may be presented ( for all the cases I . the setvalued attainability and solution tubes of the previous Sections will be further approximated by ellipsoidalvalued functions. Particularly. The ellipsoidal calculus suggested further yields. that the speciﬁc methods selected in the sequel are based on an ellipsoidal technique that would allow to approximate the setvalued solutions of the above by ellipsoidalvalued solutions. W[t] would thus be parallelized into an array of identical problems each of which would consist (β ∗ ) (α∗ ) in solving an ODE that describes an ellipsoidalvalued function E+ (t) or E+ (t). for all the cases I IV indicated in this Section). It is probably not unnatural . Moreover .14 Why Ellipsoids ? The aim of this book is to indicate some constructive techniques for solving problems of estimation and feedback control under setmembership uncertainty and state constraints with an aspiration that these techniques would allow eﬀective algorithmization and computer animation. W[t] of the convex compact valued functions X [·]. E− (t). the ellipsoidal calculus used here also allows eﬀective internal ellipsoidal approximations in the form of (1. The calculation of X [t]. (that is. therefore.IV under consideration) in the form of intersections (1. in the absence of state constraints. 90 . the ability to address the following issues: (i) The exact representation and approximation of attainability domains for linear systems with or without state constraints through both external and internal ellipsoids. the basic mathematical tool for describing the class of problems raised here is setvalued calculus. The ellipsoidal representations (1. Technically one of the basic justiﬁcations for such an approach is that crossections X [t]. ( cases I.14.14. again stands for the elements of an inﬁnite denumerable variety of ellipsoidalvalued functions described by ordinary diﬀerential equations.14.1) X [t] = ∩E+ (t) . W[t] = ∩E+ (t) (α) (β) (α) (β) over a parametrized inﬁnite variety of ellipsoidalvalued functions E+ (t).1) for X [t].1. may be calculated by solving a system of ordinary diﬀerential equations ( ODE’s). (γ) (δ) where as before the dash stands for the closure of the respective set and where (γ) (δ) E− (t).2) X [t] = ∪E− (t). in its turn. E+ (t) ( which may even be assumed denumerable). As we have seen in the above . II).14.
3.an object far more complicated than the standard setvalued ( ” Aumann” ) integral that represents similar tubes in the absence of input uncertainty. has its advantages and disadvantages. The number of available processors would then determine the accuracy of the solution. by polyhedralvalued functions. etc. (iii) The exact ellipsoidal representation or external approximation of the informational domains for guaranteed ( setmembership) state estimation under unknown but bounded errors. (1.(ii) The treatment of attainability and solvability tubes X [t].14. more generally. 91 .2). (iv) The possibility to single out ”individual” external or internal approximating ellipsoids that are optimal relative to some given optimality criterion ( trace.14.Pontryagin . ( We address the reader to [187]. my be particularly described by alternated integrals of L.1). the tubes are nondominated ( ” inclusionmaximal”) among all such ellipsoidal. [175]. given by alternated integrals or by corresponding evolution equations of the ” funnel type” . The strategies will then be given in the form of analytic designs rather than algorithms as it is in the exact case. [73].valued functions. However. that one of the options is to approximate the setvalued functions X [t]. Thus. W[t] under setmembership uncertainty ( counteraction ) in the inputs . the representations of type (1. each ellipsoidalvalued function could be treated through a single processor which solves an ODE of ﬁxed dimension. (see Sections 3. of course. the respective tubes. of course. in general . [75]).8).11 Loosely speaking. it appears that the main diﬃculty lies in the fact that computational complexity is such 11 the possibility of exact representations and of vectorvalued criteria for approximating attainability domains under state constraints by external ellipsoids was indicated in monograph [181]. It is obvious . in theory. (1. They also allow to apply vector .valued criteria to the approximation problem. The application of ellipsoidal techniques will further allow to devise relatively simple control strategies for control synthesis that would ensure guaranteed results for the related problems.5. an inﬁnite ( denumerable ) variety of ellipsoids would produce the exact relations (1.14. volume. still allow exact internal and external ellipsoidal representations. diameter.14. lies beyond the scope of the present book.S.1) . ) or a combination of such criteria. W[t] by using boxes or . These. so that . the more accurate will be the approximation. These two properties allow to demonstrate that the δ tubes E− (t) possess the property of being an ”ellipsoidal bridge” similar to the ”Krasovski bridge” of the exact solution. The important feature that allows such ”ellipsoidalbased” analytical designs is that the multivalued mappings that generate the internal ellipsoidal δ tubes E− (t) for the solvability sets W[t] satisfy a generalized semigroup property on one hand. [225].2) mean that the more ellipsoids are allowed to approximate X [t]. The minimal volume criteria for these problems was thoroughly studied in [277]. which. This approach. W[t] ( in practice this depends on the number of available processors ). and on the other. Nevertheless.
P(t). these functions are precisely the test functions used in deﬁning the generalized ”viscosity” solutions of the HJB equation. Y(t).1).1).(1.(1.(1. x) to the HJB equation ( its level set is the attainability domain) is approximated by quadratic functions whose level sets are nondegenerate ellipsoids.1.2. There the nondiﬀerentiable solution V (t. It also appears useful to remark that any convex compact set Q in IRn may be presented as an intersection of ellipsoids Q = ∩E (σ) (This fact is a consequence of an ellipsoidal separation theorem .2.2). 92 . X 0 that deﬁne the preassigned constraints (1.9.6. A natural desire will then be to parallelize the polyhedral approximation into problems of smaller dimensions. (v) Another motivation for using ellipsoids comes from Section 4.3) on the system to be ellipsoidal valued. At the same time.the property that every point x ∈ Q may be separated from Q by an ellipsoidal surface ). Q(t).that the number of elementary operations here increases exponentially with the number of steps in the sampled problem. The last fact justiﬁes that we further take all the sets X 0 .
mathematical morphology and other areas. particularly. these are the geometrical (Minkowski) sums and diﬀerences of two nondegenerate ellipsoids with the diﬀerence having a nonvoid interior.2. [263]. we come to external and internal approximations of the sums and diﬀerences (Sections 2.Part II. With the number of approximating ellipsoids increasing to inﬁnity. probability and statistics. The external representations are given by intersections and the internal by unions ( or their closures) over the respective varieties each of which is inﬁnite. Taking intersections or unions over some ﬁnite subsets of these varieties. THE ELLIPSOIDAL CALCULUS Introduction This part is a separate text on ellipsoidal calculus . The operations on ellipsoids are discussed in the following order. Moreover.both externally and internally . we make take only one element of the respective variety.3). at least.1). denumerable. A similar fact is true for the internal ellipsoidal approximations of the sums and the external ones of the diﬀerences (Sections 2. the sum of its axes or their squares (the trace of the matrix deﬁning the ellipsoid). However. A certain reciprocity consists in the fact that the external ellipsoids for the sums and the internal ones for the diﬀerences are given by the same type of parametrization which diﬀers in both cases only in some signs in the representation formula. identiﬁcation and experiment planning.2. to exact representations. [181]. The obtained representations are then propagated to ﬁnite sums of nondegenerate ellipsoids and to setvalued integrals of ellipsoidalvalued functions E[t] ( which are not obliged to be ellipsoids). The operations treated here are motivated by the requirements of part I of the present book. which is optimal in some sense ( an array of possible optimality criteria is discussed at the end of Section 2.a technique of representing basic operations on ellipsoids.5). of course. These sums and integrals are again approximated externally and internally. Then the sums and diﬀerences will be approximated (internally or externally) by an optimal ellipsoid (Section 2. A widely studied criterion is the volume of the ellipsoid ( see [277]. stabilization. [73]). ([111]. adaptive control. These criteria may include its diameter. if the upper limit of the setvalued integral X [t] = t t0 E[τ ]dτ 93 . in the limit.by a corresponding parametrized variety of ellipsoids. the approximations converge.3). Each sum and diﬀerence is approximated . 2. [225]) etc. Particularly. First of all . in optimization and approximation. the results given here may be applied. to a substantially broader class of problems that arise in mathematical modeling.2.
Let us elaborate on the ﬁrst two of these operations presuming that the sets involved are convex and compact. 2. (Also note the theory of analytical centers .1 Given sets H1 .6 . H2 ∈ comp IRn . We also believe that one should not drop the problem of ﬁnding perhaps rough. Eﬀective algorithms for estimating the errors as well as the computational complexity of these problems are among the issues that present a further challenge. Intersections of ellipsoids are the topic of Section 2. [281]).8). they are inclusion minimal for the external ellipsoids and inclusionmaximal for the internals. but simple error estimates. Deﬁniton 2.1. The important element here is that these ellipsoidalvalued functions that approximate X [t] are ”nondominated” with respect to inclusion. The construction of internal ellipsoidal approximations to polyhedral and other types of convex sets are important particularly in algorithmic problems of mathematical programming ( [152]. the geometrical (Minkowski) sum H1 +H2 is deﬁned as H1 + H2 = {h1 + h2 }. [286]).1 Basic Notions : the Ellipsoids As we have seen in Part I.varies.4 we also mention a direction towards the calculation of approximation errors ( depending on the number of approximating ellipsoids). 2. h1 ∈H1 h2 ∈H2 94 . • the intersection of convex sets. to be consistent with the approach presented here. An indication is ﬁnally given on how to construct varieties of internal ellipsoidal approximations of intersections of ellipsoids. The solution to these is usually given in the form of an algorithm. one must be able to indicate a variety of internal ellipsoids whose union (or its closure) would approximate a nondegenerate intersection of ellipsoids ( ” from inside ”) with any desired degree of accuracy. • the geometrical (Minkowski) diﬀerence of convex sets. the basic setvalued operations involved in the calculation of solutions to the control problems of the above are the following: • the geometrical (Minkowski) sum of convex sets. In Section 2. • the aﬃne transformations of convex sets. Namely. However. then the parameters of the ellipsoidal functions that approximate X [t] may be described by ordinary diﬀerential equations ( Sections 2. Several types of external ellipsoidal approximations are described here with exact representations in the limit.7.
1.3)). The following properties may happen to be useful 95 ∀l ∈ IRn .2) namely. Otherwise. for all h2 ∈ H2 }. the support function ρ( H1 + H2 ) = ρ(lH1 ) + ρ(lH2 ). the geometrical ˙ ( or Minkowski or also “internal”) diﬀerence H1 −H2 is deﬁned as ˙ H1 −H2 = {h ∈ IRn : h + H2 ⊆ H1 }. if f (l) were convex.1) ˙ H1 −H2 = h2 ∈H2 h1 ∈H1 {h1 − h2 }. What follows from here are the assertions ˙ Lemma 2. ˙ Lemma 2.1.1.1. ˙ H = H1 −H2 if and only if H + H2 ⊆ H1 and H + H2 ⊆ H 1 imply H ⊂ H.3) ρ(lH) ≤ ρ(lH1 ) − ρ(lH2 ) = f (l). Deﬁniton 2. so that. ρ(lH) = ( co f )(l) = f ∗∗ (l) where co f is the “lower envelope” of f (l).Obviously.2 The geometrical diﬀerence H = H1 −H2 is the maximal convex set (with respect to inclusion) among those that satisfy the relation (2.1. ˙ This means H1 −H2 = ∅ if there is an element h ∈ IRn .1. ρ(lH) = ρ(lH1 ) − ρ(lH2 ) (since H is the inclusionmaximal set that satisﬁes (2. Clearly ˙ H1 −H2 = {h ∈ IRn : h ∈ H1 − h2 .2 Given sets H1 . [100]. In terms of support functions the inclusion (2.1. H + H2 ⊆ H1 .2) yields (2. H2 ∈ comp IRn .1 The set H1 −H2 may be presented as (2.1. we would have . such that h + H2 ⊆ H 1 .
96 .3 With H1 . H2 . H3 ∈ compIRn we have (2.Lemma 2.1.4) (2. Let us indicate an example which would illustrate that in the last relation both an equality and a strict inclusion are possible.1.1.5) ˙ ˙ ˙ H1 −(H2 + H3 ) = (H1 −H2 )−H3 ˙ ˙ (H1 + H2 )−H3 ⊇ H1 + (H2 −H3 ).
obviously H = {(x. y = 0} Then H1 + H2 = and one may observe ˙ (H1 + H2 )−H3 = H1 . y) ∈ IR2 : (x + 1)2 + y 2 ≤ 4} {(x.1.5) is therefore a strict inclusion. The relation (2. the set H is the intersection of the sets {(x. y = 0} Then H1 + H2 = S2 (0) ˙ ˙ (H1 + H2 )−H3 = S2 (0)−H3 = H. clearly. clearly. H2 −H3 = {0}. according to the deﬁnition of the geometric ˙ diﬀerence. y) ∈ IR2 : x2 + (y + z)2 ≤ 1  z ≤ 1} . z ≤ 1}. Therefore.1 Assume Sr (0) = {(x.2 Take H1 = {(x.1.1. where. y) ∈ IR2 : x ≤ 1. ˙ H2 −H3 = {0}. y) ∈ IR2 : (x − 1)2 + y 2 ≤ 4}.Example 2. ˙ On the other hand. y ≤ 1} H2 = S1 (0) H3 = {(x. y) ∈ IR2 : −1 ≤ x ≤ 1. y) ∈ IR2 : (x + z)2 + y 2 ≤ 4. y) ∈ IR2 : x = 0. On the other hand. H1 + (H2 −H3 ) = H1 + {0} = S1 (0) and S1 (0) ⊂ H. In other words. 97 {(x. y) ∈ IR2 : x2 + y 2 ≤ r2 } H1 = S1 (0) H2 = S1 (0) H3 = {(x. Example 2.
5) is an equality. Q)) = (l.1. where its support function ρ(lE(a.1. Deﬁniton 2. Q) = {x ∈ IRn : (l. ˙ The convex compact set H = H1 −H2 is deﬁned to be inclusionmaximal relative to those convex compact sets that satisfy the relation H2 + H ⊆ H1 . Our further aim is to introduce an ellipsoidal calculus that would allow to approximate the above relations for convex compact sets through ellipsoidalvalued relations. The external diﬀerence may also be deﬁned as the class H = {H} of all sets H ∈ conv (IRn ) that satisfy the inclusion (2.and ˙ H1 + (H2 −H3 ) = H1 . x) ≤ (l. 1 1 ∀l ∈ IRn . It is not diﬃcult to prove that the internal diﬀerence is unique.1. and it is possible to demonstrate that inf{ρ(lH)H ∈ H} = ρ(lH1 ) − ρ(lH2 ) ≥ ρ( He ) (since H1 . H2 ∈ comp (IRn ) and l ∈ IRn is ﬁnitedimensional). Deﬁniton 2. Then ρ(lH2 ) + ρ(lH) ≥ ρ(lH1 ) ∀H ∈ H. Q) is deﬁned by the equation ρ(lE(a.1.1. Ql) 2 . 98 .4 An ellipsoid E(a.6) H2 + H ⊇ H1 . a) + (l.6). Ql) 2 .3 A set He will be deﬁned as an external diﬀerence He = H1 ÷ H2 if it is inclusionminimal relative to those convex compact sets that satisfy the relation (2. Q) with center a ∈ IRn and “conﬁguration” matrix Q (symmetric and nonnegative) is deﬁned as the set E(a. a) + (l. being therefore an internal diﬀerence. ∀l ∈ IRn }. In this case the inclusion (2.
E2 = E(a2 . Inclusionminimality is deﬁned similarly. E+ are not unique. 12 We will now proceed with the following basic setvalued operations. Q) could also be presented otherwise. [120]. which gives a direct. applying them to ellipsoids The geometrical (Minkowski) sum. E2 ∈ comp IRn the geometrical (internal) diﬀerence ˙ E1 −E2 = {x : x + E2 ⊆ E1 } is unique. Q1 ). it is not diﬃcult to observe that E+ . their sum E1 + E2 may obviously not be an ellipsoid (ﬁnd an example). in terms of the inequality E(a. E+ of the sum E1 + E2 where (+) E+ ⊇ E 1 + E2 is an external approximation and E+ ⊆ E 1 + E2 is an internal approximation. An ellipsoid E0 will be inclusionmaximal relative to the variety E0 . We will (+) (−) therefore be interested in ellipsoidal approximations E+ .5 Let E0 denote a certain variety of ellipsoids E. For two given ellipsoids E1 . 12 (−) (+) (−) 99 . The geometrical (Minkowski) diﬀerence. We will further indicate the inclusionminimal external approximations and the inclusionmaximal internal approximations of E1 + E2 . Q2 ). Deﬁniton 2. This representation is also true for degenerate matrices Q but then Q−1 does not exist and has to be substituted by the MoorePenrose pseudoinverse for Q. conventional description. E0 ⊆ E imply the equality E0 = E. if theinclusions E0 ∈ E0 .With Q nondegenerate.1. Q) = {x ∈ IRn : (x − a) Q−1 (x − a) ≤ 1}. Therefore we shall further describe a rather ”complete” parametrised variety of such ellipsoids. However it may not be an ellipsoid (give an example). As we shall see. Given two ellipsoids E1 = E(a1 . the ellipsoid E(a.
Obviously (−) (+) ˙ E− ⊆ E1 −E2 . (It is now the family of sets in dRn and the ordering is inclusion).1. its intersection E1 ∩ E2 in general is not an ellipsoid. We will again discover that these are not unique and will try to describe a rather complete variety of such ellipsoids. This interpretation naturally follows from Deﬁnition 2. The Intersections. ˙ while an internal ellipsoidal approximation of the diﬀerence E1 −E2 is an ellipsoid E− that satisﬁes the inclusion (−) E− + E2 ⊆ E1 .1. We will be interested ﬁrst of all in its external approximations E + ⊃ E 1 ∩ E2 seeking for example the inclusionminimal (nondominated) ellipsoids. namely a nonsymmetrical set relative to any point or plane or even a degenerate convex set in the sense that if taken in IRn it will have no interior point. the inclusionmaximal approximations E− and the (+) inclusionminimal approximations E− are not unique. A relatively simpler situation occurs when the centers of E1 and E2 coincide. E2 ∈ comp IRn .6 An external ellipsoidal estimate of the diﬀerence E1 −E2 will be deﬁned (+) as an ellipsoid E− that satisﬁes the inclusion ˙ E− ⊇ E1 −E2 . We will further be interested in the inclusionmaximal internal and the inclusionminimal (−) ˙ ˙ external ellipsoidal approximations E− . E −E2 of the diﬀerence E1 −E2 . A more diﬃcult problem is to ﬁnd an internal ellipsoidal approximation E − ⊂ E 1 ∩ E2 to the intersection.˙ Deﬁniton 2. Indeed the intersection may easily turn to be a convex set of rather general nature.5. They could also be interpreted as nondominated elements of a (partially) ordered family. Given E1 . As we shall observe in the sequel. (−) (−) 100 .
l) + (l.1. but the inclusion H ⊂ E(0.1. Q). Since H. we have ρ(lH) < (a. .8) ρ(lH) ≥ (l. l) > 0. the property ρ(lH) > (l. Then there exists a set L = ∅ such that L = {l ∈ IRn : ρ(lH) > (l. Q) implies H ⊂ E(0. l ∈ −L which contradicts with (2. Then (a) the inclusion H ⊂ E(a. To prove assertion (b) we observe (2. Q) and (b) H ⊃ E(a. Q).7). Q) does not hold. Ql) 2 . (a.1. a) + (l. Q) implies H ⊃ E(0.4 Suppose that a set H ∈ comp IRn is symmetrical.Aﬃne Transformations As an exercise one may easily check that the inclusion x ∈ E(a. l∈L l ∈ IRn . Since H ⊂ E(a. l) > 0. Proof.7) and also the inequality (a. The assertion (a) is thus proved. Q) is equivalent to the following Ax + b ∈ E(Aa + b. AQA ) Let us now indicate two useful properties of symmetrical sets. l = 1}. namely H = −H. Q) are symmetrical. Q). 101 1 1 1 1 l ∈ IRn .1. Suppose H ⊂ E(a. Ql) 2 . Ql) 2 will also hold for l ∈ −L. Due to the previous supposition this implies (2. E(0. Lemma 2. Ql) 2 .
As we have already mentioned. . so that χk = ϕn−k (Q). We will therefore indicate a class Ψ = {ψ(E(a. .Since H = −H we have ρ(lH) = ρ(−lH) for all l ∈ IRn . n 1 1 l ∈ IRn .1. Let σ(Q) = {λ1 . Out of all the possible approximating ellipsoids we will prefer to select the inclusion minimal or maximal ellipsoids observing that these “extremal” ellipsoids are the nondominated elements (relative to inclusion) of the respective varieties.1. Q)} and nonnegativevalued. Ql) 2 . Q))} of criteria functions ψ(E(a. . k = 1. may not be ellipsoidal at all. The assertion is thus proved. Q)) that would be (a) deﬁned on the set of all nondegenerate ellipsoids {E(a. Among the nondominated varieties of inclusionminimal or inclusionmaximal ellipsoids we may then want to single out some individual elements that would be optimal relative to some prescribed optimality criterion. for the coeﬃcient of the (n − k)th degree term of its characteristic polynomial. Together with (2. Q1 ) ⊇ E(a2 . λn } denote the set of eigenvalues of Q. (We shall generally also require the monotonicity property (b) to be invariant relative to aﬃne transformations of ellipsoids. Ql) 2 . χ(λ) = k=0 χk λk . In the latter case we will introduce internal and external ellipsoidal approximations of these sets. The results of these operations are convex sets which may either again be ellipsoids or what is more common.8) the latter inequality yields ρ(lH) ≥ (l.Then for all m ∈ N we have 102 . . Lemma 2. Therefore ρ(lH) ≥ (−l.5 Suppose E(a1 . . l ∈ IRn . . n. our objective is to add and subtract ellipsoids ( in the “geometrical” sense) and also to intersect them and to apply aﬃne transformations. Q2 ). (b) monotonous by increasing with respect to inclusion: ψ(E1 ) ≤ ψ(E2 ) if E1 ⊆ E2 .) Let Q stand for a symmetric positive matrix and ϕk (Q). . a) + (l. .
The latter inclusion yields Q1 ≥ Q2 . with diagonal elements λi . Given E(0. is actually the sum of the squares of the semiaxes of E(0. . n we observe that if Q1 ≥ Q2 . From Lemma 2. Ql) 2 . Having two positive n × n matrices Q1 . . where T QT is diagonal. Q2 ). Q) into E(0. . + λn . Q1 ) ⊇ E(0. (Here the transformation Q → T QT keeps the eigenvalues of T QT the same as of Q and the lengths of the semiaxes of E(0. . T QT ). n). . . Q2 ).(i) ϕk (Qm ) ≥ ϕk (Qm ). where e(i) = (e1 . Q2 with respective eigenvalues λ1 ≤ λ2 ≤ . x) ≤ 1}. Q2 ). 1 2 (iii) min{σ(Qm )} ≥ min{σ(Qm )} . Q1 ) ⊇ E(a1 . Q)) = 1 (l. Q) is ρ(e(i) E(0. 1 so that the length of the ith semiaxis of E(0. is the ith orth in the orthogonal coordinate n space of IRn . (1) (2) (2) (2) (1) (1) E(0. . Q) (a) The trace: ψ[Q] = tr(Q) = ϕn−1 (Q) = λ1 + . then. The latter property yields the assertions of the Lemma. 1 2 Proof. ej = δij . [120]. λi ≥ λi . . . . then E(a1 . 103 . a canonic orthogonal transformation T x = z (T  = 0) transforms E(0. . 1 2 (k = 1. e(i) ). with support function ρ(lE(0. T QT )) = ( i=1 2 λi li ) 2 . i = 1. Q)) for the “size” of an ellipsoid E(0. Q1 ) ⊇ E(a2 . .4 it follows that if E(a1 . ≤ λ(2) . Q). Q). . . Q)). ≤ λ(1) n and λ1 ≤ λ2 ≤ . . Let us now indicate some common type measures ψ[Q] = ψ(E(0. (ii) max{σ(Qm )} ≥ max{σ(Qm )} . . T QT )) = (i) (i) λi . Thus n ρ(lE(0. T QT ) the same as of E(0. . . Therefore tr(Q) is equal to the sum of the squares of the semiaxes of E(0. n.1. Q) = {x ∈ IRn : (Q−1 x.
5 as well as their positive combinations are also monotonous with respect to inclusion. .·λn = ϕ0 (Q) is proportional to the volume vol (E(0. Q) . tr(Q) and tr(Q2 ). Q). . vol (E(0. . Q). Here the value max{λi ∈ IR : i = 1. . (c) The product ψ[Q] = λ1 ·λ2 ·.1. Q). Q)) = Π 2 ( det Q)(Γ( n n + 1)−1 2 where Γ stands for the gammafunction. This indicates the range of cases that we are able to handle. Indeed a direct calculation yields . Q)) is the diameter of E(0. where d = d(E(0. Q)) of E(0. . . . Q)). This follows from the fact that d/2 is equal to the length of the largest semiaxis of E(0. It is obvious that monotonous functions of these appearing in Lemma 2. However we shall formulate our results primarily for vol E(0. 104 . One just has to recall that the determinant det Q of Q is equal to the product ϕ0 (Q) = λ1 · . We shall now specify some parametrized varieties of ellipsoids that allow to approximate the geometrical sums and diﬀerences of ellipsoids and even to give an exact representation of these.so that d/2 is the radius of the “smallest” ndimensional ball that includes E(0. .(b) The “trace of the square” yields a criteria ψ[Q] = tr(Q2 ) = ϕn−1 (Q2 ). [213]. n} = d 2 2 . Q). · λn . (d) The diameter : ψ[Q] = d(E(0.
Consider a parametric family of matrices Q(p) = (1 + p−1 )Q1 + (1 + p)Q2 . l)−1/2 + E1 + E2 ⊂ E(a1 + a2 . 105 .2.2) and (2.2. p > 0 is properly deﬁned and is an external approximation of the sum E1 + E2 . λ1/2 ] max + 1/2 = (1. denote the roots of the equation det(Q1 − λQ2 ) = 0 as λmin = λ1 ≤ λ2 ≤ . Q(p)).2. Conversely.e.2. Given two such ellipsoids E1 = E(a. Denote + − = [λmin . ≤ λn = λmax . (b) With vector l ∈ IRn .1 (a) The ellipsoid E = E(a1 + a2 .2. We will also be interested in the family Q(−p). given.2.2 External Approximations: the Sums. Q(p)) deﬁnes a scalar parameter p ∈ (2. (2.3) are true. with parameter p ∈ + given. (λ1 > 0. Q2 ). i. such that equalities (2. l)1/2 (Q2 l. Q1 ) + E(a2 . . the equality (2. Internal Approximations: the Diﬀerences In this section we will deal only with nondegenerate ellipsoids.3) .1) for any p > 0. l = 1. such that ρ(lE(a1 + a2 . there exists a vector l ∈ IRn . Q2 )). λmin ). λn < ∞). Q2 ∈ L(IRn . Q(p))) = ρ(lE(a1 .2) p = (Q1 l. l = 1. IRn ). Q1 ) and E2 = E(a. . These roots are also said to be the “relative eigenvalues” of the matrices Q1 .2. Lemma 2.
Q(−p)) ⊆ E1 − E2 . i. For these values of p the ellipsoid E is an internal approximation of the diﬀerence ˙ E1 − E2 .2.10. observing that p ∈ + (check the latter inclusion as an exercise.4). [120]). Adding (Q1 l.2. With a further addition of (l. Q. Lemma 2. implies the inclusion (2.2. (b) ˙ E(a1 − a2 .e. Conversely.2.2).2. using the extremal properties of matrix eigenvalues.2).2) and therefore (2.2). l) ≥ 2(Q1 l.2. Q1 )) + ρ(lE(a2 . there exists a vector l ∈ IRn .g. Q(−p)) is a nondegenerate ellipsoid if and only if (2.5) p ∈ (1. l)1/2 where Q(p) > 0 for any p > 0. Q1 ) ⊃ E(0. l)1/2 is obviously true for any p > 0.4) (Q(p)l. with p ∈ + given. we obtain (2. l) to both sides. The inequality p−1 (Q1 l. A similar reasoning passes through for geometrical diﬀerences. To prove the assertion (b).3) is therefore true for any given l with p and l related through (2. see e. l) + (Q2 l. due to the continuity in l of the righthand side of (2. l = 1.2) into (2.2.2. the latter turns into an equality for the given l (this can be veriﬁed through direct calculation). Q(p))) ≥ ρ(lE(a1 .D.2. l)1/2 + (Q2 l. we select the parameter p due to (2. l) + p(Q2 l.Proof. This follows from Theorem 7. a1 + a2 ) to both sides this implies ρ(lE(a1 + a2 . Then (a) E = E(a1 − a2 . Q2 )) for any l ∈ IRn and therefore.2). Q2 ). Chapter X of reference [120]. The equality (2.2 Suppose int E(0. such that (2. 106 . with l ∈ IRn given.2.1.E. After a substitution of (2. l)1/2 (Q2 l.3) do hold. λmin ).2. l)1/2 ≥ (Q1 l.
107 . l 1/2 (l. Q2 )) ≤ ρ(lE(a1 . the latter condition ensures that Q(−p) > 0. there exists a vector l ∈ IRn .8) together with the relation ˙ ρ(lE(a1 . (due to Theorem 7. This yields p ∈ (1. which implies (l.6) turns the latter into an equality. Q2 )). The latter inequality is further equivalent to the inclusion E(a1 − a2 . which means p < λmin . λmin ). Q1 ) − E(a2 . Q(−p)) + E(a2 .1 with p substituted by (−p).6) for the given values of l and p. (l. which is true for any l ∈ IRn and equivalent to ρ(lE(a1 − a2 .2) is fulﬁlled. due to the deﬁnition of the geometrical diﬀerence and the conditions of the lemma. implies ˙ (2. we suppose the parameter p to be deﬁned due to (2. Q1 ) ⊃ E(0. such that equality (2. Q2 ). Q2 ) for any p ∈ (1. As shown above. Q2 ) ⊆ E(a1 . λmin ). in view of the condition p > 1 that Q(−p) is positive deﬁnite if and only if Q1 p−1 − Q2 > 0. Q2 )) then yields equality (2. Q1 l) > (l. Consider the inclusion int E(0. Q1 )) − ρ(lE(a2 . Under these conditions a direct substitution of p into (2.2. Q(−p)) ⊆ E(a1 .) in l). and the continuity of the righthand side of (2.2. we come to the inequality (2.10. Q(−p))) ≤ ρ(lE(a1 . Q1 ) which. Q1 )) − ρ(lE(a2 . The inclusion (2.2) and such that p ∈ − .Proof.2. On the other hand. To prove assertion (b) with l ∈ IRn given. we observe.2. once p ∈ − is given.2. Q1 l)1/2 > 1.6) Q(−p)l. Q2 l) for any l ∈ IRn and therefore implies the condition p= Taking Q(−p) = (1 − p−1 )Q1 + (1 − p)Q2 = (p − 1)(Q1 p−1 − Q2 ). Q2 l)1/2 ≤ Q1 l.2.2. Following the proof along a scheme similar to that of Lemma 2.2.2.2.7) E(a1 − a2 . l 1/2 . Q1 ) − E(a2 . l 1/2 − Q2 l. Chapter X of [120].
k ∈ m + 1. n Suppose in addition that E1 = E(0.8) lj = 0 if j ∈ 1. C) ⊆ E1 − E2 and ˙ ρ(lE(0. Also assume the symbol 1. we have (2. i ∈ m + 1. n lj if j = k −lj if j = k.6) for the given p and l. It can be checked that vectors h(k) are linearly independent. E2 = E(0. Lemma 2. Q) ⊆ E(0. 0 = ∅.D. Q2 ) and that the matrices Q1 . symmetric matrix C with elements {cij } where i stands for the row and j for the column of C. We will now prove the assertion (a). Q. Consider a positive deﬁnite. Q2 . i ∈ m + 1. Then the following implications hold: (a) If E(0. C) ⊆ E1 + E2 and ρ(lE(0. m lj = 0 if j = m + 1. n for all i = j. Q are diagonal. Q) ⊆ E(0. l = 1 and suppose that for some m ∈ [0.2. 108 . n].2.This also yields (2. Q)) = ρ(lE1 + E2 ) then cij = 0 (b) If ˙ E(0. Q)) = ρ(lE1 − E2 ) then cij = 0 . Assertion (b) will be left as an exercise. Proof.3 Fix a vector l ∈ IRn . n. For the given vector l deﬁne an array of vectors h(k) . where hj = (k) for all i = j. Q1 ).E.2. Its proof is similar to that of (a).
i=j cij li lj . what follows is that crs = 0. (j) Qj = diag {q11 . 2. then take (2. r = s.2. r. Repeating this procedure for all the values k ∗ ∈ m + 1. h(k) )1/2 = (C h(k) . h(k) )1/2 = (Q1 h(k) . If m > 0. . j = 1.12) χ(l) = i. The respective terms may now be cancelled out from the righthand side of (2.11) we come to χ(l) = χ(h(k) ) n (2. deﬁne the function χ : IR → IR by the equality n n χ(z) = i=1 (qii − cii )zi2 . . qnn }.10) Assuming Q = diag {q11 . n except for a previously ﬁxed pair r. Q)) − ρ2 (lE(0. n.2.2.2.2.12) for the respective value of k. Here we have for all k ∈ m + 1. C))). qnn }. 2 109 .13) and taking into account the symmetry of C.13) 1 φ(l) = (ρ2 (lE(0. This directly implies crs = 0. .j=m+1. h(k) )1/2 + (Q2 h(k) . s ∈ m + 1.By diagonality of the respective matrices and the equality of supports we have (2. .9) (Qh(k) .12) in its reduced form we can cancel out similar terms for a new value k ∗ = k. s. s ∈ m + 1.2.2. . so that the “last” cancellation yields χ(l) = 0. we have n 2 j=1. (j) (Qh(k) . we ﬁnally come to χ(l) = 2crs lr ls . . .12).2.j=k ckj lk lj = 0.2. (2. Combined with the inclusion relations this implies (2. n. h(k) )1/2 .2. for any r. r = s. Substituting l for l(k) with a ﬁxed value of k into (2. Since r.11) and due to (2. s were chosen arbitrarily. h(k) )1/2 . Taking equation (2. The proof is therefore complete for m = 0. n (2. .
n this yields n j=m+1 cij hj = 0 and since the vectors h(k) .2. . qnn }. The equality (2. n. Also keep the notation of (2.2.4 Consider an ellipsoid E(0. . k ∈ m + 1. j ∈ m + 1. 110 . Q(p)) = E(0.2) yields n i=m+1 2 qii li n = i=m+1 2 cii li . where li = 0. Q(p) = diag {qn . (j) Proof. Q(p)) ⊇ E(0. E2 = E(0. Q(p))) = ρ(lE1 + E2 ). i ∈ m + 1. Denote Qj = diag {q11 . m.E. C = {cij }. . Q2 are diagonal. By diﬀerentiability we necessarily have ∂ϕ(l) ∂l = 0. m. Then the following implications hold: If E(0.2. assuming that Q1 . Further assume the vector l ∈ IRn . n. Together with (2. n.2. . 2. we have (2.14) this implies cii = qi for i ∈ m + 1. Due to the previous Lemma 2. k ∈ m + 1. C) together with ellipsoids E1 = E(0. C). then E(0. n are linearly independent we conclude that cij = 0 for any i ∈ 1. . Q(p)) ⊆ E(0. Q2 ).2) with given l. l=h(k) For all the values of i ∈ 1.14) cii ≤ qii . qnn }. C) ⊇ E1 + E2 and ρ(lE(0.2. . . Q.2. j = 1. C) (j) (k) and p∈ + .The function ϕ(l) has a local minimum at l = h(k) for any k ∈ m + 1. l = 1 to be given and the parameter p to be deﬁned due to relation (2.2. Lemma 2.9) and the deﬁnition of h(k) of the previous theorem.3 (a) and the inclusion E(0.3) with value of p from (2. n. .D.2. Q1 ).
Q(1) l)1/2 (l. qii cii ≥ − (k) . where p = (l.3. we come to the condition qii ≤ cii and therefore to the equality qii = cii . l(k) )1/2 + (Q2 l(k) . i ∈ 1. recalling that Q(p) = (1 + p−1 )Q(1) + (1 + p)Q(2) . or. l(k) )3/2 for all l = 1. k ∈ m + 1.15) cii qii qii ≥ + (k) . multiply and divide it by (Q1 l(k) .2. l(k) )1/2 . l (k) )1/2 (k) . m. In particular. l (k) )1/2 (k) . k ∈ m + 1.15) and using the equality relation for the support functions at l = l(k) . n.2. Q(2) l)−1/2 . namely ∂ ∂l and ∂ ∂l 2 ξ(l) 2 l=l(k) ≥ 0 ζ(l) l=l(k) ≥ 0. Due to the diﬀerentiability of these functions the second order necessary conditions of optimality imply that the matrices of the second order partial derivatives are nonnegative. ζ(l(k) ) = 0. l = l(k) . m.16). If we now observe that the denominators in the resulting inequality are equal (due to the condition of the lemma and being the values of the support functions at l = l(k) ) we may conclude that qii ≥ cii for all i ∈ 1. l (k) )1/2 (Cl (Q1 l (Q2 l (1) (2) Take the righthand side of (2.3 and the deﬁnition of l(k) a similar condition for the matrix of second derivatives of ζ yields the following inequality for the diagonal elements i = 1. l (k) )1/2 (k) . m: (2. C)) and ζ(l) = ρ(lE(0. 111 . this implies that the diagonal elements of the respective matrices are nonnegative. n with ξ(l(k) ) = 0. after a direct calculation. substituting the obtained relations into (2. l (k) )1/2 (Q(p)l (Cl (k) 2 n j=m+1 cij lj (Cl(k) .By the conditions of the lemma both nonnegative functions ξ(l) = ρ(lE(0. m.2.2.2. Here the second term on the righthand side disappears due to Lemma 2. Q(p))) − ρ(lE(0. C)) − ρ(lE1 + E2 ) have local minima at l = l(k) . Due to Lemma 2.
so that we would have E1 + E2 ⊆ E(0. The transformation T obviously does not violate the inclusion E(0. we will ﬁrst prove the following essential Theorem 2.17). Q) to be tangential to E1 + E2 . Q∗ ) ⊇ E(0. Then the set of inclusionminimal external estimates of the sum E1 + E2 will consist of the ellipsoids of the form E(a1 + a2 .2. Q∗ z )1/2 = (z.2. Q(p)) ⊇ E(0. 1 l) into (¯. lE Let us now select an invertible matrix T such that the matrices Q∗ = T Q1 T. Since it was just established that its diagonal elements are all equal to zero. l) l. Given an ellipsoid E(0. C) implies that the matrix Q(p) − C is nonnegative. what follows is that all the rest of the elements must also be zero. [120] ). we may assume all the centers of the ellipsoids considered here to be zero. The existence of such a transformation T follows from results in Linear Algebra and the theory of matrices (see e. Q∗ ). particularly a1 = a2 = 0. with p ∈ + .16) ρ(¯ lE(0. We may obviously consider E(0. E2 = E(a2 . Q∗ z)1/2 + (z. Q) ⊇ E1 + E2 let us indicate that there exists a value p such that the ellipsoid E(0. Q∗ z)1/2 z ¯ 1 2 112 . Q1 ). ˙ However. Q) and E1 + E2 .2. before proving an assertion similar to Lemma 2. Q2 are positive deﬁnite and that Q(p) is deﬁned due to formula (2. Q(p)). assuming the existence of a vector l = ¯ ∈ IRn . Q2 ) the matrices Q1 .1 Suppose that for the ellipsoids E1 = E(a1 . Q(p)) could be “squeezed” in between E(0. referring also to Lemma 2. which is l l. Proof. ¯ = 1. such that l l (2.2. 2 1 Taking the mapping ¯ = T z one may transform the equality (2. Q) ⊇ E1 + E2 . Q∗ ) + E(0.g.4.4. Without loss of generality.2. Q).1).The inclusion E(0. 2 l) (¯ Q¯ 1/2 = (¯ Q∗ ¯ 1/2 + (¯ Q∗ ¯ 1/2 l.2. Q(p)) ⊆ E(0. so that with Q∗ = T QT we still have E(0. but for the diﬀerences E1 −E2 . The theorem is thus proved. Q)) = ρ(¯ 1 + E2 ). Q∗ = T Q2 T 1 2 would both be diagonal.
z ∗ and Ez (0. The following proposition is similar to Lemma 2. Q∗ (p))) z z = ρ(¯Ez (0. Q∗ )).2. Q) on space Z. Q∗ (p))) > ρ(z ∗ E(0. 113 .17) z ρ(¯E(0. E2 = E(0. there would have existed a vector z ∗ such that (2. Q∗ ). Q∗ )). Q.19) we have Ez (0. Q2 ). Lemma 2.4 it then follows that Ez (0. p = (¯.2.18) and the inequality (2. but is applied to geometrical diﬀerences.where z = T −1 ¯ Following (2. Q∗ ) + Ez (0. Q∗ )) z 1 2 = ρ(¯E(0. Q∗ )) + ρ(¯E(0. Q1 ). The vector z ∗ is obviously noncollinear with z . Q∗ ) = Ez (0. Q∗ (¯)) ⊆ E(0. Q∗ (¯)) 2 1 (2.4. z 1 2 From Lemma 2.2. Further assume the vector l = ¯ l = 1 to be given and the parameter p = p to be deﬁned due to relation (2. Deﬁne Z to be the 2dimensional space ¯ generated by z .2.18) ρ(z ∗ E(0. Q∗ z )1/2 (¯. part (a) it now follows that E(0. ¯ l=¯ l.2.2) we may now select ¯ l. Q∗ ) ⊆ E(0. Q2 are diagonal. assuming that Q1 .1.2. Q∗ ) p and ρ(¯Ez (0. C) together with ellipsoids E1 = E(0.4. Q∗ (¯))) = ρ(¯E(0. Q) to be the projection of the ellipsoid E(0.2. Q∗ z )−1/2 ¯ z 1¯ z 2¯ and further take Q∗ (¯) = (1 + p−1 )Q∗ + (1 + p)Q∗ . p ¯ ¯ 2 1 We then come to the relations p E(0. Q∗ )). In ¯ view of (2.E. Q∗ (p)) is in contradiction with (2. p Indeed with if this inclusion being false.D. z p z From Lemma 2. Q∗ (¯)) ⊇ Ez (0.5 Consider a nondegenerate ellipsoid E(0.2.2.2. Q∗ )) = ρ(¯Ez (0.2). l. Q∗ ) + E(0.18).
C)) ≤ ρ(lE1 −E2 ).1 there exists an ellipsoid E(0. By Lemma 2. l) Let us further proceed with all the formal procedures.21) ρ(¯ lE(0.20).Then. Cl)1/2 ≤ (l. presuming also p < λmin = min{(l. (2. C) ⊆ E1 −E2 implies the existence of an ε > 0 such that E l ≤ (l. λmin ) and therefore Q(−¯) > 0. Q(−¯)) is properly deﬁned (Q(−¯) > 0). and consequently ˙ ρ(lQ(−¯)) ≤ ρ(lE(0. the relations p p (2. Q(−¯)) ⊆ E(0. if the ellipsoid E(0. p Due to (2.21) this yields ρ(¯ lE(0. the inclusion E(0. Q1 l)(l.2. C) ⊆ E1 −E2 implies E(0.2. Q(−¯)) + ρ(¯ 2 ) = ρ(¯ 1 ) p lE lE and (2. ¯ is the pair given in the formulation of the theorem. C)) + ρ(¯ 2 ) = ρ(¯ E1 ). Then for this pair the ¯l relations (2. ˙ Proof.2.21) are fulﬁlled and the ﬁrst of them is equivalent to the inequalities ˙ ρ(lQ(−¯)) ≤ ρ(lE(0. C) + E2 ⊆ E1 . C) ⊆ E1 −E2 p ρ(¯ lE(0. Let us start with the indication that the inclusion E(0.2. l) l. Q2 l)−1 ∈ IR : l = 1} ¯ so that altogether p ∈ (1. C(p)) with (2. lE l.19) (2. ¯ p Suppose that p. Q(−¯)) = E(0.2.2.22) C(p) = (1 + p−1 )C + (1 + p)Q2 114 . C)) ≤ ρ(lE1 ) − ρ(lE2 ) ≤ ρ(lE1 −E2 ) p for any l ∈ IRn . ˙ Further on. Q2 l)1/2 and therefore p = (¯ Q1 ¯ 1/2 (¯ Q2 ¯ −1/2 > 1.2. C) p and p∈ ¯ − ˙ E(0. Q1 l)1/2 − (l.20) imply E(0. ¯ l. Q(−¯)) = ρ(¯ 1 −E2 ) p lE ˙ .2.
2.2. C) + E2 ⊆ E(0.D.2.which satisﬁes the inclusion E(0.26). l) 1/2 1/2 = ((C(¯). p Since C is nondegenerate.27) we come to the inclusion C ⊆ Q(−¯) = (1 + p−1 )Q1 + (1 − p)Q2 p ¯ ¯ which.2. C(¯)) ⊆ E1 p Being deﬁned through (2. From (2. we obtain (2. ¯ Combining (2.23) and taking p0 = 1 + p∗ .24) the value p∗ is positive (as C > 0) and p0 > 1. l) l. lE p∗ = (C ¯ ¯ 1/2 (Q2 ¯ ¯ −1/2 l.2.E.2. C) ⊆ E(0.24) ρ(¯ lE(0.26) we also observe C(¯) ≤ Q1 and due to (2. (2. together with (2. l) ¯ so that p0 = p. Q. l) l.2.2.25) we have p p0 = p∗ + 1 = ((C ¯ ¯ 1/2 + (Q2 ¯ ¯ 1/2 )(Q2 ¯ ¯ −1/2 l. 115 . (2.2. l) l.25) and therefore ˙ E(0. l) According to Theorem 2. l) ≤ (Q1 ¯ ¯ 1/2 (Q2 ¯ ¯ −1/2 = p l. p ˙ Rearranging (2. p E(0. C(p∗ ))) = ρ(¯ lE(0.22). we have Q(−¯) > 0 and therefore p indeed lies within the domain p ¯ p ∈ (1. C)) + ρ(¯ 2 ).26) C = (1 + (p0 )−1 )C(¯) + (1 − p0 )Q2 . With l = ¯ given and p = p∗ taken as l (2.2.2. ¯ ¯ )(Q2 ¯ ¯ p l. Let us now prove the analogy of Theorem 2.2. C) + E2 ⊆ E(0.2.23) it also satisﬁes the equality (2.1 we then have (2.1 for geometrical diﬀerences. l) l.20) produces C = Q(−¯). l) l. λmin ). C(p)) for any p > 0.2. C(¯))−E2 ⊆ E1 −E2 .
l) p = p∗ + 1 = ((¯ Q¯ 1/2 + (¯ Q2 ¯ 1/2 )(¯ Q2 ¯ −1/2 l. l) and prove that Q(−¯) is positive deﬁnite. Q) ⊆ E1 −E2 let us indicate that there exists a value p such that the ellipsoid E(0. ρ(¯ lE(0. Q) and E1 ÷E2 .29) where and p = p∗ + 1. Q2 ) holds. l) ¯ Q1 ¯ 1/2 (¯ Q2 ¯ −1/2 = p.2 Suppose that int E(0. l) l. By (2. Then the set of maximal ˙ internal estimates of the diﬀerence E1 −E2 consists of ellipsoids of the form E(a1 − a2 . To do this. assuming the existence of a vector ¯ ∈ IRn . Given a nondegen˙ erate ellipsoid E(0. ¯ ¯ 1/2 (Q2 ¯ ¯ −1/2 ¯ l. p ∈ . Q) to be tangential to E1 −E2 . Q(−p)) ⊆ E1 ÷ E2 .2. so that we would have (2. Without loss of generality suppose again that a1 = 0. l) l. Q(−p)).2.2. l) l.30) or so that l.27) E(0. Q) ⊆ E(0. deﬁne the matrix p D(p∗ ) = (1 + (p∗ )−1 )Q + (1 + p∗ )Q2 or (2.Theorem 2. l) 116 Q = (1 − p−1 D(p∗ ) + (1 − p)Q2 p∗ = (¯ Q¯ 1/2 (¯ Q2 ¯ −1/2 l. ¯ = 1. Q)) + ρ(¯ 2 ) = ρ(¯ 1 ) lE lE l. l) (¯ Q¯ 1/2 + (¯ Q2 ¯ 1/2 = (¯ Q1 ¯ 1/2 l. Q)) = ρ(¯ 1 −E2 ).2. l) l.2. ˙ We may consider E(0. such that l l (2.28) ρ(¯ lE(0. a2 = 0. ¯ = (l. l) − . Q1 ) ⊇ E(0. Proof. lE ˙ Let us ﬁrst deﬁne E(0. l) . Q(−¯)) with p p = p = (Q.2.29) we have (2. l) l. Q(−p)) could be “squeezed” in between E(0.
Q(−p)) : p ∈ − E1 + E2 = ∩{E(a1 + a2 . Q(p)) : p ∈ + }. Q1 l∗ )1/2 (l∗ . (2. To conclude this section we shall summarize its results in the following Theorem 2.3 Given nondegenerate ellipsoids E1 .2. Proof.Due to Theorem 2.1 it follows E1 + E2 ⊆ ∩{E(0. in other words.31) and with int E1 ⊇ E2 . Q(p∗ )). assume the existence of a point x∗ such that (2. From Lemma 2. + }. x∗ ) > ρ(l∗ E1 + E2 ).30) and the equality p = p this yields Q = Q(−¯). Q(p)) : p ∈ + }. we come to ρ(l∗ E1 + E2 ) = ρ(l∗ E(0.2.2. Q(p)) : p ∈ x ∈ E1 + E2 .1. }.2. Together with (2.32) ˙ E1 −E2 = {E(a1 − a2 .35) Selecting p = p∗ = (l∗ . The last condition ensures the existence of a vector l = l∗ that yields (2. It is clearly suﬃcient to prove the theorem for a1 = a2 = 0. λmin ). also proving that ¯ p Q(−¯) > 0. Q2 l∗ )−1/2 and following Lemma 2. . the following relations are true (2. 117 (l∗ .33) (2. we further observe E(0. which means p ∈ (1. Q.2. E2 .2.1 (a).2(a) we come to the desired p ¯ inclusion (2.34) x∗ ∈ ∩{E(0.2.2.2. D(p∗ )) ⊆ E1 or. To prove the exact equality. ¯ where Q stands for the closure of set Q.2.D.2. that D(p∗ ) ≤ Q1 .8).E.2. Q) + E2 ⊆ E(0. Following Lemma 2.
x − x∗ ) ≤ ε2 } ⊆ int (E1 −E2 ). (2.36) x∗ ∈ int (E1 −E2 ). assume the existence of such a vector x∗ that ˙ (2.36) this implies x∗ ∈ E(0.2.31).2. The equality (2. Theorem 2. α ∈ [0. 118 . 1]} to be its largest axis.38).2.2.D.2. As E1 = −E1 .2. ˙ Since x∗ ∈ int (E1 −E2 ) there exists an ε > 0 for which ˙ Sε (x∗ ) = {(x − x∗ . E2 = −E2 (these sets are symmetrical around the origin).2. α ∈ [0. Fig.2. 1]} ˙ satisﬁes S ⊆ int (E1 −E2 ).37). What follows is that there exists a nondegenerate ellipsoid ∗ ˙ E(0. Q(−p)) : p ∈ − }. ˙ To prove (2.E.2.2. we obviously have ˙ ˙ E1 −E2 = −(E1 −E2 ) and therefore the whole set S = {x : x ∈ Sε (z).37) x∗ ∈ {E(0. To indicate that there is actually an exact equality. (2. Q. Q(−p∗ )) that satisﬁes x∗ ⊆ E(0.2 (a) we see two ellipsoids whose sum is the nonellipsoidal set that is the intersection of the ”nondominated” (inclusionminimal) ellipsoids that approximate it externally and are constructed due to formula (2.32).) From Theorem 2. z = −x∗ + 2αx∗ .2 which immediately yields ˙ {E(0. Q(p∗ )) in contradiction with (2.3 may be illustrated on a 2dimensional example.34). we recall that int (E1 −E2 ) = ∅ and follow Lemma 2. Q(−p)) : p ∈ − } ⊆ E1 −E2 . In the center of Fig.2 it now follows that for some p∗ ∈ − there ˙ exists an ellipsoid E(0.33). assuming set X ∗ = {z : z = −x∗ + 2αx∗ .2 (b) shows a nondegenerate geometric diﬀerence of two ellipsoids ( the set with two kinks) that also arrives as the ( closure of the) union of the nondominated (inclusionmaximal) ellipsoids that approximate it internally and are constructed due to formula (2.Together with (2. S ∈ Σ are chosen randomly but give a good illustration of the nature of the approximations. Q(−p)) ∈ E1 −E2 in contradiction with (2. C ) ⊆ S ⊆ int (E1 −E2 ).2.2. C) ⊆ E(0.32) is thus proved.2.2. In both examples the parameters p ∈ Π+ . (Give an example by explicit calculation of C ∗ .2.
3 Internal Approximations: the Sums. while the variety Q− [S] for approximating the diﬀerences E1 ÷ E2 (externally). E2 = E(a2 .1) and S ∈ Σ with Σ = {S ∈ L(IRn . In a similar way we deﬁne the variety (2.3.2) with S ∈ Σ. Given E1 = E(a1 . External Approximations: the Diﬀerences We shall now introduce a representation that will allow an internal approximation of the sum of two nondegenerate ellipsoids by a parametrized variety of ellipsoids. where (2. we introduce a parametric family of matrices Q+ [S]. The variety Q+ [S] will be used for approximating the sums E1 + E2 (internally). and an external approximation of the geometrical diﬀerence of these. where Q1 > 0.3. It will be demonstrated that this approximation may be exact. Let us start from the ﬁrst case Q− [S] = S −1 [(SQ1 S )1/2 − (SQ2 S)1/2 ]2 S −1 Q+ [S] = S −1 [(SQ1 S )1/2 + (SQ2 S )1/2 ]2 S −1 119 . Q1 ).!!! insert Figures 2.2. IRn ) : S = S. Q2 ).1 (a) and (b) !!! ——————– 2. S = 0}. Q2 > 0.————. The matrix S is therefore selected from the set Σ of symmetrical nondegenerate matrices.
3. for any l ∈ IRn . Q1 l)1/2 (l. that this would be possible if there existed a number λ > 0 and a vector l = l∗ . one has (2. As in the previous sections it is clearly suﬃcient to consider the case. l∗ = 1. Q1 l) + (l. by direct substitution. that would ensure the relation (2.1 The ellipsoid E = E(a1 + a2 .3.3. IRn ) we have (ρ(lE[S]))2 = (l. Q+ [S])) = ρ(lE1 ) + ρ(lE2 ) is true with l = l∗ .3.3. for any S ∈ Σ. when a1 = 0. Proof. Q1 l) + (l.3). · (ρ(lE[S]))2 ≤ (l. Conversely. Q2 l)1/2 which proves the inclusion (2. T = SQ2 1/2 120 . D = (Q2 −1/2 Q1 Q2 −1/2 1/2 ) . l = 1. Q2 l) + 2 (SQ1 S )1/2 S −1 l.3) E[S] = E(a1 + a2 .4) we observe. For any matrix S ∈ L(IRn . To prove that for a given S there exists an l = l∗ that ensures the equality (2.3. Q2 l) + + 2 (SQ1 S )1/2 S −1 l. Q2 l) + 2(l. Q+ [S]l) = = + By H¨lder’s inequality o (ρ(lE[S]))2 ≤ (l. there exists a matrix S ∗ ∈ Σ such that (2. z = Q2 −1/2 ∗ l . Q+ [S]) ⊆ E1 + E2 . such that the equality For each S ∈ Σ there exists a vector l = l∗ .Lemma 2.3. namely. Q+ [S]) is an internal approximation of the (Minkowski) sum E1 + E2 . (SQ2 S )1/2 S −1 l 1/2 1/2 (l. (SQ1 S )1/2 S −1 · or (SQ2 S )1/2 S −1 l. (2.4) is true with S = S ∗ .5) Denote (SQ1 S )1/2 − λ(SQ2 S )1/2 S −1 l∗ = 0. Q1 l) + (l. l = 1. (SQ2 S )1/2 S −1 l) . a2 = 0.4) ρ (lE(a1 + a2 .
4) to (2. Using the polar decomposition theorem.7) into (2.3. in addition that the matrix T is symmetrical.5) reduces to (T DDT )1/2 T −1 z = λ(T T )1/2 T −1 z. (2.3. of the matrix T D. U ∈ L(IRn .7) T D = U H. The proof of Lemma 2. we ﬁnd an orthogonal matrix 0 such that S ∗ = OS is symmetrical and therefore S ∗ ∈ .3. For matrix S deﬁned by the solution T obtained this way.3.3.3.1) it follows that Q[S] = Q[OS] and in such a way also that (2. there obviously exists an orthogonal matrix U such that vector U Dz is directed along with z.E.6) (T D · DT )1/2 · T −1 z = λz.6).3.3. Taking the polar decomposition ( [120]).3. With vector Dz = 0 given.9) with the U selected as above. .3. What remains is to ﬁnd a solution to the equation T B = BT for symmetrical and nonsingular T where B = U D. we ﬁnally transform the original equation (2.3.9) for λ ∈ IR.3. positive deﬁnite matrix D ∈ L(IRn . IRn ) orthogonal and T ∈ L(IRn .8) T DU −1 = U DT.5) is valid for S = S ∗ .5) will hold. The proof given in Chapter VIII of ( [120]) needs a slight modiﬁcation which we leave to the reader. IRn ) and the nonzero vector z ∈ IRn are given in advance. Then the last relation takes the form (2.then (2.9) U · Dz = λz. where the symmetrical.3.3. We now have to solve the system (2. hence there is a λ > 0 that ensures (2.8). Suppose.1 implies 121 Q.D. This can be done by using a well known result of matrix theory. we obtain an orthogonal matrix and a symmetrical (here also nonsingular) matrix H such that (2.3. The condition of symmetricity for H means that (2. IRn ) symmetrical and nonsingular. From formula (2. Substituting (2. equation (2.
This is actually (2.10) ρ( E(0. their product has simple structure.6) (using the notations of the lemma). generated respectively by the varieties of matrices Q+ [S].3.4) is true.3.2). Q+ [S]).3. (2.3. Q− [S]). for all i ∈ 1. where the matrix (T DDT )1/2 is positive deﬁnite and symmetrical and the matrix T −1 is symmetrical. Q) the inclusions E(0. B Q+ [S]B)) ≤ ρ( E(0. As previously. From this it follows that there is an invertible matrix B ∈ L(IRn .3.1 there is a condition that there exists a vector ∈ IRn . E(0. E(a1 − a2 .1). due to (2. S ∈ Σ. According to Lemma 2. n.3. It is left to the reader to prove as an exercise that.Corollary 2. Then the following assertions are true (a) the set of inclusionmaximal internal estimates of the sum E1 +E2 consists of ellipsoids of the form E(a1 + a2 . This leads to the relation (2. Q+ [S]). we assume a1 = a2 . n} of linearly independent eigenvectors (that are not necessarily orthogonal). Q+ [S]S ∈ }. S ∈ Σ. Q+ [S]) ⊆ E(0. Q− [S]. IRn ) that maps the ith unit vector ei ∈ IRn into zi ∈ IRn . Q) ⊆ E1 + E2 imply Q+ [S] = Q. In order to prove the maximality of E(0.1 Consider the parametrized varieties of ellipsoids E(a1 +a2 . under the above conditions. B QB)) ≤ ≤ ρ( E1 ) + ρ( E2 ) 122 . Q− [S]) are the inclusionmaximal internal and the inclusionminimal external estimates for E1 + E2 and E1 ÷ E2 respectively. The next step is to prove that the ellipsoids E(0. Proof. Theorem 2. Q− [S]). namely a complete set {zi ∈ IRn : i = 1.3. = 1. Q+ [S]) we shall demonstrate that for any ellipsoid E(0. Q+ [S]) where S ∈ Σ.1 The following equality is true E1 + E2 = ∪{E(a1 + a2 . (b) Assuming int E1 ⊇ E2 . the set of inclusionminimal external estimates of the diﬀerence E1 ÷ E2 consists of the ellipsoids of the form E(a1 − a2 .3. such that (2.
Substituting = ei + ej .3.3. The proof of part (b) is similar.11) E1 ÷ E2 = ∩{E(a1 − a2 .1 is given in Figures 2. The ﬁrst one shows the nonellipsoidal sum of two ellipsoids and the variety of nondominated (inclusionmaximal) ellipsoids that approximate it internally. Taken together this means B Q+ [S]B = B QB.3. and is left to the reader. The parameters 123 . The second part of the Theorem implies the following assertion Corollary 2. n.10).D.3.3.1).10) we obtain qii + 2qij (+) (+) (+) ≤ qii + 2qij + qjj for arbitrary ﬁxed i and j. we arrive at the reverse inequality.3. the equality Q+ [S] = Q. i = j into (2. This implies that the diagonal elements of B Q+ [S]B and B QB coincide. i ∈ 1. Q. Part (a) is thus proved. Carrying out the substitution of = ei −ej into (2.11). A 2dimensional illustration of Theorem 2.3.3. However. where qkr and qkr denote the element in the kth row and rth column of the matrix B Q+ [S]B and B QB1 respectively the element in the kth row and rth column of the matrix B Q+ [S]B and B QB1 respectively.2 The following representation is true (2. this implies (+) qij ≤ qij . one should have in mind that part (b) is true only if the diﬀerence E1 ÷ E2 has a nonvoid interior which implies that matrix Q− [S] > 0.for all ∈ IRn with equality holding for = ei . This diﬀerence then appears as the intersection of the external ellipsoids ( over all the variety).E. Q− [S])S ∈ Σ}. see Lemma 2.3). The second one shows a nondegenerated geometrical diﬀerence of two ellipsoids ( the set with two kinks) and the variety of nondominated (inclusionminimal) ellipsoids tha approximate it externally. The sum then arrives as the union of the internal ellipsoids ( over all the variety of these. due to formula (2. and by the invertibility of B. By the equality of the diagonal elements.3. due to formula (2.1 (a) and (b).
λmin ). λ1/2 ]. 124 . S ∈ Σ are chosen randomly but give a good illustration of the nature of the representations. Let E1 = E(a1 . Let − = + ∪ (1. Q2 ) be a pair of nondegenerate ellipsoids. ———–!!! insert ﬁgures 2.1 (The Representation Theorem). Let Q(p) be a parameterized family of ellipsoids Q(p) = (1 + p−1 )Q1 + (1 + p)Q2 . Let us once more indicate this ˙ result collecting all the facts in one proposition.p ∈ Π− . E2 = E(a2 . max 1/2 where λmin > 0. p∈ + = [λmin . int(E1 −E Theorem 2. Q2 ).4. 3 (a) and (b) ————————— — 2. Q1 ). When calculating E1 −E2 we also assume ˙ 2 ) = ∅. λmax < ∞ are the roots of the equation det(Q1 − λQ2 ) = 0 (the relative eigenvalues of Q1 .4 Sums and Diﬀerences: the Exact Representation ˙ The results of the previous Sections indicate that the sums E1 + E2 and diﬀerences E1 −E2 of ellipsoids could be exactly represented through the unions and intersections of the elements of certain parameterized families of ellipsoids.
4. Q− [S]).1)(2.4.4. Theorem 2.4. there exists a value p ∈ + such that (2. Q(−p))p ∈ + } } } − } The facts given in this theorem may be treated as being related to integral geometry.4. Q(p))p ∈ {E(a1 + a2 .9) E1 + E2 ⊆ E(a1 + a2 . Then the following inclusions are true (2.4) E1 + E2 ⊆ E(a1 + a2 . . E1 ÷ E2 . E1 ÷ E2 ⊇ E(a1 − a2 . Namely.2) (2. the following exact representations are valid: (2. Moreover.8) reﬂect a certain type of geometrical duality in treating the geometrical sums and diﬀerences of ellipsoids. particularly.4.7) (2.4. The speciﬁc properties formulated in (2.4. the external representations (2. to the representations of ellipsoidal sets (bodies) in IRn . Q(−p)). S = 0}.5) (2.2) for the sum yields (with a change of sign from (Q+ [S] to Q− [S]).2 (i) Given E1 + E2 and an ellipsoid E ⊇ E1 + E2 .4.4. ∀p ∈ ∀S ∈ ∀S ∈ ∀p ∈ + .4.1) for approximating the sum yields.8) E1 + E2 = E1 + E2 = E1 ÷ E 2 = E1 ÷ E 2 = {E(a1 + a2 . IRn ) : S = S. the external representation (2.3) for the diﬀerence.4. Q+ [S]).3) (2. As it was also demonstrated in the previous sections.4.4.6) (2.5)(2.4) and (2. Q− [S])S ∈ {E(a1 − a2 . E1 ÷ E2 ⊆ E(a1 − a2 . E1 + E2 ⊇ E(a1 + a2 .4. − .4. Q− [S] denote the following parametrized families of ellipsoids Q+ [S] = S −1 [(SQ1 S )1/2 + (SQ2 S )1/2 ]2 S −1 Q− [S] = S −1 [(SQ1 S )1/2 − (SQ2 S )1/2 ]2 S −1 where S∈ = {S ∈ L(IRn . Q(p)). .Also let Q+ [S]. Q+ [S])S ∈ {E(a1 − a2 .1 also indicates that the parametrized varieties involved are also the varieties of inclusionminimal external and inclusionmaximal internal estimates for E1 + E2 . the internal representation (2. This could be summarized in Theorem 2. Q(p)) ⊆ E 125 .4) for the diﬀerence and the internal representation (2.4.4.1) (2. with a change in the sign of the parameter p.4.
as described above. S ∈ }. there exists an S ∈ such that (2. Q2 ).1). posses a certain type of ”Pareto” property. Similarly. The variety of nondominated (inclusionmaximal) internal estimates for E1 + E2 is therefore {E(a1 + a2 . Q+ [S]) ⊆ E1 + E2 .. Q1 ) and E2 = E(a2 . where pi ∈ Π+ . Without loss of generality we further set a1 + a2 = 0. . E1 + E2 ) = ζ(p[k]) where p[k] = {p1 .. .4.. Q(−p)) ⊆ E1 ÷ E2 . we have E1 + E2 ⊂ ∩{E(a1 + a2 . The variety {E(a1 + a2 . The important fact is that the mentioned Pareto property is invariant under linear transformations.11) E1 ÷ E2 ⊆ E(a1 − a2 . . (iv) Given E1 ÷ E2 ..4..4. such that (2.k}. It is precisely this fact that allows us to propagate the static schemes of this Section to systems with linear dynamics.(ii) Given E1 + E2 and an ellipsoid E ⊆ E1 + E2 there exists an S ∈ (2. Q(pi ))i = 1. . we shall brieﬂy discuss it for the case of external approximation of the sum of two ellipsoids E1 = E(a1 . there exists an p ∈ − . The mentioned relations allow to say that the nondominated ellipsoids. Taking k arbitrary external ellipsoids E(a1 + a2 . k) h+ (∩E(0. (int (E1 ÷E2 ) = ∅) and an ellipsoid E ⊇ E1 ÷E2 . the varieties of nondominated (inclusionminimal) external and nondominated (inclusionmaximal) external estimates for E1 ÷ E2 are E(a1 − a2 .12) E ⊆ E(a1 − a2 . such that (iii) Given E1 ÷E2 . Without going into the details of this problem. This means that after a linear transformation the nondominated ellipsoids remain nondominated. Q− [S]) ⊆ E. Q(−p)).. One of the further problems in ellipsoidal approximations to be discussed would be to to estimate the number of ellipsoids that would give a desired accuracy of approximation. Q(pk )) of the type given in (1. we come to 126 . Q+ [S]. Calculating the Hausdorﬀ semidistance (i = 1.10) E ⊆ E(a1 + a2 . Q(p)).. (int (E1 ÷ E2 ) = ∅) and an ellipsoid E ⊇ E1 ÷ E2 . Q(pi )). .4.. Q− [S]) and E(a1 − a2 . p ∈ + } is therefore the set of nondominated (inclusionminimal) upper ellipsoidal estimates for E1 + E2 . pk } is a kdimensional vector.
. k ρ(l ∩i E(0.. It is not diﬃcult to observe that ζ 0 (k) ≥ ζ 0 (k + 1) and ζ 0 (k) → 0 if k → ∞. in its turn. of course. An appropriate issue is also to describe eﬀective algorithms for calculating ζ 0 (k) or its estimates.5 The Selection of Optimal Ellipsoids We shall no proceed with describing the optimal ellipsoidal estimates (external or internal) for E1 + E2 and E1 ÷ E2 . speaking in general. i = 1. but simpler than the exact one. . We thus assert the following Lemma 2. It may be also interesting to single out. The respective optimal ellipsoids are those that are generated by the vector p[k( )] = p0 [k( )]. selected through some cost function (optimality criterion). we should note that the nonsimple problem of the accuracy and computational complexity of the ellipsoidal approximation requires special treatment and in its full detail spreads beyond the scope of this book. speciﬁed by the previous Lemma.k} The best ellipsoids are those that are generated by the optimalizing vector p[k] = p0 [k] of the previous problem. Q(pi ))) = min{ i=1 (l(i) . However. a single ellipsoid that is optimal in some sense. A similar reasoning may be applied to the other approximation problems. It would be also important.1 The minimal number of ellipsoids that approximate the sum E1 + E2 with given accuracy > 0 may be determined as the smallest integer k = k( ) that satisﬁes the inequality ζ 0 (k) ≤ .ζ(p[k]) = max{ρ(l ∩i E(0. Q(pi )) − ρ(lE1 + E2 )(l. 2. we may specify them by solving the problem ζ 0 (k) = min{ζ(p[k])p[k] : pi ∈ Π+ .4. If the 127 . to obtain the estimate in a more explicit form or to obtain estimates that are perhaps less precise. among the variety of approximating ellipsoids. Q(pi )l(i) )1/2 Σk l(i) = l} i=1 Presuming that the best k ellipsoids are those that give the smallest Hausdorﬀ semidistance ζ(p[k]). l) ≤ 1} where.
By the monotonicity of ψ. where p∗∗ is the value for which the maximum of the function g : (0. because the existence of a unique stationary point cannot always be guaranteed.4.1) . this still can be done up to o(ε) in the important special case when Q2 = ε2 Q0 .1 to ﬁnd the minimal external estimates of the sum of two ellipsoids with respect to three important parameters of ellipsoids: the volume.2 (a) There exists a unique ellipsoid with minimal sum of squares of semiaxes. Then (a) The ψminimal external ellipsoidal estimate of the sum E1 + E2 is E(a1 + a2 . like Tr(Q3 ). Q(−p∗∗ )). However. Some rather simple necessary conditions of optimality could then be applied for this purpose. where p∗ is the value for which the minimum of the function f : {0. is attained. (which is Tr(Q)) that contains the sum E1 + E2 .5.1 Suppose the function ψ(E) is continuous and monotonously increasing with respect to inclusion. in our test cases we are not in such a favourable position.2. and do the same for the maximal external estimates of the diﬀerence. due to Theorem 2. the solution may be sought for only from the parametrized varieties of this theorem. then. analogous results can be obtained for other functions. Q(p∗ )). where (TrQ1 )1/2 ∗ (2.2.D. p∈ − p∈ + Proof. f (p) = ψ(Q(p)). is attained. (b) Suppose that int(E1 ) ⊃ E2 holds. Q. Then the ψmaximal internal estimate of the diﬀerence E1 ÷ E2 is E(a1 − a2 .5. Using our technique. Q(p∗ )).4. follows from Theorem 2. It is of the form E(a1 + a2 . Lemma 2.cost function ψ(E) is monotonous bby increasing with respect to inclusion (ψ(E ) ≥ ψ(E ) or ψ(E ) ≤ ψ(E ) if E ⊇ E ).5. the sum of squares of semiaxes (or trace) and Tr(Q2 ). Although the case of the diﬀerence is similar to the above. Lemma 2. etc. λmin ) −→ R g(p) = ψ(Q(−p)). We shall now apply Lemma 2.5. ∞} −→ R.E. p = (TrQ2 )1/2 128 . as we shall see later.
qn (p)}. we are allowed to suppose that Qj = diag{q (j) 1 . The function f : (0. Proof.4. Since the eigenvalues of Q do not change with nondegenerate linear transformations. Q(−p∗ )).3 (a) There exists a unique ellipsoid of minimal volume that contains the sum E1 + E2 .5. the minimal property of the estimating ellipsoid is not changed if a nondegenerateaﬃne transformation is applied. . Lemma 2. Q(·)). It is of the form E = E(a1 − a2 . and also that there exists an internally estimating ellipsoid for the diﬀerence E1 ÷ E2 such that its sum of squares of semiaxes (which is Tr(Q)) is maximal.1 we only have to ﬁnd p∗ ∈ (0. considering the function deﬁned by g(p) = f (−p). and det(Q(·)) depends on the product of the eigenvalues of Q. Because of the way Q(p) is derived from Q1 and Q2 . ∞) that minimizes the volume of E(a.6) falling into the set Π− . Q(−p∗∗ )) where p∗∗ is the single root of equation (2. Q. It is of the form E(a1 + a2 . Q2 ) ⊂ int(E(a1 . n and hence Q(p) = diag{q1 (p). . This is obviously equivalent to the minimization of log det(Q(·)). p(p + 1) i=1 λi + p For it we also have p∗ ∈ Π+ . Therefore.(b) Suppose that int(E1 ) ⊃ E2 holds. and therefore the root is the same as in the above case and unique. The matrix Q(−p∗ ) being positive deﬁnite implies p∗ < λmin . therefore it has to be an element of Π+ .13) and p∗ ∈ Π− . q (j) n }. By Lemma 2.2) = . In case (b). then there exists a unique ellipsoid E of maximal volume contained in the diﬀerence E1 ÷ E2 . (b) Suppose that E(a2 . Q1 )).5. so is the value of p∗ ∈ (0. ∞) is the unique solution of the equation n 1 n (2. .5. we have g(p) = f (p).D. j ∈ 1. Q(p∗ )) where p∗ ∈ (0. ∞) −→ R f (p) = d d Tr[Q(p)] = Tr[(1 + p−1 )Q1 + (1 + p)Q2 ] dp dp d = Tr [(1 + p−1 )Q1 + (1 + p)Q2 ] dp has one single root at p = p∗ . . ∞). . Proof. 129 .E. . Then it is of the form E(a1 − a2 .5. Here p∗ is deﬁned by the equality (2.
p−→∞ The proof of (a) now is complete. Lemma 2. a similar argument shows that there is one single local maximum. p+1 lim det(Q(p)) = lim det(Q(p)) = ∞. ∞} −→ IR f (p) = Using the relationship d d log det(Q(p)) = Tr Q(p)−1 Q(p) dp dp we obtain Tr[(p−1 Q2 −1 Q1 + I)−1 ] = By diagonality this means that p n = . i.5. Q(p∗ )). If here the parameter p corresponding to the maximal volume fell onto any of the endpoints of the interval (0. Therefore it has one single root corresponding to a minimum. 2. But this is excluded by the condition we set. It is of the form E(a1 + a2 . Q. i ∈ 1.All this means that we have to ﬁnd the roots of the function f : {0. After having established this. j ∈ 1.3) f (p) = γ22 p3 + γ12 p2 − γ12 p − γ11 γij = Tr(Qi Qj ).4 (a) There exists a unique ellipsoid with minimal Tr(Q2 ) that contains the sum E1 + E2 .5. n are the eigenvalues of the pencil Q1 −λQ2 — and they are again invariant with respect to the aﬃne transformation used for diagonalization. The left hand side of this equality strictly increases from 0 to n and the right hand side strictly decreases from n to 0 while p increases from 0 to ∞.D. e. λmin ) then the matrix would be semideﬁnite. The value p∗ ∈ + . dp n . as we have p−→0 n d log det(Q(p)). where p∗ is the unique positive root of the polynomial (2. the ”ellipsoid” would have zero volume.E. 0). for i. p+1 i=1 λi + p where λi . 130 . In the case of (b) we need the roots in (−∞.
The right hand side is positive here. Q2 ) = εE(a0 . implying the existence of a positive root. Direct calculations indicate that f (p) = d Tr[Q(p)2 ]. as we have the equality Tr(Qi Qj ) = Tr(Rj Qi Rj ) where Rj is the square root of Qj . The equality f (p) = 0 is equivalent to p2 = γ12 p + γ11 . Therefore their graphs may have no more than one intersection. because we take the trace of a positive deﬁnite matrix. a2 ) + (l. Q0 ). (2. Hence γ22 > 0 and γ11 > 0. The proof of the statement of part (b) is obvious.(b) Suppose that int(E1 ) ⊃ E2 holds. Q. Q0 )). while the right hand side −1 −1 is a continuous function with value γ1 . so that ρ(lE(a2 . Q0 l)1/2 = ερ(lE(a0 .4) ε2 = εE0 . or in other terms E(a2 . Here p∗∗ is a root of the polynomial deﬁned by the equality g(p) = f (−p) that falls into Π− .e.5. Proof. Then it is of the form E(a1 − a2 . Q2 l)1/2 = ε(l. Q(−p∗∗ )). γ22 p + γ12 Here the left hand side is a convex strictly increasing function.D. which tends to γ11 γ22 with p → ∞. Q2 ) may be presented as a2 = εa0 . Q2 )) = (l. E0 = E(a0 . dp All the coeﬃcients γij s are positive. γ12 at p = 0.E. Q0 ) i. a0 ) + ε(l. when the parameters of E2 = E(a2 . The proof of (a) is complete. and also that there exists an internally estimating ellipsoid for the diﬀerence E1 ÷ E2 with maximal Tr(Q2 ). 131 Q 2 = ε 2 Q0 . For the sake of approximating the solutions to diﬀerential inclusions we may need to treat a particular case.
5.8) now the proof of part (a) is complete. where (2. Tr1/2 (Q0 )) (b) Suppose that int(E2 ) = ∅. and calculating the optimal p = p∗ that gives Tr(Q(p)) = min. (2.8) Introducing the notation (2.5. with Qε (q) given by (2.5.5.5. Lemma 2. (which is Tr(Q)) that is contained in the diﬀerence E1 ÷ E2 is of the form E(a1 − εa0 .E. By (2. we obtain (2. Q1 and Q0 .5) and (2. however here we also need the inequality ε−1 q ∗ < λmax = ε−2 µmax that will automatically hold for small ε. (which is Tr(Q)) that contains the E1 + E2 is of the form E(a1 + a2 . Qε (−q ∗ )). Writing formally (2.4). Then the ellipsoid with the maximal sum of squares of semiaxes.6) and (2.5. The lemma is now proved for both cases.5.5. Proof.5. The proof of case (b) is similar to the above. Here µmax is the maximal relative eigenvalue of Q. Qε (q) = Q(qε−1 ).6) q∗ = Qε (q) = Q1 + ε(q −1 Q1 + qQ0 ) + ε2 Q0 Tr1/2 (Q1 ) .Lemma 2.7) Q0 (p) = (1 + p−1 )Q1 + ε2 (1 + p)Q0 .5.5 Let us have for ε > 0 the relation (2.9) ε2 (p∗ ) = TrQ1 . Qε (q ∗ )). Tr1/2 (Q0 Q−1 ) 1 132 . (a) Then the ellipsoid with minimal volume that contains the sum E0 + E2 is of the form E(a1 + εa0 .5.D. (a) Then the ellipsoid with minimal sum of squares of semiaxes.5.5.10) q∗ = n1/2 . Qε (q ∗ )).5) follows.6 Let us have for ε > 0 the relation (2.4). TrQ0 ) q = pε.
we obtain for the Taylorseries expansion in ε of equation (2. We look for the root of equation (2. Proof. Comparing them with the relative eigenvalues λ1 ≤ λ2 ≤ .5.4) to be true. n (2.5. Q. and then (a) follows.D.6) where we have to use (2. (q ∗ )2 = n n i=1 µ−1 i . Then the ellipsoid with maximal volume that is contained in the diﬀerence E1 ÷ E2 is of the form E(a1 − εa0 . .5. Lemma 2.5.5.8) and according to the equality q = pε. p+1 i=1 λi + p n and carrying out the substitutions. ≤ µn the relative eigenvalues of the matrices E1 and E0 . ≤ λn of E1 and E2 we obviously have for all i ∈ 1. where q ∗ is deﬁned by (2.2. into the form (2.11). For part (b) we have to take the negative sign.7 Assume the relation the relation (2.5. −1 = 1 + εq −1 i=1 1 + εqµi and n − εq ∗ n i=1 n µ−1 = n(1 − ε(q ∗ )−1 ) + o(ε). by the analytic dependence of the roots on the parameters.E. i From the comparison of the coeﬃcients of ε.12) λi np = . with formula Qε (q) given by (2.5. .5.5. Qε (q ∗ )).6). . (a) Then the ellipsoid with minimal Tr(Q2 ) that contains the sum E1 + E2 is of the form E(a1 + εa0 .11) ε 2 λ i = µi . .(b) Suppose that int(E2 ) = ∅.13) q∗ = Tr1/2 (Q2 ) 1 . Qε (−q ∗ )).20) 1 n . and the condition of positive deﬁniteness follows in the same way as for Lemma 2.6) (2.5.4. Rewriting equation (2. Denote by µ1 ≤ µ2 ≤ . 1/2 Tr (Q1 Q0 ) 133 .5.
Substituting this into equation (2. but this proof is diﬀerent.14) 2. Using the above scheme.5. Tr(Q2 ) 1 (q ) = . On the other hand.13. Proof. Part (a) of the following theorem is given in ( [73]). It is not diﬃcult to prove.D. Q∗ ) where − (2. where q ∗ is deﬁned by (2. + (b) There exists a unique ellipsoid of minimal volume containing the diﬀerence E1 ÷ E2 . Tr(Q1 Q0 ) ∗ 2 Q.15) The rest of the proof is analogous.5.14) εγ00 q 3 + γ10 q 2 − εγ10 q − γ11 = 0 and from the comparison of coeﬃcients in the Taylorseries expansion in ε for this equation (2.5. we look for the root again in the form of q = pε. − Proof.5.3) γij = Tr(Qi Qj ). and the external estimates of minimal volume of the diﬀerence.(b) Suppose that int(E2 ) = ∅. j = 1 or 0.8 (a) There exists a unique ellipsoid of maximal volume contained in the ˜ sum E1 + E2 . Qε (−q ∗ )).5.16) Q∗ = Q1 + Q2 + 2Q2 1/2 [Q2 −1/2 Q1 Q2 −1/2 ]1/2 Q2 1/2 . we formulate the converse statements concerning the internal estimates of maximal volume of the Minkowskisum. as in [Chern].17) Q∗ = Q1 + Q2 − 2Q2 1/2 [Q2 −1/2 Q1 Q2 −1/2 ]1/2 Q2 1/2 . Lemma 2.5. Then the ellipsoid with maximal Tr(Q2 ) that is contained in the diﬀerence E2 ÷ E1 is of the form E(a1 − εa0 .5. is now used for i. Q+ ) where (2.4) (2. It is of the form E(a1 + a2 . It is of the form E(a1 − a2 . part (b) is a new result. and appears to be more general. The notation of equation (2.5.E.5. which is proved by the technique used there. Corresponding to the above. that (a) is valid with Q∗ = S −1 [(SQ1 S )1/2 + (SQ2 S )1/2 ]2 S −1 + 134 .
5. y 1/2 By the argument of the proof of Part 2 in [84].5. n with di = (ri − pi 1/2 )2 .where S is the matrix diagonalizing both Q1 and Q2 .2. By the aﬃne invariance of the volume function. y 1/2 − P y. except that in addition to the internal estimates that indicate the 135 . this expression is independent of the choice of S. If we deﬁne di .E. pn } is a partial identity. and SQ2 S = P . This is justiﬁed because of the inclusion int(E1 ) ⊃ E2 holds. n.3(a). rn }.9 Let us have the relation (2. we can use again the matrix S to diagonalize Q1 and Q2 . D) among those with the property: a.18) Qε+ = Q1 + 2ε(Q1/2 [Q−1/2 Q1 Q0 o o −1/2 1/2 ] Q0 ) + o(ε)I 1/2 (b) the ellipsoid with minimal volume that contains the diﬀerence E1 ÷ E2 is of the form E(a1 − εa0 . y + Dy.D. .5. the existence and uniqueness of such an ellipsoid follows. and the same argument implies that a = 0 and D is diagonal. then it can be established by direct calculation that D is an external estimate. Let us now consider (b). It is also possible to observe that although S is not unique. A substitution now yields (a). where P = diag{p1 . Q. Our aim is to ﬁnd a minimal volume ellipsoid E(a. Thus Fig. . . .9) for ε > 0.5. to get SQ1 S = D1 = diag{r1 . The statement is thus true.5.1(a) is the same as 2. To illustrate the material of this Section. Substituting the unit coordinatevectors into the previous inequality we obtain di 1/2 ≥ ri 1/2 − pi 1/2 1/2 i ∈ 1. i. Let us select S = N Q2 −1/2 where N is orthogonal. we introduce several ﬁgures. The reader may verify this as an exercise.19) Qε− = Q1 − 2ε(Q1/2 [Q−1/2 Q1 Q0 o o −1/2 1/2 ] Q0 ) + o(ε)I 1/2 The proof follows the lines of the above through expansions of the respective relations in ε and the solution is found within the terms with ε of power 1. Qε+ ). e. where (2. y 1/2 ≥ D1 y. . Qε− )). Then for a suitable N the matrix S will meet the requirements. Lemma 2. i ∈ 1. where (2. Then (a) the ellipsoid with maximal volume that is contained in the sum E1 + E2 is of the form E(a1 + εa0 . .
Their intersection m i=1 E(a(i) . . .2(b) is the same as 2. i The following assertion is obvious 136 .1(b) is the same as 2.5.1(a).5. . minimal Tr(Q2 ) ( both drawn in continuous lines) and minimal volume (drawn in dotted line).2.2.2(a).2.1) Assuming A = {α ∈ IRm : take the inequality (2.(b). . m}. .(b).(a). i = 1.5. but in addition to the external estimates that indicate the sum of two ellipsoids we also indicate the internal ellipsoid of maximal volume ( shown with dotted line). Qi ) = P[m] consists of all the vectors x ∈ IRm that simultaneously satisfy the inequalities (2. . but in addition the dotted line shows the volumeminimal ellipsoid for the Minkowskidiﬀerence of two ellipsoids.2) i=1 m (x − a(i) . Q−1 (x − a(i) )) ≤ 1. Fig. αi = 1.3.(b) !!! —————– The next issue is to consider intersections of ellipsoids. Qi ). . αi ≥ 0. .6. The description of this operation is more complicated than before and does not reach the same degree of detailization being conﬁned mostly to the external ellipsoidal estimates of these intersections.sum of two ellipsoids.6. . but in addition the dotted line shows the volumemaximal internal ellipsoid for the Minkowski diﬀerence. 2. i (i = 1. Fig.(b). 2. αi (x − a(i) . i = 1. m. ——–!!! insert Figures 2. we also indicate the external estimates that are of minimal trace ( TrQ). Finally.2(a) is the same as 2. .5. m). .5. . Q−1 (x − a(i) )) ≤ 1. Fig. 2.2.6 Intersections of Ellipsoids Let us consider m nondegenerate ellipsoids Ei = E(a(i) .
.5. In other terms (2.1 is Lemma 2.Lemma 2. Each of the ellipsoids E[α] is therefore an external estimate for P[m]. By direct calculation we may observe that for a given α ∈ A the inequality (2.6. if necessary. i It is not diﬃcult to check that h2 [α] ∈ [0.1).2) for any α ∈ A (and vice versa). The intersection P[m] of m ellipsoids Ei is now presented as an intersection of a parametrized family of ellipsoids E[α]. i αi Q−1 a(i) . α ∈ A.2) deﬁnes an ellipsoid (2.3) E[α] = {x : (x − a[α]. It is clear that the variety E[α] α ∈ A contains each of the ellipsoids Ei .2 The following assertion is true. i = 1. m. Among these we may select. Q−1 a(i) i αi Q−1 a(i) . for any α ∈ A. .6. (1 − h2 [α])−1 Q[α]). A direct consequence of Lemma 2.6.4) E[α] = E(a[α].6. then x∗ satisﬁes (2.6.6. The set P[m] = {∩E[α]α ∈ A}.6. so that Ei = E[e(i) ]. 1].1 If x∗ ∈ IRm is a solution to the system (2. so that Pm ⊆ E[α]. . Q[α](x − a[α])) ≤ 1 − h2 [α]} where a[α] = i=1 m m −1 αi Q−1 i m i=1 αi Q−1 a(i) i Q[α] = i=1 m i=1 m i=1 m i=1 αi Q−1 i h2 [α] = αi a(i) . 137 . i m i=1 m i=1 −1 −1 − αi Q−1 i αi Q−1 i m i=1 m i=1 αi Q−1 a(i) i αi Q−1 a(i) . an optimal external ellipsoidal estimate for P[m]. .
α2 = 1 in (2. for example. the optimal external ellipsoidal estimate E[α0 ] for P[m] (taken. In ﬁgure 2. in general. Qi ). Here numbers 1. α2 = 0 and α1 = 0.4). among all the possible external estimates? In the sequel of this section we shall produce some examples (see Examples 2. α1 + α2 = 1). δij = 1 if i = j. 138 .1(b) shows the two intersecting ellipsoids and with dotted line the volumeminimal external estimate obtained by a onedimensional search in α1 ∈ [0.6. . Qi ) = P[m] of nondegenerate ellipsoids E(a(i) . .6. δij = 0 if i = j j = 1.2 indicate the intersecting ellipsoids whilst the unmarked ones are the external estimates E[α] calculated due to formula (2. Figure 2. One of the questions that arise here is whether the variety E[α] is suﬃciently complete in the sense of the following question: will the optimal external estimate E[α0 ] (with respect to a function Ψ(E)) chosen only among the ellipsoids E[α].6..(b) !!! —————— However.1(a) one may observe an illustration of Lemma 2.6. for one of the criteria of the above) may not be among the ellipsoids Ei . 2. ( α2 ≥ 0.1(a). The intersecting ellipsoids 1. m. . .2.2. particularly.where e(i) ∈ IRm is a unit vector (an orth) along the ith coordinate axis and its coordinates are deﬁned as (i) ej = δij .4).6.3 ) that give an answer to this question.2 correspond to values α1 = 1. Meanwhile let us introduce another formula for the intersection m i=1 E(a(i) . 1]. α ∈ A be the same as the optimal ellipsoid E (also with respect to Ψ(E)) chosen among some other class of external ellipsoids or.6.6.6. ——————!!! Insert ﬁg 2.
that satisfy the equality m i=1 M (i) = I. Qi ). = {0}. Then.8) whatever are the matrices M (2.6.9. Qi )) for the ellipsoids E(a(i) .6. Otherwise this means (2. Taking the support functions ρ( E(a(i) .9) (we may omit the transpose.6.) does hold.9) (i) ∈ IRn and any array of matrices M (i) that satisfy the last n Q⊆ i=1 M (i) Qi . We shall further proceed under this assumption. coming to m ∈ IRn .5). Moreover.1 The intersection P[m] has an interior point : intP[m] = ∅.6. provided only that (2.6.5) ρ( Q) = { inf i=1 ρ( (i) m Qi ) i=1 (i) = } that relates the support function ρ( Q) of an intersection Q = ∩m Qi with the support i=1 functions ρ( Qi ) for each set Qi . we may substitute (2. m ( i=1 M (i) − I) = 0 which should be true for any equality.7) implies Q=∩ m i=1 M (i) Qi over all the matrices M that satisfy (2. (2.6.6) (i) = M (i) (i) where the matrix M (i) ∈ IRn×n does exist for any .6.7) ρ( Q) = inf i=1 ρ( M (i) Qi ) M (i) : m i=1 (M (i) − I) = 0 or to the relations m ρ( Q) ≤ ρ(  i=1 M (i) Qi ).6) into (2. assuming (2. since M (i) are chosen arbitrarily.6.6.6.Assumption 2. (2.) 139 (i) .6. we may apply the convolution formula ( [265]) m (2.
for m = 2 we have (2. M (i) Qi M (i) ) restricted only by the equality (2. . p = {p1 . i M = {M (1) . .6. M] = ( i=1 pi ) i=1 p−1 M (i) Qi M (i) . .6.6. Qi ) satisﬁes the inclusion (2. .11) is true with the intersection taken over all M (i) of (2. As we have seen earlier.In terms of ellipsoids and in view of the formula M E(a. M]) m where a[m] = M (i) a(i) pi ≥ 0 i=1 m m Q[m. (I − M )Q2 (I − M ) )) for any n × n matrix M ∈ L(IRn .6. M (m) }. M Q1 M ) + E((I − M ) a(2) . IRn ). M (i) = I and for the same class of matrices (2.9).6. m i=1 E(a(i) . Q) = E(M a.12) P[2] ⊆ ⊆ E(M a(1) . whatever are the matrices M (i) of (2. M (i) Qi M (i) ). M QM ) this gives (2.6. p.6.9).6. Particularly.10). m i=1 M (i) = I. . p. . The intersection P[m] of n ellipsoids E(a(i) .6.6. the sum of m ellipsoids may. We therefore come to the assertion Lemma 2. . Moreover.10) P[m] ⊆ i=1 m m i=1 E(M (i) a(i) . be approximated from above by one ellipsoid.6.3 The intersection P[m] of m nondegenerate ellipsoids E(a(i) . .9) we have (2. M (i) Qi M (i) ) ⊆ E(a[m]. Q[m.10) by the sum of m ellipsoids E(a(i) . however.11) P[m] = ∩m i=1 E(M i a(i) M (i) Qi M (i) ). Qi ) is therefore approximated from above in (2. the equality (2. pm } 140 .9). Namely.
M]) which is true for any M.6. Scheme A For a matrix Q positive symmetrical we may rewrite M QM = (Q1/2 M ) (Q1/2 M ) and introduce the norm M QM 2 P[m] = p M E(a[m.13) P[m] ⊆ E(a[m.6. deﬁned at the end of Section 2. Q2 (I − M ) 1/2 + 1/2 Q2 (I 141 . L ∈ IRn×n is deﬁned as K. M].6.15) P[2] = = E(0. Q1/2 M = trM QM . L = trK L. provided M satisﬁes (2.6. taking perhaps one of the above. where the scalar product K.6.6. The present scheme is now deﬁned through minimizing M Q1 M (2. p.9).6.9). Let us ﬁrst consider two ellipsoids with centers a(1) = a(2) = 0 so that (2.16) 2 + (I − M )Q2 (I − M ) = 1/2 Q1 M 1/2 . p > 0.4.3 and Theorem 2. M]. Among the ellipsoids E(a[m].6. p > 0. we conclude that the intersection P[m] may be presented through the inclusion (2. Q1 ) ∩ E(0.9) and the equality (2. or the equality (2. (I − M )Q2 (I − M ) ) The external bounding ellipsoid may be now designed through the following schemes. M]) = Q1/2 M . Q2 ) ⊆ E(0.1 ( formula (2.6.6.6.4 The intersection P[m] satisﬁes the inclusion (2.Combining the results of Lemma 2.5)) .9). p.4. M Q1 M ) + E(0.13) over all M of (2. Q[m. Q[m.1.6. L of two n × nmatrices K.14) with M of (2.14) over all M of (2. Lemma 2. Q1 M 2 = − M ) . M) we may now select those that are optimal relative to some criteria. p.
6.20) Consider a speciﬁc case Example 2. ( check here that what one gets is precisely a minimum ) .6.19) where (2.over M which leads to the optimal M = M0 : M0 = M0 = (Q1 + Q2 )−1 Q2 .17) E(0. which is (2. Solving this problem through the equation f1 (p) = 0.1 Take the twodimensional ellipsoids E1 = E(0. 0 (1 + k 2 )−1 trM 0 Q1 M 0 = 16k 2 (1 + 4k 2 )−4 + k 2 (1 + k 2 )−2 = α2 (k) tr(I − M 0 )Q2 (I − M 0 ) = 4k 2 (1 + 4k 2 )−2 + k 4 (1 + k 2 )−2 = β 2 (k) 142 . S2 = (I − M0 )Q2 (I − M0 ) .6.6. Q2 ). S 0 ) 0 0 0 trS 0 = ((trS1 )1/2 + (trS2 )1/2 )2 .6. 4k 2 (1 + 4k 2 )−1 0. following (2. E2 = E(0. (1 + p−1 )M0 Q1 M0 + (1 + p)(I − M0 )Q2 (I − M0 )) whatever is the p > 0. The bounding ellipsoid may now be optimalized over p due to one of the criteria of the above (see Section 2.1). Further on. where Q1 = Then M0 = (Q1 + Q2 )−1 Q2 = 1 0 0 k2 .13) we have (2.6.we observe 0 0 p2 = trS1 /trS2 . (I − M0 )Q2 (I − M0 ) ) ⊆ E(0. Q2 = 4k 2 0 0 1 0 0 P[m] ⊆ E(0. Q1 ).18) 0 0 S1 = M0 Q1 M0 . Let us for example select an optimal p = p0 . minimizing over p the trace 0 0 tr((1 + p−1 )S1 + (1 + p)S2 ) = f1 (p) where (2. (1 + p−1 )S1 + (1 + p0 )S2 ) = E(0. 0 The ﬁnal calculation gives an upper bound for P[m]. M0 Q1 M0 ) + E(0.
Scheme B The next option is not to minimize (2. Following (2.6. M (p > 0. M ]) and to minimize E(0.6. M ]) given by the inclusion E(0. S[p. M ]) over p. S[p. After a minimization of trE(0. M ]) = min as the criterion. but to take the bounding ellipsoid E(0.19). Since f2 (M ) is strictly convex and f2 (M ) → ∞ with M → ∞.16) ﬁrst over M . S 0 ) The one of Scheme A is when S 0 is taken due to (2. S[p. The value M 0 for this case is calculated through the minimum of (2. Q2 (I − M ) = = trM Q1 M + tr(I − M )Q2 (I − M ) 143 . this leads to the problem of minimizing the function f2 (M ) = ((trM Q1 M )1/2 + (tr(I − M )Q2 (I − M ) )1/2 )2 over M .16) which is M. (2.20). Q1 M + (I − M ). M ∈ IRn×n ) having trE(0. (1 + p−1 )M QM + (1 + p)(I − M )Q2 (I − M ) ) = E(0. M ]) directly over the pair p. then over p. (I − M )Q2 (I − M ) ) ⊆ ⊆ E(0. We have thus indicated two options for the bounding ellipsoid P[m] = E(0. S[p.p0 = α(k)β −1 (k).6. We also gather that p∗ = (trM ∗ Q1 M ∗ )1/2 (tr(I − M ∗ )Q2 (I − M ∗ ) )−1/2 so that the optimal ellipsoid E ∗ = E(0. there exists a unique minimizer M ∗ . M QM ) + E(0.6. S[p∗ .6. S[p. M ∗ ]).18). we have trS 0 = (α(k) + β(k))2 0 0 S 0 = (1 + p0 )(p0−1 S1 + S2 ).
2 .2(b)) At the same time.3 were calculated by S.BA). In all the consecutive ﬁgures the intersecting ellipsoids are marked by numbers 1. while the approximating ellipsoids are marked as A (AD.2 (a).Fefelov.2 are centerd at 0.which is equivalent to the minimization of (trM Q1 M )1/2 + (tr(I − M )Q2 (I − M ) )1/2 We shall now illustrate the given schemes on twodimensional examples. 144 . for each of these schemes the ellipsoids AA.2.2 (a) The ellipsoids 1.6. one may observe. E2 with diﬀerent centers.6. that scheme AD gives a larger one than AA .2 (a)). At the same time.6. S[p∗ .B are now applied to ellipsoids E1 . Here both schemes AA and BA give the same external ellipsoid (Fig. Consider ﬁrst the case when the centers of E1 and E2 coincide.AA) and B(BD. comparing on them the results given by Schemes A and B.2 are centered at 0. Example 2.On the other hand. However. Apart from distinguishing these two cases. we have P[m] = E(0.6.BD (which are not shown). we shall also distinguish for each case a minimization over diagonal matrices M only (cases AD and BD respectively) from a minimization over all possible matrices M (cases AA and BA ). scheme BD does not give any other ellipsoid except 1. (b) The ellipsoids 1. 2.(b) !!!————————— The schemes A. Here schemes A and B give diﬀerent external ellipsoids AA and BA ( Fig.6.2.BA are smaller (by inclusion) than AD. 13 —!!! place ﬁgures 2.2.6.2. 13 Examples 2. M ∗ ]) where M ∗ is calculated by minimizing f2 [M ] . in Scheme B.
but AA coincides with AD and BA with BD (Fig.3(b).6.3 (a)).3(c)). which turns out to be the optimalizer and therefore may be substituted into f3 (p) allowing to write f3 (p∗ ) = min f (p).B give diﬀerent external ellipsoids. 0 0 a1 = a2 = trS1 S2 0 0 a3 = trS1 S1 It may be checked.3(c) !!!—— Scheme C This one is similar to Scheme 1. but instead of minimizing the trace f1 (p).BD give worse results in both cases.2.2 ( Fig.3(b). ——!!! place ﬁgures 2. 2.6.BA give the same external ellipsoid which is close to optimal by trace or volume ( Fig. that the given polynomial has a unique positive root p = p∗ > 0. p > 0.6.(b1). 2.3 (a). Note that AD. S2 = S2 ) 0 0 a0 = trS2 S2 .BA give the same external ellipsoid which clearly is not optimal by either trace or volume. 2.3 (a) Here schemes AA.(b1)).6. p 145 . Note that scheme BD gives nothing more than ellipsoids 1.Example 2. without diﬃculty. (b0 Here schemes A.6. we minimize 0 0 0 0 f3 (p) = tr(((1 + p−1 )S1 + (1 + p)S2 )((1 + p−1 )S1 + (1 + p)S2 ) )l Equation f3 (p) = 0 now yields a cubic polynomial a0 p3 + a1 p2 + a2 p + a3 = 0 0 0 0 0 where (S1 = S1 . 2.6. (c) Here schemes AA.6.
i m i=1 c2 = trM (i) Qi M (i) i and (2. pi > 0. to the equations pi i=1 −c2 γi2 i + i=1 γi c2 = 0. by ci = pi (i = 1.6. i=1 we come to the system ∂ϕ = 0. . . . or otherwise. Let us now return to the case of an arbitrary ﬁnite number m of intersecting ellipsoids and select the external circumscribed ellipsoid as a traceminimal set.22) M (i) = I.21) where γi = trQ[m. . S 0 ) ⊇ E(0. .6. . .The optimal circumscribed (external) ellipsoid 0 0 E(0. . pm } and assuming m pi = 0. m. . S1 ) + E(0. . m. . p] over p = {pi . . i i = 1. . ∂pi m m i = 1. Q1 ) 146 . p] = i=1 m i=1 γi c2 = ϕ(p) i pi p−1 . . We have m (2. . m) so that the optimal value ϕ(p0 ) = ( m i=1 ci )2 Further on we shall brieﬂy describe a possible approach to the calculation of internal ellipsoidal approximations of an intersection of two nondegenerate ellipsoids E1 = E(a(1) . Minimizing trQ[m. the solution to which is given by γi0 m = i=1 ci c−1 i and therefore. S2 ) is given by 0 0 S 0 = (1 + p∗−1 )S1 + (1 + p∗ )S2 . .
a(2) = ∗ . Q−1 [S](z − a)) ≤ 1} = {z : Σi. Q− ∈ L(IRn . ∀S ∈ Σ 147 . = 1. Q[S]) with the hyperplane x(1) ≡ x(2) = L. Q[S]) = {z : (z − a. Then (2. where Q[S] = S −1 [(SQ(1) S )1/2 + (SQ(2) S )1/2 ]2 (S )−1 and S is a symmetrical matrix of dimension 2n × 2n. H ∈ L(IR2n . We assume that this intersection has an interior point : intE1 ∩E2 = ∅ ( Assumption 2. E2 are nondegenerate in IRn and the set H is assumed to have an interior point in IR2n . 2} i.j Then (2.23) E(a. IRn ).3) and Corollary 2. IR2n ) The set H is the sum of two degenerate ellipsoids in IR2n . Let us denote a = a(1) + a(2) and Q− Q− 11 12 Q− Q− 21 22 x(1) x(2) Q−1 [S] ≡ z= where x(i) ∈ IR2n . j. i.3 ( where one just has to take the closure of the approximating variety).24) H ∩ L ⊃ E(a. a∗ . a∗ ∈ IR2n .j Let us now intersect sets H and E(a. j = 1. Q− (x(i) − a(j) ))i. Q(1) ) + E(a(2) . Q(2) .6.1 ). Q[H]). 2. it may still be approximated internally according to formula (2.6. Q2 ). Nevertheless. Q[S]) ∩ L. Consider the direct product E1 ⊗ E2 = E(a(1) .3. Q(2) ) = H. Q(2) ≡ 0 a(2) 0 0 0 Q2 Clearly. We may therefore write H ⊃ E(a(1) + a(2) . Q(1) . i. since E1 .and E2 = E(a(2) .j (x(i) − a(i) .3. ∗ ∗ where a(1) = ∗ and Q(1) ≡ (1) (2) a(1) 0 Q1 0 0 0 .6.
j=1 i.24). Namely. (1 − h2 [S])−1 Q∗ [S]).26) where Q∗ [S] = (Σ2 Q− )−1 .6.6. Q[S]) ∩ L = {x : Σi.j 2 The previous reasoning results in Lemma 2.27) E1 ∩ E2 = ∪{E(q[S].( Here Σ is the set of all symmetrical matrices in IR2n ).23) as (2. The last relation follows from (3.j=1 i. b[S] = Σ2 (Q− a(i) + Q− a(j) ) i.j=1 i. taking x(1) = x(2) = x. due to an extension of Corollary 2. The following relation is true (2. Q− (x − a(j) )) ≤ 1} i. we will have (2.6. i. x ∈ IRn .1 .j i.j i. E1 ∩ E2 ⊃ E(q[S]. (1 − h2 [S])−1 Q∗ [S])S ∈ Σ}. Then the internal ellipsoidal approximation of E1 ∩ E2 may be described due to formula (2.6.1 holds: the intersection E1 ∩ E2 of two nondegenerate ellipsoids has an interior point ( intE1 ∩ E2 = ∅).3. Q− a(j) − (q[S]. q[S]).6. 148 .25) H ∩ L = ∪{E(a.25).j ((x − a(i) . Moreover. Q[S]) ∩ LS ∈ Σ} The obtained relations may be now rewritten in IRn .5 Suppose Assumption 2. we may observe that then H ∩ L = E1 ∩ E2 and E(a. h2 [S] = Σ2 (a(i) .6.j and 1 q[S] = (Q∗ )1/2 [S] · b[S].6. where S is any symmetrical matrix in IR2n .j We may now rearrange the previous relation and rewrite (3.6.
6. qi ∈ IRn . . The internal ellipsoids..6. are unmarked ( Fig. .4 (a) and (b)). i = 1.6.6. calculated due to relations (2. . Let us ﬁnd the external estimates of their Minkowski sum m (2.6.2.6. by deﬁnition. 149 . This example was calculated by D. Qi > 0. (3.1) which is.2. An interesting exercise here would be to specify some types of optimal or extremal internal ellipsoids and also to describe some ”minimal ” variety of internals that would nevertheless ”wipe out” the set E1 ∩ E2 ” from inside”. as before.26). Qi ∈ L(IRn .25). E= p(1) ∈E1 14 S[m] = i=1 Ei . We leave this to the interested reader. .4 Here we demonstrate some internal ellipsoids for an intersection E1 ∩ E2 of two ellipsoids where these are marked by numbers 1.4 relations analogous to ( 3. .27) are true for intersections of a ﬁnite number n of ellipsoids.7 Finite Sums and Integrals: External Approximations Consider m nondegenerate ellipsoids Ei = E(qi . p(m) ∈Em p(1) + . Qi ).4(a) and (b) !!! —————— 2.. IRn ).Remark 2. p(m) . However we shall ﬁnalize this section with yet another illustration Example 2. . m.1 Under nondegeneracy conditions similar to those of Lemma 3. 14 ——!!! insert ﬁgures 2.7.6.Potapov.6.
Q(p[3]) = (p1 + p2 + p3 )(p−1 Q1 + p−1 Q2 + p−1 Q3 ).7.7.2. Q(p[3])). E2 = E(0.7.6) Q(p[m]) = i=1 m q[m] = i=1 m qi m pi · i=1 p−1 Qi i for each set of pi > 0.1 The external estimate (2. p2 > 0.2. . i = 1.5) and (2. Applying formula (2. 150 . where Q(p[2]) = (p1 + p2 )(p−1 Q1 + p−1 Q2 ). with parameter p > 0 taken as p= This gives (2.1) once more (now to E(p[2]) and E3 ). E3 = E(0. .2) p1 + p2 p3 p3 > 0. p1 > 0.7.7. Let us ﬁrst take three ellipsoids: E1 = E(0. Q3 ).4) E(p[m]) = E(q[m]. Applying (2.1) is presented in the form p = p1 /p2 . where Q(p[3]) = (1 + p−1 )Q(p[2]) + (1 + p)Q3 ).7. Qi ) is given i=1 by (2. m. Q1 ). Q2 ).1) ﬁrst to E1 + E2 . one comes to E1 + E2 ⊆ E(p[2]) = E(0.7. Q(p[m])) where (2. 1 2 3 Now the general assertion is as follows: Theorem 2. 1 2 and parameter p = p[2] of (2. one obtains E(p[3]) = E(0. Q(p[2])).We shall try to get a hint at the type of formula required. .3) E(p[m]) of the Minkowski sum S[m] = m Ei of m nondegenerate ellipsoids Ei = E(qi .
9) (2. . Direct calculations yield the following Lemma 2. ∗ ) 2 . .E.7. m.7.7.7) (2.7. Q(p[m]) = ρ( ∗ S[m]). The proof is given by induction.D. Assuming (2.1). .4) . In the form of recurrence relations. i Q.7.12) .7. k = 1. 151 (2. For m = 2 the statement clearly follows from (2. p= this gives pi pm+1 m+1 m+1 Q(p[m + 1]) = i=1 pi i=1 p−1 Qi . ∗ ) = 1 ﬁxed. .2. . 1 i = 1. . .8) After taking p > 0 as m i=1 q ∗ [m + 1] = q ∗ [m] + qm+1 Q(p[m + 1]) = (1 + p−1 )Q[m] + (1 + p)Qm+1 .7.7.7.2.10) Q(p[k + 1]) = (1 + pk+1 p−1 [k])Q[k] + (1 + p−1 p[k])Qk+1 k+1 p[k + 1] = p[k] + pk+1 .(2.1) to E(p[m]) + Em+1 one comes to (2.6) to be true for given m and applying (2. pk > 0. one has (2.10) is selected as (2.11) with pi = (Qi ∗ . pm } of (2. then ρ( ∗ E(q ∗ [m]. m ∈ IRn . . .1 If the parameter p[m] = {p1 .Proof.7. ( ∗ .
Bi ] ⊂ IR2 . 2. 2.13) is true. the family {E(0. the ﬁnite sum S[m] may be presented as an intersection of ellipsoids. π2 On the other hand min 1 · Vol2 (E(0. Q(p[m]) contains all the inclusionminimal ellipsoids circumscribed around S[m] The following example illustrates that in the case of adding three. i = 1. Example 2. 3 where A1 = (−1. 0) √ 1 3 A2 = − . p3 ] = Q(p[3]). that is covered by the ball of radius 2 around the origin. p2 . 3 } .14) B1 = (1. with 1 · Vol2 (S(0. 2 2 3 (2. Q[p1 . Q[p1 .7. Q(p[m]). i = 1. this does not mean that the variety E(q ∗ [m]. 2)) = 16.12) implies Lemma 2. does not contain the covering ellipsoid of minimal volume. 2. i = 1. 2 2 √ 1 3 A3 = .7. which now belong to the parametrized variety E(q ∗ [m]. p3 ]) : pi > 0.13) S[m] = { E(q ∗ [m].7. p2 . 2 2 The Minkowski sum F= Fi i=1 is the regular hexagon.Formula (2. p3 ]) : pi > 0. As in the case of two ellipsoids. S(0.2 The following relation is true (2.1.7.− 2 2 √ 1 3 B3 = − . − . 0) √ 1 3 B2 = .7. Consider the segments Fi = [Ai . (or in general more than two) ellipsoids. 3 π2 = 81 . Although the equality (2. Q(p[m])  p[m] ∈ IRm } . Q[p1 . 2) ⊂ IR2 . p2 .7. 4 152 .
A further step is to approximate setvalued integrals. The well known inequality between the arithmetic and geometric mean completes the proof. Q. Q(p[m])) by vectors p[m] > 0. t1 ] −→ IRn Q : [t0 . p3 ]) = 2 · Vol2 (E(0.D. Assume an ellipsoidal valued function P(t) = E(q(t). p2 . Q[p1 . 1 p1 + p2 + p3 3 det(Q[p1 .1. 3 1 0 0 0 − √ 3 4 1 4 3 4 √ − 43 3 4 √ 3 4 √ 3 4 1 4 . p2 .E. we obtain: 1 1 1 + 4p2 + 4p3 p√ 1 1 3 − p13 4 p2 √ 3 1 − p13 4 p2 1 3 + p13 4 p2 . t1 ] given.1. Q(t)) t ∈ [t0 .1 Consider the variety E[m] = E(q ∗ [m]. t1 ] with the functions q : [t0 . p2 . Select an optimal ellipsoid among those of the form E ∈ E[m] relative to the criterion ψ(Q(p[m])) = min where the function ψ is one of those given in Section 2. Q[p1 .7. Exercise 2. p3 ] = (p1 + p2 + p3 ) Calculating the determinant. We have Fi = E(0. t1 ] −→ L(IRn .Proof. 2. p3 ])) = · 1 π 4 (p1 p2 p3 ) 3 3 . Qi ) with Q1 = Q2 = Q3 = Consider the matrix i = 1. IRn ) continuous and Q(t) > 0 for all t ∈ [t0 . What would be its setvalued integral 153 .
t1 ] in the Hausdorﬀmetric h for any subdivision ΣN (2.15) with I(ΣN ) = i=1 E(q(τi ). Q(ΣN )) q (ΣN ) = i=1 ∗ q(τi )σi N i=1 and Q(ΣN ) = N i=1 p∗ (N ) i 2 σi (p∗ )−1 (N )Q(τi ). N − 1. i with p∗ i > 0.7. N } converging to I[t0 . τi = τi−1 + σi − 1. . I[t0 . N } and σ(N ) = max{σi : i = 1.16) lim h(I(ΣN ). t1 ] = t1 t0 E(q(t).Q(·) continuous.7. Q(τi ))σi ΣN = {τ0 = t0 . .I[t0 .7. this.17) where N I(ΣN ) ⊆ E(q ∗ (ΣN ). . obviously is all the more true for the integral of an ellipsoidal valued function P(·). .7. the integral I[t0 . σi >= 0. . t1 ] can be treated as a setvalued Riemannintegral with integral sums N (2. Q(t))dt ? Since the sum of a ﬁnite number of ellipsoids is not obliged to be an ellipsoid. . Applying Theorem (2.7.1) to (2. . 154 . i = 1. t1 ]) = 0. . With the functions q(·). σN →0 N →∞ In the sequel assume σi = (t1 − t0 )/N = σ(N ) for i = 0. .15) we have (2.
21) follows from propositions similar to Lemmas (2.18) and (2. t1 ] is given by relation (2.7. i Assuming p : [t0 . Equality (2. The last argument allows to formulate Theorem 2.7.7.1). (2.19) N →∞ lim Q(ΣN ) = p(τ )dτ p−1 (τ )Q(τ )dτ = Q+ (t1 p(.7. (2. taking pi (N ) = p(t0 + σ(N )) and having in view the continuity of Q we observe t1 t0 t1 t0 (2. t1 ].7.16).))) whatever is the function p(·) > 0. Making a limit transition in (2. namely from 155 . we arrive at the inclusion (2.7.17) in view of (2.20) I[t0 .13).7.7.7. t1 ] denotes the open cone of continuous. the following equality holds (2.7.18) while (2. where C+ [t0 .7.19). t1 ]. Q+ (t1 p(. Moreover. t1 ] = E(qt0 (t1 ). Q+ (t1 p(·)))p(·) ∈ C+ [t0 .2 An external ellipsoidal estimate for the integral I[t0 .20). t1 ] −→ IR to be a continuous function with positive values.)) N →∞ lim q ∗ (ΣN ) = t1 t0 q(τ )dτ = qt0 (t1 ). t1 ] ⊆ E(qt0 (t1 ).21) I[t0 .7.7.After substitution pi (N ) = σ −1 (N )p∗ (N ) the last equality transforms into i N N Q(ΣN ) = i=1 pi (N )σ(N ) i=1 σi p−1 (N )Q(τi ). positive valued functions over the interval [t0 .
The proof follows from direct substitution.Lemma 2.7. t1 ] of (2. 156 .E.27). ∗ ) 2 . Recalling (2. taking p(·) to be ﬁxed.22) Proof.7.7.28) E(q 0 . Prove that for the sum (2.7.23) the diﬀerential equation π(t) = p(t) t t0 p(τ )dτ.20) is selected as p(t) = (Q(t) ∗ . Q0 ) + t1 t0 E(q(t). Let us ﬁnally indicate some diﬀerential relations for qt0 (t) and Q(t) = Q+ (tp(·)).26) (2.7.25) complemented by (2.24).7.2.25) and (2. t1 ]) = ρ( ∗ E(qt0 (t1 ). then the respective support function verify the equality: ρ( ∗ I[t0 . ( ∗ . and (2.7. or.26).27) ˙ Q+ (t) = π(t)Q+ (t) + π −1 (t)Q(t) Q+ (t0 ) = 0 q˙0 (t) = q(t) t qt0 (t0 ) = q(t0 ). with ∗ 1 t ∈ [t0 .D.7. ∗ ) = 1 ﬁxed.7.7. Exercise 2.7.24) (2. Q+ (t1 )) the external ellipsoidal representation is still given by equations (2. (2. Q. the change appearing only in the initial conditions (2.7. Q(t))dt ⊆ E(qt0 (t1 ).3 If the function p(·) ∈ C+ [t0 .7. after diﬀerentiating both sides by t and introducing the notation (2.7. Q+ (t1 p(·)))) (2. so that Q0 and q 0 have to be added on the respective right hand side.18) we have the representation Q+ (t) = t t0 p(τ )dτ t t0 p−1 (τ )Q(τ )dτ . t1 ] ∈ IRn .7.
Q+ (t). Q+ (t). We shall discuss two ways of selecting these.Before ending this Section. in view of (2.7.31) (b) π(t) = Tr1/2 ((Q+ (t))2 ) Tr1/2 (Q(t)Q+ (t)) (2.30) (a) π(t) = (2.32) (c) π(t) = n1/2 Tr1/2 (Q(t)(Q+ (t))−1 ) Summarizing these results. Q(t)) = π(t)Q+ (t) + π −1 (t)Q(t).25). we come to 157 . let us single out some ”individual” external ellipsoids. τ ] ). for example. Integrating relation (2.7.7. we have Q+ (τ ) = where (2. (c) ψ[F[t]] = detF[t].29) F(π(t). Due to the results of Section 2.7. (b) ψ[F[t]] = Tr(F 2 [t]). Let us now minimize the matrix F[t] = F(π. Q(t)) over π ( at each instant t ∈ [t0 . taking. Q(t))dt. the following ”local” optimality criteria ( see Section 2.7. τ t0 F(π(t).5 the respective optimalizers are Tr1/2 (Q+ (t)) Tr1/2 (Q(t) (2.4). Q+ (t).1) (a) ψ[F[t]] = Tr(F[t]).7.
τ ] has to be selected due to equalities (2.7. t ∈ [t0 .(c) taken for each t ∈ [t0 . E0 + E1 = E(0.7.7. assuming again . One should be aware. without loss of generality. may be calculated due to equation (2.30) . in view of Example 2.7. Q0 ) + E(0. Q1 ) ⊇ E(0.24) with control π(t).24). τ ].(2. that qi = 0.27) is nondominated with respect to inclusion.7.4.7. as speciﬁed in Lemma 2.7. .3 Compare the solutions of the optimal terminal control problem for system (2.7.24) the functions π(t) may treated as ( positivevalued ) controls.33) accordingly.33)..(c). We shall introduce the internal ellipsoidal approximation of these.7. E1 .7.7.. where the nonlocal optimality criteria to be minimized over π(t). (In the sense that for each t the respective set E + [t] is an inclusionminimal external ellipsoidal estimate of I[t0 .Lemma 2. m. t ∈ [t0 .7. 15 158 . We shall also calculate the internal ellipsoidal approximations for ﬁnite sums of ellipsoids and for integrals of ellipsoidalvalued functions. Applying formula (2. where the function π(t).(b).8 Finite Sums and Integrals: Internal Approximations m Consider again the sum S[m] = Ei i=1 of m nondegenerate ellipsoids Ei = E(qi .4 (a)The parameters Q+ (τ ) of the external ellipsoids E(qt0 (τ. det(Q+ (τ )) Exercise 2.32) respectively. i = 1.7. 15 TrQ+ (τ ).24).3) to E0 .. could be (2. Q[S1 ]).(b).1. due to optimality criteria (2. t0 ≤ t ≤ τ . One may observe that in equation (2. 2. generated by equations (2.24)(2. The problem of selecting optimal ellipsoids may then be reduced to an openloop terminal control problem. that these criteria would be minimized only in the class of ellipsoids described by formula (2. τ ] . Qi ). we have. t]).7.4. Tr(Q+ (τ ))2 . with the solutions obtained due to local criteria (a). (b) Each of the ellipsoidal ”tubes” E + [t]. Q+ (τ )) = E + [τ ] singled out through the local optimality criteria (a).
2) E(0.4) S[m] = i=1 Ei ⊇ E(0.1 to (2. Q(S[m − 1])) + E(0.where S[1] = S1 and (2.Sk }. Qm ) ⊇ E(0. 1 1 Further on.. k 1 1 S[k] = {S1 .8. Q(S[m]).8. the representation of Corollary 2. Q(S[1]))S1 } + E(0.. Q2 ). Continuing this procedure. assuming that the last relations are true for S[m − 1].8. Q(S[0]) = Q0 . m (2. Si ∈ Σ. Q2 ) = ∪{E(0.1 yields (2. Sm }. Applying the representation of Corollary 2. ∀S[2] = {S1 . Q(S[1])) + E(0. 1 Moreover . Q(S[2])). Q1 ) = ∪{E(0.8.3) S[2] = E0 + E1 + E2 ⊇ E(0. ∀S[m] = {S1 . . we have. Q(S[1]))S[1] ∈ Σ}. . S2 }. where −1 −1 Q(S[2]) = S2 [(S2 Q(S[1])S2 ) 2 + (S2 Q2 S2 ) 2 ]2 S2 ..4. 159 . due to the same representations (2. we come to E0 + E1 + E2 = ∪{E(0.8.8. we have. where (2.4. Q0 ) + E(0.1) −1 Q(S[1]) = S1 [(S1 Q0 S1 ) 2 + (S1 Q1 S1 )]2 S −1 . Q(S[2]))S2 }.5) −1 Q(S[k]) = Sk [(Sk Q(S[k − 1])Sk ) 2 + (Sk Qk Sk ) 2 ]2 (S )−1 .3). E(0.
. Q[m])S[m]}. is treated similarly. by induction. Concluding the discussion. Qi ).. Qi ) = S[m]. Q(S[m − 1]))S[m − 1]} + E(0. S2 }. 160 . m.6) S[m] = ∪{E(0. Similarly.8. i = 1.8. m q[m] = i=1 qi . The following representation is true (2.8) S[m] = ∪{E(q[m].8. Q(S[2])S1 .which gives S[2] = ∪ ∪ {E(0. holds for any sequence S[m].8. The general case. we are now able to formulate Theorem 2.6). Q(S[m]))S[m]}. This allows Corollary 2. (2. ..1 The internal ellipsoidal estimate m E − [m] ⊆ E(0.8. Qm ) = = ∪{E(0. where the union is taken over all the sequences S[m] of symmetrical matrices Si ∈ Σ. i=1 for the sum S[m] of m+1 nondegenerate ellipsoids E(0. Qi ) is given by the inclusion (2.7) S[m] ⊇ E − [m] = E(q[m].8. Q(S[m])). with S[m] = i=0 m E(qi .8.4) with exact representation (2.1 The inclusion (2.
i = 1. τi = τi−1 + σi−1 .10) I(Σk ) = ∪{E( k σi qi .15) with convergence property (2.Remark 2.. and −1 Qσ (S[k]) = Sk (Sk Qσ (S[k − 1])Sk ) 2 + σk (Sk Qk Sk ) 2 ]2 Sk−1 1 1 The last relations are equivalent to Qσ (S[k]) − Qσ (S[k − 1]) = −1 −1 = σk Sk ((Sk Qσ (S[k − 1])Sk ) 2 (Sk Qk Sk ) 2 + σk (Sk Qk Sk ) 2 (Sk Qσ (S[k − 1])Sk ) 2 )Sk + 2 + σk (Sk Qk Sk ). . Qσ (S[k]) and (2. 1 1 1 1 Denoting τk = t.1. Σk = {τ0 = t0 . 161 .. .9.15). However. Qσ (S[k]))S[k]} i=1 where S[k] is such that Si ∈ Σ.m.8.7. k. Applying Theorem 2.8.7. Q(t))dt.. σ 2 Qk ) ⊇ k ⊇ E( i=1 σi qi . j = 0. Exercise 2. Qσ (S[k − 1]) + E(qk . S(τj ) = Sj .8. we observe k−1 (2. .. Its Riemannean integral sum is the one given in (2..8. i = 1... i = 1. k} and S[k] = S[t].8) has to be substituted by its closure.9) I(Σk ) ⊇ E( i=1 qi ..8. the union in the righthand side of (2.. Let us now pass to the internal approximation of the setvalued integral t1 t0 I[t0 .1 to (2..1 The last assertions were proved for the sum S[m] of m + 1 nondegenerate ellipsoids Ei .16). S[i] = S[τi ] = {S(τj ). Prove the assertion of the previous remark. The basic relations turn to be also true if these are degenerate. . σi > 0. i}.7.. t1 ] = E(q(t).
14) dQ− (t)/dt = S −1 (t)((S[t]Q− (t)S [t]) 2 (S[t]Q(t)S [t]) 2 + + (S[t]Q(t)S [t]) 2 (S[t]Q− (t)S [t]) 2 )S −1 (t).k q(i) = q(τi ).15) dq0 (t)/dt = q(t).8.8. Let us assume that.11). i = 1.13) max{σi .8.8. ( for an arbitrary t ) and denoting lim qσ (t) = q0 (t). 1 1 1 1 (2. t] . The inclusion (2. k → ∞.8. Passing to the limit in (2.. . qσ (t) = i=1 q(i)σi we observe. Q− (t0 ) = 0 . lim Qσ (S[t]) = Q− (t) we arrive at the diﬀerential equations (2.8. that the previous relations may be rewritten as (2.8. q0 (t0 ) = 0.12) with (2. t] with values in Σ..9) and the relation (2.11) Qσ (S[t]) − Qσ (S[t − σ]) = σk S −1 [t](S[t]Qσ (S[t − σk ])S [t]) 2 (S[t]Q(t)S [t]) 2 + 2 + σk (S[t]Q[t]S [t]) 2 (S[t]Qσ (S[t − σk ])S [t]) 2 )S −1 [t] + σk (S[t]Q[t]S [t]).12) qσ (t) − qσ (t − σk ) = q(t)σk .(2. matrixvalued function S[τ ]. 1 1 1 1 (2.16) imply 162 k→∞ lim I[Σk ] = I[t0 .8. τ ∈ [t0 .. the values S[τi ] in the above are generated by a measurable. k} → 0.8.
Q(t) are the solutions to the diﬀerential equations (2. where q0 (t).8. This Section ﬁnalizes Part II. t1 ] allows an internal approximation (2.8. (2.8.1 The inclusion (2.2 The sum E(q 0 . t] ⊇ E(q0 (t).8.2 The integral I[t0 . The representation (2.16) yields (2.17) I[t0 . Q− (t))S[t]}} over all measurable functions S[·] of the type considered above. t] = cl{∪{E(q0 (t). Q− (t) satisfy the diﬀerential equations (2.18) I[t0 .8. Further.8.(2.8.18) is true .14). is true.17) where q0 (t).8. where the union is taken over all measurable functions S[t] with values in Σ.8. We shall now apply the results of this part to the problems of Part I.13).8.8.16) is true with (2. Q− (t)). t] allows an internal approximation (2.17) and a representation of type (2.19) q0 (t0 ) = q 0 . An obvious consequence of this Theorem is Corollary 2.8.10) is true for any value of k and since (2. 163 . since (2.1.Lemma 2.8. Q(t))dt = I[t0 .8.18) .10)(2. the limit transition in (2. Q(t0 ) = Q0 . but taken for internal ellipsoids.(2.15) with zero initial conditions .8. The result may be summarized in Theorem 2.8.8. We ﬁnally oﬀer the reader to formulate and solve a problem similar to Exercise 2. whatever is the measurable function S[t] with values in Σ.8.14).8. Q0 ) + t1 t0 E(q(t).10) .8.8.15) with initial conditions (2.
This result is similar to those achieved for static situations in Sections 2. Namely. constructing external and internal ellipsoidal.4.2.valued approximations of the attainability (reachability) domains and tubes. it is again possible. the ellipsoidal solvability tubes constructed here are such that they retain the property of being ”Krasovski bridges”.3).4. Moreover. in its general setting. however. Each of these evolution equations generates a respective variety of ellipsoidalvalued tubes that approximate the original attainability tubes externally or internally and ﬁnally yield. It is shown that the scheme of Section 1. there exists an 164 . 3.5 ( under conventional nondegeneracy conditions). however. but is now given for a dynamic problem ( Sections 3. an exact representation of the approximated tube. In order to achieve these results we introduce two corresponding types of evolution funnel equations with ellipsoidalvalued solutions.6 ). once the starting position is in a speciﬁc internal ellipsoidal solvability tube. through their intersections or unions. We start from systems with no uncertainty.Part III. The problem is substantially more diﬃcult if the system is under the action of uncertain ( unknown but bounded) inputs. This allows to us the internal approximations for synthesizing the control strategies in the target control problem.2.8). The same result is given in backward time (Section 3. ( now following the schemes of Section 1. The important point is that these ellipsoidal approximations that reﬂect the evolution dynamics of uncertain or conﬂictcontrol systems are again described through the solutions of ordinary diﬀerential equations. is that the timevarying coeﬃcients of the approximating ellipsoidal tubes are further described through ordinary diﬀerential equations with righthand sides depending on parameters.3 is substituted for its internal ellipsoidal approximation and the control strategy is constructed accordingly ( Section 3. The main point. to incorporate both internal and external approximations of sums or geometrical (Minkowski) diﬀerences of ellipsoids. The approximation of the domains of attainability under counteraction or of the solvability domains for uncertain systems requires.2.4 remains true except that the the solvability tube of Deﬁnition 1. The speciﬁc advantage of such solutions is that the strategies are given ( relative to the solution of a simple algebraic equation) in the form of an analytical design.4). The external and internal ellipsoidal approximations of the solvability tubes for uncertain systems are derived in Section 3. ELLIPSOIDAL DYNAMICS : EVOLUTION and CONTROL SYNTHESIS Introduction In this part we apply the calculus of Part II to the problems of Part I. One should realize. that attainability domains for linear systems are among the relatively simpler constructions in control theory.. to implement an ” ellipsoidal ” control synthesis in the form of an analytical design ( relative to the solution of an algebraic equation). Once the internal approximation of the solvability tubes are known.
2) (3. 3. 165 . (f − q(t). their crossections being inclusionmaximal at each instant of time and – second – that the respective ellipsoidalvalued mappings satisfy a generalized semigroup property which we call the lower and upper semigroup property – for internal and external tubes accordingly. where A(t) ≡ 0.1.1. In terms of inclusions we have. 3.7. taking (3. −1 (x0 − x∗ .1.analytical control design that keeps the trajectory within this tube despite the unknown disturbances.1. where one may observe some examples on solvability tubes and ellipsoidal control synthesis for 4dimensional systems animated through computer windows. X0 (x0 − x∗ )) ≤ 1. We should emphasize the key elements that allow to use the ellipsoidal tubes introduced here for designing synthesizing control strategies ( both with and without uncertainty).1).9. we shall further presume the constraints on u.1.1 EllipsoidalValued Constraints Let us again consider system (1. Namely. (3. f. P (t)) . x0 to be ellipsoidal–valued: (3.4) (u − p(t). P −1 (t)(u − p(t)) ≤ 1. t0 ≤ t ≤ t1 . to be the ”ellipsoidalvalued bridges”. Q−1 (t)(f − q(t)) ≤ 1.1. q(t) and the vector x∗ are given together with continuous matrix functions P (t) > 0.14. Q(t) > 0 and matrix X0 > 0.5) u ∈ E(p(t).1) x = u + f (t) ˙ x(t0 ) = x0 . speciﬁcally.1) and pass to its transformed version (1.3) (3. The techniques of this part are illustrated in Sections 3. These are – ﬁrst – that the approximating ( internal) ellipsoidal tubes are nondominated with respect to inclusion. It is these two elements that allow the internal ellipsoidal approximations to retain the property of being bridges. where the continuous functions p(t).
1. x0 ) ≤ (l.12) σ→0 lim σ −1 h(X [t + σ].1.1. the attainability domain X [t] = X (t.13) M = E(m. f ) ≤ (l. X0 ) with boundary condition for the attainability tube X [·]. M ) for the solvability tube . u) ≤ (l. (l. P (t)l) 2 .1. X0 )) is deﬁned by the setvalued integral (1.1. the setvalued function X [t] satisﬁes the evolution equation. With f (t) continuous.7) f ∈ E(q(t).1. 166 . with terminal set M being an ellipsoid.1. (l.3. based on (1. m ∈ IRn .8) (3. in terms of support functions. q(t)) + (l.11) X [t] = E(x . X0 ) + t0 ∗ E(p(t) + f (t).6) (3.1. On the other hand.10) (l.14) σ→0 lim σ −1 h(W [t − σ]. the inequalities (3. P (t)) = 0 W [t1 ] = E(m. P (t))dt. x∗ ) + (l.1. (3. X [t] + σE(p(t) + f (t).9) (3. M > 0. Q(t)l) 2 . With f (t) given. W [t] − σE(p(t) + f (t).3) or (1. M ∈ L(IRn .6) (3. x0 ∈ E(x∗ . t0 .14.3. we have an evolution equation (3. X0 ) or. X0 l) 2 ∀l ∈ IRn . P (t)) = 0 X [t0 ] = E(x∗ . p(t)) + (l.1). M ). which is now t 1 1 1 (3. IRn ). E(x∗ . Q(t)) .(3.
(3. t1 ] contains an internal tube β(t)S.15) After the introduction of an additional ellipsoidalvalued state constraint (3. y(t) ∈ IRk . under unknown but bounded disturbances f (t) ∈ E(q(t).1. Q(t)). y(t) to be continuous).1 or 1. and the constraint (3. bounded by restriction (3. the equation for the solvability tube under both uncertainty and state constraints is as follows σ→0 lim σ −1 h+ (W [t − σ] + σE(q(t).6). P (t))) = 0 W [t1 ] = E(m.1.1. 16 167 . which is 16 lim σ −1 h(W [t − σ] + σE(q(t). propagating them to continuoustime dynamic processes. We shall therefore introduce external and internal ellipsoidal approximations of these within a scheme that would generalize the results of Part II.Passing to an uncertain system with f (t) measurable. K(t)) − σE(p(t). As we have observed in Part II.1. M ) ∩ E(y(t1 ).1.1. β(t) > 0 . If the function u(t) is given. the sets X [t].3). Our further subject is therefore the one of ellipsoidalvalued dynamics.1. (X [t] is the maximal solution to this equation).16) is the corresponding information domain which satisﬁes the evolution equation (3. y(t).1. we come to the equation for the solvability tube under uncertainty. (3. IRk ).1.19) y(t) ∈ G(t)x(t) + E(0.2 which imply that the tube W [t]. K(t) continuous. then the attainability domain X [t] for system (3.1.7). Q(t)). The set X [t] gives a guaranteed estimate of the state space vector x(t) of system (3. W [t] − σE(p(t). Q(t))) = 0. X0 ). W [t1 ] = E(m. σ→0 (3. through the measurement of vector (3. W [t] generated by the solutions to the evolution equations of this Section.17)).1. K(t1 )).7. K(t)) + σE(q(t) + u(t).11) (u(t) given) . are not obliged to be ellipsoids. P (t))) = 0.7. W [t] ∩ E(y(t). K(t)). (3. we start from the simplest attainability problem. Q(t)). We recall that this equation was introduced under nondegeneracy Assumptions 1.1). t ∈ [t0 . K(t) > 0. X [t] ∩ E(y(t).1. Following the sequence of topics of Part I.1.17) (The solvability tube is the maximal solution to (3. K(t)) with K(t) ∈ L(IRk . M ).18) σ→0 lim σ −1 h+ (X [t + σ]. X [t0 ] = E(x∗ .16) Gx(t) ∈ E(y(t). (3.16) is due to a measurement equation with observed values y(t) (assuming u(t).
so that if necessary. and to approximate its attainability domain X [t] = X(t.7. we have (3.3) where (3.7.2 Attainability Sets and Attainability Tubes: the External and Internal Approximations Our ﬁrst subject is to consider the diﬀerential inclusion (3.2. P (s))ds.2).2. X0 )). X + (t0 ) = X 0 . (3. X + (t)).2 and the substitution (2.2. Applying these relations to the present situation and changing the notations to those of (3. Here X + (t) actually depends on π(·).24).23) that the inclusion 168 . ˙ X [t] ⊆ E(x∗ (t). t0 ≤ t ≤ t1 . (3.2. where t (3. (2.7. X0 ). through relations (2.1) x ∈ E(p(t).20).3.2) X [t] = E(x . x0 ∈ E(x∗ .4) x∗ (t) = p(t) + f (t). It follows from Theorem 2.2. we shall also use the notation X + (t) = X + (tπ(·)). t0 .6) x∗ (t0 ) = x∗ .22). (2.7. X0 ) + t0 ∗ E(p(s) + f (s). The external ellipsoidal approximation for such a sum had been indicated in Section 2.7.2. π(t) > 0.2. ˙ x(t0 ) = x0 . particularly.7. P (t)) + f (t).5) ˙ X + (t) = π(t)X + (t) + π −1 (t)P (t). E(x∗ .
(3.2.7)
X [t] ⊆ E(x∗ (t), X + (tπ(·))
is true, whatever is the function π(t) > 0 that allows representation (2.7.20), (2.7.23) with p(t) > 0. Moreover, the equality (3.2.8) X [t] = ∩{E(x∗ (t), X + (tπ(·))π(·)}
is true if the intersection is taken over all the functions π(·) of the type indicated above. We leave to the reader to observe that (3.2.7) remains true if the intersection is taken over all piecewise continuous or even continuous functions π(t) > 0. This ﬁnally leads to the proof of the following assertion Theorem 3.2.1 The external ellipsoidal approximation to the attainability domain X [t] = E(x∗ (t), X + (t)) of the diﬀerential inclusion (3.2.1) is given by the inclusion (3.2.7) with exact representation (3.2.8), where the intersection may be taken over all piecewise– continuous (or even continuous) functions π(t) > 0. Let us now return to the last theorem, approaching it through another scheme  the technique of funnel equations. Following Sections 1.4, 3.1, we observe that the tube X [t] satisﬁes the funnel equation (3.1.12). This allows to write X [t + σ] ⊆ X [t] + σE(p(t), P (t)) + o(σ)S, where σ −1 o(σ) → 0 if σ → 0, and S is a unit ball in IRn , as before. With X [t] being an ellipsoid of type X [t] = E(x∗ (t), X + (t)), we may apply the expansion (2.5.6), so that the external approximation to X [t + σ] would be (3.2.9) where (3.2.10) X + (t + σ) = X + (t) + σπ −1 (t)X + (t) + σπ(t)P (t) + σ 2 P (t), X [t + σ] ⊆ E(x∗ (t + σ), X + (t + σ)) ,
with π(t) > 0 continuous. Relations (3.2.9) , (3.2.10) are true for any σ > 0 and any π(t) of the indicated type. 169
Dividing the interval [t0 , t] into subintervals with subdivision = {σ1 , ..., σs },
s
τ0 = t0 , τs = t0 +
i=0
σi , t = τs ,
where
s
σ > 0,
i=0
σi = t − t0 ,
we have : X (τ1 ) = E(x∗ , X + (t0 )) + σ1 E(p(t0 ) + f (t0 ), P (t)) ⊆ ⊆ E(x∗ (τ1 ), X + (τ1 )) = E + [τ1 ], where (3.2.11) x∗ (τ1 ) = x∗ + σ1 (p(t0 ) + f (t0 ))
(3.2.12)
2 X + (τ1 ) = (1 + σ1 π −1 (t0 ))X + (t0 ) + σ1 π(t0 )P (t0 ) + σ1 P (t0 ).
We further have: X + (τk ) ⊆ E(x∗ (τk−1 ), X + (τk−1 )) + σk E(p(τk−1 + f (τk−1 ), P (τk−1 )) ⊆ ⊆ E(x∗ (τk ), X + (τk )), (k = 1, ..., s), where (3.2.13) x∗ (τk ) = x∗ (τk−1 ) + σk (p(τk−1 ) + f (τk−1 )),
(3.2.14)
X + (τk ) = (1 + σk π −1 (τk−1 ))X + (τk−1 ) + σk π(τk−1 )P (τk−1 ).
Dividing relations (3.2.13), (3.2.14) by σk and passing to the limit, with 170
max{σk k = 1, ..., s} → 0 , s → ∞, and t being ﬁxed as the endpoint of the interval [t0 , t), whatever is the subdivision and the integer s, we again come to equations (3.2.4), (3.2.5) with initial condition (3.2.6). This gives an alternative proof for the relation (3.2.7) of Theorem 3.2.1. Let us now assume A(t) ≡ 0. Then Theorem 3.2.1 transforms into Corollary 3.2.1 For every t ∈ [t0 , t1 ] the following equality is fulﬁlled X (t) = ∩{E(x(t), X + (tπ(·))π(·))}, where X + (t) = X + (tπ(·)), are the solutions of the following diﬀerential equations x = A(t)x + p(t); x(t0 ) = x∗ , ˙
˙ X + = A(t)X + + X + A (t) + π −1 (t)X + (t) + π(t)P (t);
X + (t0 ) = X0 .
We shall now indicate that with a certain modiﬁcation this result remains true for the special case when E(p(t), P (t)) is a product of ellipsoids and therefore does not fall under the requirements of Section 3.1. Let us start from a generalization of Lemma 2.2.1. Lemma 3.2.1 Suppose E1 = E(q1 ; Q1 ), E2 = E(q2 ; Q2 ) where Q1 = A1 0 0 0 , Q2 = 0 0 0 A2 ,
A1 (A2 ) is a symmetric positively deﬁned m × m (respectively, k × k−) matrix, m + k = n. Then E1 + E2 = ∩{E(q1 + q2 , Q(p))p > 0}, Q(p) = (1 + p−1 )A1 0 0 (1 + p)A2 .
171
Proof. The upper estimate E1 + E2 ⊆ E(q1 + q2 ; Q(p)), p > 0, can be obtained along the lines of the proof of Lemma 2.3.1. Consider now an arbitrary vector v = {l, b} ∈ IRn , l ∈ IRm , b ∈ IRk such that l = 0, b = 0. It is not diﬃcult to demonstrate that ρ(vE1 + E2 ) = v (q1 + q2 ) + (l A1 l) 2 + (b A2 b) 2 = = v (q1 + q2 ) + (v Q(p)v) 2 for p = (l A1 l) 2 /(b A2 b) 2 . This yields ρ(vE1 + E2 ) = ρ(vE(q1 + q2 , Q(p))), for every direction v = {l, b} with l = 0, b = 0. From the continuity of the support functions of the convex compact sets E1 + E2 and of the set ∩{E(q1 + q2 , Q(p))p > 0} we conclude that equality ρ(vE1 + E2 ) = ρ(v ∩ {E(q1 + q2 ; Q(p))p > 0}) is true for all v ∈ IRn . The last relation implies the assertion of this Lemma Q.E.D. Denote symbol Π+ [t0 , t1 ] to stand for the set of all continuous positive functions from [t0 , t1 ] to IR1 . Combining the last Lemma with Corollary 3.2.1, we come to the conclusion Corollary 3.2.2 Consider the diﬀerential inclusion x ∈ A(t)x + P(t), ˙ x(t0 ) ∈ E(x∗ , X0 ), t0 ≤ t ≤ t1 , with 172
1 1 1 1 1
P(t) = Ek (s(t), S(t)) × Em (q(t), Q(t)), where Ek (s(t), S(t)) ⊂ IRk , Em (q(t), Q(t)) ⊂ IRm , k + m = n. For every t ∈ [t0 , t1 ] the following equality is true X(t) = ∩{E(z(t), Z(tπ(·), χ(·))π(·), χ(·)} where {π(·), χ(·)} ∈ Π+ [t0 , t1 ], and z : [t0 , t1 ] → IRn , Z : [t0 , t1 ] → L(IRn , IRn ) are the solutions to diﬀerential equations z = Az + v(t), ˙ v(t) = {s(t), q(t)}, z(t0 ) = x∗ .
˙ Z = A(t)Z + ZA (t) + χ−1 (t)Z + χ(t)Q[t],
Q[t] = Q(t, π(t)) = and Z(t0 ) = X0 .
(1 + π −1 (t))S(t) 0 0 (1 + π(t))Q(t)
In order to deal with internal approximations, we will follow the last scheme, dealing now with funnel equation (3.1.12). This time we have (3.2.15) where σ −1 o1 (σ) → 0 , σ → 0. With X [t] being an ellipsoid of type X [t] + σE(p(t), P (t)) ⊆ X [t + σ] + o1 (σ)S,
173
X − (t + σ)).15) X [t + σ] ⊇ E(x∗ (t + σ). namely. we come to ordinary diﬀerential equations which are equation (3. 1 1 −1 (t).2. 174 1 1 −1 . taking Q1 = X − (t) .2. similar to the one for the external approximations of the above . where X − (t + σ) = H −1 (t)[H(t)X − (t)H (t) + σ(H(t)X − (t)H (t)) 2 (H(t)P (t)H (t)) 2 + 1 1 + σ(H(t)P (t)H (t)) 2 (H(t)X − (t)H (t)) 2 + σ 2 H(t)P (t)H (t)]H and x∗ (t + σ) = x∗ (t) + σ(p(t) + f (t)). in view of (3.X [t] = E(x∗ (t). X − (t0 ) = X0 . we have.2.3. Q2 = σ 2 P (t) . Changing the respective notations. After a discretization and a limit transition in the last equations. X − (t)) .3) with (2. S = H(t) .1).16) dX − (t)/dt = H −1 ((H(t)X − (t)H (t)) 2 (H(t)P (t)H (t)) 2 + 1 1 + (H(t)P (t)H (t)) 2 (H(t)X − (t)H(t)) 2 )H with initial conditions x∗ (t0 ) = x0 .4) and (3. .3. we may apply formula (2.
X 0 )) of the diﬀerential inclusion (3. where B(t) is continuous. where the union is taken over all measurable matrixvalued functions with values in Σ. the given relations still remain true under relaxed assumptions that allow degeneracy in the following sense.2. (3.2.2.2. However.2.1 . P (t)).19) x ∈ A(t)x + B(t)E(p(t).where x∗ (t).2. where x∗ (t). E(x∗ . One may remark.1. t)B(τ )E(0.6).17) . q0 (t). Relation (3. m < n. x∗ (t). Consider system (3.2.the variety of symmetric matrices. X − (t) satisfy the equations (3.18) is a direct consequence of Corollary 2. where one just has to change notations S(t).16) and H(t) is a continuous function of t with values in Σ . P (τ ))dτ.4).What follows from here is the inclusion (3. P (t)). X − (t)). ˙ x(t0 ) = x0 .8.2 The internal approximation of the attainability domain X [t] = X(t.4). where matrix S(τ. 175 . that all the earlier conclusions of this Section were made under the assumptions that ellipsoids E(p(t).1) is given by the inclusion (3.18) X [t] = ∪{E(x∗ (t).1. P (t) ∈ L(IRm .2.20) X [t] = t0 ∗ S(τ. t0 . p(t) ∈ IRm . Moreover. X − (t) respectively.2.9. the following representation is true (3.2. X(t) satisfy (3. x0 ∈ E(x∗ . E(x∗ . The parameters of this sytem allow to generate the setvalued integral t (3. t) is deﬁned in Section 1. X0 ) are nondegenerate. see (1. IRm ). A detailed proof of the same inclusion follows from Theorem 2. Q− (t) to H(t). (3. X − (t))H(·)}.16). X0 ). of course.2.1. The given reasoning allows to formulate Theorem 3.2.17) X [t] ⊇ E(x∗ (t).
(3.2.16) is then substituted by (3.2. Namely.2.2 Suppose Assumption 3. [147]. .2. on every ﬁnite time interval [t0 . t].21) ˙ X + (t) = A(t)X + (t) + X + (t)A (t) + π(t)X + (t) + π −1 (t)B(t)P (t).7 and the reasoning of the present Section.1 for system (3.2.21) . It is actually equivalent to the requirement that system ( 3.2. The details of the proof follow the lines of Section 2.2.2.2.2 is also true for internal ellipsoidal approximations ( under the same Assumption 3.1). ˙ we have the following assertion.2.22) x∗ (t) = A(t)x∗ + B(t)p(t) + f (t).2). Under the last Assumption the analogies of Theorems 3. 3.2. [212].(3.19) has a nonempty interior ( intX [t] = ∅). System (3.23) dX − (t)/dt = = A(t)X− (t) + X− (t)A(t) + H −1 ((H(t)X − (t)H (t)) 2 (H(t)P (t)H (t)) 2 + + (H(t)P (t)H (t)) 2 (H(t)X − (t)H(t)) 2 )H 176 1 1 1 1 −1 .Assumption 3. for all t > t0 .19) is true.1 There exists a continuous scalar function β(t) > 0.4) substituted by ( 3.1. Lemma 3. (3.2.2. l)1/2 . Then the results of Theorems 3.2.2.2.1 (for the attainability domain X [t] of this system) remain true with equations (3. taking equations (3.1 and Corollary 3. This assumption implies that the attainability domain X [t] of system (3.19) with unrestricted control u(t) would be completely controllable.2 for system (3.5).2.2.2.1 An assertion similar to Lemma 3. t > t0 such that the support function ρ(lX ∗ [t]) ≥ β(t)(l.2. Remark 3.19) still remain true.
16) may be interpreted as controls which. we shall introduce some evolution equations for ellipsoidalvalued mappings. X + (τ )) or E(x∗ (τ ). some evolution equations with ellipsoidalvalued solutions ? Let us investigate this issue. This.5). may be selected on an interval [t0 . Prior to that.2.2.4 (due to a ”local” optimality condition) or due to nonlocal criteria. τ ] so as to optimalize the terminal ellipsoid E(x∗ (τ ). however. Prove the statement of Remark 3.2 It is now possible to single out individual ellipsoidal tubes that approximate X [t] externally or internally. X − (τ )) in the classes of ellipsoids determined by equations (3. namely. We emphasize once again that that functions π(t). in their turn.2.3.7. X 0 ).1. (1. The next natural step would be to introduce ellipsoidal approximations for solvability tubes. X [t] + σE(p(t).7.21)(3. P (t) + f (t))) = 0. for example ( see Exercise 2.1.3 Evolution Equations with Ellipsoidal . we come to what seems to be a natural question : do the ellipsoidal mappings that approximate X [t] satisfy.23).1.1.2. may be done as described in Lemma 2. Writing down the evolution equation (3.3).Valued Solutions Having found the external and internal ellipsoidal approximations for the attainability domains X [t] and recalling that X [t] satisﬁes an evolution ” funnel” equation.12) for X[t] with the ellipsoidal data of Section 3.2.33).1) lim σ −1 h(X [t + σ]. it should be clear that in general the solution to (3. H(t) in equations (1.3. for example. Remark 3.2. As indicated in the above. X [t0 ] = E(x0 . 177 σ→0 . particularly. we have (3. 3.2. of types (2.7. Let us now introduce another equation.1) is not ellipsoidalvalued.Exercise 3.
P (t)). (3.3.2. Lemma 3.3.2) is not unique.3. This follows. it follows that always E + [t] ⊇ X [t].(ii)). t1 ] = E(x0 . (taken for π(t) selected due to (3. E[t] + σE(p(t).3. E(t0 ) = E(x0 .1 The external approximation E + [t] = E(x∗ (t).3. P (τ )l∗ (τ )) 2 .2) with respect to inclusion. l∗ (τ )) = 1. I0 [t0 .1). The last lemma indicates that the solution to equation (3.2) σ→0 lim h− (E[t + σ].7). X + (t + σπ(·)) + o(σ)S ⊇ E(x∗ (t).3. Deﬁniton 3. t1 ]).3. X 0 )) + t1 t0 E(p(t).3. P (t))dt. X 0 ).3.2) if (i) E + [t] satisﬁes (3.3)) that ensure the minimality property (iii) of Deﬁnition 3. From the deﬁnition of the semidistance h− and of the solution E + [t] (points it (i). t0 ≤ τ ≤ t.3.3) π(τ ) = (l∗ (τ ). 1 where l∗ (τ ) is a measurable function of t. (iii) E + [t] is the minimal solution to (3. (l∗ (τ ).2) and from the equalities ρ(lE + [t]) = ρ(lI0 [t0 .1. due to (3. X + (tπ(·))) is a solution to the equation (3. so that E + [t] is an external approximation of X [t].1.3. P (t) + f (t)l)) = 0.1 A function E + [t] is said to be a solution to the evolution equation (3. X + (tπ(·))) + σE(p(t).2) in the sense of Deﬁnition 3.3. that ensures the ellipsoidalvalued function E[t] to satisfy (3. This is all the more true due to 178 .2) almost everywhere. (ii) E + [t] is ellipsoidalvalued.3.3. from the inclusion E + (x∗ (t + σ).(3. provided π(τ ) > 0 is selected as (3.
X ∗ ) + and a direct substitution leads to Lemma 3.3. X 0 ))) ⊇ E + (t. τ.16) with x∗ (τ. E(x0 . whatever is the measurable function π(τ ) > 0.2. x∗ ) = x∗ . we shall also denote E + (t. P (τ )l)(l. 179 t τ E(p(s).2.2 The following inclusions are true (3. X + (τ. P (τ )l)(l. (iii) E − [t] is the maximal solution to (3. obviously E + (t. X ∗ )) ⊇ E(x∗ . t0 . X ∗ )). X 0 ). τ.7. X ∗ )) = E(x∗ (t.5) almost everywhere.5) lim σ −1 h+ (E[t + σ].2 A function E − [t] is said to be a solution to the evolution equation (3. τ.1 The ellipsoidal function E + [t] = E(x∗ (t).5) if (i) E − [t] satisﬁes (3.2) consider equation (3. l) = 1} ≤ π(τ ) ≤ max{(l. X + (t. E(x∗ . (ii) E − [t] is ellipsoidalvalued.3. E(x0 .3.3. E(x∗ . X ∗ )) .3.4) E + (t. X ∗ .3. x∗ ). E + (τ. t0 .3.3.4) describe the dynamics of the external ellipsoidal estimates E + (t. where x∗ (t. P (s))ds σ→0 . l) = 1}. τ. τ. x∗ ).(3. P (t))) = 0.3. E[t] + σE(p(t). The proof of this corollary follows from the results of Sections 2.5) with respect to inclusion. Relations (3.3. For a given function π(t) and given initial pair x∗ . 2. τ. E(x∗ . t0 ≤ τ ≤ t1 selected due to the inequalities min{(l. The sets E + (t.3. Together with (3. Then. E[t0 ] = E(x0 . X ∗ ) satify (3. t0 ≤ τ ≤ t. They thus deﬁne an ”upper” semigroup property of the respective mappings. X ∗ ) = X ∗ .2) . τ. X + (t. E(x0 . Deﬁniton 3. t0 .4). X 0 )). X 0 )) are sometimes referred to as supperattainability domains .3.Corollary 3. τ. X + (tπ(·))) is a solution to (3. τ. τ.
t0 ≤ τ ≤ t.16) with initial conditions x∗ (τ ) = x∗ . For a given function H(t) and a given initial pair x∗ .This leads to Theorem 3. X 0 )).6) E − (t.3. clearly. The last relation describes the dynamics of the internal estimates E− (t. X − (t) are the solutions to (3. X ∗ denote E− (t.2. (3. τ. Moreover.(3. t0 .16).3.4). we have Lemma 3.3. E(x∗ . E(x∗ .3 for any function E− [t] = E(x∗ (t). P (s))ds. X − (t)) are the solutions to diﬀerential equations (3. whenever x∗ (t). X0 )) are sometimes referred to as subattainability domains. E− (t. t0 . X ∗ )) + and a direct substitution leads to Lemma 3.(ii) of Deﬁnition 3.3. X 0 ))) ⊆ E − (t. where x∗ (t). Thus. E(x∗ .2. X − (·) to equations (3.2. t0 . τ. Then.5) that satisﬁes points (i).3.From the deﬁnition of the semidistance h+ and of the solution E − [t] (points (i). τ. X − (t)) is a solution to the evolution equation (3.16).2 is an internal approximation for X [t]. t τ E(p(s). X ∗ )) for the attainability domains X [t]. E(x0 . representation (3. X ∗ )) ⊆ E(x∗ . deﬁning thus a ”lower” semigroup property for the respective mappings.3.2.3. X ∗ )) = E(x∗ (t). X − (t)). X − (τ ) = X ∗ . X − (t)) generated by the solutions x∗ (·).2. (3.18) yields the fulﬁllment of the requirement of point (iii) of Deﬁnition 3. 180 .2.4).5). A similar type of descriptions may now be introduced for solvability tubes.2.3 Any solution E − [t] to (3.1 The internal approximation E − [t] = E(x∗ (t). E(x0 . τ.4 The following relation is true (3.2. E(x0 . The sets E− (t.(ii)) it follows that always E − [t] ⊆ X [t]. E − (τ.4).
9.4.8.4) with boundary conditions (3.4.4.2)(3.4) + (H(t)P (t)H (t)) 2 (H(t)X− (t)H(t)) 2 ]H with boundary conditions (3. x(t1 ) = m . X+ (t). 3.3.4. we recall that in our case t1 t (3. whatever are the solutions to diﬀerential equations (3.4. with obvious changes of signs.2.5) 1 1 −1 (t). 1 1 ˙ X− (t) = −H −1 (t)[(H(t)X− (t)H (t)) 2 (H(t)P (t)H (t)) 2 + (3.4. P (t))dt.5) with π(t) > 0.6) E− (x(t).4. 3. we come to the diﬀerential equations (3. Returning to relation (1. Our aim is to approximate these tubes by ellipsoidalvalued functions.3.1.1.4).2)(3. 3.2. ˙ (3. we then come to Theorem 3.1. X+ (t1 ) = M.3. the last assertion develops into exact representations 181 .4. X− (t)) ⊆ W [t] ⊆ E+ (x(t). Denote the solutions to (3.5) as x(t).4.2) x = p(t) + f (t).17). following the approximation schemes of Sections 2. Then. 2. M ) − E(p(t) + f (t). X− (t1 ) = M. X− (t) respectively.4.1 The following inclusions are true (3. ( 3. X+ (t)).1) W [t] = E(m. Similarly to (3. H(t) ∈ Σ.3 .3). As in the previous Sections 3.1.4 Solvability in Absence of Uncertainty We shall now pass to the treatment of solvability tubes for the simplest case of systems without uncertainty and state constraints.3) ˙ X+ (t) = −π(t)X+ (t) − π −1 (t)P (t).
(3. P (t)) = 0.1. with the same boundary condition (3. P (t)) = 0.14) σ→0 lim h+ (E[t − σ]. E+ (t).4. The next issue is to write down an evolution equation with ellipsoidalvalued solutions for each of the approximating functions E− (t). namely.Theorem 3.8) min{(l.11) σ→0 lim σ −1 h− (W [t − σ]. σ→0 (3. σ→0 lim σ −1 h+ (W [t − σ].2). 182 .4. P (t)) = 0.4. l) = 1} ≤ π(t) ≤ max{(l. Consider equations (3. The solution to these equations are not obliged to be ellipsoidalvalued.10) (3. Therefore.9) W [t] = ∪{E− (x(t).15) E[t1 ] = E(m.4. M ). where the union is taken over all measurable functions H(t) with values in Σ. P (t)) = 0 with boundary condition (3. E[t] − σE(p(t) + f (t). we introduce another pair of equations.13) lim h− (E[t − σ].4.2 (i) The following ”external” representation is true (3. where the intersection is taken over all measurable functions π(t) that satisfy the inequalities (3. P (t)l)(l. W [t] − σE(p(t) + f (t).7) W [t] = ∩{E+ (x(t).4. in analogy with (3. l) = 1}. W [t] − σE(p(t) + f (t).4. X+ (t))π(·)}.4. M ). X− (t))H(·)}.4. (ii) The following ”internal ” representation is true (3. E[t] − σE(p(t) + f (t).12) W [t1 ] = E(m. P (t)l)(l.4.
(properties (i).E+ [t]) satisﬁes (3. is said to be a solution to the evolution equation (3. M ∗ )) = E(x(t. It also follows that the last relations are true for any functions E+ [t].4.4.4.4).4.16) E− [t] ⊆ W [t] ⊆ E+ [t].2). M ∗ )). X− (t.(ii) of Deﬁnition 3.E− [t]) is ellipsoidalvalued.13) (resp. M ∗ )). (3. M ∗ we shall denote (3. (3.13) (resp. whatever is the measurable function H(t) with values in Σ.17) E+ (t. X− (t. (3. M ∗ ) satisfy (3. (3.4. τ. τ.5) is a solution to the evolution equation (3.8). E(m∗ .4. (3. τ.4.4). τ. (ii) E+ [t] (resp.if (i)E− [t] (resp. m). M ∗ )) = E(x(t.4.4. E− [t].3 (i) The ellipsoidalvalued function E+ [t] = E(x(t).4.4. X+ (τ. X+ (t. τ.2).2.3. H(·) and a given pair of boundary values m∗ .1.4.3). X+ (t. (3. (3. τ. From the deﬁnitions of the semidistances h− . (ii) The ellipsoidalvalued function E− [t] = E(x(t). maximal ) solution to (3. τ. whatever is the measurable function π(·) selected due to the inequalities (3.4.4. τ.13) (resp.18) E− (t. M ∗ ).4. τ.1 A function E+ [t] ( respectively.This gives Theorem 3. The minimality and maximality properties of the respective solutions are described similarly to Sections 3. m∗ ) = m∗ .4. m∗ ). that satisfy properties (i).(ii)) it follows that always (3.14) ) . X+ (t) to the diﬀerential equations (3. where x(t. E− [t]).2). 3. X+ (t)) generated by the solutions x(t). h+ and of the solutions E+ [t]. For a given pair of functions π(·). (3.4. E− [t].4. m∗ ). 183 .14)) with respect to inclusion.4. M ∗ ) = X− (τ. τ.14) ) almost everywhere.5) is a solution to the evolution equation (3. .4.E− [t]) is the minimal ( resp. τ. X− (t)) generated by the solutions to the diﬀerential equations (3.14). (iii) E+ [t] (resp.4.4. with boundary condition x(τ. τ.4.13). (3. E(m∗ .Deﬁniton 3. M ∗ ) = M ∗ . (3.4.3).
t1 .21) describe the dynamics of the ellipsoidal estimates for W [t]. M ∗ )).4. τ.4. t1 . Relations (3.20).4.1. t1 . E(m∗ .4.4. (3.4. E(m. obviously. E(τ. M ∗ )) ⊆ E(x∗ .1. τ.1). τ t E(p(s). E(m.20) E+ (t.4.21) x = A(t)x + p(t) + f (t). ˙ (3. τ. Exercise 3.4.Then.1 The following inclusions are true with t ≤ τ ≤ t1 : (3. 184 . M )) ⊆ E+ (t.4) are then substituted by (3. By direct calculations prove that equations (3.1 Assume that the original system (3.(3.4. with same boundary condition (3. E(m.2) . E(m.19) E− (t. t1 . (3. M ))). P (s))ds ⊆ E+ (t. (3. (now . E+ (τ. E− (t. E(m∗ .22) ˙ X+ (t) = A(t)X+ (t) + X+ (t)A(t) − π(t)X+ (t) − π −1 (t)P (t).4.4. M ))) ⊆ E− (t.5) as before. τ. that is with A(t) = 0. M ∗ )) − and a direct substitution leads to Lemma 3.23) ˙ X− (t) = A(t)X− (t) + X− (t)A(t) − H −1 (t)[(H(t)X− (t)H (t)) 2 (H(t)P (t)H (t)) 2 − + (H(t)P (t)H (t)) 2 (H(t)X− (t)H(t)) 2 ]H 1 1 1 1 −1 (t). respectively deﬁning. The next step is to proceed with the approximations of solvability tubes for systems with uncertainty. M )). however) the ”lower” and ”upper” semigroup properties of the corresponding mappings.4.1) is given in the form (1. in backward time.
as shown in Section 1.15). this would give (3. t1 ]. would form an appropriate variety of external and internal ellipsoidal approximations to the solution W [t] of (3. Q(t)) ⊆ W [t] − σE(p(t). the one that gives precisely the solvability set.2) W [t − σ] + σE(q(t). H −1 [(H(t)X(t − σ)H (t)) 2 + σ(H(t)Q(t)H (t)) 2 ]2 H and (3. X(t − σ)) + σE(q(t). as in the previous Section. −1 ).5. that their solutions. We shall now look for an ellipsoidalvalued function E(x(t). Due to (2. see footnote for formula (3.15) 185 .5. we will try.5) 17 E(x(t). P (t))) ⊆ This equation is treated under nondegeneracy Assumptions 1.5. (3. Taking equation (3.1.7.4) W [t] − σE(p(t).1) and an external approximation for the righthand side.1. X(t)) − σE(p(t). (3. X(t)) = E[t] that would ensure an internal approximation for the lefthand side of (3.15).1. t ∈ [t0 . constructing them such. would be ellipsoidalvalued and on the other.3. P (t)) + o(σ)S. that f (t) is unknown but bounded ? Since we do not expect ellipsoidalvalued functions to be produced by solving (3.15). Q(t)) ⊇ ⊇ E(x(t − σ). on one hand.3) E(x(t − σ).1.2).4. Being interested in the solvability set (under uncertainty) we shall further presume that W [t] is inclusionmaximal namely. P (t)) ⊆ E(x(t).1). Q(t)). X(t − σ)) + σE(q(t).1.1) 17 W [t − σ] + σE(q(t).5. Q(t)) ⊇ 1 1 ⊇ E(x(t − σ) + σq(t). we again observe that in general its setvalued solution W [t] is not ellipsoidalvalued.4.6 1. 1.1. relation ( 3. to introduce other evolution equations than (3.(2. P (t)).5. (3.16) for the solvability tube of such a system. (3.15).5.15). yields (3.16). How should we construct the ellipsoidal approximations for W [t] now. Indeed. X(t)) − σE(p(t).5 Solvability Under Uncertainty In this section we discuss solvability tubes for uncertain systems with unknown input disturbances.1.2.7.1.7.1.
X[t − σ] + σH −1 (t)[(H(t)X(t − σ)H (t)) 2 (H(t)Q(t)H (t)) 2 + +(H(t)Q(t)H (t)) 2 (H(t)X(t − σ)H (t)) 2 )]H (3.8) X(t − σ) + σH −1 (t)[(H(t)X(t − σ)H (t)) 2 (H(t)Q(t)H (t)) 2 + +(H(t)Q(t)H (t)) 2 (H(t)X(t − σ)(H (t)) 2 ]H 1 1 1 1 x(t − σ) + σq(t) = x(t) − σp(t).5.10) ˙ X+ (t) = −π(t)X+ (t) − π −1 (t)P (t) + x = p(t) + q(t).7) and (3.5.9) and (3.⊆ E(x(t) − σp(t).5.5.6) 1 1 1 1 −1 (t) = = E(x(t) − σp(t). (1 + σπ(t))X(t) + σ 2 (1 + (σπ(t))−1 )P (t)). X(t) satisfy the following equalities (3. ˙ +H −1 (t)[(H(t)Q(t)H (t)) 2 (H(t)X+ (t)H (t)) 2 + 1 1 +(H(t)X+ (t)H (t)) 2 (H(t)Q(t)H (t)) 2 ]H which have to be taken with boundary conditions 186 1 1 −1 (t). Combining (3. (3. we come to the diﬀerential equations (with further notation X = X+ ) (3. −1 (t) = = (1 + σπ(t))X(t) − σ 2 (1 + (σπ(t))−1 )P (t). (1 + σπ(t))X(t) + σ 2 (1 + (σπ(t))−1 )P (t)).3). which is ensured if x(t).5.5.5. we require the equality E(x(t − σ) + σq(t). . Dividing both parts by σ and passing to the limit (σ → 0).5) and requiring that the righthand parts of these inclusions are equal (within the terms of ﬁrst order in σ).
5.5. l) 2 (X+ (t)l.11) x(t1 ) = m.10) this implies (3.12).5. A solution to the evolution equation (3.14) E[t − σ] + σE(q(t). Let us introduce an evolution equation (3. M ) Deﬁniton 3.13) E[t1 ] = E(m.12) σ→0 lim σ −1 h− (E[t − σ] + σE(q(t).(3.11) satisﬁes the inclusion (3.12) almost everywhere together with boundary condition (3.5.5. P (t)) + o(σ)S Lemma 3. (H(t)X+ (t)H(t)) 2 H −1 l). 2 By equation (3.5. Q(t))l ⊇ ⊇ E[t] − σE(p(t). l)− 2 + π −1 (t)(P (t)l.12) obviously satisﬁes the inclusion (3.1 The ellipsoid E(x(t). l)(X+ (t)l. l)− 2 )+ 2 +(X+ (t)l. l))− 2 + (l. X+ (tπ(·). X+ (t1 ) = M. we have 1 1 ˙ ∂ρ(lE+ [t])/∂t = (X+ (t)l. p + q).5.15) ∂ρ(lE+ [t])/∂t = (l.5.9)(3.5.1 A solution to (3.14).5. E[t] − σE(p(t). H(·))) = E+ [t] given by equations (3. with boundary condition (3.5.13). l) 2 ((X+ (t)l. Q(t)). 1 1 1 187 .5. Introducing the support function ρ(lE+ [t]) and calculating its derivative in t .5. (3. l) 2 π(t)(P (t)l.13) will be deﬁned as an ellipsoidalvalued function E[t] that satisﬁes (3. P (t))) = 0.5.5. p + q) − 1 1 1 1 1 1 − (P (t)l.5. l)− 2 ((H(t)Q(t)H(t)) 2 H −1 l.
X+ (tπ(·).5. t1 ]. ∀l. this gives (3. l) 2 .5. M ) = E+ [t1 ].18) −∂(ρ(lW [t]) − ρ(lE+ [t]))/∂t ≤ 0. in other terms.5. the same inclusion for all t ∈ [t0 . Q(t))).5. For a given vector l. l)− 2 (P (t)l.5.4) . the value of the respective derivative ∂ρ(lE+ [t])/∂t = ρ(lE(q(t). H(t) ∈ Σ. Following the scheme of Section 1. for a given vector l . or .2 The ellipsoid E+ [t] is an external estimate for W [t] .5. to the proof of the Lemma.5. Q(t))) − ρ(−lE(p(t). W [t] ⊆ E+ [t] = E(x(t).5.19) ρ(lW [t]) ≤ ρ(lE+ [t]) = ρ(lE(x(t). Q(t)l) 2 .5.4. whatever are the respective functions π(·). H(·). we have (3. take (3.20) π(t) = (X+ (t)l.5. 1 1 Integrating this inequality within the interval [t − σ. H selected due to (3. t] and having in view the continuity of P (t).1). and since the boundary conditions are W [t1 ] = E(m.16) ∂ρ(lE+ [t])/∂t ≤ (l. P (t))) − ρ(lE(q(t). H(·))).17) −∂ρ(lW [t])/∂t ≤ ρ(−lE(p(t).21)(X+ (t)l. (3.5. l)− 2 ((H(t)Q(t)H(t)) 2 H −1 l. Q(t)l) 2 .the solvability set under uncertainty. this yields relation (3.21). H(·)).17). Indeed. so as to ensure (3. Lemma 3. X+ ((tπ(·). p + q) − (l.5. Together with (3. whatever are the parametrizing functions π(t) > 0. 1 1 1 1 1 1 and take H(t) due to a relation similar to (2. P (t))) 188 .20). Q(t). for t1 − σ ≤ t ≤ t1 and consequently. We shall now indicate the inclusionminimal solutions to (3.12). with π.6 and incorporating relation (3.5. we come to (3. P (t)l) 2 + (l.14) and therefore. (H(t)X+ (t)H(t)) 2 H −1 l) = (l.From inequality a + a−1 ≥ 2 and the inequality of CauchyBuniakowski it then follows ( for all l ∈ IRn ) (3.
5. we come to ρ(lE+ [t]) ≤ ρ(lE ∗ [t]). we come to the conclusion that for every vector l there exists a pair π(·).5. l).18) and arriving at (3. M ) . H.20).21).10) that could be ”squeezed” between W [t] and E+ [t] if the last one is chosen due to (3. H[t] selected due to (3.(3.5.5.22) and also ρ(lE(x(t). Indeed. The explicit relations for (3. Namely. if for each l the ellipsoid E+ [t] = E(x(t). the respective functions π(·). H(·)) is inclusionminimal in the class of all external ellipsoids governed by equations (3. due to (3. so as to ensure a ”tightest” external bound for W [t].20).5. H(·) are continuous or at least measurable in t. (3.15) ensures an equality in (3.(3. continuous or piecewise continuous depending on the properties of l(t). H(t) may be treated as feedback controls selected.21).5.5.5. in view of the indicated relations. Integrating the last inequality from t to t1 and having in view that E+ [t1 ] = E ∗ [t1 ] = E(m.20).5.(3.is clearly the largest among all feasible π. H(·)) is selected to be inclusionminimal.l ∈ IRn . where E ∗ is any other external estimate.(3. H(·) are still measurable . we now observe that the inclusionminimal external ellipsoids E + [t] ensure that the approximation of W [t] is as ”tight” as possible. as we shall see. 18 We now note that the maximal solution W [t] to (3. the ellipsoid E+ [t] = E(x(t). 189 .9).5. H : ∗ ∂ρ(lE+ [t])/∂t ≥ ∂ρ(lE+ [t])/∂t.1. H(·))) ≤ ρ(lW [t]) + o(σ.1.(3.19). This implies Lemma 3.10). Combining this fact with the previous assertions. X+ (tπ(·). X+ (tπ(·).21) indicate that π(·).21). we observe this after integrating (3.5. X+ (tπ(·).5.5.5.5. where for each l there is its own pair of functions π.20). H(·). With l = l(t) = 0 in (3.5.21).5. This means that along the direction l there is no other external ellipsoid governed by equations (3. H(·))). X+ (tπ(·). The selected functions π(t).20).(3.3 With π[t].5.9). such that (3. 18 ρ(lW [t]) ≤ ρ(lE(x(t).19).
(3. H(·)} + o∗ (σ)S We may consequently repeat this procedure indicating (see (1. (3. E(x(t). Continuing the procedure yet further . X+ (tπ(·).5.7.24).5. (3.11) thus deﬁne an array of external ellipsoidal estimates for the solvability set W [t].5. Q(t))dt and for a given vector l the existence of a pair π(·). we may observe. we have Theorem 3. t1 ].24) (o∗ (σ) ≡ 0) is true for t ∈ [t0 . H(·))π(·). t ∈ [t0 .23).2 and (1.24) W [t] = ∩{E(x(t). (3. X+ (tπ(·).1 there exists.5. H(·) that again yield (3.5. a pair π(·). X+ (tπ(·).8). t1 − σ1 . (ii) The relation (3. t1 − σ1 ].5. that with σ → 0 the equalities (3. for every vector l we have (3. for every vector l ∈ IRn . Relations (3. H(·))). X+ (tπ(·).7.10) satisfy the evolution equation (3.7. H(·). are true for all t ∈ [t0 .7.5.5.24). Therefore. X+ (t))π(·).9)(3.5.5. now for the sets deﬁned by (1. under Assumption 1. This yields (3.15) that are generated by solutions x(t).5.for t1 − σ ≤ t ≤ t1 .7.5.5. to the respective limit transition of Lemma 1.12). t1 ]. for the indicated values of t one has ρ(lW [t]) = inf{ρ(l.1.7.(o∗ (σ) = 0).1 Under Assumption 1. H(·)} + o(σ).7) and passing. X+ (t) to the diﬀerential equations (3. H(·))) for some π(·). t1 ]. P (t))dt − t1 −σ1 t1 −σ1 −σ2 E(q(t).25) ρ(lW [t]) = ρ(lE(x(t).5. and moreover.13) and are minimal with respect to inclusion among all solutions to (3. H(·)) for the solutions W [t] to the evolution equation (3. now for t ∈ [t1 − σ1 − σ2 .1.12). particularly.13). Summarizing the results of the above. H(·) of measurable functions that ensure the following : (i) The support function (3. 190 .23) ρ(lW [t]) = ρ(lE(x(t).5. (iii) The external estimates E+ [t] = E(x(t).9). W [t1 − σ1 ]) = = W [t1 − σ1 ] + t1 −σ1 t1 −σ1 −σ2 ˙ E(p(t).6)) for W (t1 − σ1 − σ2 .5.
The next stage is to arrive at a similar theorem for internal estimates.1). E[t] − σE(p(t).1. E(x(t) − σp(t). X(t − σ)) + σE(q(t).5.26) W [t − σ] + σE(q(t). P (t)).31) x(t1 ) = m. X− (t1 ) = M.30) ˙ X− (t) = π(t)X− (t) + π −1 (t)Q(t) − −H −1 (t)[(H(t)P (t)H (t)) 2 (H(t)X− (t)H (t)) 2 + +(H(t)X− (t)H (t)) 2 (H(t)P (t)H (t)) 2 ]H 1 1 1 1 −1 (t).5. X(t)) = E[t] that ensures an external approximation for its lefthand side and an internal approximation for its righthand side. (3.5. together with (3.5.2) this would give (3. and (3.29) ⊆ E(x(t). Q(t)). Let us introduce an evolution equation (3. P (t))) = 0 191 .29) .5.9) and (3.27) E(x(t − σ).32) σ→0 lim σ −1 h+ (E[t − σ] + σE(q(t). which have to be taken with same boundary conditions as before. namely.5.5.5.6) .8)).( 3.1). P (t)) ⊆ W [t] − σE(p(t).(2.15). H −1 (t)[(H(t)X(t)H (t)) 2 + σ(H(t)P (t)H (t)) 2 ]2 H (3.5. Q(t)) ⊆ ⊆ E(x(t − σ).4. Due to (2. dividing both parts by σ and passing to the limit with σ → 0 (similarly to (3. 1 1 −1 (t)) ⊆ Equalizing the righthand side of (3.5. we shall look for an ellipsoidal function E(x(t). Q(t))) ⊆ ⊆ E(x(t − σ) + σq(t).28) with the lefthand side of (3.5. X(t)) − σE(p(t). we come (with further notation X = X− ) to equations (3.5. Returning to (3. X(t − σ)) + σE(q(t). P (t)). (1 + σπ(t))X(t − σ) + σ 2 (1 + (σπ(t))−1 )Q(t))).28) E(x(t).5. Q(t)). and considering the relation opposite to (3.4. X(t)) − σE(p(t).
2 Under Assumption 1. ( 3. A further reasoning is similar to the one that preceded Theorem 3.33) are to be taken instead of (3. W [t] = ∪{E(x(t).36) ρ(lW [t]) = sup{ρ(lE(x(t).4 The ellipsoid E(x(t). H(·))) = E− [t] given by equations (3. X− (tπ(·).33) E[t1 ] = E(m.1.9).5. (3.5.35) (ii) (3. where π(t) > 0.33) is a solution to the evolution equation (3.9).15) that are generated by the solutions x(t). A solution to the last equation is to be considered due to Deﬁnition 3. M ).5.with boundary condition (3. (3.32).5.5.34) E[t − σ] + σE(q(t).3.35).5. are also the maximal solutions to the evolution equation (3. Q(t))+ ⊆ ⊆ E[t] − σE(p(t).1. H(·)}. the following relations are true: (i) (3.5. H(·)}. H(·)))π(·). π(·). 192 .5.5. X− (t) to the diﬀerential equations (3.5.5.5. H(·))π(·).15) and that the respective representations are true relative to closures of the corresponding sets (as in Theorem 2.12).32). P (t)) + o(σ)S.1. H(t) are measurable functions.5.5.1.1).5. though) to the reader.13).5. (3. (3.5. X− (tπ(·). where (3. Similarly to Lemma 3.5.5.5. We leave the details ( which are not too trivial.1 one may prove the following: Lemma 3. for every vector l ∈ IRn . except that the ellipsoid E− [t] is now an internal estimate for the maximal solution W [t] to the evolution equation (3.33). conﬁning ourselves to the formulation of Theorem 3. (3.34)(3. H(·)) for the maximal solutions W [t] of the evolution equation (3. X− (t.7. X− (tπ(·). The internal estimates E− [t] = E(x(t).30) .1.32). It obviously satisﬁes the inclusion (3.5.5.
equations (3. P (t)))−σE(q(t).39) ˙ X+ (t) = A(t)X+ (t) + X+ (t)A(t) − π(t)X+ (t) − π −1 (t)P (t) + H −1 (t)[(H(t)Q(t)H (t)) 2 (H(t)X+ (t)H (t)) 2 + +(H(t)X+ (t)H (t)) 2 (H(t)Q(t)H (t)) 2 ]H 1 1 1 1 x = A(t)x + p(t) + q(t).1).5.5.4. Q(t)) + o(σ)S and the ellipsoidal estimates (external and internal) are taken accordingly. P (t)))−σE(q(t). as in the above.1. Remark 3. (3. E− [t] of the solvability tubes W [t] (” under uncertainty ”) may be interpreted as approximations of Krasovski’s ”bridges” or Pontryagin’s ”alternated” integrals. σ ≤ δ.1. Exercise 3.5.5.5.5.11).5.5.5.4.5. 2. (Theorems 2.1) given in the form (1.9).5. 193 . The nondegeneracy assumption implies in this case that there exist numbers > 0.1.31) as in the above.5.Maximality is treated here with repect to inclusion . (3. δ > 0 such that ˙ (W [t] − σE(p(t).10).1) prove that equations (3.1) is by deﬁnition substituted by (3. Exercise 3. t1 ].9). (3. ˙ −1 (t). Q(t)) ⊇ S.38) and (3.7.5. (3.2).5.40) ˙ X− (t) = A(t)X− (t) + X− (t)A(t) + π(t)X− (t) + π −1 (t)Q(t) − −H −1 (t)[(H(t)P (t)H (t)) 2 (H(t)X− (t)H (t)) 2 + +(H(t)X− (t)H (t)) 2 (H(t)P (t)H (t)) 2 ]H 1 1 1 1 −1 (t).30) will be substituted by (3.5.37) ˙ W [t − σ] ⊆ (W [t] − σE(p(t). with same boundary conditions (3.2 By direct calculation prove that with A(t) = 0 .1 The external and internal ellipsoidal approximations E+ [t].2 ( or 1.1. Under a nondegeneracy Assumption 1. (3.30) may also be derived through the results of Part II.7. t ∈ [t0 . when the evolution equation (3. that is with system (3.5.10). (3.
5.4. E+ [t]. (3.(3.39).21) if E(p(t).5.5. 3. ( W[t] is the respective solvability set described on the same Section ) would satisfy the inclusion x[t] ∈ W[t].19).8 We are now prepared to deal with problems of control synthesis with the aim of using the described relations as a basis for constructive techniques in analytical controller design. Q(t)) = E(p(t) − q(t)). E− [t] are nondominated ( inclusionminimal and maximal respectively).17)(3. Lemma 3.5.4. Remark 3.5. with initial state x[τ ] = xτ ∈ W[τ ]. Q(t)) satisfy the matching condition of Remark 1.1. X+ (t.4.40) will be simpler and reduce to those of type (3. M ∗ ).9)(3. the sets E− [t] turn to be ellipsoidalvalued ”bridges” as will be indicated in Section 3. one may verify that the external and internal ellipsoidal approximation mappings E+ (·).4. M ∗ ) satisfy (3. E− (·) satisfy the external and internal semigroup properties respectively.4.2 Relations (3. γP (t)). X− (t. τ. x(t)) + f (t).20). (3. (3.4. 0 ≤ γ < 1 We ﬁnally mention an important property of the estimates E− [t]. P (t))−E(q(t).4.30) and satisfy the upper and lower semigroup properties (3. P (t)).9).4. Due to this and to the semigroup property of Lemma 3.30) with boundary conditions (3.20). E(q(t). τ ≤ t ≤ t1 194 .(3.10).5.5. Let us return to the Control Synthesis problem 1.18).6. Using the notations of (3.5 The external and internal approximation mappings (for the solvability tube under counteraction W [t]).4 of Section I.2 which means ˙ E(p(t).4.3 The ellipsoids E+ [t].5.10).5.5. ˙ τ ≤ t ≤ t1 .5. τ. are deﬁned through relations (3.21).(3. (3. xτ ) to equation x(t) ∈ U(t.19).4. There the idea of constructing the synthesizing strategy U(t. τ. τ.6 Control Synthesis Through Ellipsoidal Techniques In this Section we shall apply the results of the previous paragraphs to the analytical design of synthesizing control strategies through ellipsoidal techniques developed in the previous Sections. x) for this problem was that U(t. x) should ensure that all the solutions x[t] = x(t.17). where now x(t.5.(3.5.Remark 3. m∗ ).
constructing a strategy U− (t.6. E− [t]) = min{ x − s s ∈ E− [t]} d[t. X− (t1 )). In order to obtain a simpler scheme. we should just copy the scheme of Section 1. x) − ρ(lE− [t]) l ≤ 1} One may readily observe that relations (3. x) may be constructed again according to the scheme of Section I. The conjecture is that once W[t] is substituted by E− [t].9) in section 1. the desired strategy U− (t.4. X(t)).1) x[t] = U− (t.6. M ) = M = E(x∗ (t1 ).3) coincide with (1. P (t). we will now substitute W[t] by one of its internal approximations E− [t] = E(x∗ . x[t] ∈ E− [t].9).(3. x) .4).6. x[t]) + f (t).4. x) such that for every solution x[t] = x(t. x) is the unit vector that solves the problem (3. let us start with the maximization problem of (3. namely E(p(t). It will be proven that once the approximation E− [t] is selected “appropriately”. x] = (l0 .4. x[τ ] = xτ .3) U− (t. Indeed. x) − ρ(l0 E− [t]) = max{(l. the following inclusion would be true (3. x] = h+ (x. It may be solved in more detail than its analogue (1.4.2) and therefore x[t1 ] ∈ E(m. xτ ) that satisﬁes equation (3. (3.4.6.This strategy is thus actually deﬁned as an algorithm.4).4) a nd (3. as we have seen. xτ ∈ E− [τ ]. This exact solution requires. for each instant of time t. P (t)) if x ∈ E− [t] p(t) − P (t)l0 (l0 .(1. τ ≤ t ≤ t1 . to calculate the tube W[t] and then.4 (since E− [t] is an ellipsoid). to solve an extremal problem of type (1.6. P (t)l0 )−1/2 if x ∈ E− [t].4.6.4.12) if set W[t] is substituted for for E− [t] and P(t) for E(p(t). 195 .9) whose solution ﬁnally yields the desired strategy (1.6.12) U(t.5) d[t.and would therefore ensure x[t1 ] ∈ M. τ. ˙ τ ≤ t ≤ t1 . x) = where l0 = l0 (t.6. except that W[t] will now be substituted by E− [t].
6) then we can take (3. where λ > 0 is the unique root of the equation h(λ) = 0. (a = 0). Q).4) or (3.5) in x). Since it is assumed that x ∈ E(0. s0 = arg min{ (x − s) s ∈ E− [t].6. Q) and a vector x ∈ E(a.7) l0 = k(x(t) − s0 ).6. (3. k > 0. With λ → ∞ we also have (I + λQ−1 )−1 x. (This can be veriﬁed by diﬀerentiating either (3. Then the necessary conditions of optimality for the minization problem x − s = min. Q−1 (I + λQ−1 )−1 (x − a) − 1 Proof.6.1 Consider a nondegenerate ellipsoid E = E(a. Assume a = 0. Lemma 3.4). are reduced to the equation −x + s + λQ−1 s = 0 where λ is to be calculated as the root of the equation h(λ) = 0. E(a. x = x(t)}. E− [t]) with t ﬁxed.6.If s0 is the solution to the minimization problem (3. Q))/∂x may be expressed through l0 = (x − s0 )/ x − s0 . with h(λ) = (I + λQ−1 )−1 (x − a).6.8) s0 = (I + λQ−1 )−1 (x − a) + a.6. Then the gradient l0 = ∂d(x. so that l0 will be the gradient of the distance d(x. Q). Q−1 (I + λQ−1 )−1 x → 0 196 (s. Q−1 s) ≤ 1 . in (3. we have h(0) > 0.6.
P (t))} = U− (t. namely 1 ρ(lE(p. x(t)) − ρ(l  E− [t]) l ≤ 1}. Relation (3. u)u ∈ E(p(t).6. for d[t] ≥ 0.4.e. Corollary 3. The case a = 0 can now be given through a direct shift x → x − a.1).6. such that u[t] ∈ E(p(t).6. Q.D. x). we will always have the inclusion (3. l = 0.2 Given ellipsoid E(p. P )} = (l. We will demonstrate that once x[t] is a solution to equation (3.1 With parameters a. P (t)) and problem (1.4. P l)− 2 We will now prove that the ellipsoidal valued strategy U− (t.6. the maximizer u∗ for the problem max{(l. the multiplier λ may be uniquely expressed as a function λ = λ(x) Let us now look at relation (1. The root λ0 is unique since direct calculation gives h (λ) < 0 with λ > 0. λ ≥ λ∗ for some λ∗ > 0. x) of (3.This yields h(λ) < 0. P ) .6. Indeed. P l)− 2 This Lemma is an obvious consequence of the formula for the support function of an ellipsoid. is the vector u∗ = p + P l(l. p) + (l. P (t)). τ. P )) = (l.3) does solve the problem of control synthesis. ˙ Calculating d[t] = d[t. In the present case we have P(t) = E(p(t). x[t]] = max{(l. provided we start from a point xτ = x(τ ) ∈ E− [τ ]. assume xτ ∈ E− [τ ] and x[t] = x(t.3) now follows from the following assertion: Lemma 3.12) therefore reduces to (3.6. xτ ) . we observe 197 1 .2).6. Q given and x varying. u)u ∈ E(p. (With isolated trajectory x[t] given. it is clearly driven by a unique control u[t] = x(t) − f (t) a.E. τ ≤ t ≤ t1 to be the respective trajectory.12). u∗ ) . The equation h(λ) = 0 therefore has a root λ0 > 0.9) arg max{(−l0 .
P (t)). x(t)) + (l0 . Integrating dd2 [t]/dt = dd2 [t. X− (t)). dt .4. x[t]) − ρ(l0 E− [t]) = ˙ dt ∂t d = l(l0 .(3. where x(t).10) and remembering the rule for diﬀerentiating a maximum over a variety of functions. ∂ d d[t] = (l0 . Substituting these relations into (3. x[t]) − ρ(l0 E− [t])] dt dt and since l0 = 0 is a unique maximizer. X− (t) satisfy the system (3. we have d 1 d[t] = (l0 . P (t)l0 )1/2 + (l0 . due to the BunyakovskySchwartz inequality. X− (t)) . For a ﬁxed function H(·) we have E− [t] = E(x(t). u[t]) − [(l0 .6.10) where E− [t] = E(x(t). u[t]). x). p(t)) + (l0 . X(t)l0 )1/2 ] dt (3. X− (t)l0 )−1/2 · dt 2 ·(l0 . x) ⊆ E(p(t). provided x(τ ) ∈ X− (τ ). M) = 0 + which means x(t1 ) ∈ M. we come to the equality d2 [τ. ( see notations of Section 1.11) where u[t] ∈ U− (t. x(t1 )] = h2 (x(t1 ). x[t]]/dt from τ to t1 . with equality d d[t] = 0 dt attained if u[t] ∈ U− (t.2).6. p(t)) + (l0 . x(τ )] = d2 [t1 .4 ). What follows is the assertion 198 d d[t] ≤ −(l0 .6. (3.4. u[t]) − (l0 .4). H −1 (t)([H(t)X− (t)H(t)]1/2 [H(t)P (t)H(t)]1/2 + +[H(t)P (t)H(t)]1/2 [H(t)X− (t)H(t)]1/2 )H −1 (t)l0 ) or.d d d[t] = [(l0 .
τ ≤ t ≤ t1 .6.5. With x ∈ E− [t].2 (see (1. x).(3.5. (iii) The internal ellipsoids E− [t] satisfy the evolution equation (3.1.6. the following inclusion is true: x[t] ∈ E− [t].4. and therefore x[t1 ] ∈ E(m.1.4. With this speciﬁcity. with the only unknown being the multiplier λ of Lemma 3. x) may be considered as an analytical design.4.6. x) of (3. x) = E− [t]). the suggested strategy U− (t. x) for any internal approximation E− [t] = E− (x(t).1 (i) Due to Theorem 3.1). 199 . x) is singlevalued. 3. The given facts are particularly due to the lower semigroup property of the respective mappings (see Lemma 2.6.3). Remark 3. te = 5.8. the function U− (t.8)).3) . The ”ellipsoidal” synthesis described in this section thus gives a solution strategy U− (t. X− (t)) with given parametrization H(t) of (3.6.3). X− (t)).(3.7.1).1. te ]. We will seek for graphical representations of the solutions.6.Theorem 3.6. Due to the assumptions of the Section 3. ts = 0. being uppersemicontinuous in x.1 which can be calculated as the only root of equation h(λ) = 0.7). We shall now proceed with numerical examples that demonstrate the constructive nature of the solutions obtained above. But in view of Corollary 2. whilst with x ∈ E− [t] it is multivalued (U− (t. depending on the parameters of the internal approximation E− [t] ( which may also be calculated in advance).1 and its ”ellipsoidal” version.7 Control Synthesis: Numerical Examples Let us take system (1.1 the function λ = λ(x) may be calculated in advance. measurable in t and ensuring the existence of a solution to the diﬀerential inclusion (3.6.32) and therefore the equation (1.8) indicate that strategy U− (t. Theorem 2. Once x[τ ] ∈ E− [τ ] and the synthesizing strategy is U− (t. (ii) Relations (3. each element x ∈ intW [t] belongs to a certain ellipsoid E− [t] and may therefore be steered to the terminal set M by means of a certain ”ellipsoidalbased” strategy U− (t. (3.9) which implies Theorem 1.6. M ) .4). x) is given explicitly. and study it throughout the time interval [ts .6.1 intW [t] = ∅.2) to be 4dimensional.1 Deﬁne an internal approximation E− [t] = E− (x(t). and the ”inclusionmaximal” property of the ellipsoids E− [t].
0 0 0 200 . x4 ). we we shall present them through their two dimensional projections. x0 } be given by 1 x0 = . where we show the graph of solutions and of solvability set. the third. ﬁrst and third (x1 . In some of the ﬁgures. is also deﬁned by time independent constraints: 0 p(t) ≡ . x3 ). skew axis corresponds to time and ranges from 0 to 5. and each shows projections of the original ellipsoids onto the planes spanned by the ﬁrst and second ( x1 . Let the initial position {0. x4 ) coordinate axes.And as the ellipsoids appearing in this problem are four dimensional. and second and fourth (x2 . The restriction u(t) ∈ E(p(t). x2 ). third and fourth(x3 . P (t)) on the control u. 0 4 0 the target set M = E(m. We consider a case when the right hand side is constant: 0 −1 A(t) ≡ 0 0 1 0 0 0 0 0 0 −4 0 0 . The drawn segments of coordinate axes corresponding to state variables range from −10 to 10 according to the above scheme. 1 0 describing the position and velocity of two independent oscillators. in a clockwise order. starting from bottom left. The ﬁgures below are therefore divided into four windows. M ) by 5 m= 5 0 0 and 1 0 M = 0 0 0 1 0 0 0 0 1 0 0 0 0 1 at the ﬁnal instant t1 = 5.
0 1 so that the controls do couple the system.2342 x(0) = .23) to be used here by selecting H(t) = P −1/2 (t).1385 0 0 0 0 31. M): 4. P (t)l ) if x ∈ E− [t]. and use the discretized version of (3. We divide the interval [0.3) we apply a single valued selection: p(t) if x ∈ E− [t] 0 0 0 −1/2 p(t) − P (t)l (l .1385 0 0 X− (0) = 0 0 12.4. (3.4.1 0 P (t) ≡ 0 0 0 1 0 0 0 0 1 0 0 0 .4.22).4. x) = again in its discrete version. The use of a singlevalued strategy in the discrete version does not aﬀect the existence of solutions to the respective recurrence equations. 5] into 100 subintervals of equal lengths.3) should steer every solution of 201 .1) u(t.6. P (t)).23) implemented through a standard ﬁrstorder scheme ( see.1 with an implication that the control strategy U− (t.22) and (3. The results to be presented here we obtain by way of discretization.(3.3611 44. We shall specify the internal ellipsoid E− [t] = E(x(t). U(t.1370 and 31. 0 ≤ t ≤ 5 in (3. The calculations give the following internal ellipsoidal estimate E− [0] = E(x(0). the class of feasible strategies is such that c UP = {U(t. Therefore.1236 Now.3611 0 0 2. X− (t)) of (3.6043 −3. x)}. as it is easy to check. X− (0)) of the solvability set W[0] = W(0. −2.4.1845 2. for example [63]. [272]. x0 ∈ E− [0] and therefore we may apply Theorem 3. [273] for technical details).2371 1. x) ⊆ P = E(p(t). x) of (3.18).6. Instead of the set valued control strategy (3. 5.4.7.
the solvability set E− [0] = E(x(0). Figure 3.. but will interpret the limit (σ → ∞) of solutions to (3. 3. There is actually no such problem. into M. k = 0.(3. x(tk ))).(3.3) x(tk+1 ) − x(tk ) = σ −1 (A(tk )x(tk ) + u(tk .7.0264 4.7.2) constructed for the same tube E− [t] as in (3. . Figure 3.3) as a solution to the diﬀerential inclusion (3. For the discrete version this produces 0. x) + f (t). ˙ x0 = x(0). 100.7. M ).7..4). however.71.3) serves as a discretetime version of the diﬀerential equation (3. ———————!!! insertFigures 3.1).7.4) x[t] = A(t)x[t] + u(t. x[t]) ˙ However. ts = t0 = 0 ≤ t ≤ 5 = t100 = te . t ∈ [0. x[0] = x0 : σ = tk+1 − tk > 0.7. where we use u(t. which.2).9512 x[5] = 4.2 !!! ———— 202 . Equation (3.7.7. and the trajectory of the solution of (3..3). within the accuracy of the computation. the last equation has a singlevalued but discontinuous righthand side which leads to additional questions on the existence of solutions to this equation.2 shows the target set M = E(m.7.7.7.2).7.3).2) x = A(t)x + U− (t. x) of (3.7. X− (0)) at the initial instant t = 0.7. for the discretetime system (3. may be treated as a solution of (3.1 shows the graph of the ellipsoidal valued map E− [t].7.1). (projections appearing as circles).0830 as a ﬁnal state.7.We will therefore avoid the singlevalued equation (3.0457 −0. 5] and of the solution of (3.
9528 x[5] = 4.7.3). 5. let us still apply formulae (3. except one.(3. that is actually x0 ∈ W(0.6. 203 .2. (upper left window). 0 4 2 ——————!!! insert Fig.0255 4.7.1658 as a ﬁnal state. After this happens.0215 −0. and the trajectory remains inside for the rest of the time interval. at one point in time the trajectory enters the tube E− [t]. Leaving the rest of the data to be the same.1 cannot be used. As the illustration shows. The trajectory of the solution to (3.7.3).3 shows the phase portrait of the result. The drawn projections of the initial state are inside. Figure 3. and with a thin line if it is inside.3 !!!—————– Though Theorem 3.In the next example we show by way of numerical evidence. as it was possible to steer the solution to (3. M) \ E− [0]. In this way we obtain 0. taking 1 x0 = . but “not very much”.7. The above phenomenon indicates that • the initial state must be inside the solvability set W[0] = W(0. what can happens if the initial state x0 ∈ E− [0]. the line changing into thin. 5.1) into the target set M.7.1) and (3. M).6. Analogously to Figure 3.7. • in this particular numerical example the control rule works beyond the tube E− [t].3) is drawn with a thick line.7. 3.7. we change the initial state x0 in such a way that the inclusion x0 ∈ E− [0] is hurt. as long as it is outside of the respective ellipsoidal solvability set. Theorem 3.1 does take eﬀect.
0536 ——————!!! insert ﬁgures 3.5 and 3.4): 1 x0 = .5540 —————!!! insert Fig. (Figure 3. 0 0 4 0 0 0 0 4 resulting in a ﬁnal state 0. M ) by setting the radius to be 2: 4 0 0 0 0 4 0 0 M = .7.7.5.5875 4.3668 −0.7.0460 4.6 show the eﬀect of changing the target set.7. We take the data of the ﬁrst example except for the matrix M in the target set M = E(m.6 !!!————— 204 .7. 3. 3. 3.In the third example.8914 x[5] = .0158 −0. we move the initial state x0 further away. 3. 0 4 3 and obtain as ﬁnal state 0.9150 x[5] = .4 !!!———————— Figures 3. so that the control rule does not work any more.7.
7.7.9565 x[5] = .The switching of the control.0536 −0.8. changing the matrix P (t) in the bounding set P = E(p(t).7. we allow more freedom for the controls.1).1308 Numerical simulations were made on a SUN SparcStation.3.7.8 !!!————– Finally we shall consider two coupled oscillators. 4.7.7. M ) deﬁned by 10 0 m= 0 10 and 1 0 M = 0 0 0 1 0 0 205 0 0 1 0 0 0 0 1 . represented by a system with parameters −5 0 x0 = .0235 4. and later in Figure 3. Taking again the data of the ﬁrst example. −10 10 with target set M = E(m.6. due to the speciﬁc form of (3. is clearly seen in Figure 3. P (t)) again by setting the radius to be 2: 4 0 P (t) ≡ 0 0 with a ﬁnal state 0 4 0 0 0 0 4 0 0 0 0 4 0. ————————!!! insert ﬁg. 3.
7.at ﬁnal instant t1 = 3.9. The fourdimensional ellipsoidal tubes and the synthesized control trajectory in phase space are shown in an appropriate scale in Figures 3. 0 0 0 .10.7. 206 .25 0 A(t) ≡ 0 0 0 16 0 −16 and the constraint on the controls is deﬁned by 0 p(t) ≡ . The system matrix A is constant: 0 1 0 −1 0. ( here note the relatively ”small size” of the target set) .10 !!!— 3.8 ”Ellipsoidal” Control Synthesis for Uncertain Systems In this Section we shall further apply the results of the previous paragraphs to the analytical ”ellipsoidal” design of synthesizing control strategies. 0 0 9 0 0 0 0 0. in 100steps. 3.1 0 0 P (t) ≡ .7. 3.1) through a diﬀerence scheme similar to the above.9. this time constructing them for uncertain systems.1 The target control problem is solved as before. ———————!!! insert ﬁgures 3.7.7. 1 0 0 0 9 0 0 0 0 0. with synthesizing strategy calculated due to (3.
(3. x) should ensure that all the solutions x[t] = x(t.8. x[t]) + f (t).2 which.8. The conjecture is that once W ∗ [t] is substituted by E− [t]. Q(t). ˙ τ ≤ t ≤ t1 . due to the presence of uncertain items. To obtain a simpler scheme.7. τ ≤ t ≤ t1 . to calculate the tube W ∗ [t] and then. ˙ would be true. 207 . for each instant of time t.7. xτ ∈ E− [τ ]. This would ensure the terminal condition x[t1 ] ∈ E(m.10). is more complicated.9). The strategy U 0 (t. x) is again actually deﬁned as an algorithm which. x[τ ] = xτ .8. There the idea was that the respective synthesizing control strategy U(t. except that W [t] should be substituted by E− [t]. due to relations (3. x(t)) + E(q(t).8. xτ ) to the diﬀerential inclusion x(t) ∈ U(t.1) the inclusion (3.31).8. Namely.5. with initial state x[τ ] = xτ ∈ W ∗ [τ ].2) x[t] ∈ E− [t]. (3.5. xτ ) to the system (3. could be speciﬁed through the Alternated Integral (1.8.7. so that the setvalued function W ∗ [t] would satisfy the evolution equation (1.30). x) such that for every solution x[t] = x(t.7.7. we should just copy the scheme of Section 1. M ) = M = E[t1 ]. Namely.8. It will be proven again that once the approximation E− [t] is selected “appropriately’. x) = U 0 (t.2 presumed here.9).9).Let us consider the Problem of Control Synthesis Under Uncertainty of Section 1. (Deﬁnition 1. X− (t)). x) may be ∗ constructed as in Section 1. Q(t)). to solve an extremal problem of type (1.1).(1. we shall substitute W ∗ [t] by one of its internal ellipsoidal approximations E− [t] = E(x∗ . than in the absence of these. we should construct the new 0 approximate strategy U− (t. would satisfy the inclusion x[t] ∈ W ∗ [t]. 0 namely.8. whatever is the function f (t) ∈ E(q(t).8.10) whose solution ﬁnally allows to specify the desired strategy U(t. The exact solution scheme requires.8). of course. Here W ∗ [t] = W∗ [t] is the solvability set of Deﬁnition 1. τ. under Assumptions 1. x) according to (1.1 or 1.5. τ ≤ t ≤ t1 0 x[t] = U− (t. τ. the desired strategy U− (t. τ ≤ t ≤ t1 and would therefore ensure the desired terminal condition x[t1 ] ∈ M. as we have seen.
(3.8) and in Lemma 3. Calculating 208 . x) = E(p(t).5. Indeed. x) the maximizer for problem (3.6.5. provided we start from a point xτ = x(τ ) ∈ E− [τ ]. x) = U− (t. P (t)) if x ∈ E− [t] p(t) − P (t)l0 (l0 .(3. The respective relations may now be obtained in more detail than in the general case of Section 1.5) d[t.1 We emphasize again that the given scheme follows the lines of Section 3.2).8. u)u ∈ E(p(t).9) reduces to (3.10) of Section 1.5.6 . Further on. 0 The desired solution strategy U− (t. (3. This reﬂects the uncertainty (3. x).4) as in Section 3.6) in the inputs f of the system.6.6. we have to underline that the main new point here is the calculation of the derivative dd[t.9) which depends on vector l0 = l0 (t.6) 0 arg max{(−l0 . P (t)l0 )−1/2 if x ∈ E− [t].8.a direct analogue of problem (1.8. We will demonstrate that once x[t] is a solution to (3.30) rather than by (3. x) of (3. and as before (3.4.8. x] = d(x.8. Without repeating the similar elements in the scheme.8. τ ≤ t ≤ t1 to be the respective 0 trajectory.4) d[t.based strategy U− (t.8.8. x] = (l0 .1.2. x) − ρ(l0  E− [t]) = max{(l.8. (3.8. x) = ∂d(x. E− [t]) = h+ (x. x) − ρ(lE− [t]) l ≤ 1}. The further reasoning is analogous to that of Section 3.5.8. where l0 = l0 (t. x]/dt due to the diﬀerential inclusion (3. and therefore relation (3.30). we notice that again P(t) = E(p(t). it will always satisfy (3.6. τ. E− [t]) = min{ x − s s ∈ E− [t]} Remark 3. so that problem (1.7). P (t))} = U− (t.3) 0 U− (t.1).8.6.6.8.3) follows from Lemma 3.4). U(t.8. xτ ) . The properties of l0 are similar to those described in (3.4. P (t)) is an ellipsoid.9).8. 0 We will now prove that the ”ellipsoidal” . X− (t)) deﬁned by (3. x). E− [t])/∂x is the unit vector that solves the problem (3.2).9).(3. x) taken here is deﬁned by relations (3.1. assume xτ ∈ E− [τ ] and x[t] = x(t.1) with E− [t] = E− (x∗ .8. but the tube E− [t] = E− (t.1.8. since E− [t] is an ellipsoid. x) must satisfy a relation of type (1.3) does solve the problem of control synthesis of Deﬁnition I.
7) and diﬀerentiating the respective function of the ”maximum” type due to equation d x[t] = u[t] + f (t). H(·) we have E− [t] = E(x(t). H −1 (t)([H(t)X− (t)H(t)]1/2 [H(t)P (t)H(t)]1/2 + +[H(t)P (t)S(t)]1/2 [H(t)X− (t)H(t)]1/2 )H −1 (t)l0 )}. x(t)) − ρ(l  E− [t]) l ≤ 1}.d[t] = d[t. x[t]) is a realization of the feedback control strategy U and f (t) is an input disturbance . p(t) + q(t)) + (l0 .30). (π(t)X− (t) + π −1 (t)Q(t))l0 ) − (l0 .8) . For ﬁxed functions π(·). 209 (3.8. dt where E− [t] = E(x(t). (3. u[t] + f (t)) − (l0 . x[t]] = max{(l. P (t)l0 )1/2 − (l0 . we come to d d[t] ≤ dt ≤ (l0 . p(t) + q(t)) − (l0 . d ∂ d[t] = (l0 . where u[t] ∈ E(p(t). x(t)) + (l0 . Q(t)l0 )1/2 .8.8. X− (t)). f (t) ∈ E(q(t). X− (t)l0 )1/2 ].8. u[t] + f (t)) − (l0 . where x(t). x[t]) − ρ(l0 E− [t]) = ˙ dt ∂t d = (l0 . u[t] + f (t)l) − [(l0 . X− (t)) . P (t)).9). we observe d d d[t] = [(l0 .(3. X− (t)l0 )−1/2 · dt 2 ·{(l0 . Q(t). X− (t) satisfy the system (3. x[t]) − ρ(l0 E− [t])] dt dt 0 and since l is a unique maximizer.8. Applying inequality a2 + b2 ≥ 2ab and the the BunyakovskySchwartz inequality to the righthand part of the previous formula.7) Substituting this into (3. we have d 1 d[t] = (l0 . dt where u[t] ∈ U (t.
8. This also gives d V (t. M ).In other terms we have d d[t] ≤ (l0 . dt 210 d 0 x ⊆ U− (t. dt .8.6). x) + E(q(t). τ ≤ t ≤ t1 . with given parametrization H(t). see notation W of (1. x) + f (t). x) and any feasible f (t) this yields (almost everywhere) d d[t] ≤ 0. Q(t)) in the synthesized system d 0 x = U− (t. x(t1 )] ≤ d2 W ∗ [τ. Q(t))) dt 0 With u[t] ∈ U− (t.9) and therefore x[t1 ] ∈ E(m. dt due to (3.19) ) h2 (x(t1 ).8. x[t]]. u[t] + f (t)) + ρ(−l0 E(p(t). x ∈ E− [t]. Q(t)). (3.8. M) = d2 ∗ [t1 . x) = d2 [t]. x(τ )] = h2 (x(τ. the following inclusion is true: x[t] ∈ E− [t]. x) of (3.1 Deﬁne an internal ellipsoidal approximation E− [t] = E− (x(t).8. whatever is the disturbance f (t) ⊆ E(q(t).8. x) ≤ 0 V (t.3) . whatever is the solution x[t] to the diﬀerential inclusion (3. + W + so that x(t1 ) ∈ M if x(τ ) ∈ X− (τ ). dt Integrating dd2 [t]/dt from τ to t1 . Once x[τ ] ∈ 0 E− [τ ] and the synthesizing strategy is selected as U− (t. What follows is the assertion Theorem 3. X− (t)) to the solvability set W∗ [t]. P (t))) − ρ(l0 E(q(t).30).8. we come to the inequality (d[t] = d2 ∗ [t. X− (τ )).6). π(t) in (3.
x) = E− [t]).8. measurable in t and ensuring the existence of a solution to the diﬀerential inclusion (3. t1 ] are singletons ) to ”more disturbance” allowed.8.2 (i)Due to Theorem 3. With this reservation.7. x) is singlevalued. For the calculations we use a standard ﬁrst . (ii) We emphasize once more that the constructions given in Sections 3. whilst with x ∈ E− [t] it is multivalued (U− (t.order discrete scheme for equations (3. t1 .9 Control Synthesis for Uncertain Systems: Numerical Examples In this Section our particular intention ﬁrst is to illustrate through simulation the eﬀect of introducing an unknown but bounded disturbance f (t) into the system.5. The result is that in the ﬁrst case we obtain a “large” internal ellipsoidal estimate E− [t] of the solvability set W ∗ [t0 ] = W ∗ (t0 . while in the last it shrinks to be “small”.5. (3. ˜ We shall now proceed with further numerical examples ( this time for uncertain systems) that demonstrate the constructive nature of the suggested solution schemes. x) for any internal approxi∗ mation E− [t] = E− (x(t).8. X− (t)) of the solvability tube W [t] . We also indicate the behaviour of isolated trajectories of system (3. Remark 3.5.36)). so that the problem still remains solvable.3). We shall do this by considering a sequence of three problems where only the size of the bounding sets for the disturbances f (t) increases from case to case. which may be calculated in advance. the function 0 0 U− (t. which implies that there exists an internal curve x(t) such ˜ that x(t) + (t)S1 (0) ⊆ W∗ [t] for all t with continuous (t) > 0.1) in the presence of various speciﬁc feasible disturbances f (t) ∈ E(q(t).9). With x ∈ E− [t].1 indicates that each of the tubes E[t] is an ellipsoidalvalued bridge ( see Remark 3.2.2 (see (3. 5] — into 100 subintervals of 211 . M) .each element x ∈ intW ∗ [t] belongs to a certain ellipsoid E− [t] and may therefore be steered to the terminal set M by means of a certain ”ellipsoidal strategy” U− (t. Then clearly intW∗ [t] ≡ ∅. Q(t)). 3.8. Q(t)). t ∈ [t0 . as indicated 0 in Remark 3.8.1 . starting from no disturbance at all (that is where the sets Q(t) = E(q(t).9). Such a strategy may be speciﬁed in explicit form except for a scalar multiplier λ = λ(x) .3.8 are derived here under Assumption 1.0 The ”ellipsoidal” synthesis thus gives a solution strategy U− (t. Theorem 3.8.6.30) by dividing the time interval — chosen to be [0. the suggested strategy U− may be interpreted as an analytical design. being therefore uppersemicontinuous in x.5. x).
9. (3.5. again in its discrete version. x0 } given by 2 −10 x0 = . (The discrete version obviously does not require any additional justiﬁcation for using the singlevalued selection).1) u(t.4) with the initial position {0.1. the system becomes coupled. x) = p(t) if x ∈ E− [t] 0 0 0 −1/2 p(t) − P (t)l (l . [272].) The 212 . We calculate the parameters of the ellipsoid E− [t] = E(x(t).1). M ) deﬁned by 10 0 m= 0 10 and 100 0 0 0 0 100 0 0 M ≡ 0 0 100 0 0 0 0 100 at the ﬁnal moment t1 = 5. We consider a 4dimensional system of type (1.equal length ( the details of such schemes may found in [63]. We suppose the right hand side to be constant: 0 −1 A(t) ≡ 0 0 1 0 0 0 0 0 0 −4 0 0 .3) we apply a single valued selection: (3.8. 1 −6 at the initial moment t0 = 0 and target set M = E(m. however.30). (Through the constraints on the control and disturbance. P (t)l ) if x ∈ E− [t]. X− (t)) by chosing a speciﬁc parametrization which is H(t) = P −1/2 (t) and Tr1/2 (X− (t)) π(t) = Tr1/2 (Q(t)) in equation (3.2)(3. 1 0 describing the position and velocity of two independent oscillators.1. [273]. Instead of the set valued control strategy (3.1.
P (t)) on the control and v(t) ∈ E(q(t).5. 0 0 0 9 0 P (t) ≡ 0 0 0 1 0 0 0 0 0 9 0 0 0 0 1 The center of the disturbance is the same in all cases: 0 q(t) ≡ 0 The diﬀerence between the three cases i = 1. 3 the data are chosen in such a way that neither the controls. 2.restriction u(t) ∈ E(p(t).8. 3 appear in the matrices: 0 0 Q(1) (t) ≡ 0 0 1 0 Q(2) (t) ≡ 0 0 0 0 0 0 0 9 0 0 0 0 0 0 0 0 1 0 0 0 . in these cases Assumption I. At the same time.1 that allows such a reduction is not fulﬁlled. but note that in the cases i = 2. the solvability set W∗ [t] contains an internal trajectory so that intW∗ [t] = ∅ ( see Remark 3.7. Obviously. Its internal ellipsoidal approximations E− [t] exist and may be calculated due to schemes of Section 3. More precisely. case i = 1 is the one treated in Section 3. in these cases the problem can not be reduced to simpler situations without disturbances.1 Clearly. nor the disturbances dominate the ˙ ˙ other.6.1 0 0 Q(3) (t) ≡ 0 0 1 0 0 0 0 13. both P − Q = ∅ and Q − P = ∅. 0 9 1 0 0 0 0 13. Q(t)) on the disturbance is also deﬁned by time independent constraints: 0 p(t) ≡ . 213 . that is. 0 0 0 0 .2(ii)).
3911 0 0 22.8. X− (0)) ⊂ E(x(0). ranging from 0 to 5. we present their two dimensional projections. and of the solutions to equation (3.1 is applicable.8930 0 0 (3) X− (0) = .3397 132.2735 341. Since the ellipsoids appearing in this problem are four dimensional.5488 46. The drawn segments of coordinate axes corresponding to the state variables range from −30 to 30.1 .2197 37.9. x2 }).3397 0 0 28.3911 98.8. Figures 3. x3 }). Also.0094 61. 0 0 147.5685 −5.1 .The calculations give the following internal ellipsoidal estimate E− [0] = E(x(0). 2. as it can be proved on the basis of their construction.3) steers the solution of (3. The ﬁgures are therefore divided into four windows. analogously to the corresponding inclusions between the original (nonellipsoidal) solvability sets W (i) (0. 5]. i = 1. 1. 5.5502 66.2087 and 323. and second and fourth ({x2 .3047 28.7509 12. M).4685 −8.4742 x(0) = . 3.1077 0 0 61. 3 and therefore Theorem 3. x4 }). 3: 2. showing projections of the original ellipsoids onto the planes spanned by the ﬁrst and second ({x1 . X− (0)) of the solvability set W (i) (0. X− (0)) ⊂ E(x(0). 2.5502 0 0 25.9.6241 22.9.2) x(tk+1 ) − x(tk ) = σ(A(tk )x(tk ) + u(tk .2197 0 0 21. t ∈ [0. 0 0 45. implying that the control strategy of (3.9.1) into M under any admissible disturbance f (t) ∈ E(q(t). we have the inclusions E(x(0). X− (0)) holding. in a clockwise order starting from bottom left.7732 Now.2863 21.1077 469. x(tk )) + f (tk )) 214 (3) (2) (1) (i) (i) (i) . as is easy to check. and since the objective is to describe the solutions also through graphical representations. respectively. M).3.3 is time.9377 30.9. The skew axis in Figures 3. ﬁrst and third({x1 .4791 0 0 (2) X− (0) = .3.3661 25. 2.3 show the graph of (i) the ellipsoidal valued maps E− [t]. x0 ∈ E(x(0). X− (0)) for i = 1. i = 1.8. Q(i) (t)) in all three cases. third and fourth ({x3 . 0 0 33.4382 0 0 (1) X− (0) = . x4 }) coordinate axes.2735 0 0 30.
and trajectories of the same solutions of (3. . is drawn in a thin line if it is inside the current ellipsoidal solvability set. ˙ (There may be problems with deﬁning the existence of solutions to the last equation .2).3. W− (0)) at t = 0. 3.9. (3.9. k = 1.9. ——————–!!! insert ﬁgures 3. that is the solution to (3.9. The time interval [0. x) is deﬁned by (3.1). however.5. Then a new value for f is selected and the above procedure is repeated for the next pair of intervals. and by a thick line if it is outside.3 !!!———— Figures 3.x[0] = x0 .9.1).2) in phase space.. A value f is chosen randomly at the boundary of E(q(t). 5] is divided into subintervals of constant lengths. which is a discrete version of the equation x[t] = A(t)x[t] + u(t. We will therefore avoid this last equation and refer only to (3. the solvability set E− [0] = E(w(0). 215 .9.7. So the statement of Theorem 3. The controlled trajectory.1 is that the control ensures that a thin line cannot change into thick.9. (3.4.9. etc. x[t]) + f (t). x) may turn to be discontinuous in x . 100.2) and (3. see analogous situation in Section 3. M ). one being f (t) ≡ 0 and two other — so called extremal bangbang type — feasible disturbances.9.6 show the target set M = E(m.1) and we consider three diﬀerent choices of the disturbance f (t)...9.1 . 0 = t0 ≤ t ≤ t100 = 5. since function u(t. (projections appearing as (i) (i) circles of radius 10). Here u(t.1).7).9.8. 3. Q(i) (t)) and the disturbance is then deﬁned by f (t) = f over all the ﬁrst interval and f (t) = −f over the second. The construction of these disturbances is the following. σ = tk+1 − tk > 0.
it does not succeed.3. In one case the control rule deﬁned using the ellipsoidal tube E− [t] steers the trajectory into the target M. t1 .3) x0 ∈ E− [0] \ E− [0] holds.6 !!! —————– The ellipsoids E− [0] are only subsets of the respective solvability sets W ∗ (0.9. then it may occur that such a behaviour can be illustrated on the ellipsoidal approximations.9.9. However. 5. There may of course be other control rules.7 successfully hits the target set M at t = 5. like the one based on the exact (nonellipsoidal) solvability sets W ∗ [0] = W ∗ (t. that could be successful.3) holds indeed. ——————!!! insert ﬁgures 3. (This is case i = 1. (One thick trajectory changing into thin is clearly seen in the right hand side windows. so there is no disturbance.9.9.7. like if the initial state is not contained in E− [t0 ]. 3 0 In Figures 3.9. 3. The trajectory in Figure 3.9. we return to the parameter values of the previous examples and change the initial state only. then it is not true that the trajectory can be steered into the target set M under any disturbance f (t) ∈ Q(t).7). while under the other disturbance.) Figure 3. taking −12 0 x0 = . 5.1 there does not follow a negative statement.8.9.—————!!! insert ﬁgures 3.8 shows two trajectories under two simulated feasible disturbances f (t) ∈ (2) E(q(t). Compare these examples with those of Section 3. by moving it in such a way that (2) (1) (3.8 it can be seen that relation (3.4 .7 and 3. M).9. To show this. M) is ”appropriate”. and the projection of the endpoint of the other is outside in the lower left window.8 !!!—————– 216 . M). Q(t)). once x(0) ∈ W ∗ [0].9. if the ellipsoidal approximation E− [0] ⊂ W ∗ (0. therefore from Theorem 3.
Our new problem with free terminal time t will then be solvable for a given position {τ. For time interval [τ.Finally we again consider a system that describes two coupled oscillators with matrix 0 −1 A(t) ≡ 0 −1 1 0 0 0 0 0 0 −9 0 0 .10. not later than at t1 rather than at ﬁxed t1 . P.1.11 !!! —— 3. we shall require that the terminal inclusion x(t) ∈ M could be reached at any instant t ∈ (t0 . M.8.9.8. We have in view that the constraints on u. t.9.9. —————!!! insert ﬁgures 3. M). Recall the solvability set of Section 1. according to the respective notations.9. Taking the disturbances to be restricted by Q(1) . Q(3) of the above and simulating the respective target control synthesis problem. 3. Q(2) . 1 0 and with the other parameters ( x0 . t] it should be denoted. 3.8.11 accordingly. We shall now brieﬂy describe this problem without going into speciﬁc details with the main aim to demonstrate a numerical example of a nonconvex solvability set.8. M) : t ∈ [τ.1. p. as in Sections 3.9.9.10 Target Control Synthesis within Free Time Interval Considering again the Problem of Control Synthesis Under Uncertainty of Deﬁnition 1. we come to results shown in Figures 3. where Wf (τ. q) same as in the previous ﬁgures. We shall look for an ”ellipsoidal” control synthesis solution to this problem within a scheme similar in nature to the one of Section 3. 3. 3. as W ∗ (τ.9. M) = {W ∗ (τ. Thus. M). x} if and only if x ∈ Wf (τ. f and the target set M are all ellipsoidalvalued. t1 ] (namely. we shall modify this deﬁnition by deleting the requirement that the terminal instant t1 is ﬁxed. 217 . m. t1 ]}. 3.10. t. as before).
M] = E(x(τ ).3. 19 218 . M)) of W(τ. t1 ] and deﬁne an internal approximation E− [τ. The results of the previous Sections allow to formulate the following assertion.10. set Wf (τ. M] . H(·)).Clearly. τ. t1 ] and H(t) ∈ Σ[t0 . t1 ] that symbolizes the interval where these functions are deﬁned).5. M ).41). M]t ∈ [τ. X− (τ π(·).1). τ. M)) and the pair π(·). t. t.5 for the classes of functions π(·). where E− [τ. t. if xτ ∈ Ef (τ.3) taken for E− [τ ] = E[τ.8. H(·)). Then . ( Here the earlier symbols Π+ . Σ of Sections 3. xτ ) is a solution to ( 3. Denote {E− [τ.5. M] = E(x(τ ). H(·)). M).E. X− (τ π(·). the strategy (3.1 Fix continuous functions π(t) ∈ Π+ [t0 . Proof. X− (τ π(·).8. H(·) are complemented by [t0 .3). where E− [τ ] is substituted by E[τ. t1 ]} ∈ Wf (τ. Q.2 . there exists a minimal value t = t∗ among those t that ensure xτ ∈ E[τ. whatever is the disturbance f (t) ∈ E(q(t). t.10. 19 Once (3. the inclusion x[t ] ∈ E(t . Q(t). M]t ∈ [τ. Mπ(·). H(·) is ﬁxed. t. M] = E(x(τ ). Theorem 3.(3. t. M)) stands for the internal ellipsoid described by equations (3. This is due to the continuity of the distance function Here symbol E− [τ.30) or (3. H(·)). t. M] for some τ ∈ [t0 . H(·)).D. Follows from the fact that {E− [τ. t. M]. t] and x[t ] = x(t . t. t. Mπ(·).31) taken at instant t instead of t1 . t1 ]} = Ef (τ. t. M) shall be true for all t ∈ [τ. but with boundary condition (3.8. t] and in particular x[t] ∈ E(m. t.1) xτ = x(τ ) ∈ E− [τ. M) is not bound to be convex. Hence. M] solves the terminal control problem by time t. t. t].5. see [305]. M) for τ ∈ [t0 .
q are the same as in the examples of ﬁgure 3. t. M]) = d[xτ .10. t] in t. M) of W(τ. t. M] that e generates the value t∗ . t] and H(·) ∈ Σ[t0 .9. t]. ( dividing the time interval — chosen to be [0. W ∗ (τ. But since 0 ≤ de [xτ . H(·)). M)) = de [xτ . M] = E+ (x(τ ). M. X+ (tπ(·). t1 ] and deﬁne an external approximation E+ [τ. t.10.1).1 One should be aware that in general the functions d[xτ . except that the initial position is given by 0 −20 x0 = . t]. t] ≤ d[xτ .10. then the problem of target control synthesis of this Section (under uncertainty. M]) = h+ (xτ . with free target time). t. t] . We shall now proceed with numerical examples. t. t] for all t ∈ [t0 .9. Check the the following assertion.3) de [xτ . p. Once (3.9.10.9. we further come to the following fact Lemma 3..1) is found here through the same parametrization. For the calculations we use the same discrete scheme as in Section 3. 5] — into 100 subintervals of equal lengths) and the control strategy oftype (3. E[τ. M] / for all t ∈ [t0 . (the latter function de is also continuous in t). ( check this assertion). Remark 3. P.1 The ”optimal time” t∗ ≤ t∗ .10. so that t∗ is the minimal root of the equation (3.10. Exercise 3. 0 5 219 .10. t. t. M) for τ ∈ [t0 .2) d[xτ . Fix continuous functions π(·) ∈ Π+ [t0 .d(xτ . Time t∗ shall then be the exact ”optimal time ”. t] and t∗ to be the minimal root of equation e (3.4) xt0 = x[t0 ] ∈ E+ [t0 .2) may lead to unstable numerical procedures that require additional regularization.13. so that the practical calculation of the roots of equations (3. t] are not monotonous in t. (3.10. Denote de (xτ . t. t] = 0. de [xτ . whatever is the internal tube E[τ. The parameters A.3. t] = 0.1. E[τ. t1 ].cannot be solved.
M). The numerical calculation on the basis of Theorem 3. as well as the controlled trajectories under two simulated disturbances f resulting in that the trajectories arrive to the target set M at time t = t∗ = 4.6 i. M). So t∗ = 4.10.2 !!!———— 220 . In Figure 3. that is. e. 4.1 shows the internal estimate of the set Wf (τ. Note that the data are chosen in such a way that neither the controls. t∗ ].1 then we keep the trajectory in the ellipsoidal valued map starting from the above ellipsoid E(0.6. t∗ . Q) nor Q ⊃ P holds. —————!!! insert ﬁgures 3. M]t ∈ [0. the ellipsoidal valued map E− [t. M] = E[0. In such a way we obtain that this relation holds for t = t∗ = 4. The layout of the two last Figures is the same as before. According to Theorem 3. t1 ]. Figure 3. M ) we have: 20 0 m= 0 −20 For the constraint E(q.10.the closest time instant by which the set M can e be hit for any disturbance f .10. t ∈ [0.10. M) at τ = 0 in the form of {E− [0. is carried out in the following way: after creating the internal estimate Ef (t0 . in this case the problem cannot be reduced to simpler situations without disturbances. M]. M].at the initial instant t0 = 0 and for the target set M = E(m. we check whether x(t0 ) = x0 ∈ E− (t0 . P ) = P ⊃ Q = E(q. t∗ . 3. t∗ . t1 ]}. taking increasing values of t ∈ [t0 .1. nor the disturbances dominate the other. t. x0 ∈ E− [0.9.6 is an upper estimate of t∗ . t. with the drawn segments of coordinate axes corresponding to the state variables ranging from −40 to 40.1.2 we see again the above set.10. M). Obviously.10.6. Q) on the adversary f here the matrix Q = Q(2) of Section 3. neither E(p.
[181]. the realization of the measurement and the constraints on the uncertain items. The calculation of information sets.Part IV. system inputs and measurement ”noise” are assumed to be unkown in advance. A detailed description of the bounding approach could be found in monographs [276]. The ﬁrst one is when the bounds on the unknowns are speciﬁed in advance. for example). One may try to approximate them by ﬁnitedimensional elements however. for example. [73]. A key element here is the notion of information set of states consistent with the system equations. among the objectives of this book 221 . is not a simple problem. ELLIPSOIDAL DYNAMICS: STATE ESTIMATION and VIABILITY PROBLEMS . this rather rough approximation may be misleading. We emphasize that here the uncertain items . [186].the initial states. as the center of the smallest ball that includes the information set ( which is the socalled the Chebyshev center of this set). though. in sophisticated applications ( to some types of pursuitevasion games. emphasizing constructive techniques for their solution worked out in the spirit of the earlier parts. even for the ”linearconvex” problems of this book. Introduction This last Part IV of the present book is concentrated around state estimation and viability problems. A natural move in this setting is to use the setmembership (”bounding”) approach. by ellipsoids. This leads to the problem of ”guaranteed state estimation” introduced in Section 1. which actually are inﬁnitedimensional elements. [225]. as in the present book. The problem may then be further treated in two possible settings. As mentioned above. particularly. . with no statistical information on them being available. it requires to describe more or less arbitrary types of convex compact sets. and reviews [226].12. [187]. It may also be useful to ﬁnd a single vectorvalued state estimator. This approximation may turn to be useful in applied problems where computational simplicity stands above accuracy of solution. On the other hand. The information set always includes the unknown actual state of the system and thus gives a setvalued guaranteed estimate of this state. which may be selected. Indeed. One of the main problems here is to give an appropriate description of the evolution of the information sets in time and of the dynamics of the vectorvalued estimators. The approximation of information sets by only one or few ellipsoids was described in [277].
The DP techniques allow to link the bounding approach with another deterministic approach to state estimation. but are rather approximated by classical solutions to systems of HJB equations constructed for adequate classes of linearquadratic extremal problems. in the limit. Among the problems of viability and state estimation are those. The estimators are then calculated through the knowledge of the information state .Baras and M.is to produce an ellipsoidal approximation by a parametrized variety of ellipsoids. Given is a ”measure of uncertainty” for the uncertain items and the vectorvalued estimator is generated through a system which realizes the minimal norm of a certain inputoutput map or a saddle point of an appropriate dynamic game. gives an exact representation of the information sets. if one deals with magnitude constraints on the inputs. Two types of such equations are given in Sections 4.2.6. [289]).3.and 4. The H∞ approach to estimation and feedback control was studied in many papers.1. Since systems with magnitude constraints on the inputs generate HJB equations with no classical solutions. This second scheme is often referred to as the socalled H∞ approach. Section 4. where the viability restriction or the state constraint induced by the measurement equation are not continuous The idea of such representations was indicated in [181]. [30].James who introduced the notion of information state. as described in Sections 2. ( Section 4. Particularly. The DP approach may as well be applied to the calculation of attainability domains. for example.7 and 3.5. In terms of level sets the last construction is again an ellipsoidal approximation. in the incorporation of the same DP equations to both settings. §§ 12. 21 20 222 .the solutions to the HJB equation of the H∞ approach. [231] and especially those of J. [82]. [15]. In this book these generalized solutions are not calculated explicitly.4 indicates that the respective ellipsoids could be transformed to be the same as those obtained through the purely geometrical considerations of Parts II and III.the value function of a certain problem in dynamic optimization calculated as a ” forward” solution of an appropriate HJB equation. However.2. 20 The parameters of the approximating external ellipsoids are described here as solutions to systems of ordinary diﬀerential equations. the latter equations could be analyzed within the notions of generalized solutions ( of the ”viscosity” or ”minmax” types. while the former follows from Dynamic Programming (DP) considerations. basically. The latter is derived through the relations of Section 2. see also [32]).3. 15. Here we mention [94]. which. 21 The important connection between the two approaches is that the information sets are the level sets for the information states . Similar assertions are also proved for the calculation of viability kernels . It is thus observed that the connection between the two approaches to the deterministic treatment of uncertainty in dynamics lies. This second approach to state estimation presumes that no bounds on the uncertain items are known. then the ellipsoidal approximations to these domains may again be achieved through the construction of level sets for value functions of appropriate linear quadratic extremal problems.
when the noise in the observations is modelled by discontinuous functions. The new problem is constructed such that it is free of the inadequacies of the original problem on one hand. [192].6 presents an ”ellipsoidal ” version of the technique. A detailed description of this scheme for state estimation and viability problems of general type is given in references [191].14. ( This particularly happens. for example). Section 4.in time. ( Section 4. and allows an approximation of the original one. The ﬁrst three Sections ( 4. A possible scheme for handling such situations lies in imbedding the original problem into one with singular perturbations. that may turn to be only Lebesguemeasurable.6).3) 223 . on the other.
it will be indicated that both approaches may be handled through one and the same equation of the HJB type. f (t). treating them in the context of the problem of state estimation with a further aim on using ellipsoidal techniques.3) Ψ(τ. f (t). t0 ≤ t ≤ τ.12.12. which was described by funnel equations (1. v(t). Here the uncertain items are taken to be unknown but bounded with given bounds on the performance range. The ﬁrst of these . as mentioned in the previous Introduction.4. Although formally somewhat diﬀerent. Namely.12.12.5) with u(t) ≡ 0. The second one is the socalled H∞ approach based in its linear version on the calculation of the minimalnorm inputoutput map for the investigated system and the error bound for the system performance expressed through this norm.1.2) y(t) = G(t)x(t) + v(t).10) or (1. these two approaches appear to have close connections. The technique is based on an approximation of the original problem with magnitude constraints by a parametrized variety of problems with quadratic integral constraints. ˙ (4.11). Such a scheme shall then allow a turn to ellipsoidal approximations of attainability domains. we shall indicate an approximation technique for solving the respective HJB equation. For the case of problems with ellipsoidal magnitude constraints on the uncertain items that are treated in the next Section and are among the the main points of emphasis in the present book. as we have seen in Section 1.1 Guaranteed State Estimation: a Dynamic Programming Perspective We shall begin this Section by discussing the two basic approaches to the deterministic treatment of uncertainty in the dynamics of controlled processes.1. t0 ≤ t ≤ τ } are now bounded by the inequality (4. rewriting it as (4.1). 224 .12. Let us start with a slightly more general problem than in Section 1. v(t))dt + φ(x0 ) ≤ µ2 . x(t0 ) = x0 .1. The estimate is then sought for in the form of a set .(1. These may be demonstrated particularly through the techniques of Dynamic Programming that are the topic of this Section. We shall assume that the unknown items ζ(·) = {x0 . is the bounding approach based on setmembership techniques. Consider again the system (1.1) x(t) = A(t)x(t) + f (t). ζ(·)) = τ t0 ψ(t.12.” the informational domain”.
stands for the scalar product in the respective space IRk . 225 .6) (4.4 we shall synchronise these notations with the earlier ones. squareintegrable in t ∈ [t0 . N (t) are positive deﬁnite. I1 . emphasizing that the treatment of the state estimation (ﬁltering) problem .7) (4. L(x0 − a)).8) I0 (x0 ) = (x0 − a.1.1. f (·)) = esssupt (f (t) − f ∗ (t). As we shall observe in the sequel. τ ]. t ∈ [t0 .5) ψ(t. q ∈ IRk ). I2 (τ.1.9) Ψ(τ.1. where (p. M (t)(f (t) − f ∗ (t)) ≤ µ2 . In Section 4. ζ(·)) = max{I0 . the bounds may be of the quadratic integral type. v(·)) = esssupt (v(t) − v ∗ (t). N (t)(v(t) − v ∗ (t)) ≤ µ2 .1. the measurement y(t). N (t)(v(t) − v ∗ (t))). a particular case of which is described by ellipsoidal–valued constraints – the inequalities 22 (4. Another common type of restriction is given by magnitude bounds. v(t)) = (f (t) − f ∗ (t). f (t). f ∗ (t). M (t)(f (t) − f ∗ (t))) + +(v(t) − v ∗ (t).1. and L > 0. Particularly. t ∈ [t0 .14. the aim of the state estimation (”ﬁltering”) problem could be described as follows: (a) determine an estimate x0 (τ ) for the unknown state x(τ ) on the basis of the available information: the system parameters.where Ψ(τ.and the restrictions on the uncertain items ζ(·) (if these are speciﬁed in advance).1) may be or may not be given in advance and the corresponding solution will of course depend on this speciﬁcity of the problem. namely . τ ]. ζ(·)) reﬂects the accepted uncertainty index for the unknown items. v ∗ (t) are given vector functions of respective dimensions. Despite of the latter fact. a ∈ IRn is a given vector. as given here. M (t). continuous. such that (4.3 the notations for the bounds on the unknowns are independent of those introduced earlier. (4.1. I2 }. L(x0 − a)) ≤ µ2 I1 (τ.4) φ(x0 ) = (x0 − a. In this case the functional (4. the number µ in the restriction (4. 22 In the coming Sections 4. q) (p. is independent of the earlier material. τ ].
provided the bound on the uncertain items ( the number µ ) is given in advance . t ∈ [t0 . when the conﬁguration of X (τ ) may be quite complicated).1. merely through the smallest number σ 2 that ensures the inequality (4. where e(τ ) = T (ζ(·)) with y = y(t) given.1. if possible. for example. however.12.The inclusion x(τ ) ∈ X (τ ) ˆ will be secured as X (τ ) is convex.1) with speciﬁed µ are given together with the available measurement y = y(t). the value µ for the bound on the uncertain items is not presumed to be known. The latter ”worstcase” estimate is less precise than in 226 . Suppose that the constraints (4. ζ(·)) under restrictions (4. z ∈ X (τ ). deﬁned through the relation (4. one may be certain that for the unknown actual value x(τ ) we have : x(τ ) ∈ X (τ ) . the smallest number σ 2 is clearly the square of the minimal norm of the inputoutput mapping T .5). Let us discuss the problem in some more detail.11) e2 (τ ) ≤ σ 2 Ψ(τ. The set X (τ ) gives an unimprovable estimate of the statespace variable x(τ ) .(This may not be the case for the general nonlinear problem.10) min max(x − z. x(τ ) − x(τ )) ˆ ˆ is then estimated. It obviously depends on the type of norm ( the type of functional Ψ(ζ(·) ) selected to evaluate η(·) ). Since we deal with the linear case. and may therefore ﬁnd a certain point x(τ ) ∈ X (τ ) that would serve as the required estimate x0 (τ ). As mentioned above.2).(b) calculate the error bounds for the estimate x0 (τ ) on the basis of the same information. With X (τ ) calculated. at the ˆ end of the previous Section. an ordinary diﬀerential equation. preferably through a dynamic recurrencetype relation.12.1.1. this point x(τ ) may be particularly selected as the ”Chebyshev ˆ center” for X (τ ) . x ˆ x z z and is obviously the center of the smallest ball that includes the set X (τ ) . in its turn.(1. in the second or H∞ approach. x − z). τ ]. (c) describe the evolution of the estimate x0 (τ ) and the error bound in τ . The bounding approach then requires that the solution would be given through the information domain X (τ ) of Deﬁnition 1. x − z) = max(ˆ − z.1. On the other hand. while the value of the estimation error e2 (τ ) = (x(τ ) − x(τ ).
t0 .5).12) Deﬁne (4.13) V (τ.1.1.14) X (τ ) = {x : V (τ. x) will be referred to as the information state of system (4. τ ]}.1. it actually indicates a larger error bound). An obvious assertion is given by Lemma 4. through a Dynamic Programming (DP) technique. t0 . x) = inf {Φ(τ. to start with. f (·)) = x(t. η(·))) − v ∗ (t).3 to emphasize the connections between the two approaches and to indicate . t0 .11).1. producing either of them. depending on the ”a priori” information. this may sometimes suﬃce for the speciﬁc problem under discussion. relative to measurement y(·) and criterion Φ.1. η(·)) x(τ. as well as on the required accuracy of the solutions.the ﬁrst approach ( since. ζ(·))v(t) ≡ y(t) − G(t)x(t. the respective function V (τ. η(·)) = x} η(·) Φ(τ. M (t)(f (t) − f ∗ (t))+ +(y(t) − G(t)x(t. N (t) > 0 the operation ” inf” in the line above may be substituted for ”min”.(4.1.1. (4. τ ] and functional Φ(τ. Denote η(·) = {x0 . η(·))) − v ∗ (t)))dt Clearly.1 The informational domain X (τ ) is the level set (4. However. 227 .1 Given measurement y(t). as one may observe. x(t. x). presuming y(·) to be given and restriction (4. N (t)(y(t) − G(t)x(t. t0 .1.1. Deﬁniton 4. We shall use the upcoming discussion in Section 4. η(·)) = (x0 − a. x(t. η(·)) = {Ψ(τ. P0 (x0 − a)) + τ t0 ((f (t) − f ∗ (t).2). η(·)) − v ∗ (t)} With L. x0 .3)(4. η(·)) of (4.1). Let us start by introducing a scheme for describing the information domains X (τ ). a general framework that incorporates both of these. x(t. t0 .1. η(·)) and Φ(τ. t ∈ [t0 .3) to be of the quadratic integral type (4. f (t). x) ≤ µ2 } for the information state V (τ. t0 .1. t ∈ [t0 .
M −1 (t) )− ∂t ∂x 4 ∂x ∂x (y(t) − G(t)x. N (t)(y(t) − G(t)x))} = 0 so that. The respective function Φ(τ.of the estimate x∗ (τ ) for the unknown state x(τ ).1.17) V (t0 .type of equation. x) and further on . x)  x ∈ Rn } ≤ µ2 Since Lemma 4.1. x). As already emphasized above.16) with boundary condition (4. the linearquadratic solutions will be important in organizing ellipsoidal approximations for systems with magnitude constraints. N (t)(y(t) − G(t)x)) = 0 (4. Let us therefore introduce a DP . after an the elimination of f . (ii)the information state depends both on y(·) and on the type of functional Φ . L(x − a)) 228 .It should be emphasized here that both V (τ. The crucial diﬃculty here is the calculation of the sets X (τ ).15) V 0 (τ ) = inf{V (τ.1. apart from their separate signiﬁcance. x) ≥ 0 and X (τ ) depend on the given measurement y(t) as well as on the type of functionals Ψ. the function V (τ. (A(t)x + f (t) − (f − f ∗ (t). Ax + f ∗ ) + ( .5). We emphasize once more the main conclusions: (i)the information domain X (τ )is the level set for the informational state V (τ.1. the knowledge of V (τ.1 indicates that the X (τ ) is a level set for V (τ.(4. Φ and that X (τ ) = ∅. x) will thus allow to calculate the sets X (τ ). Applying standard techniques . The calculations are relatively simple for an exceptional situation .1. provided (4.the linearquadratic case.1. M (t)(f − f ∗ (t)))− v ∂τ ∂x −(y(t) − G(t)x. η(·)) then obviously satisﬁes the Optimality Principle of Dynamic Programming [109]. taking V (τ.13) when Ψ(τ. [53]. ζ(·)) is given by (4. x) that corresponds to the given number µ. that the respective HJB equation is as follows ∂V 1 ∂V ∂V ∂V +( . [109]. x) = (x − a. we may observe is ∂V ∂V + max{( .3).1. x) to be the value function for the linearquadratic problem (4.
being an ellipsoid E(z(τ ). P(τ )(x − z(τ ))) ≤ µ2 − k 2 (τ )}.1.20) z = A(t)z + P −1 G (t)N (t)(y(t) − G(t)z − v ∗ (t)) + f ∗ (t). where z(τ ). (4.23 An obvious consequence of the given reasoning is the following assertion Lemma 4.1 Note that the matrixvalued function P(t) does not depend on the measurement y(·).1. k 2 (t) are the solutions to the following wellknown equations [149]. x) into equation (4.21).1. 23 229 . P(τ )) given by the relation (4. (4.9).1.18) V (τ. is not present in stochastic theory. x) . P(τ ) > 0.1.Its solution is a quadratic form (4. However.1.1. k 2 (t0 ) = 0 Equations (4.1.21) .2 Under restrictions (4.21) ˙ k 2 = (y(t) − G(t)z − v ∗ (t).21) are derived by direct substitution of V (t.1. (4.1.20) are the same as in stochastic ”Kalman” ﬁltering theory. while the scalar function k 2 (t) depends on the measurement.19)(4.19) ˙ P = −PA(t) − A (t)P − PM −1 (t)P + G (t)N (t)G(t).16). z(t). P(τ )(x − z(τ )) + k 2 (τ ) where P(t).5) on the uncertain inputs ζ(·) = {η(·). [276]. k 2 (τ ) are deﬁned through equations (4. ˙ z(t0 ) = a.(4.1. P(t0 ) = L. [57]. It is speciﬁc for the setmembership approach and reﬂects the dynamics of the ”size” of the information set. P −1 (τ )) = = {x : (x − z(τ ). the third one.19)(4. Remark 4.1. The estimation error is given by an errror set R(τ ) = X (τ )−‡(τ ) which therefore depends only on k 2 (τ ). (4. [181].(4. v(·)} the informational domain X (τ ) for the system (4.3) . One may easily observe that the ﬁrst two equations (4.1.14) for the informational state V (τ. x) = (x − z(τ )). (4.1). N (t)(y(t) − G(t)z − v ∗ (t))).2) is the level set (4.1.1.1.1.1.1.1.22) X (τ ) = E(z(τ ).
v ∗ (·)}. so that in this case the informational set X (τ )) is ¯ a singleton and X (τ ) = {x(τ.1. we come to (4. P(τ )e) ≤ µ2 } The other extreme situation is when the measurement is the best possible. h(t0 ) = 0. so that x(t) = x(t.4 There exists a function ( measurement ”noise”) v = v (t).1.21) immediately indicates the worstcase realization y ∗ (t) of the measurement y(t) which yields the ”largest” set X (t) ( with respect to inclusion). and the realization of the measurement ”noise” is v(t) ≡ v ∗ (t).20). such that the ¯ ∗ triplet ζ(·) = {a. Subtracting (4. x(t0 ) = x0 . The worstcase error set is the ellipsoid R(τ ) = X (τ ) − z(τ ) = E(0.1. (u(t) ≡ 0).1.21) yield h(t) ≡ 0.23) x = A(t)x + v.1).24) where A(t) = A(t) + P −1 (t)G (t)N (t)G(t). (among other possible triplets).1.3 The worstcase realization y(t) = y ∗ (t) of the measurement is a function that ( among other possible triplets ζ(·)) ) may be generated by the triplet {x0 = a. due to system (4. v(·) = v ∗ (·) .1. t0 . f (t) ≡ f ∗ (t).(4. If the actual realization x(t) is generated by x0 = a.1. η ∗ (·)).1. Namely.1. We therefore come to Lemma 4. η ∗ (·))} 230 . f (t) ≡ f ∗ (t). v(t) ≡ v ∗ (t)} which yields k 2 (τ ) = 0 . (4. v (t)} generates . t0 .24). µ2 P(τ )) = {e : (e.1.23) from (4. K(t) = P −1 (t)G (t)N (t). then (4. f (t). then y ∗ (t) = G(t)x(t. ˙ ˙ h = A(t)h(t) + K(t)(v(t) − v ∗ (t)) + f (t) − f ∗ (t). a mea¯ surement y (·) that ensures k 2 (τ ) = µ2 . k 2 (t) ≡ 0. where x(t) is the realization of the actual trajectory . if it possible to obtain the speciﬁc measurement y ∗ (t) through the triplet η(·) = η ∗ (·) = {a. η ∗ (·)) + v ∗ (t) is the worstcase realization of the measurement and the respective value V ∗ (τ ) = V 0 (τ )y(·)=y∗ (·) = 0 In order to check the last assertions. Lemma 4.2). generated due to equation (4. let us introduce an equation for the function h(t) = x(t)−z(t).1. t0 .Formula (4.
1.1. P(τ )(x − z(τ ))).26) ˙ h(t) = A(t)h(t) + K(t)(v(t) − v (t)) ¯ we shall require that k(t). ¯ ensuring particularly. x) = (x − z(τ ).1.1.5). x(t0 ) = x0 .1. {x(τ )} = X (τ ). h(τ ) = 0.1.1.Returning to equation (4.25) ˙ k 2 (t) = (G(t)q(t) + v (t) − v ∗ (t).N (t) ≡ 0 on the inputs η(·) = {x0 . ¯ ¯ and (with f (t) ≡ f ∗ (t) ) (4. we come to (4. z(t) are the solutions to the equations 231 . f (·)} the attainability X (τ ) for system (4. so that we simply have the standard system (4. N (t)(G(t)q(t) + v (t) − v ∗ (t))). where P(t).1.21) in view of measurement equation y(t) = G(t)x(t) + v∗ (t).24) and rewriting (4.25) for the function V (τ.3)(4. We may therefore follow the calculations of the above. We leave to the reader to verify that such a solution v (t) does exist . h(t0 ) = 0. at instant τ . the equalities x(τ ) = x(τ.3)(4.1. η ∗ (·)) = z(τ ).5 Under restrictions (4.N (t) ≡ 0. The procedure then ”automatically” gives the following result Lemma 4. t0 .27) k 2 (t0 ) = 0.1.28) under the constraint (4. h(t) satisfy the following boundaryvalue problem (4.28) x = A(t)x + f (t). ˙ with quadratic constraint (4. ¯ Finally .N (t) ≡ 0 Then the set X (τ ) is merely the attainability domain for system (4.21) is the level set (4.1. let us assume that there is no measurement equation.1.(4. The solution v (t) to this problem obviously satisﬁes the requirements of the last Lemma.1.1.3) .5). setting N (t) ≡ 0.5).1. k 2 (τ ) = µ2 .1.
t ∈ [t0 . x) of (4.1.2 From Dynamic Programming to Ellipsoidal State Estimates Let us now specify the information state V (τ. β(t) ≥ 0. β(·). N (t). P(τ )) given by relation (4. β(·). P(τ )) = = {x : (x − z(τ ).1.1. presuming Φ is deﬁned through relations (4.7).1l.1. the restrictions (4. if necessary. M (t)(f (t) − f ∗ (t))) + γ(t)(v(t) − v ∗ (t). γ(t) ≥ 0 232 Λ(τ.2.1 Assuming M (t).1. η(·).2.13) for the case of magnitude constraints.8) are equivalent to the system of inequalities (4. One may observe that Φ(τ.1.2. α. being an ellipsoid E(z(τ ). γ(·)) = α(x0 − a.(4. We are now prepared to extend the results of this Section to problems with magnitude constraints.(4.31) X (τ ) = E(z(τ ). ˙ z(t0 ) = a.2) α ≥ 0. P(t0 ) = L.29) (4. γ(·)) ≤ 1 . through its generalized versions that deal with nondiﬀerentiable functionals (see [ooo]. 4. We shall not pursue the last direction.4) . η(·). η) again satisﬁes the Optimality Principle (and is thus a quasipositional functional in terms of [170]). τ ] continuous. α. L(x0 − a))+ + τ t0 (β(t)(f (t) − f ∗ (t).[ooo]). Denote Λ(τ. N (t)(v(t) − v ∗ (t))))dt Lemma 4.6) (4. but shall rather apply yet another scheme which will be of direct further use in this book. One may therefore again calculate V (τ.1) whatever are the parameters (4. P −1 (τ )(x − z(τ ))) ≤ µ2 }.30) z = A(t)z + f ∗ (t). ˙ P + PA(t) + A (t)P + PM −1 (t)P = 0. x) through the HJB equation or.1.
2.1. assume for example.D.E.2. η(·)) = x}. then multiply (4.4).2.2 The function Φ(τ. the third condition (4.2. η(·). take any triplet ω(·) .1. Taking α = 1.2. 24 .2.9) may be expressed as (4.2) being true for almost all t. t ∈ e . β(t) ≡ (mes(e)−1 .1 may as well be proved if β(t).1) due to (4.Then taking α = 0.2.2.2.2).(4. as Ω = {ω(·)}. With (4. Conversely.3) α+ τ t0 (β(t) + γ(t))dt = 1 The functions β(·).7) by β(t) and (4. η(·)). (4. β(t) ≡ 0. With slight modiﬁcations the present Lemma 4. Similarly.2.1.(4.2.2.4) Φ(τ. (4. β(t) ≡ 0. f (·)} on the set of elements η(·) restricted by the equality {x : x(τ.2. that (4. f (·)) ≥ > 0 on a set e of measure mes(e) > 0. ω(·))  ω(·) ∈ Ω} The proof of an analogous fact may be also found in [181]. we obtain (4. τ ]. assume (4.7) turns to be true. η(·)) = sup{Λ(τ. For further calculations we emphasize the following obvious property Lemma 4. Q. γ(·) are taken to be measurable.1. γ(·)} = ω(·) and the variety of triplets ω that satisfy (4.3 The functional Λ(τ.2. β(·).3).1) .8) also follows from (4.1) . with inequalities (4. ω(·)) is convex in η(·) = {x0 . γ(t) are taken to be continuous 24 233 .2. We further denote the triplet {α.7) is false and therefore that I1 (t.4) and thus (4.2. Proof.1. Further on .2.where (4. one comes to a contradiction with (4.3) given .1). one comes to (4.1) to be true for any ω(·) ∈ Ω.12).1. then integrate the last two relations over t ∈ [t0 . t ∈ e. η(·)) of (4.1. Using a similar reasoning the reader may now verify the following assertion Lemma 4. γ(t) ≡ 0. Adding the results .2.8) by γ(t). γ(t) ≡ 0.6) by α .1) (4.
k = k(t.10). x) = sup{V (τ. We therefore come to the relation (4.1.8) V (τ. concave) in ω and convex in η . γ(·)).1. x. γ(·)) satisfy the equations 234 . under boundary condition (4. P(τ. according to Lemma 4.2. ω) is the solution to equation (4.2. η(·).1. x. ω(·)) and M (t).2 we have (4. the order of operations inf. where P = P(t.2. γ(·)).2. Solving problem (4.2.2. we observe (4.2 ) with V (τ.9) . x) = sup min Λ(τ. γ(·)). L(x − a)) This leads to Lemma 4. η(·).6) x(τ. ω(·)) = = (x − z(τ.7) is given by (4. x(τ.2. γ(·)) + k 2 (τ. with M (t).6). η(·)) = x} may be solved through equation (4. ω(·)) ω η(·) under restrictions ω(·) ∈ Ω and (4.7) V (τ. γ(t)N (t). η(·).Due to Lemma 4.2. ω(·)) = min{Λ(τ. γ(t)N (t) respectively with boundary condition being (4. The internal problem of ﬁnding (4. ω(·)) is linear ( therefore. (4. N (t) for β(t)M (t). ω(·))} η ω(·) under restriction ω(·) ∈ Ω.4 The information state (4.2.2. x) = inf {Φ(τ.9) V (t0 .3. x.10) V (τ. ω(·)). sup may be interchanged. ω(·))  η(·).2.16).11) V (τ. η(·).16) ( see Remark 4. x) substituted for V (τ.2. (see [101]. N (t) substituted for β(t)M (t). x. η(·)) = x The functional Λ(τ.5) V (τ. ω(·))  ω(·) ∈ Ω} where V (τ. [86]). ω(·))(x − z(τ.2. η(·)) = x} = η = inf sup{Λ(τ. z = z(t. ω(·)) = α(x − a. x. x. η(·))  x(τ.. In view of minmaxtype theorems.2.
(4.2.12) (4.2.13)
˙ P = −PA − A P − β −1 (t)PM −1 (t)P + γ(t)G (t)N (t)G(t), z = A(t)z + γ(t)P −1 (t)G (t)N (t)(y(t) − G(t)z − v ∗ (t)) + f ∗ (t), ˙
(4.2.14) dotk 2 (t) = γ(t)(y(t) − G(t)z − v ∗ (t), N (t)(y(t) − G(t)z − v ∗ (t))) (4.2.15) Pt0 = αL, z(t0 ) = x0 , k(t0 ) = 0
Finally this develops into the assertion Theorem 4.2.1 For the system (4.1.1),(4.1.2) the information state V (τ, x) relative to measurement y(·) and nonquadratic ( ”magnitude”) criterion (4.1.9), is the upper envelope (4.2.16) V (τ, x) = sup{V (τ, x, ω(·))  ω(·) ∈ Ω}
of a parametrized family of quadratic forms V (τ, x, ω(·)) of type (4.2.11) over the functional parameter ω(·) = {α, β(·), γ(·)} , where ω(·) ∈ Ω. As we have observed in the previous sections, the informational domain X (τ ) = E(z(τ ), P −1 (τ )) is deﬁned by V (t, x) through inequality (4.2.13) with µ given. Moreover, for each of the ellipsoidal level sets (4.2.17) X (τ, ω(·)) = = E(z(τ, ω(·)), P −1 (τ, ω(·)) = {x : V (τ, x, ω(·)) ≤ µ2 } where V (τ, x, ω(·)) is a nondegenerate quadratic form(!), we obviously have X (τ ) ⊆ X(τ, ω(·)) = E(z(τ, ω(·)), (µ2 − k 2 (τ ))P −1 (τ, ω(·))), ∀ω(·) ∈ Ω, so that (4.2.16) yields the following fact Theorem 4.2.2 For the system (4.1.1),(4.1.2), with criterion (4.1.9), the informational set X (τ ) is the intersection of ellipsoids X (τ, ω(·)) = E(z(τ, ω(·)), (µ2 − k 2 (τ ))P −1 (τ, ω(·))) namely, 235
(4.2.18) where
X (τ ) = {∩E(z(τ, ω(·)), (µ2 − k 2 (τ ))P(−1 τ, ω(·)))  ω(·) ∈ Ω}
z(t) = z(t, γ(·)), P(t) = P(t, ω(·)), k 2 (t) = k 2 (t, γ(·)) are deﬁned through equations (4.2.12)(4.2.15). The worst case measurement y(t) = y ∗ (t) is a function that may be generated ( among other possible triplets) by triplet ζ ∗ (·) , where x0 = a, f (t) = f ∗ (t), v(t) = v ∗ (t) .This yields k 2 (τ ) = 0 and V ∗ (τ ) = V 0 (τ )y(·)=y∗ (·) where V 0 = inf {V (τ, x)x ∈ IRn } = 0 The last part of the theorem that deals with the worstcase measurement y ∗ (·) may be checked by substituting ζ(·) into (4.2.13), (4.2.14) and following the reasoning of the previous Section. Remark 4.2.1 Observe that again function k 2 (t) depends upon the measurement y(·), while P(t) does not depend upon y(·). Remark 4.2.2 The fact that functions β(t), γ(t) are taken measurable does not forbid to use equation (4.1.16) and the further schemes of Section 4.1 for the function V (t, x, ω(·)). This particularly is due to the unicity of solution to the extremal problem (4.2.8). Besides that, α(t), β(t) may be assumed continuous ( see footnote after Lemma 4.2.1). In the absence of state constraints induced by the measurement (N (t) ≡ 0), one should simply delete the restriction (4.1.8) and set γ(·) = 0 in the previous Theorem. This also gives k(t) ≡ 0. Corollary 4.2.1 In the absence of the state constraint (4.1.8) relations (4.2.17),(4.2.18) generated by equations (4.2.12)  (4.2.15) remain true , provided γ(t) ≡ 0. The set X (t) is then the attainability domain of Section 1.2 for system (4.1.1) under ellipsoidal magnitude constraints (4.1.6),(4.1.7). Further, in Section 4.4, we shall rearrange the results obtained here in terms of earlier notations and compare them with those obtained Parts IIIII. But prior to that we shall discuss the calculation of error bounds for the estimation problems. 236
4.3
The State Estimates, Error Bounds and Error Sets
Let us now pass to the discussion of the estimates and the error bounds. Consider the informational domain X (τ ) to be speciﬁed. Under the assumptions of Sections 4.1, 4.2 set X (τ ) will be closed and bounded. Let us calculate the Chebyshev center of X (·). Following formula (4.1.10), we have to minimaximize the function min max (x − z, x − z) = max (x − x, x − x) ˆ ˆ
z x x
under the restriction V (τ, x) ≤ µ2 . Applying the conventional generalized Lagrangian technique [69], [260], [265], we have (4.3.1) min max{(x − z, x − z) − λ2 µ V (τ, x)} z x
Since x(τ ) is the center of the smallest ball that includes X (τ )  a convex and compact set, ˆ the inclusion x(τ ) ∈ X (τ ) is always true. ˆ Here the number λ2 is the Lagrange multiplier which generally depends on µ as also does µ x(τ ) = xµ (τ ). With V (τ, x) being a quadratic form of type (4.1.6), the solution to (4.3.1) ˆ ˆ is the center of the ellipsoid (4.1.22), namely, x(τ ) = z(τ ), whatever is the value of µ. ˆ Summarizing the results, we have Lemma 4.3.1 The minmax estimate x(t) (the Chebyshev center) for the informational ˆ domain X (τ ) of Section 1.12, satisﬁes the property x(τ ) ∈ X (τ ) ˆ and in general depends on µ : x(τ ) = xµ (τ ). ˆ ˆ In the linearquadratic case (4.1.3)(4.1.5) the vector x(τ ) = z(τ ) ˆ is the center z(τ ) of the ellipsoid E(τ, P −1 (τ )) described by the (4.1.22) and does not depend on the number µ .
237
In order to compare the setmembership (bounding) and the H∞ approaches, let us ﬁnd the estimate x(τ ) for the H∞ approach to state estimation. Then we have to solve the ¯ following problem: Find the smallest number σ 2 that ensures min max{(x − z, x − z) − σ 2 Ψ(τ, ζ(·))} ≤ 0
z ζ(·)
under the conditions x(τ, η(·)) = x; G(t)x(t, t0 , η(·)) + v(t) ≡ y(t); t0 ≤ t ≤ τ.
2 This, h