Professional Documents
Culture Documents
Editor-in-Chief
Ralph C. Smith, North Carolina State University
Editorial Board
Athanasios C. Antoulas, Rice University Michel C. Delfour, University of Montreal
Siva Banda, Air Force Research Laboratory Max D. Gunzburger, Florida State University
Belinda A. Batten, Oregon State University J. William Helton, University of California, San Diego
John Betts, The Boeing Company (retired) Arthur J. Krener, University of California, Davis
Stephen L. Campbell, North Carolina State University Kirsten Morris, University of Waterloo
Series Volumes
Osmolovskii, Nikolai P. and Maurer, Helmut, Applications to Regular and Bang-Bang Control: Second-Order
Necessary and Sufficient Optimality Conditions in Calculus of Variations and Optimal Control
Biegler, Lorenz T., Campbell, Stephen L., and Mehrmann, Volker, eds., Control and Optimization with
Differential-Algebraic Constraints
Delfour, M. C. and Zolésio, J.-P., Shapes and Geometries: Metrics, Analysis, Differential Calculus, and
Optimization, Second Edition
Hovakimyan, Naira and Cao, Chengyu, L1 Adaptive Control Theory: Guaranteed Robustness with Fast Adaptation
Speyer, Jason L. and Jacobson, David H., Primer on Optimal Control Theory
Betts, John T., Practical Methods for Optimal Control and Estimation Using Nonlinear Programming, Second
Edition
Shima, Tal and Rasmussen, Steven, eds., UAV Cooperative Decision and Control: Challenges and Practical
Approaches
Speyer, Jason L. and Chung, Walter H., Stochastic Processes, Estimation, and Control
Krstic, Miroslav and Smyshlyaev, Andrey, Boundary Control of PDEs: A Course on Backstepping Designs
Ito, Kazufumi and Kunisch, Karl, Lagrange Multiplier Approach to Variational Problems and Applications
Xue, Dingyü, Chen, YangQuan, and Atherton, Derek P., Linear Feedback Control: Analysis and Design
with MATLAB
Hanson, Floyd B., Applied Stochastic Processes and Control for Jump-Diffusions: Modeling, Analysis, and
Computation
Michiels, Wim and Niculescu, Silviu-Iulian, Stability and Stabilization of Time-Delay Systems: An Eigenvalue-
Based Approach
Ioannou, Petros and Fidan, Barıs,¸ Adaptive Control Tutorial
Bhaya, Amit and Kaszkurewicz, Eugenius, Control Perspectives on Numerical Algorithms and Matrix Problems
Robinett III, Rush D., Wilson, David G., Eisler, G. Richard, and Hurtado, John E., Applied Dynamic Programming
for Optimization of Dynamical Systems
Huang, J., Nonlinear Output Regulation: Theory and Applications
Haslinger, J. and Mäkinen, R. A. E., Introduction to Shape Optimization: Theory, Approximation, and Computation
Antoulas, Athanasios C., Approximation of Large-Scale Dynamical Systems
Gunzburger, Max D., Perspectives in Flow Control and Optimization
Delfour, M. C. and Zolésio, J.-P., Shapes and Geometries: Analysis, Differential Calculus, and Optimization
Betts, John T., Practical Methods for Optimal Control Using Nonlinear Programming
El Ghaoui, Laurent and Niculescu, Silviu-Iulian, eds., Advances in Linear Matrix Inequality Methods in Control
Helton, J. William and James, Matthew R., Extending H1 Control to Nonlinear Systems: Control of Nonlinear
Systems to Achieve Performance Objectives
Nikolai P. Osmolovskii
Systems Research Institute
Warszawa, Poland
University of Technology and Humanities in Radom
Radom, Poland
University of Natural Sciences and Humanities in Siedlce
Siedlce, Poland
Moscow State University
Moscow, Russia
Helmut Maurer
Institute of Computational and Applied Mathematics
Westfälische Wilhelms-Universität Münster
Münster, Germany
10 9 8 7 6 5 4 3 2 1
All rights reserved. Printed in the United States of America. No part of this book may be reproduced,
stored, or transmitted in any manner without the written permission of the publisher. For information,
write to the Society for Industrial and Applied Mathematics, 3600 Market Street, 6th Floor, Philadelphia,
PA 19104-2688 USA.
No warranties, express or implied, are made by the publisher, authors, and their employers that the
programs contained in this volume are free of error. They should not be relied on as the sole basis to
solve a problem whose incorrect solution could result in injury to person or property. If the programs are
employed in such a manner, it is at the user’s own risk and the publisher, authors, and their employers
disclaim all liability for such misuse.
Trademarked names may be used in this book without the inclusion of a trademark symbol. These names
are used in an editorial context only; no infringement of trademark is intended.
Figure 8.3 reprinted with permission from John Wiley & Sons, Ltd.
Figure 8.5 reprinted with permission from Springer Science+Business Media.
Figure 8.8 reprinted with permission from Elsevier.
is a registered trademark.
List of Figures xi
Notation xiii
Preface xvii
Introduction 1
vii
viii Contents
7.5 Equivalence of the Quadratic Forms in the Basic and Induced Opti-
mization Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
Bibliography 367
Index 377
List of Figures
5.1 Optimal production and maintenance, final time tf = 0.9. (a) State vari-
ables x1 , x2 . (b) Regular production control v and bang-bang maintenance
control m. (c) Adjoint variables ψ1 , ψ2 . (d) Maintenance control m with
switching function φm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
5.2 Optimal production and maintenance, final time tf = 1.1. (a) State vari-
ables x1 , x2 . (b) Regular production control v and bang-singular-bang
maintenance control m. (c) Adjoint variables ψ1 , ψ2 . (d) Maintenance
control m with switching function φm . . . . . . . . . . . . . . . . . . . . 252
6.1 Minimal fuel consumption of a car. (a) State variables x1 , x2 . (b) Bang-
bang control u. (c) Adjoint variable ψ2 . (d) Switching function φ. . . . . 274
6.2 Time-optimal solution of the van der Pol oscillator, fixed terminal state
(6.120). (a) State variables x1 and x2 (dashed line). (b) Control u and
switching function ψ2 (dashed line). (c) Phase portrait (x1 , x2 ). (d) Ad-
joint variables ψ1 and ψ2 (dashed line). . . . . . . . . . . . . . . . . . . 287
6.3 Time-optimal solution of the van der Pol oscillator, nonlinear boundary
condition (6.129). (a) State variables x1 and x2 (dashed line). (b) Control
u and switching function ψ2 (dashed line). (c) Phase portrait (x1 , x2 ).
(d) Adjoint variables ψ1 and ψ2 (dashed line). . . . . . . . . . . . . . . 289
6.4 Time-optimal control of the Rayleigh equation. (a) State variables x1
and x2 (dashed line). (b) Control u and switching function φ (dashed
line). (c) Phase portrait (x1 , x2 ). (d) Adjoint variables ψ1 and ψ2 (dashed
line). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
xi
xii List of Figures
·
X : norm in the normed space X.
[a, b] ⊂ X : [a, b] := {x ∈ X | x = αa + (1 − α)b,
where α ∈ [0, 1]}, closed interval in the linear
space X with endpoints a, b.
A : closure of the set A.
X∗ : dual space to the normed space X.
x ∗ , x : value of the linear functional x ∗ ∈ X∗ at the
point x ∈ X.
K∗ : K ∗ := {x ∗ ∈ X∗ | x ∗ , x ≥ 0 for all x ∈ K},
dual cone of K.
g
(x0 ) : Fréchet derivative of the mapping g : X → Y
at x0 .
J (x) → min : J (x) → min , fi (x) ≤ 0, i = 1, . . . , k, g(x) = 0
(g : X → Y ); abstract optimization problem in
the space X.
σ (δx) : σ (δx) = max{m(δx),
g(x0 + δx)
},
: m(δx) = maxi=0,...,k fi (x0 + δx), f0 (x) :=
J (x) − J (x0 ), violation function in the
abstract optimization problem at x0 .
λ = (α0 , . . . , αk , y ∗ ) : Lagrange multipliers in the abstract optimiza-
tion problem.
xiii
xiv Notation
k
L(λ, x) : L(λ, x) = α0 J (x) + i=1 αi fi (x) +
y ∗ , g(x) , Lagrange function in the abstract
optimization
problem.
0 : 0 = λ | αi ≥ 0, (i = 0, . . . , m), αi fi (x̂) =
0 (i = 1, . . . , k),
k ∗
: i=0 αi +
y
= 1, Lx (λ, x̂) = 0 ,
Lx = ∂L/∂x,
set of the normed tuples of Lagrange multipli-
ers at x̂.
K : K = {x̄ ∈ X | fi
(x0 ), x̄ ≤ 0, i ∈ I ∪
{0}; g
(x0 )x̄ = 0}, cone of critical directions
(critical cone) at the point x0 ,
I = {i ∈ {1, . . . , k} | fi (x0 ) = 0}, set of active
indices at x0 .
z(·) : element of a function space; z(t) is the value
of z(·) at t.
x(t) : state variable, x ∈ Rd(x) .
u(t) : control variable, u ∈ Rd(u) .
= [t0 , tf ] : time interval.
w := (x, u) : state x and control u.
p = (x0 , xf ) : ∈ R2d(x) , if is a fixed time interval, x0 :=
x(t0 ), xf := x(tf ).
p := (t0 , x0 , tf , xf ) : ∈ R2+2d(x) , if is a variable time interval.
ψ(t) : adjoint variable, ψ ∈ Rd(x)∗ .
ẋ = f (t, x, u) : control system, where ẋ:= dx dt .
H (t, x, u, ψ) : = ψf (t, x, u), Pontryagin function or Hamil-
tonian
Hx , Hu : partial derivatives of Hwith respectto x and
u, e.g., Hx := ∂H ∂x = ∂x1 , . . . , ∂xn ∈ R ,
∂H ∂H n∗
n = d(x).
Hux : second partial derivative of H with respect to
x and u,
⎛ ⎞
∂2H 2H
. . . ∂u∂ 1 ∂x
⎜ ∂u1 ∂x1 n ⎟
Hux = ⎝ . . . ... . . . ⎠,
2
∂ H 2
∂um ∂x1 . . . ∂u∂mH ∂xn
m = d(u), n = d(x).
w 0 (t) = (x 0 (t), u0 (t)) : pair satisfying the constraints of an optimal
control problem.
δx(t) or x̄(t) : variation of the state x 0 (t).
δu(t) or ū(t) : variation of the control u0 (t).
n ∂2H
Hux x̄, ū : = m i=1 j =1 ∂ui ∂xj ūi x̄j , m = d(u), n =
d(x).
Notation xv
The book is devoted to the theory and application of second-order necessary and sufficient
optimality conditions in the Calculus of Variations and Optimal Control. The theory is
developed for control problems with ordinary differential equations subject to boundary
conditions of equality and inequality type and mixed control-state constraints of equality
type. The book exhibits two distinctive features: (a) necessary and sufficient conditions
are given in the form of no-gap conditions, and (b) the theory covers broken extremals,
where the control has finitely many points of discontinuity. Sufficient conditions for regular
controls that satisfy the strict Legendre condition can be checked either via the classical
Jacobi condition or through the existence of solutions to an associated Riccati equation.
Particular emphasis is given to the study of bang-bang control problems. Bang-bang
controls induce an optimization problem with respect to the switching times of the control.
It is shown that the classical second-order sufficient condition for the Induced Optimization
Problem (IOP), together with the so-called strict bang-bang property, ensures second-order
sufficient conditions (SSC) for the bang-bang control problem. Numerical examples in
different areas of application illustrate the verification of SSC for both regular controls and
bang-bang controls.
SSC are crucial for exploring the sensitivity analysis of parametric optimal control
problems. It is well known in the literature that for regular controls satisfying the strict
Legendre condition, SSC allow us to prove the parametric solution differentiability of opti-
mal solutions and to compute parametric sensitivity derivatives. This property has lead to
efficient real-time control techniques. Recently, similar results have been obtained for bang-
bang controls via SSC for the IOP. Though the discussion of sensitivity analysis and the
ensuing real-time control technques are an immediate consequence of the material presented
in this book, a systematic treatment of these issues is beyond the scope of this book.
The results of Sections 1.1–1.3 are due to Levitin, Milyutin, and Osmolovskii. The
results of Section 6.8 were obtained by Milyutin and Osmolovskii. The results of Sections
2.1–3.4, 5.1, 5.2, 6.1, 6.2, and 6.5 were obtained by Osmolovskii; some important ideas
used in these sections are due to Milyutin. The results of Sections 4.1 and 4.2 (except for
Section 4.1.5) are due to Lempio and Osmolovskii. The results of Sections 5.3, 6.3, 6.6,
and 7.1–7.5 were obtained by Maurer and Osmolovskii. All numerical examples in Sections
4.1, 5.4, 6.4, and Chapter 8 were collected and investigated by Maurer, who is grateful for
the numerical assistance provided by Christof Büskens, Laurenz Göllmann, Jang-Ho Robert
Kim, and Georg Vossen. Together we solved a lot more bang-bang and singular control
problems than could be included in this book. H. Maurer is indebted to Yalçin Kaya for
drawing his attention to the arc-parametrization method presented in Section 8.1.2.
xvii
xviii Preface
where the second derivatives Fxx , Fẋx , and Fẋ ẋ are calculated along the extremal x(·) that
is of interest to us. It is considered in the space W 1,2 of absolutely continuous functions
x̄ with square integrable derivative x̄. ˙ By using ω for a given extremal, one formulates
a necessary second-order condition for a weak minimum (the positive semidefiniteness of
the form), as well as a sufficient second-order condition for a weak minimum (the positive
definiteness of the form). As is well known, these quadratic conditions are equivalent (under
the strengthened Legendre condition) to the corresponding Jacobi conditions.
The simplest problem of the calculus of variations can also be considered in the
space W 1,∞ of Lipschitz-continuous functions x(·), and then, in particular, there arises
the problem of studying the extremality of broken extremals x(·), i.e., extremals such that
the derivative ẋ(·) has finitely many points of discontinuity of the first kind. What are
second-order conditions for broken extremals, and what is the corresponding quadratic
form for them? A detailed study of this problem for the simplest problem was performed by
the first author in the book [79], where quadratic necessary and sufficient conditions for the
so-called “Pontryagin minimum” (corresponding to L1 -small variations of the control u = ẋ
under the condition of their uniform L∞ -boundedness) were obtained, and also the relation
between the obtained conditions and the conditions for the strong minimum (and also the
so-called “-weak” and “bounded strong” minima) was established. For an extremal x(·)
with one break at a point t∗ , the corresponding form becomes (cf. [39, 85, 86, 107]):
tf
= a ξ̄ 2 + 2[Fx ]x̄av ξ̄ + Fxx x̄, x̄ + 2Fẋx x̄, x̄ ˙ x̄
˙ + Fẋ ẋ x̄, ˙ dt,
t0
where ξ̄ is a numerical parameter, x̄(·) is a function that can have a nonzero jump [x̄] :=
x̄(t∗ +) − x̄(t∗ −) at the point t∗ and is absolutely continuous on the semiopen intervals
1
2 Introduction
Here, [Fx ] and [ẋ] denote the jumps of the gradient Fx (t, x(t), ẋ(t)) and the derivative ẋ(t)
of the extremal at the point t∗ , respectively (e.g., [ẋ] = ẋ(t∗ +) − ẋ(t∗ −)); a is the derivative
in t of the function
Fẋ (t, x(t), ẋ(t))[ẋ] − F (t, x(t), ẋ(t∗ +)) + F (t, x(t), ẋ(t∗ −))
at the same point (its existence is proved); and x̄av is the average value of the left-hand and
the right-hand values of x̄ at t∗ , i.e., x̄av = 12 (x̄(t∗ −) + x̄(t∗ +)). The Weierstrass condition
(the minimum principle) implies the inequality a ≥ 0, which complements the well-known
Weierstrass–Erdmann conditions for broken extremals.
In [96] and [97] it was shown (see also [90]) that the problem of the “sign” of the
quadratic form can be studied by using methods analogous to the classical methods. The
Jacobi conditions and criteria formulated by using the corresponding Riccati equation are
extended to the case of a broken extremal. In all these conditions, a new aspect consists
only of the fact that the solutions of the corresponding differential equations should have
completely certain jumps at the point of break of the extremal. Moreover, it was shown in
[97] that, as in the classical case, the quadratic form reduces to a sum of squares solving,
by the corresponding Riccati equation satisfying (this is the difference from the classical
case), a definite jump condition at the point t∗ , which also gives a criterion for positive
definiteness of the form and, therefore, a sufficient extremality condition for a given
extremal.
In book [79], quadratic extremality conditions for discontinuous controls were also
presented in the following problem on a fixed time interval [t0 , tf ]:
On the other hand, this problem, even being referred to as the calculus of variations, is
sufficiently general and its statement is close to optimal control problems, especially owing
to the local relation g(t, x, u) = 0. In [79], it was shown how, by using quadratic conditions
for the general problem of calculus of variations, one can obtain quadratic (necessary and
sufficient) conditions in optimal control problems in which the controls enter linearly and the
constraint on the control is given in the form of a convex polyhedron under the assumption
that the optimal control is piecewise-constant and (outside the switching points) belongs
to vertices of the polyhedron (the so-called bang-bang control). To show this, in [79],
we first used the property that the set V of vertices of a polyhedron U can be given by a
nondegenerate relation g(u) = 0 on an open set Q consisting of disjoint open neighborhoods
of vertices. This allows us to write quadratic necessary conditions for bang-bang controls.
Further, in [79], it was shown that a sufficient minimality condition on V guarantees (when
the control enters linearly) the minimum on its convexification U = co V . In this way, the
quadratic sufficient conditions were obtained for bang-bang controls.
However, in [79], there is a substantial gap stemming from the fact that, to avoid
making the book too long, the authors decided to omit the proofs of quadratic conditions
for the general problem of the calculus of variations and restricted themselves to their
formulation and the presentation of proofs only for the simplest problem. Although the
latter gives the idea of the proofs in the general case, there are no formal proofs of quadratic
conditions for the general problem of the calculus of variations in [79]. Part I of the
present book is devoted to removing this gap. Therefore, Part I can be considered as a
necessary supplement to the book [79]. At the same time, the material contained in Part I
is independent and is a complete theory of quadratic conditions for smooth problems of
the calculus of variations and optimal control that are covered by the statement presented
above.
Part I is organized as follows. First, in Chapter 1, we present a fragment of the abstract
theory of higher-order conditions of Levitin, Milyutin, and Osmolovskii [54, 55, 56], more
precisely, a modification of this theory for smooth problems on the set of sequences
determining one or another concept of minimum. In this theory, by a higher order, we
mean a nonnegative functional determining a growth estimate of the objective function on
admissible sequences of variations from . The main result of the abstract theory is that for
a given class of problems, a given higher order γ , and a given set of sequences , one defines
a constant Cγ (by using the Lagrange function) such that Cγ ≥ 0 is a necessary minimality
condition (corresponding to ), and Cγ > 0 is a sufficient condition. The constant Cγ is
said to be basic. In each concrete class of problems (for given and γ ), there arises the
problem of “decoding” the basic constant. By decoding we mean the simplest method for
calculating its sign. We illustrate the decoding of the basic constant by two simple examples,
obtaining conditions of the order γ , where γ is a certain quadratic functional.
In Chapter 2, on the basis of the results of Chapter 1, we create a quadratic theory
of conditions for Pontryagin minimum in the general problem of the calculus of variations
without local mixed constraint g(t, x, u) = 0. We perform the decoding of the basic constant
for the set of the so-called “Pontryagin sequences” and a special higher order γ , which is
characteristic for extremals with finitely many discontinuities of the first kind of control. We
first estimate the basic constant from above, thus obtaining quadratic necessary conditions
for Pontryagin minimum, and then we estimate it from below, thus obtaining sufficient
conditions. After that (in Section 2.7), we establish the relation of the obtained sufficient
conditions for Pontryagin minimum with conditions for strong minimum.
4 Introduction
Vossen [67] satisfies the second-order test derived in Chapter 5 while existing sufficiency
results fail to hold.
In Chapter 6, we investigate the pure bang-bang case. We obtain second-order nec-
essary and sufficient optimality conditions for this case as a consequence of the conditions
obtained in Chapter 5. In the pure bang-bang case, the conditions amount to testing the
positive (semi)definiteness of a quadratic form on a finite-dimensional critical cone. Nev-
ertheless, the assumptions are appropriate for numerical verification only in some special
cases. Therefore, again we study various transformations of the quadratic form and the
critical cone which will be tailored to different types of control problems in practice. In
particular, by means of a solution to a linear matrix differential equation, the quadratic form
can be converted to perfect squares. We demonstrate by practical examples that the obtained
conditions can be verified numerically.
We also study second-order optimality conditions for time optimal control problems
with control appearing linearly. More specifically, we consider the special case of time
optimal bang-bang controls with a given initial and terminal state. We aim at showing that
an approach similar to the above-mentioned Riccati equation approach works as well for
such problems. Again, the test requires us to find a solution of a linear matrix differen-
tial equation which satisfies certain jump conditions at the switching points. We discuss
three numerical examples that illustrate the numerical procedure of verifying positive defi-
niteness of the corresponding quadratic forms. Finally, following [79], we study second-
order optimality conditions in a simple, but important, class of time optimal control problems
for linear systems with constant entries.
Second-order optimality conditions in bang-bang control problems have been derived
in the literature in two different forms. The first form was discussed above. The second
form belongs to Agrachev, Stefani, and Zezza [1], who first reduce the bang-bang control
problem to a finite-dimensional Induced Optimization Problem (IOP) and then show that
well-known sufficient optimality conditions for the induced problem supplemented by the
strict bang-bang property furnish sufficient conditions for the bang-bang control problem.
In Chapter 7, we establish the equivalence of both forms of sufficient conditions. The
proof of this equivalence make extensive use of explicit formulas for first- and second-
order derivatives of the trajectory with respect to variations of the optimization variable ζ
comprising the switching times, the free initial and final time, and the free initial state.
We formulate the IOP with optimization variable ζ which is associated with the bang-bang
control problem. We give formulas for the first- and second-order derivatives of trajectories
with respect to ζ which follow from elementary properties of ordinary differential equations
(ODEs). The formulas are used to establish the explicit relations between the multipliers of
Pontryagin’s minimum principle and the Lagrange multipliers, critical cones and quadratic
forms of the original and IOPs. In our opinion, the resulting formulas seem to have been
mostly unknown in the literature. These formulas provide the main technical tools to obtain
explicit representations of the second-order derivatives of the Lagrangian. The remark-
able fact to be noted here is that by using a suitable transformation, these derivatives are
seen to involve only first-order variations of the trajectory with respect to ζ . This prop-
erty facilitates considerably the numerical computation of the Hessian of the Lagrangian.
Thus, we arrive at a representation of the quadratic form associated with the Hessian of the
Lagrangian.
Finally, Chapter 8 is devoted to numerical methods for solving the IOP and testing
the second-order sufficient conditions in Theorem 7.10. After a brief survey on numerical
6 Introduction
methods for solving optimal control problems, we present in Section 8.1.2 the arc-
parametrization method for computing bang-bang controls [44, 45, 66] and its extension to
piecewise feedback controls [111, 112, 113]. Arc parametrization can be efficiently imple-
mented using the code NUDOCCCS developed by Büskens [13, 14]. Several numerical
examples illustrate the arc-parametrization method and the verification of second-order
conditions.
Chapter 1
Here, we present the general theory of higher-order conditions [78] which will be used in
what follows in obtaining quadratic optimality conditions in the canonical problem of the
calculus of variations. In Section 1.1, we formulate the main result of the general theory in
the smooth case. Section 1.2 is devoted to its proof. In Section 1.3, we present two simple
applications of the general theory.
Definition 1.1. We say that the minimum is attained at a point x0 on (or x0 is a point of
-minimum) if there is no sequence {δxn } ∈ such that for all n,
J (x0 + δxn ) − J (x0 ) < 0, fi (x0 + δxn ) ≤ 0 (i = 1, . . . , k), g(x0 + δxn ) = 0.
9
10 Chapter 1. Abstract Scheme for Obtaining Higher-Order Conditions
In a similar way, the strict minimum on at x0 is defined. We need only replace the strict
inequality J (x0 + δxn ) − J (x0 ) < 0 in the previous definition with nonstrict inequality and
require additionally that the sequence {δxn } contains nonzero terms.
Obviously, the minimum on 0 is a local minimum. If 0 ⊂ , then the minimum
on is not weaker than a local minimum. The inclusion 0 ⊂ holds iff contains a
zero sequence. This condition holds in all applications of the general theory.
In what follows, the point x0 is fixed, and, therefore, as a rule, it will be omitted
in the definitions and notation. By δ, we denote the set of variations δx ∈ X such that
x0 + δx ∈ . Note that 0 ∈ δ. We set f0 (x) = J (x) − J (x0 ) for x ∈ . Denote by S
the system consisting of the functionals f0 , f1 , . . . , fk and the operator g. The concepts
of minimum and strict minimum on are naturally extended to the system S. The con-
cepts introduced below can also be related to problem (1.1), as well as to the system S.
Sometimes, it is more convenient to speak about the system and not about the problem.
For δx ∈ δ, we set m(δx) = max0≤i≤k fi (x0 + δx). Let g be the set of sequences
{δxn } ∈ such that g(x0 + δxn ) = 0 for all sufficiently large n. Consider the condition
m ≥ 0 | g . By definition, this condition means that for any sequence {δxn } ∈ g , there
exists a number starting from which m(δxn ) ≥ 0. In what follows, such a notation will be
used without additional explanations. The following proposition follows from the defini-
tions directly.
The definition easily implies the uniqueness of the operator H . If contains the zero
sequence (and hence 0 ⊂ ), then H is the Frechét derivative of the operator h at the
point x0 , and the strict -differentiability implies the strict differentiability. In what fol-
lows, we set H = h
(x0 ). The function of two variables, (δx, x̄) −→ h(x0 + δx) + h
(x0 )x̄,
that maps δ × X into Z, is called a fine linear approximation of the operator h at the
point x0 on . These concepts are used for operators, as well as for functionals.
It is assumed that all functionals J , f1 , . . . , fk and the operator g in problem (1.1) are
-continuous at x0 . Introduce the set of active indices:
I = {i ∈ {0, 1, . . . , k} fi (x0 ) = 0}, where f0 (x) = J (x) − J (x0 ). (1.2)
Obviously, 0 ∈ I . It is assumed that the functionals fi , i ∈ I , and the operator g are strictly
-differentiable at x0 . Also, it is assumed that either g
(x0 )X = Y (in this case, we say
that for g at x0 on , the Lyusternik condition holds) or the image g
(x0 )X is closed in Y
and has a direct complement which is a closed subspace in Y . Precisely, these assumptions
define a smooth problem (smooth system) on at the point x0 . In this chapter we will
consider only this type of problem.
Let us explain that, in calculating this quantity, for each sequence {δxn } ∈ g that does not
vanish, we first calculate the limit inferior (lim inf ) of the ratio m(δxn )/γ (δxn ) as n → ∞,
and then we take the greatest lower bound of the limits inferior over the whole set of
sequences from g that do not vanish. An analogous notation will be used for other
functions and other sets of sequences. Proposition 1.2 implies the following assertion.
The condition Cγ (m, g ) ≥ 0 is called the γ -necessity on . It is easy to see that the
γ -necessity on is equivalent to the following condition: there are no ε > 0 and sequence
{δxn } ∈ + such that
It is convenient to compare the concept of γ -necessity in this form with the concept of
minimum on . Further, we set
σ
Cγ (σ , ) = inf lim inf .
γ
Propositions 1.3 and 1.4 imply the following assertion.
Proposition 1.6. Each of the two conditions Cγ (m, g ) > 0 and Cγ (σ , ) > 0 is sufficient
for the strict minimum on .
Theorem 1.7. (a) If g
(x0 )X = Y , then the inequality Cγ (0 , σ γ ) ≥ 0 is equivalent to the
inequality Cγ (m, g ) ≥ 0. If g
(x0 )X = Y , then 0 ≥ 0, and , therefore, Cγ (0 , σ γ ) ≥ 0.
(b) The inequality Cγ (0 , σ γ ) > 0 is always equivalent to the inequality Cγ (σ , ) > 0.
In the case where g
(x0 )X = Y , the following three inequalities are pairwise equivalent to
each other:
Cγ (0 , σ γ ) > 0, Cγ (σ , ) > 0, and Cγ (m, g ) > 0.
Proposition 1.9. Let g (x0 )X = Y . Then there exists a number a, 0 < a ≤ 1, such that
0 ≤ max{aL
0 , 0 },
L
(1.4)
1
L
0 ≤ max 0 , 0 . (1.5)
a
Proof. We first prove inequality (1.4). If 0 is empty, then 0 = −∞, and hence in-
equality (1.4) holds. Suppose that 0 is not empty. By Proposition
1.8, there exists
a, 0 < a ≤ 1, such that for any λ ∈0 , the inequality a ≤ αi holds. Moreover, the
condition αi +
y ∗
= 1 implies αi ≤ 1. Let λ = (α, y ∗ ) ∈ 0 . We set ν := αi .
Then λ̂ := λ/ν ∈ L0 , a ≤ ν ≤ 1. Therefore, for any δx ∈ δ,
Proposition 1.10. Let g (x0 )X = Y . Then there exists a number a, 0 < a ≤ 1, such that
1
0 , σ γ ) ≤ max{Cγ (0 , σ γ ), Cγ (0 , σ γ )}.
Cγ (L (1.7)
a
theory of higher order for smooth problems, and for the reader’s convenience they are
formulated in this section.
The compatibility criterion of a set of linear inequalities and equations has the following
form.
is compatible iff
m
∗
sup αi ξi + y , y < 0.
λ∈ i=1
(By definition, sup∅ = −∞.) Along with this lemma, in studying problems with
inequality-type constraints, an important role is played by the estimate of the distance to
the set of solutions of a set of linear inequalities and equations [41], which is presented
below.
Lemma 1.12 (Hoffman). There exists a constant C = C(l, A) with the following property:
if for certain ξ ∈ Rm and y ∈ Y the system
li , x + ξi ≤ 0 (i = 1, . . . , m), Ax + y = 0
x
≤ C max{ξ1 , . . . , ξm ,
y
}.
In the case where there is no equation Ax + y = 0, Lemma 1.12 holds with the
estimate
x
≤ C max{ξ1+ , . . . , ξm+ }.
Finally, we present the Lyusternik-type theorem on the estimate of the distance to the
level of the equality operator in the form which is convenient for us (see [78, Theorem 2]).
Let a set of sequences satisfy the same conditions as in Section 1.1. The following
theorem holds for an operator g : X → Y strictly -differentiable at a point x0 .
Theorem 1.13. Let g(x0 ) = 0 and let g
(x0 )X = Y . Then there exists C > 0 such that for
any {δxn } ∈ , there exists {x̄n } ∈ 0 satisfying the following conditions for all sufficiently
large n: g(x0 + δxn + x̄n ) = 0,
x̄n
≤ C
g(x0 + δxn )
.
1.2. Proof of the Main Theorem 17
Lemma 1.14 (Main Lemma). Let a sequence {δxn } and a sequence of numbers {ζn } be
such that δxn ∈ δ for all n, ζn+ → 0, and L 0 (δxn ) + ζn < 0 for all n. Then there exists a
sequence {x̄n } ∈ 0 such that the following conditions hold :
(1)
x̄n
≤ O(σ (δxn ) + ζn+ );
(2) fi (x0 + δxn + x̄n ) + ζn ≤ o(
x̄n
), i ∈ I ;
(3) g(x0 + δxn + x̄n ) = 0 for all sufficiently large n.
Proof. For an arbitrary n, let us consider the following set of conditions on x̄:
fi
(x0 ), x̄ + fi (x0 + δxn ) + ζn < 0, i ∈ I ; g
(x0 )x̄ + g(x0 + δxn ) = 0. (1.12)
Let (αi )i∈I and y ∗ be a tuple from the set of system (1.12). We set αi = 0 for i ∈ / I.
Then λ = (α0 , . . . , αk , y ∗ ) ∈ L ∗
0 . The converse is also true: if λ = (α0 , . . . , αk , y ) ∈ 0 ,
L
then αi = 0 for i ∈ / I and the tuple ((αi )i∈I , y ∗ ) belong to the set of system (1.12).
Therefore,
∗
max αi (fi (x0 + δxn ) + ζn ) + y , g(x0 + δxn )
I
k
∗
= max αi (fi (x0 + δxn ) + ζn ) + y , g(x0 + δxn )
L
0 i=0
= 0 (δxn ) + ζn < 0.
L
The latter relation is implied by the definition of the function L
0 and the normalization
αi = 1. According to Lemma 1.11, system (1.12) is compatible. Then by Hoffman’s
lemma (Lemma 1.12), there exist C > 0 and a sequence {x̄n
} such that for all n,
fi
(x0 ), x̄n
+ fi (x0 + δxn ) + ζn ≤ 0, i ∈ I ; (1.13)
g
(x0 )x̄n
+ g(x0 + δxn ) = 0; (1.14)
x̄n
≤ C max max{fi (x0 + δxn ) + ζn },
g(x0 + δxn )
. (1.15)
i∈I
We set {x̄n } = {x̄n
+ x̄n
}. Condition (1.17) implies g(x0 + δxn + x̄n ) = 0 for all sufficiently
large n, and conditions (1.16) and (1.18) imply
x̄n
≤ O(σ (δxn ) + ζn+ ). Further, we obtain
from conditions (1.13) and the property of strong -differentiability of the functionals fi
at the point x0 that
fi (x0 + δxn + x̄n ) + ζn
= fi (x0 + δxn ) + fi
(x0 ), x̄n
+ fi
(x0 ), x̄n
+ ζn + o(
x̄n
)
≤ fi
(x0 ), x̄n
+ o(
x̄n
) = o1 (
x̄n
), i ∈ I.
The latter relation holds because of (1.18). The lemma is proved.
Now, we prove a number of assertions from which the main result (Theorem 1.7) will
follow. Below we assume that the order γ is strict and higher on , and all assumptions of
Section 1.1 hold for the set of sequences and problem (1.1) at the point x0 .
0 , σ γ ) ≤ Cγ (m, g ). Indeed,
Proof. We first show that Cγ (L
Here, the first inequality is obvious, and the second inequality follows from the obvious
estimate L0 ≤ m | g . The equality is proved in the same way as relation (1.9).
Now let us prove inequality Cγ (m, g ) ≤ Cγ (L 0 , σ γ ), which will finish the proof
of the theorem. If Cγ (L 0 , σγ ) = +∞, the inequality holds. Let Cγ (L 0 , σ γ ) < +∞,
and let C be such that
L
Cγ (L
0 , σ γ ) := inf lim inf
0
< −C.
σ γ γ
Then there exists a sequence {δxn } ∈ +σ γ at which 0 (δxn ) + Cγ (δxn ) < 0, and, moreover,
L
δxn ∈ δ for all n. We set ζn = Cγ (δxn ). According to the main lemma, there exists a
sequence {x̄n } such that the following conditions hold:
(α)
x̄n
≤ O(σ (δxn ) + C + γ (δxn ));
(β) fi (x0 + δxn + x̄n ) + Cγ (δxn ) ≤ o(
x̄n
), i ∈ I ;
(γ ) g(x0 + δxn + x̄n ) = 0 for all sufficiently large n.
Since {δxn } ∈ σ γ , the first condition implies
x̄n
≤ O(γ (δxn )). We set {δxn
} = {δxn + x̄n }.
Then condition (γ ) implies {δxn
} ∈ g , and condition (β) implies
fi (x0 + δxn
) + Cγ (δxn ) ≤ o(γ (δxn )), i ∈ I.
From this we obtain
m(δxn
)
lim inf ≤ −C.
γ (δxn )
1.2. Proof of the Main Theorem 19
Therefore,
m(δxn
)
lim inf ≤ −C.
γ (δxn
)
Taking into account that {δxn
} ∈ g , we obtain from this that Cγ (m, g ) ≤ −C. There-
fore, we have proved that the inequality Cγ (L0 , σ γ ) < −C always implies the inequality
Cγ (m, g ) ≤ −C. Therefore, Cγ (m, g ) ≤ Cγ (L 0 , σ γ ). The theorem is completely
proved.
Theorem 1.16. Let g
(x0 )X = Y . Then the following three inequalities are pairwise equiv-
alent:
Cγ (m, g ) ≥ 0, Cγ (L 0 , σ γ ) ≥ 0, and Cγ (0 , σ γ ) ≥ 0.
Now, consider the sequence of relations
The first of these relations was proved in Section 1.1 (inequality (1.11)), and the other
relations are obvious. The following assertion follows from (1.19), Theorem 1.15, and
inequality (1.7).
Corollary 1.17. Let g
(x0 )X = Y . Then for 0 < a ≤ 1 the following inequalities hold :
1
Cγ (0 , σ γ ) ≤ Cγ (σ , ) ≤ Cγ (m, g )+ ≤ Cγ (L +
0 , σ γ ) ≤ Cγ (0 , σ γ )+ .
a
This implies the following theorem.
Theorem 1.18. Let g
(x0 )X = Y . Then the following four inequalities are pairwise equiv-
alent:
Therefore, in the case where the Lyusternik condition holds, we have proved all the
assertions of Theorem 1.7.
Proof. Since the image Y1 := g
(x0 )X is closed in Y , the condition Y1 = Y implies the
existence of y ∗ ∈ Y ,
y ∗
= 1 such that y ∗ , y = 0 for all y ∈ Y1 , and hence λ
= (0, y ∗ ) ∈ 0
and λ
= (0, −y ∗ ) ∈ 0 . From this we obtain that for any δx ∈ δ,
max (λ, δx) ≥ max{(λ
, δx), (λ
, δx)} = |y ∗ , g(x0 + δx) | ≥ 0.
0
Now let us prove assertion (b). Since the inequality Cγ (0 , σ γ ) ≤ Cγ (σ , ) always
holds by (1.11), in order to prove (b), we need to prove the following lemma.
Lemma 1.20. Let g
(x0 )X = Y . Then there exists a constant b = b(g
(x0 )) > 0 such that
Cγ (σ , ) ≤ b Cγ (0 , σ γ )+ . (1.20)
Proof. The proof uses a special method for passing from the system S = {f0 , . . . , fk , g} to a
certain auxiliary system Ŝ. We set Y1 = g
(x0 )X. According to the definition of the smooth
problem, Y = Y1 ⊕ Y2 , where Y2 is a closed subspace in Y . Then Y ∗ = W1 ⊕ W2 , where W1
and W2 are such that any functional from W1 is annihilated on Y2 , and any functional from
W2 is annihilated on Y1 . Without loss of generality, we assume that if y = y1 + y2 , y1 ∈ Y1 ,
and y2 ∈ Y2 , then
y
= max{
y1
,
y2
}. Then for y ∗ = y1∗ + y2∗ , y1∗ ∈ W1 , y2∗ ∈ W2 , we
have
y ∗
=
y1∗
+
y2∗
.
Let P1 : Y → Y1 and P2 : Y → Y2 be projections compatible with the decomposition
of Y into a direct sum Y = Y1 ⊕ Y2 . Then P1 + P2 = I , P1 P2 = 0, and P2 P1 = 0. We set
g1 = P1 g and g2 = P2 g. Then g = g1 + g2 , g1
(x0 )X = Y1 , and g2
(x0 )X = {0}. Introduce
the functional fg (x) =
g2 (x)
. The condition g2
(x0 )X = {0} implies that fg is strictly
-differentiable at the point x0 and fg
(x0 ) = 0. Consider the system Ŝ consisting of the
functionals f0 , . . . , fk , fg and the operator g1 . All subjects related to this system will be
endowed with the sign ∧. Since g = g1 + g2 , we have
g
= max{
g1
,
g2
}. Therefore,
σ (δx) := max{f0 (x0 + δx), . . . , fk (x + δx),
g(x0 + δx)
}
= max{f0 (x0 + δx), . . . , fk (x0 + δx), fg (x0 + δx),
g1 (x0 + δx)
} =: σ̂ (δx).
This implies
Cγ (σ , ) = Cγ (σ̂ , ). (1.21)
Further, since the Lyusternik condition g1 (x0 )X = Y1 holds for the system Ŝ, by Corollary
1.17, there exists â > 0 such that
1
Cγ (σ̂ , ) ≤ ˆ 0 , σ̂ γ )+ .
Cγ ( (1.22)
â
Now let us show that ˆ 0 is empty, then
ˆ 0 ≤ 0 . If ˆ 0 = −∞, and hence the inequality
ˆ
holds. Let 0 be nonempty, and let λ̂ = (α0 , . . . , αk , αg , y1∗ ) be an arbitrary element of the
ˆ 0 . Then
set
αi ≥ 0 (i = 0, . . . , k), αi fi (x0 ) = 0 (i = 1, . . . k), αg ≥ 0; y1∗ ∈ W1 ,
k k
αi + αg +
y1∗
= 1, αi fi
(x0 ) + αg fg
(x0 ) + y1∗ g1
(x0 ) = 0.
i=0 i=0
1.3. Simple Applications of the Abstract Scheme 21
y2∗
= αg , y2∗ , g2 (x0 + δx) = αg
g2 (x0 + δx)
.
We set y ∗ = y1∗ + y2∗ , and λ = (α0 , . . . , αk , y ∗ ). As is easily seen, then we have λ ∈ 0 and
ˆ λ̂, δx) = (λ, δx). Therefore, for arbitrary δx and λ̂ ∈
( ˆ 0 , there exists λ ∈ 0 such that
ˆ
the indicated relation holds. This implies 0 (δx) ≤ 0 (δx). Also, taking into account that
σ̂ = σ , we obtain
Cγ ( ˆ 0 , σ̂ γ ) ≤ Cγ (0 , σ γ ). (1.23)
It follows from (1.21)–(1.23) that
1
Cγ (σ , ) ≤ Cγ (0 , σ γ )+ . (1.24)
â
It remains to set b = 1/â. The lemma is proved.
Therefore, we have shown that in the case of violation of the Lyusternik condition, the
inequalities Cγ (σ , ) > 0 and Cγ (0 , σ γ ) > 0 are equivalent. Thus, we have completed
the proof of Theorem 1.7.
condition for a local minimum and is equivalent to γ -sufficiency. Below we transform the
expression for
0 (δxn )
Cγ (0 , σ γ ) := inf lim inf
{δxn }∈σ γ n→∞ γ (δxn )
k
(λ, δx) := αi fi (x0 + δx) + y ∗ , g(x0 + δx) = L(λ, x0 + δx),
i=0
0 (δx) := max (λ, δx),
λ∈0
If K = {0}, then as usual we set Cγ (0 , K) = +∞. Obviously, K = {x̄ ∈ X | σ
(x̄) = 0}.
We call K the critical cone.
Proof. Evidently,
0 (δx) = 0 (δx) + o(γ (δx)) as δx → 0, (1.26)
σ (δx) = σ
(δx) + O(γ (δx)) as δx → 0. (1.27)
We set
Then we set
0 (δxn )
σ
:= {δxn } ∈ 0 | σ
(δx) = 0 , Cγ (0 , σ
) := inf lim inf .
{δxn }∈σ
n→∞ γ (δxn )
Since, obviously, σ
⊂ σ
γ , we have Cγ (0 , σ
) ≥ Cγ (0 , σ
γ ). We show that, in
fact, equality holds. Suppose that Cγ (0 , σ
γ ) = +∞. Take any ε > 0. Let {δxn } ∈ σ
γ
be a nonvanishing sequence such that
0 (δxn )
lim ≤ Cγ (0 , σ
γ ) + ε.
n→∞ γ (δxn )
1.3. Simple Applications of the Abstract Scheme 23
By applying the Hoffman lemma (see Lemma 1.12) for each δxn (n = 1, 2, . . . ) to the system
f
(x0 ), δxn + x̄ ≤ 0, i ∈ I , g
(x0 )(δxn + x̄) = 0,
regarded as a system in the unknown x̄, and by bearing in mind that σ
(δxn ) ≤ O(γ (δxn )),
we obtain the following assertion: we can find an {x̄n } such that σ
(δxn + x̄n ) = 0 and
x̄n
≤ O(γ (δxn )). Consequently,
0 (δxn + x̄n ) 0 (δxn )
lim = lim ≤ Cγ (0 , σ
γ ) + ε,
n→∞ γ (δxn + x̄n ) n→∞ γ (δxn )
and {δxn + x̄n } ∈ σ
. This implies that Cγ (0 , σ
) ≤ Cγ (0 , σ
γ ). Consequently, the
equality Cγ (0 , σ
) = Cγ (0 , σ
γ ) holds, from which we also obtain Cγ (0 , σ γ ) =
Cγ (0 , σ
). But since 0 and γ are positively homogeneous of degree 2, by applying
the definition of the cone K, in an obvious way we obtain that, in turn, Cγ (0 , σ
) =
Cγ (0 , K). Thus, Cγ (0 , σ γ ) = Cγ (0 , K), and the theorem is proved.
Clearly, a local minimum in this space is weak. This is the minimum on the set of sequences
0 := {{δwn } |
δwn
→ 0 (n → ∞)}. Again we set = 0 .
We denote the argument of the functions J , Fi , and K by p = (x0 , xf ), where x0 ∈ Rn
and xf ∈ Rn . All relations containing measurable sets and functions are understood with
accuracy up to a set of measure zero. We assume the following.
Assumption 1.24. The functions J , Fi , and K are twice continuously differentiable with
respect to p; the function f is twice differentiable with respect to w; the function f and its
second derivative fww are uniformly bounded and equicontinuous with respect to w on any
bounded set of values (t, w) and are measurable in t for any fixed w.
Evidently, f satisfies these conditions if f and fww are continuous jointly in both
variables. Let w 0 (·) ∈ W be a trajectory satisfying all constraints that is being inves-
tigated for an optimal situation. We set p 0 = (x 0 (t0 ), x 0 (tf )), F0 (p) = J (p) − J (p 0 ),
I = {i ∈ {0, 1, . . . , k} | Fi (p0 ) = 0}. Obviously, problem (1.28)–(1.29) is smooth on 0
at the point w0 . The set 0 consists of aggregates λ = (α, β, ψ) for which the local form of
Pontryagin’s minimum principle holds:
ψ̇ = −Hx (t, w0 , ψ), ψ(t0 ) = −lx0 (p0 , α, β), ψ(tf ) = lxf (p0 , α, β), (1.32)
Hu (t, w 0 , ψ) = 0, (1.33)
where
H = ψf , l = αF + βK. (1.34)
The notation Rn∗ stands for the space of n-dimensional row vectors. We emphasize the
dependence of the Pontryagin function H and the endpoint Lagrange function l on λ that
is defined by (1.34) by writing H = H λ (t, w) and l = l λ (p). Under our assumptions 0 is
a finite-dimensional compact set, each point of which is uniquely determined by its projec-
tion (α, β).
1.3. Simple Applications of the Abstract Scheme 25
(Let us note that the sign of the constant Cγ (ω0 , K) will not change if we replace, in its
t
definition, the functional γ with the functional γ̄ (w̄) = x̄(t0 ), x̄(t0 ) + t0f ū(t), ū(t) dt.) We
define Condition ℵγ as in Section 1.3.1.
This result among others, is given, in [84]. But it is not difficult to derive it directly by
following the scheme indicated in our discussion of the finite-dimensional case. Because
the problem is smooth, Theorem 1.25 follows from Cγ (0 , σ γ ) = Cγ (ω0 , K), which is
established in the same way as in Section 1.3.1.
Notes on SSC for abstract optimization problems. Maurer and Zowe [75] consid-
ered optimization problems in Banach spaces with fully infinite-dimensional equality and
inequality constraints defined by cone constraints and derived SSC for quadratic function-
als γ . Maurer [62] showed that the SSC in [75] can be applied to optimal control problems
by taking into account the so-called “two-norm discrepancy.”
Chapter 2
In this chapter, on the basis of the results of Chapter 1, we create the quadratic theory of
conditions for a Pontryagin minimum in the general problem of the calculus of variations
without local mixed constraint g(t, x, u) = 0. Following [92], we perform the decoding
of the basic constant for the set of the so-called “Pontryagin sequences” and a special
higher order γ , which is characteristic for extremals with finitely many discontinuities of
the first kind of control. In Section 2.1, we formulate both necessary and sufficient quadratic
conditions for a Pontryagin minimum, which will be obtained as a result of the decoding.
In Sections 2.2 and 2.3, we perform some preparations for the decoding. In Section 2.4,
we estimate the basic constant from above, thus obtaining quadratic necessary conditions
for Pontryagin minimum, and in Section 2.5, we estimate it from below, thus obtaining
sufficient conditions. In Section 2.7, we establish the relation of the obtained sufficient
conditions for Pontryagin minimum with conditions for strong minimum.
where P ⊂ R2d(x) and Q ⊂ R1+d(x)+d(u) are open sets. By d(a) we denote the dimension
of vector a. Problem (2.1)–(2.4) also will be called the canonical problem. For the sake of
brevity, we set
x(t0 ) = x0 , x(tf ) = xf , (x0 , xf ) = p, (x, u) = w.
27
28 Chapter 2. Quadratic Conditions in the Calculus of Variations
We seek the minimum among pairs of functions w = (x, u) such that x(t) is an absolutely
continuous function on = [t0 , tf ] and u(t) is a bounded measurable function on . Recall
that by W 1,1 (, Rd(x) ) we denote the space of absolutely
t continuous functions x : →
Rd(x) , endowed with the norm
x
1,1 := |x(t0 )| + t0f |ẋ(t)| dt, and L∞ (, Rd(u) ) denotes
the space of bounded measurable functions u : → Rd(u) , endowed with the norm
u
∞ :=
ess supt∈[t0 ,tf ] |u(t)|. We set
Definition 2.1. We say that t∗ ∈ (t0 , tf ) is an L-point (or Lipschitz point) of a function
ϕ : [t0 , tf ] → Rn if at t∗ , there exist the left and right limit values
lim ϕ(t) = ϕ(t∗ −), lim ϕ(t) = ϕ(t∗ +)
t→t∗ t→t∗
t<t∗ t>t∗
on S if there is no sequence {δwn } ∈ S such that the following conditions hold for all its
members:
J (p 0 + δpn ) < J (p0 ), F (p0 + δpn ) ≤ 0, K(p 0 + δpn ) = 0,
ẋ 0 + δ ẋn = f (t, w0 + δwn ), (p0 + δpn ) ∈ P , (t, w0 + δwn ) ∈ Q,
where p 0 = (x 0 (t0 ), x 0 (tf )), δwn = (δxn , δun ), and δpn = (δxn (t0 ), δxn (tf )). In a similar
way, we define the strict minimum on S: it is necessary to only replace the strict inequality
J (p0 + δpn ) < J (p 0 ) by nonstrict in the previous definition and additionally assume that
the sequence {δwn } contains only nonzero members.
We can define any local (in the sense of a certain topology) minimum as a minimum
on the corresponding set of sequences. For example, a weak minimum is a minimum
on the set of sequences {δwn } in W such that
δxn
C +
δun
∞ → 0, where
x
C =
maxt∈[t0 ,tf ] |x(t)| is the norm in the space of continuous functions. Let 0 be the set of
sequences {δwn } in W such that
δwn
=
δxn
1,1 +
δun
∞ → 0. (We note that 0
corresponds to the set of sequences 0 introduced in Section 1.1.) It is easy to see that
a minimum on 0 is also a weak minimum. Therefore, we can define the same type of
minimum using various sets of sequences. We often use this property when choosing a set
of sequences for the type of minimum considered which is most convenient for studying.
In particular, this refers to the definition of Pontryagin minimum studied in this chapter.
We introduce an analogous notation for partial derivatives (except for the derivative with
respect to t) lxλ0 = ∂x
∂l
0
(p, α0 , α, β), Hxλ (t, w) = ∂H
∂x (t, w, ψ(t)), etc. Denote by 0 the set of
tuples λ such that
)
d(F
d(K)
α0 ≥ 0, α ≥ 0, αF (p ) = 0,0
α0 + αi + |βj | = 1, (2.5)
i=1 j =1
ψ̇ = −Hxλ (t, w 0 ), ψ(t0 ) = −lxλ0 (p0 ), ψ(tf ) = lxλf (p 0 ), Huλ (t, w0 ) = 0. (2.6)
Here, αi are components of the row vector α and βj are components of the row vector β.
If a point w0 yields a weak minimum, then 0 is nonempty. This was shown in [79, Part 1].
We set U(t, x) = {u ∈ Rd(u) | (t, x, u) ∈ Q}. Denote by M0 the set of tuples λ ∈ 0
such that for all t ∈ [t0 , tf ]\, the condition u ∈ U(t, x 0 (t)) implies the inequality
where L2 (, Rd(u) ) is the space of Lebesgue square integrable functions ū(t) : [t0 , tf ] →
Rd(u) . Let IF (w 0 ) = {i ∈ {1, . . . , d(F )} | Fi (p0 ) = 0} be the set of active subscripts of the
constraints Fi (p) ≤ 0 at the point w0 . Denote by K the set of z̄ = (ξ̄ , x̄, ū) ∈ Z2 () such
that
where p̄ = (x̄(t0 ), x̄(tf )), w̄ = (x̄, ū), and [ẋ 0 ]k is the jump of the function ẋ 0 (t) at the point
tk , i.e.,
[ẋ 0 ]k = ẋ 0k+ − ẋ 0k− = ẋ 0 (tk +) − ẋ 0 (tk −).
[H λ ]k = H λk+ − H λk− ,
where
H λk+ = H (tk , x 0 (tk ), u0k+ , ψ(tk )), H λk− = H (tk , x 0 (tk ), u0k− , ψ(tk )).
0 and any z̄ ∈ K:
Proposition 2.2. The following conditions hold for any λ ∈
Also, the following question is of interest: in which case can one of the inequalities
in the definition of K be omitted not changing K? For example, in which case can we omit
the inequality Jp (p0 )p̄ ≤ 0?
imply Jp (p0 )p̄ = 0, i.e., conditions (2.11) and (2.12) determine K as before.
where
Hxλk+ = ψ(tk )fx (tk , x 0 (tk ), u0k+ ), Hxλk− = ψ(tk )fx (tk , x 0 (tk ), u0k− ),
Hψk+ = f (tk , x 0 (tk ), u0k+ ) = ẋ 0k+ , Hψk− = f (tk , x 0 (tk ), u0k− ) = ẋ 0k− ,
ẋ 0k+ = ẋ(tk +), ẋ 0k− = ẋ(tk −),
and [Htλ ]k is the jump of the function Htλ = ψ(t)ft (t, x 0 (t), u0 (t)) at the point tk , i.e.,
[Htλ ]k = Htλk+ − Htλk− = ψ(tk )ft (tk , x 0 (tk ), u0k+ ) − ψ(tk )ft (tk , x 0 (tk ), u0k− ).
We note that D k (H λ ) depends on λ linearly, and D k (H λ ) ≥ 0 for any λ ∈ M0 and any tk ∈ .
Let [Hxλ ]k be the jump of the function Hxλ = ψ(t)fx (t, x 0 (t), u0 (t)) = −ψ̇(t) at the point
tk ∈ , i.e., [Hxλ ]k = Hxλk+ − Hxλk− . For λ ∈ M0 , z̄ ∈ Z2 (), we set (see also Henrion [39])
1 k λ 2
s
λ (z̄) = D (H )ξ̄k + 2[Hxλ ]k x̄av k
ξ̄k
2
k=1 tf (2.13)
1
+ lpp
λ
(p 0 )p̄, p̄ + Hww
λ
(t, w0 )w̄, w̄ dt ,
2 t0
Theorem 2.4. If w 0 is a Pontryagin minimum point, then M0 is nonempty and the function
0 (· ) is nonnegative on the cone K.
1 In the book [79], the value D k (H λ ) was defined by the formula D k (H λ ) = d ( H λ )(t ), since there
dt k k
the Pontryagin maximum principle was used.
2.1. Statements of Quadratic Conditions for a Pontryagin Minimum 33
We say that Condition A holds for the point w 0 if the set M0 is nonempty and the
function 0 is nonnegative on the cone K. According to Theorem 2.4, the condition A is
necessary for a Pontryagin minimum.
We say that Condition B holds for the point w0 if there exist a nonempty compact set
M ⊂ Leg+ (M0+ ) and a constant C > 0 such that
max λ (z̄) ≥ C γ̄ (z̄) ∀ z̄ ∈ K.
λ∈M
In conclusion, we note that the space P W 1,2 (, Rd(x) ) with the inner product
s tf
¯ = x̄(t0 ), x̄(t
x̄, x̄ ¯ 0 ) + ¯ k +
[x̄]k , [x̄] ˙¯
˙ x̄(t)
x̄(t), dt,
k=1 t0
respectively; moreover, here, p̄ = (x̄(t0 ), x̄(tf )). The verification of this assertion is also
very elementary.
We prove only that if ĝ
(w 0 )W = Y , then the image ĝ
(w 0 )W is closed in Y and has
a direct complement that is a closed subspace in Y . Note that the operator
w̄ = (x̄, ū) ∈ W −→ x̄˙ − fw (t, w0 )w̄ ∈ L1 (, Rd(x) ) (2.20)
is surjective. Indeed, for an arbitrary function v̄ ∈ L1 (, Rd(x) ), there exists a function x̄
satisfying x̄˙ − fx (t, w0 )x̄ = v̄ and x̄(t0 ) = 0. Then under the mapping given by this operator,
the pair w̄ = (x̄, ū) yields v̄ in the image. Further, we note that the operator
w̄ = (x̄, ū) ∈ W −→ Kp (p0 )p̄ ∈ Rd(K) (2.21)
is finite-dimensional. The surjectivity of operator (2.20) and the finite-dimensional prop-
erty of operator (2.21) imply the closedness of the image of operator (2.19). The following
more general assertion holds.
Lemma 2.6. Let X, Y , and Z be Banach spaces, let A : X → Y be a linear operator with
closed range, and let B : X → Z be a linear operator such that the image B(Ker A) of the
kernel of the operator A under the mapping of the operator B is closed in Z. Then the
operator T : X → Y × Z defined by the relation Tx = (Ax, Bx) for all x ∈ X has a closed
range.
Proof. Let a sequence {xn } in X be such that Txn = (Axn , Bxn ) → (y, z) ∈ Y × Z. Since
the range of AX is closed in Y , by the Banach theorem on the inverse operator there exists
a convergent subsequence xn
→ x in X such that Axn
= Axn → y and, therefore, Ax = y.
Then A(xn − xn
) = 0 for all n and B(xn − xn
) → z − Bx. By the closedness of the range
B(Ker A), there exists x
∈ X such that Ax
= 0 and Bx
= z − Bx. Then B(x + x
) = z
and A(x + x
) = y as required. The lemma is proved.
Since the image of any subspace under the mapping defined by a finite-dimensional
operator is a finite-dimensional and, therefore, closed subspace, Lemma 2.6 implies the
following assertion.
This implies that operator (2.19) has a closed range. Since the range of operator
(2.19) is of finite codimension, this range has a direct complement that is finite-dimensional,
and, therefore, it is a closed subspace.
Therefore, all the conditions defining a smooth problem on at the point w 0 hold
for the problem (2.1)–(2.4). Hence we can apply the main result of the abstract theory of
higher-order conditions, Theorem 1.7.
2.2.2 Set 0
In what follows, for the sake of brevity we make the following agreement: if a point at which
the derivative of a given function is taken is not indicated, then for the functions J , F , K,
36 Chapter 2. Quadratic Conditions in the Calculus of Variations
and l λ , p0 is such a point, and for the functions f and H λ , it is the point (t, w0 (t)). Using
the definition given by relations (1.3), we denote by 0 the set of tuples λ = (α0 , α, β, ψ)
such that
α0 ∈ R, α ∈ (Rd(F ) )∗ , β ∈ (Rd(K) )∗ , ψ(· ) ∈ L∞ (, (Rd(x) )∗ )
and the following conditions hold:
α0 ≥ 0, αi ≥ 0 (i = 1, . . . , d(F )), αi Fi (p0 ) = 0 (i = 1, . . . , d(F )), (2.22)
)
d(F
d(K)
α0 + αi + |βj | +
ψ
= 1, (2.23)
i=1 j =1
)
d(F tf
α0 Jp p̄ + αi Fip p̄ + βKp p̄ − ψ(x̄˙ − fw w̄) dt = 0 (2.24)
i=1 t0
for all x̄ ∈ W 1,1 . Let us show that this implies the conditions
ψ ∈ W 1,∞ , −ψ̇ = Hx , ψ(t0 ) = −lx0 , ψ(tf ) = lxf . (2.29)
Let ψ
satisfy the conditions ψ
∈ W 1,∞ , −ψ̇
= ψ
fx , and ψ
(t0 ) = −lx0 . Integrating by
parts, we obtain
tf tf tf
˙
tf
ψ x̄ dt = ψ x̄ |t0 − ψ̇ x̄ dt = lx0 x̄0 + ψ (tf )x̄f + ψ
fx x̄ dt
t0 t0 t0
for any x̄ ∈ W 1,1 . (Hereafter x̄0 := x̄(t0 ), x̄f := x̄(tf )). Therefore,
tf tf
˙
ψ x̄ dt − ψ
fx x̄ dt − lx0 x̄0 − ψ
(tf )x̄f = 0. (2.30)
t0 t0
and x̄(tf ) = ā. Then c̄ā + t0 ψ̄ v̄ dt = 0. Since this relation holds for an arbitrary ā ∈ R ,
v̄ ∈ L1 , we obtain that c̄ = 0 and ψ̄ = 0, i.e., ψ = ψ
and ψ
(tf ) = lxf . Therefore, conditions
(2.29) hold. Conversely, if conditions (2.29) hold for ψ, then applying the integration-
t
t
by-parts formula t0f ψ x̄˙ dt = ψ x̄ |tf0 − t0f ψ̇ x̄ dt, we obtain condition (2.28). From (2.28)
t
and the condition Hu = 0, condition (2.27), and therefore, condition (2.24) follow. There-
fore, we have shown that the set 0 defined in this section coincides with the set 0 of
Section 1.1.4.
We note that the set 0 is a finite-dimensional compact set, and the projection λ =
(α0 , α, β, ψ) −→ (α0 , α, β) is injective on 0 . Indeed, according to (α0 , α, β), the vector lxλ0
is uniquely defined, and from the conditions −ψ̇ = ψfx and ψ(t0 ) = lx0 the function ψ is
uniquely found. The same is also true for the set co 0 .
We set
δFi = Fi (p0 + δp) − Fi (p 0 ), i = 1, . . . , d(F ),
δK = K(p 0 + δp) − K(p 0 ) = K(p 0 + δp), δf = f (t, w 0 + δw) − f (t, w0 ).
Then
)
d(F tf
(λ, δw) = αi δFi + βδK − ψ(δ ẋ − δf ) dt
i=0 t0
tf tf
= δl λ − ψδ ẋ dt + δH λ dt,
t0 t0
where
)
d(F
lλ = αi Fi + βK, H λ = ψf , δl λ = l λ (p 0 + δp) − l λ (p 0 ),
i=0
δH λ = H (t, w 0 + δw, ψ) − H (t, w0 , ψ) = ψδf .
Note that in contrast to the classical calculus of variations, where δJ stands for the first
variation of the functional, we denote by δJ , δf , etc. the full increments corresponding to
the variation δw.
Let γ = γ (δw) : δW −→ R1
be an arbitrary strict higher order on whose definition is
given in Section 1.1.3. We set
σ γ = {{δwn } ∈ | σ (δwn ) ≤ O(γ (δwn ))},
where σ (δw) = max{F0 (p 0 + δp), . . . , Fd(F ) (p0 + δp), |δK|,
δ ẋ − δf
1 } is the violation
function of the problem (4.1)–(4.4). In what follows, we shall use the shorter notation lim
for the limit inferior instead of lim inf . We set
0
Cγ (0 , σ γ ) = inf lim .
σ γ γ
Theorem 1.7 implies the following theorem.
Further, there arises the problem of the choice of a higher order γ and decoding the
constant Cγ (0 , σ γ ) corresponding to the chosen order. The constant Cγ (0 , σ γ ) is
said to be basic, and by decoding of the basic constant, we mean the simplest method for
calculating its sign. In what follows, we will deal with the choice of γ and decoding of
the basic constant. As a result, we will obtain theorems on quadratic conditions for the
Pontryagin minimum formulated in Section 2.1.
2.3. Local Sequences, Representation of the Lagrange Function 39
Sequences from loc are said to be local. Obviously, 0 ⊂ loc ⊂ . Although the set
of local sequences loc is only a part of the set of Pontryagin sequences, all main
considerations in obtaining quadratic conditions for the Pontryagin minimum are related
namely to the set loc .
Let us consider the structure of local sequences. Denote by loc u the set of se-
quences {δun } in L∞ (, Rd(u) ) such that for any neighborhood V of the compact set u0
there exists a number starting from which condition (2.32) holds. We briefly write the
condition defining the sequences from loc u in the form (t, u + δun ) → u . Therefore,
0 0
= {{δwn } |
δxn
1,1 → 0, (t, u + δun ) → u }. In what follows, in order not to abuse
loc 0 0
the notation, we will omit the number n in sequences.
Let Qtu be the projection of Q under the mapping (t, x, u) → (t, u). Then Qtu is an open
set in Rd(u)+1 containing the compact set u0 . Denote by u0 (tk−1 , tk ) the closure in Rd(u)+1
of the intersection of the compact set u0 with the layer {(t, u) | u ∈ Rd(u) , t ∈ (tk−1 , tk )},
where k = 1, . . . , s + 1, and ts+1 = tf . In other words, u0 (tk−1 , tk ) is the closure of the graph
of the restriction of the function u0 (· ) to the interval (tk−1 , tk ). Obviously, u0 is the union
k
of disjoint compact sets u0 (tk−1 , tk ). For brevity, we set u0 (tk−1 , tk ) = u0 .
k
Let Vk ⊂ Qut be fixed disjoint bounded neighborhoods of the compact sets u0 ,
k = 1, . . . , s + 1. We set
s+1
V= Vk . (2.33)
k=1
6u '
∗ u1+ u(t)
V1−
V2
&
$
u(t) u1−
- V∗
1+
V1
%
t0 t1 tf t
-
While the superscripts k−, k+, and k were used for designation of left and right limit values
and ordinary values of functions at the point tk ∈ , the same subscripts will be used for
enumeration of sets and functions related to the point tk ∈ . The notation vraimaxt∈M u(t)
will be often used to denote the essential supremum (earlier denoted also by ess sup) of a
function u(·) on a set M.
Further, let {δu} ∈ loc ∗
u , i.e., (t, u + δu) → u , and let k ∈ I . For a sequence {δu},
0 0
∗ ∗
introduce the sequence of sets Mk− = {t | (t, u (t) + δu(t)) ∈ Vk− } and assume that χk−
0 ∗ is
∗ ∗ ∗
the characteristic function of the set Mk− . We set {δuk− } = {δuχk− }. Then
vraimax
∗
|u0 (t) + δu∗k− (t) − u0k+ | → 0, where u0k+ = u0 (tk +).
t∈Mk−
we define
∗ ∗
Mk+ = {t | (t, u0 (t) + δu(t)) ∈ Vk+ } and {δu∗k+ } = {δuχk+
∗
},
∗ is the characteristic function of the set M ∗ . Then (u0 +δu∗ ) | ∗ → u0k− , i.e.,
where χk+ k+ k+ Mk+
vraimax
∗
|u0 (t) + δu∗k+ (t) − u0k− | → 0, where u0k− = u0 (tk −).
t∈Mk+
The sequence {δu∗k− } belongs to the set of sequences ∗uk− defined as follows: ∗uk− consists
of sequences {δu} in L∞ (, Rd(u) ) such that
(a) vraimaxt∈M t ≤ tk , where M = {t | δu(t) = 0}, i.e., the support M of each member
δu of the sequence {δu} is located to the left from tk (here and in what follows, by the
support of a measurable function, we mean the set of points at which it is different from
zero); for brevity, we write this fact in the form M ≤ tk ;
(b) vraimaxt∈M |t − tk | → 0, i.e., the support M tends to tk ; for brevity, we write this
fact as M → tk ;
(c) vraimaxt∈M |u0 (t) + δu(t) − u0k+ | → 0, i.e., the values of the function u0 + δu on
the support M tend to u0k+ ; in short, we write this fact in the form
(u0 + δu) |M → u0k+ .
2.3. Local Sequences, Representation of the Lagrange Function 41
Analogously, we define the set ∗uk+ consisting of sequences {δu} such that
By definition, the sum of sets of sequences consists of all sums of sequences belonging to
these sets.
As before, let {δu} ∈ loc ∗ ∗ ∗ ∗
u . Let the sets Mk− , Mk+ and the functions χk− , χk+
correspond to the members of this sequence. We set
Mk∗ = Mk− ∗ ∪ M∗ , M∗ =
k+ Mk∗ ,
k∈I ∗
∗ ∗ ∗
χk = χk− + χk+ , χ = ∗ χk∗ ,
∗
k∈I
δu∗k = δuχk∗ = δu∗k− + δu∗k+ , δu∗ = δuχ ∗ = δu∗k .
k∈I ∗
We now introduce the set ∗ consisting of sequences {δw ∗ } = {(0, δu∗ )} in the space
W such that {δu∗ } ∈ ∗u (the component δx of these sequence vanishes identically).
Proposition 2.9. We have the relation loc = 0 + ∗ . Moreover, we can represent any
sequence {δw} = {(δx, δu)} ∈ loc in the form {δw} = {δw0 } + {δw∗ }, where {δw0 } =
{(δx, δu0 )} ∈ 0 , {δw∗ } = {(0, δu∗ )} ∈ ∗ , and , moreover, for all members of sequences
with the same numbers, the condition δu0 χ ∗ = 0 holds; here, χ ∗ is the characteristic
function of the set {t | δu∗ (t) = 0}.
The representation of a sequence from loc in the form of a sum of sequences from
0 and ∗ with the condition {δu0 χ ∗ } = {0}, which was indicated in Proposition 2.9, will
also be called canonical. Obviously, the canonical representation is unique. It will play an
important role in what follows.
We introduce one more notation related to an arbitrary sequence {δw} ∈ loc . For
such a sequence, we set
δvk− = (u0 + δu − u0k+ )χk−∗ = (u0 + δu∗ − u0k+ )χ ∗ ,
k− k−
δvk+ = (u + δu − u )χk+
0 0k− ∗ = (u0 + δu∗ − u0k− )χ ∗ ,
k+ k+
{δvk } = {δvk− } + {δvk+ }, {δv} = {δvk }.
k∈I ∗
Then the supports of δvk− , δvk+ , δvk , and δv are the sets Mk− ∗ , M ∗ , M ∗ , and M ∗ ,
k+ k
∗ ∗
respectively. Moreover, it is obvious that
δvk−
∞ → 0,
δvk+
∞ → 0,
δvk
∞ → 0, and
δv
∞ → 0, i.e., the sequences {δvk−∗ }, {δv ∗ }, {δv }, and {δv} belong to 0 .
k+ k u
fww (t, w 0 + δw ∗ )δw 0 , δw0 = fww (t, w0 )δw0 , δw0 + (δ ∗ fww )δw0 , δw0
= fww (t, w0 )δw0 , δw0 + (δ ∗ fxx )δx, δx .
Proposition 2.10. Let {δw} ∈ loc . Then the following formula holds for the canonical
representation {δw} = {δw 0 } + {δw∗ }:
1
δf = fw δw 0 + fww δw 0 , δw0 + δ ∗ f + (δ ∗ fx )δx + r̃,
2
where
tf
r̃
1 = o
δx
2C + |δu | dt ,
0 2
δ ∗ f = f (t, w 0 + δw ∗ ) − f (t, w0 ),
t0
Here, we have used the relation Huλ = 0 for all λ ∈ co 0 . As above, all derivatives
whose argument is not indicated are taken for w = w0 (t).
tf
2.3.3 Representation of the Integral t0 (δ ∗ Hxλ )δx dt on Local
Sequences
Proposition 2.11. Let two sequences {δx} and {δu∗ } such that
δx
C → 0 and {δu} ∈ ∗u
be given. Let λ ∈ co 0 . Then
tf tf
(δ ∗ Hxλ )δx dt = ∗
[Hxλ ]k δx(χk− ∗
− χk+ ) dt + ρ ∗λ , (2.40)
t0 k∈I ∗ t0
44 Chapter 2. Quadratic Conditions in the Calculus of Variations
where supλ∈co 0 |ρ ∗λ | = o(
δx
2C + I ∗ Mk∗ |δtk | dt), δtk = tk −t, [Hx ]
λ k = Hxλk+ −Hxλk− =
ψ(tk )(fxk+ − fxk− ) = ψ(tk )(fx (tk , x 0 (tk ), u0k+ ) − fx (tk , x 0 (tk ), u0k− )).
Proof. Since χ ∗ = χk∗ , we have δ ∗ Hxλ = δk∗ Hxλ , and, therefore,
tf tf
δ ∗ Hxλ δx dt = δk∗ Hxλ δx dt, δk∗ Hxλ = ψδk∗ f = ψ(f (t, w 0 + δwk∗ ) − f (t, w0 )).
t0 I∗ t0
Further, for ψ = ψ(t) we have ψ(t) = ψ(tk ) + k ψ, where k ψ = ψ(tk + δtk ) − ψ(tk ),
and, moreover, supco 0 |k ψ| ≤ const |δtk |, since −ψ̇ = ψfx , and supco 0
ψ
∞ < +∞.
Consider δk∗ fx . Since χk∗ = χk−
∗ + χ ∗ , we have δ ∗ f = δ ∗ f + δ ∗ f , where the in-
k+ k x k− x k+ x
crements δk− fx and δk+ fx correspond to the variations δu∗k− and δu∗k+ , respectively. For
∗ ∗
∗ f , we have
δk− x
∗ ∗ ∗ ∗
δk− fx = [fx ]k χk− + (δk− fx − [fx ]k )χk− ,
where
[fx ]k = fxk+ − fxk− , fxk− = fx (tk , x 0 (tk ), u0k− ), fxk+ = fx (tk , x 0 (tk ), u0k+ ).
∗ f −[f ]k )χ ∗ = (f (t, x 0 , u0 +δu∗ )−f k+ )χ ∗ −(f (t, x 0 , u0 )−
Further, let ηk− := (δk− x x k− x k− x k− x
k− ∗ ∗
fx )χk− . Then
ηk−
∞ → 0, since u + δuk− |Mk− → u
0 ∗ 0k+ and u |Mk− → u0k− . There-
0 ∗
∗ ∗ ∗
fore, δk− fx = [fx ] χk− + ηk− , ηk− χk− = ηk− ,
ηk−
∞ → 0. Therefore,
k
∗ Hλ
δk− = ∗ f = (ψ(t ) + ψ)([f ]k + η )χ ∗
ψδk−
x x k k x k− k−
= ∗ + ηλ = [H λ ]k χ ∗ + ηλ ,
ψ(tk )[fx ]k χk− k− x k− k−
λ
→ 0, ηλ χ ∗ = ηλ . This implies
where supco 0
ηk− ∞ k− k− k−
tf tf
∗ ∗
δk− Hxλ δx dt = [Hxλ ]k δxχk− dt + ρk−
λ
, (2.41)
t0 t0
λ =
tf λ δx dt. Let us estimate ρ λ . We have
where ρk− t0 ηk− k−
1
∗ ∗ 2
|ρk−
λ
| ≤ sup
ηk−
λ
∞
δx
C meas Mk− ≤ sup
ηk−
λ
∞
δx
2C + (meas Mk− ) .
co 0 2 co 0
∗ )2 ≤ 2
We use the obvious estimate (meas Mk− ∗ |δtk | dt.
Mk− Then
sup |ρk−
λ
| = o
δx
2C + |δtk | dt . (2.42)
co 0 ∗
Mk−
where
sup |ρk+
λ
|=o
δx
2C + |δtk | dt . (2.44)
co 0 ∗
Mk+
Taking into account that δk∗ Hxλ = δk−∗ H λ + δ ∗ H λ , we obtain from (2.41)–(2.43) that
x k+ x
tf tf
δk∗ Hxλ δx dt = ∗
[Hxλ ]k δx(χk− ∗
− χk+ ) dt + ρkλ ,
t0 t0
where supco 0 |ρkλ | = o
δx
2C + Mk∗ |δtk | dt . This and the relation δ ∗ Hxλ = δk∗ Hxλ im-
ply (2.40). The proposition is proved.
tf
2.3.4 Representation of the Integral t0 δ ∗ H λ dt on Local Sequences
t
We consider the term t0f δ ∗ H λ dt on an arbitrary sequence {δu∗ } ∈ ∗u . Since χ ∗ = χk∗ ,
we have δ ∗ H λ = δk∗ H λ , where δk∗ H λ = ψδk∗ f . In turn, δk∗ H λ = δk−
∗ H λ + δ ∗ H λ , which
k+
∗ ∗ ∗
corresponds to the representation δuk = δuk− + δuk+ . Consider the increment δk− ∗ Hλ =
∗ ∗ ∗
ψδk− f . By definition, δk− f = f (t, x , u + δuk− ) − f (t, x , u ), and, moreover, δk−
0 0 0 0 ∗ f =
∗ ∗
δk− f χk− . Recall that we have introduced the function
∗
δvk− = (u0 + δu∗k− − u0k+ )χk−
∗
.
where f k+ , ftk+ , fxk+ , fuk+ , and (f
)k+ are values of the function f and its derivatives
at the point (tk , x 0 (tk ), u0k+ ). Taking into account that k x 0 = ẋ 0k− δtk + o(δtk ), we obtain
from this that on Mk− ∗ ,
where f k− , ftk− , fxk− , and fuk− are the values of the function f and its derivatives at the
point (tk , x 0 (tk ), u0k− ). Subtracting (2.46) from (2.45), we obtain the following on Mk− :
δk∗ f = [f ]k (χk−
∗ ∗
− χk+ ) + fuk+ δvk− + fuk− δvk+ + O(|δtk | + |δvk |2 )χk∗ . (2.50)
ψ
ψ(t) = ψ(tk + δtk ) = ψ(tk ) + ψ̇ k− δtk + ηk− |δtk |
ψ (2.51)
= ψ(tk ) − ψ̇ k− |δtk | + ηk− |δtk |,
ψ
where ψ̇ k− = ψ̇(tk −) and supco 0 |ηk− | → 0 as |δtk | → 0, since ψ̇ = −Hxλ is left continuous
at the point tk . We obtain from (2.48) and (2.51) that
∗ ∗ ∗
δk− H λ = [H λ ]k χk− − [Htλ ]k + [Hxλ ]k ẋ 0k− + ψ̇ k− [Hψ ]k |δtk |χk−
1 λk+
+ Huu δvk− , δvk− + η̃k−
λ
(|δtk | + |δvk− |2 ), (2.52)
2
λ χ ∗ = η̃λ and sup
λ∈co 0
η̃k−
∞ → 0. Here, we have taken into account that
where η̃k− λ
k− k−
Hu = Hu = 0 for all λ ∈ co 0 and [f ]k = [Hψ ]k .
λk+ λk−
∗ Hλ =
We now turn to formula (2.49) and use it to obtain the expression for δk+
∗ ∗
ψδk+ f . On Mk+ , we have
ψ
ψ(t) = ψ(tk + δtk ) = ψ(tk ) + ψ̇ k+ δtk + ηk+ δtk
ψ (2.53)
= ψ(tk ) + ψ̇ k+ |δtk | + ηk+ |δtk |,
2.3. Local Sequences, Representation of the Lagrange Function 47
ψ
where ψ̇ k+ = ψ̇(tk +), and supco 0 |ηk+ | → 0. Analogously to formula (2.52), we obtain
from (2.49) and (2.53) that
∗ ∗ ∗
δk+ H λ = −[H λ ]k χk+ − [Htλ ]k + [Hxλ ]k ẋ 0k+ + ψ̇ k+ [Hψ ]k |δtk |χk+
1 λk−
+ Huu δvk+ , δvk+ + η̃k+
λ
(|δtk | + |δvk+ |2 ), (2.54)
2
λ χ ∗ = η̃λ and sup
λ∈co 0
η̃k+
∞ → 0.
where η̃k+ λ
k+ k+
A remarkable fact is that the coefficients of |δtk | in formulas (2.52) and (2.54)
coincide and are the derivative at the point t = tk of the function {k H λ }(t) introduced
in Section 2.1.6. Let us show this. Let k ∈ I ∗ and λ ∈ co 0 . Recall that by definition,
(k H λ )(t) = H (t, x 0 (t), u0k+ , ψ(t)) − H (t, x 0 (t), u0k− , ψ(t)) = ψ(t)(k f )(t),
where (k f )(t) = f (t, x 0 (t), u0k+ )−f (t, x 0 (t), u0k− ). In what follows, we will omit the su-
perscript λ of H . The conditions ψ̇(t) = −ψ(t)fx (t, x 0 (t), u0 (t)), ẋ 0 (t) = f (t, x 0 (t), u0 (t)),
and the property of the function u0 (t) implies the existence of the left and right derivatives
of the functions ψ(t) and x 0 (t) at the point tk , and, moreover, the left derivative is the left
limit of the derivatives, and the right derivative is the right limit of the derivatives. On
each of the intervals of the set [t0 , tf ]\, the derivatives ψ̇, ẋ 0 are continuous. Analogous
assertions hold for the function (k H )(t). Its left derivative at the point tk can be calculated
by the formula
d
d
(k H )(tk ) = H (t, x 0 (t), u0k+ , ψ(t)) − H (t, x 0 (t), u0k− , ψ(t))
dt − dt −
t=tk
= [Ht ]k + [Hx ]k ẋ 0k− + ψ̇ k− [Hψ ]k . (2.55)
∗ H . The right derivative
This derivative is the coefficient of |δtk | in expression (2.52) for δk−
of the function {H } (t) at the point tk can be calculated by the formula
k
d
d
(k H )(tk ) = H (t, x (t), u , ψ(t)) − H (t, x (t), u , ψ(t))
0 0k+ 0 0k−
dt + dt +
t=tk
= [Ht ]k + [Hx ]k ẋ 0k+ + ψ̇ k+ [Hψ ]k . (2.56)
∗ H . We show that
This derivative is the coefficient of |δtk | in expression (2.54) for δk+
d d
(k H )(tk ) = (k H )(tk ); (2.57)
dt − dt +
i.e., the function (k H )(tk ) is differentiable at the point tk . Indeed,
d
(k H )(tk ) = [Ht ]k + [Hx ]k ẋ 0k− + ψ̇ k− [Hψ ]k
dt −
= [Ht ]k + (Hxk+ − Hxk− )Hψk− − Hxk− (Hψk+ − Hψk− )
We denote the derivative of the function −(k H λ )(t) at the point tk by D k (H λ ). Therefore,
we have proved the following assertion.
Lemma 2.12. The function (k H )(t) is differentiable at each point tk ∈ . Its derivative
at this point can be calculated by the formulas
d
(k H )(tk ) = [Htλ ]k + Hxλk+ Hψλk− − Hxλk− Hψλk+
dt
= [Htλ ]k + [Hxλ ]k ẋ k− + ψ̇ k− [Hψλ ]k = [Htλ ]k + [Hxλ ]k ẋ k+ + ψ̇ k+ [Hψλ ]k .
tf
2.3.5 Representation of the Integral t0 δH λ dt on Local Sequences
Propositions 2.10, 2.11, and 2.13 imply the following assertion.
Proposition 2.14. Let {δw} ∈ loc be represented in the canonical form: {δw} = {δw 0 } +
{δw∗ }, where {δw 0 } = {(δx, δu0 )}, {δw∗ } = {(0, δu∗ )}, {δu0 χ ∗ } = {0}. Then the following
formula holds for any λ ∈ co 0 :
tf tf
1 tf λ
δH dt =
λ
Hx δx dt +
λ
Hww δw 0 , δw0 dt
t0 t0 2 t0
∗ ∗
+ [H ] (meas Mk− − meas Mk+ ) + D (H )
λ k k λ
|δtk | dt
k∈I ∗ Mk∗
tf
1
+ Huu
λk+
δvk− , δvk− + Huu
λk−
δvk+ , δvk+ dt
2 t0
tf
∗ ∗
+[Hxλ ]k δx(χk− − χk+ ) dt + ρH λ,
t0
λ = ελ (
2 ) dt + tf
k∈I ∗ Mk∗ (|δtk |+|δvk | |δu0 |2 dt +
δx
2C ), supλ∈co 0 |εH
where ρH λ | → 0.
H t0
For the sequence {δw} ∈ loc represented in the canonical form, we set
tf s
tf
γ (δw) =
δx
2C + |δu | dt + 2
0 2
|δtk | dt + |δv|2 dt. (2.60)
∗
t0 k=1 Mk t0
Proposition 2.15. For any sequence {δw} ∈ loc represented in the canonical form,
{δw} = {δw 0 } + {δw∗ }, {δw0 } = {(δx, δu0 )} ∈ 0 , {δw ∗ } = {(0, δu∗ )} ∈ ∗ , {δu0 χ ∗ } = {0},
and for any λ ∈ co 0 , we have the formula
where
1 λ 1 tf λ
(δw) = lpp δp, δp +
1λ
Hww δw 0 , δw0 dt
2 2 t0
s
∗ ∗
+ [H λ ]k (meas Mk− − meas Mk+ )
k=1
tf
∗ ∗
+D (H ) k λ
|δtk | dt + [Hxλ ]k δx(χk− − χk+ ) dt
∗
Mk t0
tf
1
+ Huu
λk+
δvk− , δvk− + Huu
λk−
δvk+ , δvk+ dt (2.61)
2 t0
∗ −
In expression (2.61) for 1λ (δw), all terms, except for sk=1 [H λ ]k (meas Mk−
meas Mk+ ∗ ), are estimated through γ on any sequence {δw} ∈ loc starting from a certain
Recall that by 0 we denote the set consisting of those λ ∈ 0 for which the condi-
tions [H λ ]k = 0 for all k ∈ I ∗ hold. Proposition 2.15 implies the following assertion.
Proposition 2.16. Let the set 0 be nonempty. Then there exists a constant C > 0 such
that the following estimate holds at any sequence {δw} ∈ loc represented in the canonical
form, starting from a certain number:
'
?
(t, v) = |v − u1+ |2 + 2|t − t1 | - u1+
u(t)
V2
&
$
u(t) u1−
- (t, v) > 0
6
6 V1
%
Definition 2.17. A function (t, u) : Qtu → R is said to be admissible (or an order function)
if it is continuous on Qtu and there exist disjoint neighborhoods Vk ⊂ Qtu of the compact
sets u0 (tk−1 , tk ) such that the following five conditions hold:
(1) (t, u) = |u − u0 (t)|2 if (t, u) ∈ Vk , t ∈ (tk−1 , tk ), k = 1, . . . , s + 1;
(2) (t, u) = 2|t − tk | + |u − u0k− |2 if (t, u) ∈ Vk , t > tk , k = 1, . . . , s;
(3) (t, u) = 2|t − tk | + |u − u0k+ |2 if (t, u) ∈ Vk+1 , t < tk , k = 1, . . . , s;
(4) (t, u) > 0 on Qtu \V, where V = s+1 k=1 Vk ;
(5) for any compact set F ⊂ Qtu \V, there exists a constant L > 0 such that |(t, u
) −
(t, u
)| ≤ L|u
− u
| if (t, u
) and (t, u
) belong to F .
Let us show that there exists at least one admissible function . Fix arbitrary disjoint
neighborhoods Vk ⊂ Qtu of the compact sets u0 (tk−1 , tk ) and define on V = Vk by
conditions (1)–(3). We set Vε = {(t, u) ∈ V | (t, u) < ε}. For a sufficiently small ε = ε0 > 0,
the set Vε0 is a neighborhood of u0 contained in V together with its closure. For the above ε0 ,
we set
(t, u) if (t, u) ∈ Vε0 ,
0 (t, u) =
ε0 if (t, u) ∈ Qtu \Vε0 .
Then the function 0 is admissible. An admissible function is not uniquely defined, but
any two of them coincide in a sufficiently small neighborhood of the compact set u0 .
52 Chapter 2. Quadratic Conditions in the Calculus of Variations
This functional is defined for pairs δw = (δx, δu) ∈ W such that (t, u0 + δu) ∈ Qtu a.e. on
[t0 , tf ]. Such pairs are said to be admissible with respect to Qtu (also, in this case, the
variation δu is said to be admissible with respect to Qtu ). It is easy to see that for any local
sequence {δw} ∈ loc , the values of γ (δw) can be calculated by formula (2.60) starting
from a certain number, and, therefore in the definition of γ and in formula (2.60), we have
used the same notation.
Let us verify that γ is a strict higher order on , where is the set of Pontryagin
sequences. Obviously, γ ≥ 0, γ (0) = 0, and for any variation δw admissible on Qtu , the
condition γ (δw) = 0 implies δw = 0.
Let us show that the functional γ is -continuous at zero. It is required to show
that γ (δw) → 0 for any Pontryagin sequence {δw}. Since the condition
δx
1,1 → 0 holds
for {δw} ∈ and
δx
C ≤
δx
1,1 , it suffices to show that for {δw} ∈ , the condition
tf
t0 (t, u + δu) dt → 0 holds. Let Uε (u ) be an ε-neighborhood of the set u in R
0 0 0 1+d(u) .
Assume that ε > 0 is chosen so that Uε (u0 ) ⊂ Qtu . Represent δu in the form δu = δuε + δuε ,
where
δu(t) if (t, u0 (t) + δu(t)) ∈ Uε (u0 ),
δuε (t) =
0 otherwise.
Therefore, for any fixed ε, we have meas Mε → 0. But then we can choose a subsequence
ε → +0 such that meas M ε → 0. (Recall that M ε is defined by a member of the sequence
{δu} and the corresponding member of the sequence {ε}; when defining Mε , we take the
members of the sequences {δu} and {ε} with the same numbers.) Fix such
a sequence
t
{ε}. Since
δuε
∞ ≤ O(1), we have
(t, u0 + δuε )
∞ ≤ O(1). Therefore, t0f (t, u0 +
δuε ) dt ≤
(t, u0 +δuε )
∞ meas Mε → 0. Moreover, the condition
ε → +0 implies {δuε } ∈
loc , and therefore,
(t, u 0 + δu )
→ 0, which implies tf (t, u0 + δu ) dt → 0. Then
u ε ∞ t0 ε
tf tf tf
(t, u0 + δu) dt = (t, u0 + δuε ) dt + (t, u0 + δuε ) dt → 0;
t0 t0 t0
this is what was required to be proved. Therefore, we have shown that the functional γ is
-continuous at zero. Therefore, γ is an order. Moreover, γ is a strict order.
Let us verify that γ is a higher order. Let {δw} ∈ , and let {w̄} ∈ 0 . We need to
show that γ (δw + w̄) = γ (δw) + o(
w̄
), where
w̄
=
x̄
1,1 +
ū
∞ . Since
δx + x̄
2C =
δx
2C + o(
x̄
C ) and
x̄
C ≤
x̄
1,1 , it suffices to show that
tf tf
(t, u0 + δu + ū) dt = (t, u0 + δu) dt + o(
ū
∞ ).
t0 t0
As above, represent {δu} in the form {δu} = {δuε } + {δuε }, where ε → 0, meas M ε → 0.
2.3. Local Sequences, Representation of the Lagrange Function 53
Then
tf
(t, u0 + δu + ū) dt = (t, u0 + δuε + ū) dt + (t, u0 + δuε + ū) dt
t0 Mε Mε
tf
= (t, u0 + δu) dt + ((t, u0 + δuε + ū)
t0 Mε
− (t, u0 + δuε )) dt
+ ((t, u0 + δuε + ū) − (t, u0 + δuε )) dt.
Mε
Here, we have used the relations
(t, u0 + δu) = (t, u0 + δu)(χε + χ ε ) = (t, u0 + δuε ) + (t, u0 + δuε ),
where χε and χ ε are the characteristic functions of the sets Mε and M ε , respectively. By
property (5) of Definition 2.17, we have
(t, u0 + δuε + ū) − (t, u0 + δuε ) dt ≤ const(meas Mε )
ū
∞ = o(
ū
∞ ).
Mε
Therefore, it suffices to show that
((t, u0 + δuε + ū) − (t, u0 + δuε )) dt = o(
ū
∞ ) (2.63)
Mε
or, which is the same,
tf
((t, u0 + δuε + ūε ) − (t, u0 + δuε )) dt = o(
ū
∞ ),
t0
Proposition 2.18. The following estimate holds for any {δu} ∈ loc u and {ū} ∈ u :
0
tf tf
(t, u0 + δu + ū) dt = (t, u0 + δu) dt + o(
ū
∞ ).
t0 t0
Proof. Represent {δu} in the canonical form {δu} = {δu0 } + {δu∗ }, {δu0 } ∈ 0u , {δu∗ } ∈ ∗u ,
|δu0 |· |δu∗ | = 0. The latter property holds for all members of the sequences {δu0 } and {δu∗ }
with the same numbers. According to the definition of the function (t, u), we have
tf tf
(t, u + δu) dt =
0
|δu | dt +
0 2
(2|δtk | + |δvk∗ |2 ) dt.
t0 t0 k Mk∗
= (t, u + δu) dt + o(
ū
∞ ).
0
t0
We will estimate Cγ from above. Since Cγ ≥ 0 is a necessary condition for the Pontryagin
minimum, the nonnegativity of any upper estimate for Cγ is also a necessary condition for
the Pontryagin minimum. Therefore, this stage of decoding can be considered as obtaining
a necessary condition for the Pontryagin minimum.
Let t
∈ (t0 , tf )\, ε > 0, and let [t
− ε, t
+ ε] be entirely contained in one of the
intervals of the set (t0 , tf )\. Let a point u
∈ Rd(u) be such that (t
, x 0 (t
), u
) ∈ Q, u
=
u0 (t
). Define the needle-shaped variation
u
− u0 (t), t ∈ [t
− ε, t
+ ε],
δu = δu (t; t , ε, u ) =
0 otherwise.
σ γ = {{δw} ∈
loc | σ ≤ O(γ )}.
loc
tf
(λ, δw loc ) + t0 δ
H λ dt
lim max ≥ Cγ ,
0 γ loc + γ
Proposition 2.20. Let ϕ(t, w) : Q → Rd(ϕ) be a continuous function. Let {δwloc } ∈ loc ,
{δw
} ∈
, {δw} = {δw loc + δw
}. Then δϕ = δ loc ϕ + δ
ϕ + rϕ , where
rϕ
1 = o(γ
),
rϕ
∞ → 0, δϕ = ϕ(t, w 0 + δw) − ϕ(t, w0 ), δ loc ϕ = ϕ(t, w0 + δwloc ) − ϕ(t, w 0 ), δ
ϕ =
ϕ(t, w0 + δw
) − ϕ(t, w0 ).
where
rϕ = (δ̄ 0 ϕ − δ 0 ϕ)χ
,
δ̄ 0 ϕ = ϕ(t, w 0 + δw
+ δw 0 ) − ϕ(t, w 0 + δw
),
δ0ϕ = ϕ(t, w 0 + δw 0 ) − ϕ(t, w0 ).
Therefore, δϕ = δ loc ϕ + δ
ϕ + rϕ . Since
δ̄ 0 ϕ
∞ → 0,
δ 0 ϕ
∞ → 0, meas M
= O(γ
), we
have
rϕ
∞ ≤
δ̄ 0 ϕ
∞ +
δ 0 ϕ
∞ → 0,
rϕ
1 ≤
rϕ
∞ meas M
= o(γ
). The proposition
is proved.
Proposition 2.21. Let {δwloc } ∈ loc , {δw
}
∈
, {δw} = {δwloc + δw
}. Then for any
λ ∈ 0 , we have (λ, δw) = (λ, δw loc ) + t0f δ
H λ dt + ρ λ , where sup0 |ρ λ | = o(γ
).
t
= γ loc + γ
+ o(γ
).
56 Chapter 2. Quadratic Conditions in the Calculus of Variations
Finally,
δ ẋ − δf
1 =
δ ẋ − δ loc f − δ
f − rf
1 =
δ ẋ − δ loc f
1 + O(γ
), since
δ
f
1 ≤
const meas M
= O(γ
) and
rf
1 = o(γ
). The proposition is proved.
Proof of Lemma 2.19. Let {δw} = {δw loc } + {δw
}, where {δw loc } ∈ loc
σ γ , {δw } ∈
and γ
≤ O(γ loc ). Then, according to Proposition 2.21, γ = γ loc + γ
+ o(γ
). However,
γ
≤ O(γ loc ). Therefore, γ ≤ O(γ loc ). On the other hand, since γ = γ loc + (1 + o(1))γ
,
γ
≥ 0, we have γ loc ≤ O(γ ). Therefore, γ and γ loc are of the same order of smallness.
Obviously, {δw} ∈ . Let us show that {δw} ∈ σ γ . Indeed, by Proposition 2.21,
δ ẋ − δf
1 =
δ ẋ − δ loc f
1 + O(γ
). Therefore,
σ (δw) = max{Fi (p 0 + δp), |δK|,
δ ẋ − δf
1 }
≤ σ (δw loc ) + O(γ
) ≤ O1 (γ loc ) ≤ O2 (γ ).
Thus, {δw} ∈ σ γ . Further, according to Proposition 2.21,
tf
(λ, δwloc ) + δ
H λ dt + ρ λ = (λ, δw),
t0
We can now use the results of Section 2.3. Lemma 2.19 and Proposition 2.15 imply
the following assertion.
Since
tf
fww δw 0 , δw0
1 ≤ O(
δx
C
2 + |δu0 |2 dt) ≤ O1 (γ ),
t0
1
(δ ∗ fx )δx
1 ≤ const
δx
C meas M∗ ≤ const(
δx
2C + (meas M∗ )2 ).
2
But, as was already mentioned, (meas Mk∗ )2 ≤ 4 M∗ |δtk | dt ≤ 2γ . Therefore,
(δ ∗ fx )δx
1 ≤
k
O(γ ). Further, δ ∗ f = k δk∗ f . According to formula (2.50), we have
δk∗ f = [f ]k (χk−
∗ ∗
− χk+ ) + fuk+ δvk− + fuk− δvk+ + O(|δtk | + |δvk |2 )χk∗ .
In what follows, in the formulation of Lemma 2.22, we narrow the set loc
σ γ up to its subset
defined by the following conditions:
(a) δv = 0;
(b) for any λ ∈ 0 ,
∗ ∗
[H λ ]k (meas Mk− − meas Mk+ ) ≤ 0. (2.70)
k
These conditions should hold for each member δw of the sequence {δw}. We denote by
σ γ the set of sequences {δw} ∈ σ γ satisfying these conditions.
loc 1 loc
∗
σ γ , only δw and Mk participate and,
We note that in the definitions of 2 , γ1 , and loc1 0
moreover, the variation δw is uniquely reconstructed by Mk∗ and δw0 by using the conditions
δu0 χ ∗ = 0, δv = 0. We denote the pairs (δw, M ∗ ) by b. Introduce the set of sequences of
pairs {b} = {(δw, M ∗ )} such that
{δw} = {(δx, δu)} ∈ 0 , M∗ = Mk∗ , Mk∗ → tk (k ∈ I ∗ ), δuχ ∗ = 0,
k∈I ∗
Fip δp ≤ O(γ1 ) (i ∈ I ), |Kp δp| ≤ O(γ1 ),
∗ ∗
δ ẋ − fw δw − [f ]k (χk− − χk+ ) ≤ O(γ1 ),
1
k
∗ ∗
[H ] (meas Mk− − meas Mk+ ) ≤ 0 ∀ λ ∈ 0 .
λ k
Then,
tf
2λ + t0 δ
H λ dt
lim max ≥ Cγ ,
0 γ1 + γ
where
1
2λ = 2λ (b) :=
(λ, δw)
2
tf
∗ ∗
+ k
D (H )λ
|δtk | dt + [Hx ]
λ k
δx(χk− − χk+ ) dt ,
k Mk∗ t0
Since |δx(t0 )| ≤
δx
C , we have γ2 ≤ γ1 . Let us show that the following estimate also holds
σ γ : γ1 ≤ const γ2 , where const > 0 is independent of the sequence.
on the sequence from loc1
For this purpose, it suffices to show that
δx
2C ≤ const γ2 on a sequence from loc1 σ γ . Let
us prove the following assertion.
Proposition 2.24. There exists a const > 0 such that for any sequence {b} = {δw, M ∗ }
satisfying the conditions
Since √
δu
1 ≤ ∗ − χ ∗
≤ meas M ∗ ,
tf − t0
δu
2 ,
χk−
k+ 1 k
∗ )2 + (meas M ∗ )2 ≤ 2
(meas Mk− k+ |δtk | dt ≤ γ2 , k ∈ I ∗ ,
Mk∗
we have
δx
2C ≤
δx
21,1 ≤ const
γ2 + o(γ1 ). This implies what was required. The propo-
sition is proved.
Lemma 2.25. The following inequality holds for any {b} ∈ loc1
σ γ and {δw } ∈ such that
γ ≤ O(γ2 ):
t
2λ (b) + t0f δ
H λ dt
lim max ≥ min{Cγ , Cγ }.
0 γ2 (b) + γ
∗ ∗
δ ẋ − fw δw − [f ]k (χk− − χk+ ) ≤ O(γ2 ). (2.75)
k 1
∗ − χ ∗ ) on δx in this condition? We
What is the influence of the terms k [f ]k (χk− k+
show below that the variations δx ∈ W (, Rd(x) ) can be replaced by variations
1,1
∗ − meas M ∗ , k ∈ I ∗ ,
x̄ ∈ P W 1,1 (, Rd(x) ) such that [x̄]k = [f ]k ξk , where ξk = meas Mk− k+
and, moreover, (2.75) passes to the condition
x̄˙ − fx x̄ − fu δu
1 ≤ O(γ2 ).
We will prove a slightly more general assertion, which will be used later in estimating Cγ
from below.
Therefore, we assume that there is a sequence {b} = {(δw, M ∗ )} such that {δw} ∈ 0 ,
M = ∪Mk∗ , Mk∗ → tk , k ∈ I ∗ . Moreover, let the following condition (which is weaker
∗
For each member b = (δx, δu, M∗ ) of the sequence {b}, let us define the functions δxk∗ and
x̄k∗ by the following conditions:
∗ − χ ∗ ), δx ∗ (t ) = 0, x̄˙ ∗ = 0,
δ ẋk∗ = [f ]k (χk− x̄k∗ (t0 ) = 0,
k+ k 0 k (2.77)
∗
[x̄k ] = [f ] ξk , ξk = meas Mk−
k k ∗ − meas M ∗ .
k+
2.4. Estimation of the Basic Constant from Above 61
Therefore, x̄k∗ is the jump function: x̄k∗ (t) = 0 if t < tk and x̄k∗ (t) = [f ]k ξk if t > tk , and,
moreover, the value of the jump is equal to [f ]k ξk . We set
∗ ∗ ∗ ∗
δx k = x̄k∗ − δxk∗ , k ∈ I ∗ , δx = δx k , x̄ = δx + δx = δx + (x̄k∗ − δxk∗ ).
k k
Note that x̄ ∈ P W 1,1 (, Rd(x) ). Since the functions x̄k∗ and δxk∗ coincide outside Mk∗ , we
∗ ∗ ∗ ∗ ∗ ∗
have δx k χk∗ = δx k for all k. Hence δx χ ∗ = δx . Let us estimate
δx
∞ and
δx
1 .
We have
∗
δx k
∞ ≤
x̄k∗
∞ +
δxk∗
∞ ≤ [f ]k · ξk + [f ]k meas Mk∗ .
Moreover,
1
2
ξk ≤ meas M ∗ ≤ 4 δtk dt ≤ 2γ2 .
k
Mk∗
∗ √ ∗ ∗ ∗ ∗
Hence
δx
∞ ≤ const γ2 . Since δx χ ∗
= δx , we have
δx
1 ≤
δx
∞ meas M∗ ≤
const γ2 . What equation does x̄ satisfy? We obtain from (2.76) and (2.77) that
∗
x̄˙ = fx δx + fu δu + r = fx x̄ + fu δu − fx δx + r.
∗
Since
δx
1 ≤ O(γ2 ), we have
x̄˙ − fx x̄ − fu δu
1 ≤ O(γ2 ) +
r
1 . Note that the replace-
ment of δx by x̄ does not influence the value of γ2 , since x̄(t0 ) = δx(t0 ) and M ∗ and δu
are preserved. Now let us show that
tf
∗ ∗
δx(χk− − χk+ ) dt = x̄av
k
ξk + o(γ2 ), (2.78)
t0
∗ − meas M ∗
where ξk = meas Mk− k = 1 (x̄ k− + x̄ k+ ) = 1 (x̄(t −) + x̄(t +)). Recall
and x̄av
k+ 2 2 k k
that δx satisfies equation (2.76). Represent δx in the form δx = δx 0 + δx ∗ + x , where
r
ẋr = r, xr (t0 ) = 0, and δx ∗ = δxk∗ . Then δ ẋ 0 = fx δx + fu δu and δx 0 (t0 ) = δx(t0 ). This
and the conditions
δx
C →
tf 0,
δu
∞ → 0 imply
δ ẋ
∞ → 0.
0
∗ ∗
Now let us consider t0 δx(χk− − χk+ ) dt. We set δxr0 := δx 0 + xr . Since
δx = (δx − δx ∗ ) + δx ∗ = δx 0 + xr + δx ∗ = δxr0 + δx ∗ ,
we have
tf tf tf
∗ ∗ ∗ ∗
δx(χk− − χk+ ) dt = δxr0 (χk− − χk+ ) dt + δx ∗ (χk−
∗ ∗
− χk+ ) dt. (2.79)
t0 t0 t0
since
δ ẋ 0
∞ → 0. Relation (2.80) follows from (2.81) and (2.82). Further, the conditions
δ ẋr0 = fx δx + fu δu + r, x̄˙ = fx δx + fu δu + r,
δxr (t0 ) = δx(t0 ),
0 x̄(t0 ) = δx(t0 )
imply δxr0 = x̄ − k x̄k∗ outside , and hence δxr0 (tk ) = x̄ k− − j <k [x̄]j . We obtain from
this and (2.80) that
tf
∗ ∗
δxr0 (χk− − χk+ ) dt = x̄ k− − [x̄]j ξk + o(γ2 ). (2.83)
t0 j <k
as required.
Finally, let us show that
tf tf
Hww
λ
w̄, w̄ dt = Hww
λ
δw, δw dt + ρ λ , (2.85)
t0 t0
∗ ∗
where w̄ = (x̄, ū), ū = δu, and sup0 ρ λ = o(γ2 ). Indeed, x̄ = δx + δx , where
δx
∞ →
∗
0,
δx
1 = O(γ2 ). Hence
tf tf
Hww
λ
w̄, w̄ dt = Hxx
λ
x̄, x̄ + 2Hux
λ
x̄, ū + Huu
λ
ū, ū dt
t0 t0
tf
= Hxx
λ
δx, δx + 2Hux
λ
δx, δu + Huu
λ
δu, δu dt
t0
tf
∗ ∗ ∗ ∗
+ λ
2Hxx δx, δx + 2Hux
λ
δx , δu + Hxx
λ
δx , δx dt
t0
tf
= Hww
λ
δw, δw dt + ρ λ , supρ λ = o(γ2 ),
t0 0
since
t
f λ ∗ ∗
Hxx δx, δx dt ≤ sup
Hxx
λ
∞
δx
C
δx
1 = o(γ2 ),
t0 0
t
f λ ∗ ∗
Hux δx , δu dt ≤ sup
Hux
λ
∞
δu
∞
δx
1 = o(γ2 ),
t0 0
t
f λ ∗ ∗ ∗ ∗
Hxx δx , δx dt ≤ sup
Hxx
λ
∞
δx
∞
δx
1 = o(γ2 ).
t0 0
Since p̄ = δp for the entire sequence {b}, we obtain from (2.78) and (2.85) that 2λ (b) =
3λ (b̄) + ρ λ , where sup0 |ρ λ | = o(γ2 ). We have proved the following assertion.
Lemma 2.26. Let a sequence {b} = {(δw, M∗ )} be such that {δw} = {(δx, δu)} ∈ 0 ,
M∗ = ∪Mk∗ , Mk∗ → tk , k ∈ I ∗ , and , moreover, let
∗ ∗
δ ẋ = fx δx + fu δu + [f ]k (χk− − χk+ ) + r,
k
√
where
r
1 = o( γ2 ). Let a sequence {b̄} = {(w̄, M∗ )} be such that w̄ = (x̄, ū), x̄ =
∗ ∗
δx + δx , and ū = δu, where δx = x̄ ∗ − δx ∗ = x̄k∗ − δxk∗ , and x̄k∗ and δxk∗ are defined
by formulas (2.77). Let δp = (δx(t0 ), δx(tf )) and p̄ = (x̄(t0 ), x̄(tf )). Then
∗
{δp} = {p̄}, γ2 (b) = γ2 (b̄), x̄˙ = fx x̄ + fu ū − fx δx + r,
[x̄]k = [f ]k ξk ∀ k, ξk = meas Mk− ∗ − meas M ∗ ;
δx ∗
≤ O(γ );
k+ 1 2
2λ (b) = 3λ (b̄) + ρ λ , sup0 |ρ λ | = o(γ2 ),
We will need this lemma in estimating Cγ from below. We now use the corollary of
Lemma 2.26, which is formulated below.
Proof. Indeed, by the condition of Corollary 2.27, it follows that x̄˙ = fx x̄ + fu ū + r̃,
∗ ∗
r̃
1 ≤ O(γ2 ). Then x̄˙ = fx x̄ + fu ū − fx δx + r, where r = r̃ + fx δx , and, moreover,
∗
r
1 ≤
r̃
1 +
fx δx
1 ≤ O(γ2 ). We obtain from this that
∗ ∗
δ ẋ = fx δx + fu δu + [f ]k (χk− − χk+ ) + r.
k
∗ −χ ∗ )
≤ O(γ ). The other assertions of Corol-
Consequently,
δ ẋ −fw δw − k [f ]k (χk− k+ 1 2
lary 2.27 follow from Lemma 2.26 directly.
2.4. Estimation of the Basic Constant from Above 65
Then Lemma 2.25 and also Corollary 2.27 imply the following assertion.
Lemma 2.28. The following inequality holds for any sequences {b̄} ∈ S 2 and {δw
} ∈
such that γ
≤ O(γ2 (b̄)):
t
3λ (b̄) + t0f δ
H λ dt
lim max ≥ min{Cγ , Cγ }.
0 γ2 (b̄) + γ
¯ = {(x̄, ū,
where sup0 |ρiλ | = o(γ2 ), i = 1, 2. We set {b̄} ¯ M∗ )}. Then
¯ = γ (b̄) + o(γ ),
γ2 (b̄) ¯ = 3λ (b̄) + ρ λ ,
3λ (b̄) sup |ρ λ | = o(γ2 ). (2.87)
2 2
0
¯ ∈ S 2 . Indeed,
Moreover, it is easy to see that {b̄}
¯ 1
x̄˙ − fx x̄ − fu ū
=
x̄˙ − fx x̄ − fu ū − fu ū∗
1
≤
x̄˙ − fw w̄
1 +
fu ū∗
1 ≤ O(γ2 ).
Further, we narrow the set of sequences S 3 up to the set S 4 by adding the following
conditions to the definition of the set S 3 :
(i) Each set Mk∗ is a segment adjusting to tk , i.e., Mk∗ = [tk − ε, tk ] or Mk∗ = [tk , tk + ε],
where ε → +0. In this case (see formula (2.13)),
2 |δtk | dt = ξk2 , where ξk = meas Mk− ∗ − meas M ∗
k+
Mk∗
tf
γ2 (b̄) = ξk2 + |x̄(t0 )|2 + |ū|2 dt = γ̄ ;
t0
1
1
3λ (b̄) = (λ, w̄) + D k (H λ )ξk2 + [Hxλ ]k x̄av
k
ξk =: λ .
2 2
k
max{0, ξk }. Moreover, we note that and γ̄ depend on ξ̄ , x̄, and ū. Therefore, S 4 can be
identified with the set of sequences {z̄} = {(ξ̄ , w̄)} such that
Lemma 2.29. The following inequality holds for any sequences {z̄} ∈ S 4 and {δw
} ∈
such that γ̄ (z̄) ≤ O(γ
):
t
λ (z̄) + t0f δ
H λ dt
lim max ≥ min{Cγ , Cγ }.
0 γ̄ (z̄) + γ
Now let us show that the condition k [H λ ]k ξ̄k ≤ 0 for all λ ∈ 0 holds automatically
for the elements of the critical cone K, and therefore, it is extra in the definition of the set of
sequences S 4 , i.e., it can be removed. We thus will prove that S 4 consists of the sequences
of elements of the critical cone K that satisfy the condition |ξ̄ | +
x̄
∞ +
ū
∞ → 0.
Proposition 2.31. Let λ = (α0 , α, β, ψ) ∈ 0 , and let (ξ̄ , x̄, ū) ∈ K. Then
[H λ ]k ξ̄k = lpλ p̄ ≤ 0.
k
z̄ ∈ K, ∗
λ ∈ 0 , [H ] = 0 for all k ∈ I imply
Proof. By Proposition 2.31, the conditions λ k
lp p̄ = 0, where lp p̄ = α0 (Jp p̄) + αi (Fip p̄) + βj (Kjp p̄). This and the conditions
λ
α0 ≥ 0, Jp p̄ ≤ 0, αi ≥ 0, Fip p̄ ≤ 0 ∀ i ∈ IF (w 0 ),
αi = 0 ∀ i ∈
/ I := IF (w 0 ) ∪ {0}, Kp p̄ = 0
imply what was required. The proposition is proved.
follows from the definition of the sequence {δw
} ∈
that γ
> 0 on it, and, therefore, this
definition is correct. Let A0 be the set of all limit points of the sequences {l(λ)} obtained by
the above method from all sequences {δw
} ∈
. Clearly, A0 is a closed subset in the finite-
dimensional space of linear functions l(λ) : co 0 → R1 . We note that each convergent
sequence {l(λ)} converges to its limit uniformly on co 0 . Further, to each number C, we
associate the set of affine functions AC obtained from A0 by means of the shift by (−C):
AC = {a(λ) = l(λ) − C l(· ) ∈ A0 }.
Denote by C := con AC the cone spanned by the set AC .
Fix an arbitrary element z̄ ∈ KZ , z̄ = 0 (recall that KZ := K ∩ Z()).
for any pair of sequences {ε}, {δw } such that ε → +0, {δw } ∈ , γ ≤ O(ε 2 ). Then
Proof. Let a(· ) ∈ C , i.e., a(λ) = ρ(l(λ) − C), where ρ > 0, l(· ) ∈ A0 . The latter means
that there exists {δw
} ∈
such that
tf
λ
t0 δ H dt
→ l(λ) ∀ λ ∈ co 0 . (2.92)
γ
To the sequence {δw
}, we write the sequence {ε} = {ε(δw
)} of positive numbers such that
tf
γ
ε 2 = , where γ
= (t, u0 + δu
) dt. (2.93)
ρ t0
Then γ
= O(ε2 ), and, therefore, inequality (2.90) holds. This inequality and conditions
(2.92) and (2.93) easily imply
λ (z̄)
ρ + l(λ)
max γ̄ (z̄)
≥ C.
ρ +1
0
Clearly, the maximum over 0 in this inequality can be replaced by the maximum over
co 0 , since λ and l(λ) are linear in λ. Therefore, multiplying the inequality by γ̄ (z̄) + ρ,
we obtain
max(λ (z̄) + ρl(λ)) ≥ C(γ̄ (z̄)) + ρ).
co 0
It remains to recall that a(· ) is an arbitrary element of C , and hence (2.94) implies (2.91).
The proposition is proved.
Lemma 2.35. Let Cγ > −∞. Then for any C ≤ min{Cγ , Cγ } and for any z̄ ∈ KZ , we
have the inequality
inf max {λ (z̄) + a(λ)} ≥ C γ̄ (z̄). (2.95)
a∈C λ∈co 0
In what follows, we will need the convexity property of the cone C . It is implied by
the following assertion.
where
= (δw
, t). This and (2.97) imply (2.96). Therefore, the convexity of A0 will be
proved if we ensure the existence of a sequence {α(t)} satisfying conditions (ii) and (iii).
The existence of such a sequence is implied by the Blackwell lemma, which is well known
in optimal control theory and is contiguous to the Lyapunov theorem on the convexity of
the range of a vector-valued measure. However, we note that we need not satisfy conditions
(ii) and (iii) exactly: it suffice to do this with an arbitrary accuracy; i.e., given an arbitrary
sequence ε → +0 in advance, we need to ensure the fulfillment of conditions (ii) and (iii)
with accuracy up to ε for each serial number of the sequence. Also, in this case, condition
(2.96) certainly holds. With such a weakening of conditions (ii) and (iii), we can refer to
Theorem 16.1 in [79, Part 2]. The proposition is proved.
in the left-hand side of inequality (2.91). Denote by A the set of all affine functions
a(λ) : co 0 → R1 . As was already noted, A is a finite-dimensional space. In this space,
we consider the sublinear functional
F : a(· ) ∈ A → max a(λ).
λ∈co 0
Since co 0 is a convex compact set, we can identify it with the set of support functionals
of F ; more precisely, there is a one-to-one correspondence between each support functional
a∗ ∈ ∂F and the element λ ∈ co 0 such that a∗ , a = a(λ) for all a ∈ A. Moreover,
according to this formula, a certain support functional a∗ ∈ ∂F corresponds to every element
λ ∈ co 0 .
Further, let C be a number such that the cone C defined above is nonempty. Then
what was said above implies that ∂F ∩ ∗C can be identified with the set
def
M(C ; co 0 ) = λ ∈ co 0 a(λ) ≥ 0 ∀ a ∈ C .
72 Chapter 2. Quadratic Conditions in the Calculus of Variations
In other words, for λ = λ̂, inequality (2.99) holds for arbitrary u
and t
satisfying conditions
(2.101). This implies λ̂ ∈ M co (C). The proposition is proved.
(In fact, we have the relation M(C ; co 0 ) = M co (C), but we need only the inclu-
sion for decoding the constant Cγ .) We obtain from relation (2.98) and inclusion (2.100)
that
inf max {λ (z̄) + a(λ)} ≤ max λ (z̄).
a∈C λ∈co 0 λ∈M co (C)
Proposition 2.40. The closure of the cone KZ in the space Z2 () coincides with the
cone K.
The proof of this proposition uses the Hoffman lemma on the estimation of the distance
to the solution set of a system of linear inequalities, i.e., Lemma 1.12. More precisely, we
use the following consequence of the Hoffman lemma.
Indeed, system (2.103) is compatible, since it admits the solution x̄ = −x0 , and then
by Lemma 1.12, there exists a solution satisfying estimate (2.104) (clearly, the surjectivity
condition for the operator A : X → Y in Lemma 1.12 can be replaced by the closedness of
the range of this operator considering the image AX as Y ).
is the space of absolutely continuous functions x(t) : [t0 , tf ] → Rd(x) having the Lebesgue
square integrable derivatives; it is endowed with the norm
tf 1/2
x
1,2 = x(t0 ), x(t0 ) + ẋ, ẋ dt .
t0
Then
x̄ N
1,2 → 0, and hence
x̄ N
C → 0. We set z̄N = (0, x̄ N , ūN ) and z̄N = z̄ − z̄N .
Then
z − z̄N
Z2 () =
z̄N
Z2 () → 0. The conditions z̄ ∈ K and
x̄ N
C → 0 imply
(Fip p̄N )+ + |Kp p̄N | → 0, (2.105)
i∈I
where p̄N corresponds to the sequence z̄N . Moreover, {z̄N } belongs to the subspace T2 ⊂
Z2 () defined by the conditions
Applying Lemma 2.41 on T2 , we obtain that for a sequence {z̄N } in T2 satisfying condition
} in K such that
z̄ − z̄
max λ (z̄) ≥ 0 ∀ z̄ ∈ K.
λ∈M0co
Therefore, we have obtained the quadratic necessary condition for the Pontryagin
minimum, which is slightly weaker than Condition A of Theorem 2.4. It is called Condition
Aco . Using Condition Aco , we will show below that Condition A is also necessary for
the Pontryagin minimum. We thus will complete the proof of Theorem 2.4. Section 2.6
is devoted to this purpose. But first we complete (in Section 2.5) the decoding of the
constant Cγ .
Denote by CK the least upper bound of C such that M co (C) is nonempty and
condition (2.106) holds. Then Lemma 2.42 implies
i.e., the constant CK estimates the constant Cγ from above with accuracy up to a constant
multiplier. We will prove that the constant CK estimates the constant Cγ from below with
2.5. Estimation of the Basic Constant from Below 75
accuracy up to a constant multiplier. This will allow us to obtain a sufficient condition for
the Pontryagin minimum.
Remark 2.44. Lemma 2.42 also implies the following assertion: if Cγ ≥ 0, then M0co is
nonempty and
maxco λ (z̄) ≥ 0 ∀ z̄ ∈ K.
λ∈M0
Consequently,
0 M
Cγ := inf lim ≥ inf lim . (2.108)
σ γ γ σ γ γ
76 Chapter 2. Quadratic Conditions in the Calculus of Variations
We set
√
o(√γ ) = {{δw} ∈ | σ = o( γ )}.
M M
inf lim ≥ inf√ lim . (2.109)
σ γ γ o( γ ) γ
Lemma 2.45. The following inequality holds for an arbitrary C such that
M co (C) is nonempty:
Cγ ≥ Cγ (C ; o(√γ ) ). (2.111)
In what follows, we will fix an arbitrary C such that the set M co (C) is nonempty.
Our further goal consists of passing from the constant Cγ (C , o(√γ ) ) defined by the
set of sequences o(√γ ) to a constant defined by the set of sequences loc
√
o( γ ) . For such a
passage, we need the estimate
i.e., |C (δw)| ≤ O(γ (δw)) for any sequence {δw} ∈ loc . To prove estimate (2.113), we
need a certain property of the set M co (C) analogous to one of the Weierstrass–Erdmann
conditions of the classical calculus of variations. Let us formulate this analogue.
[H λ ]k = 0 ∀ tk ∈ .
2.5. Estimation of the Basic Constant from Below 77
δε H λ = H (tε , x 0 (tε ), u0 (tε ) + [u0 ]k , ψ(tε )) − H (tε , x 0 (tε ), u0 (tε ), ψ(tε )).
Then for a small ε > 0, the condition λ ∈ M co (C) implies δε H λ ≥ Cδε , where δε =
(tε , u0 (tε ) + [u0 ]k ) − (tε , u0 (tε )) = (tε , u0 (tε ) + [u0 ]k ). Taking into account that
δε H λ → [H λ ]k and δε → 0 as ε → +0, we obtain [H λ ]k ≥ 0. Constructing an analo-
gous sequence t ε = tk + ε to the right from the point tk , we obtain −[H λ ]k ≥ 0. Therefore,
[H λ ]k = 0. The proposition is proved.
In Section 2.1.5, we have defined 0 as the set of tuples λ ∈ 0 such that [H ] = 0
λ k
for all tk ∈ . Proposition 2.46 means that M (C) ⊂ co 0 . This and Proposition 2.16
co
imply estimate (2.113). By the way, we note that under the replacement of M co (C) by
co 0 , estimate (2.113) does not hold in general.
Recall that by u0 we have denoted the closure of the graph of the function u0 (t)
assuming that u0 (t) is left continuous. By Qtu we have denoted the projection of the set
Q under the mapping (t, x, u) → (t, u). Denote by {V } an arbitrary sequence of neighbor-
hoods of the compact set u0 contained in Qtu such that V → u0 . The latter means that for
any neighborhood V
⊂ Qtu , of the compact set u0 , there exists a number starting from
which V ⊂ V
.
Let {δw} ∈ be an arbitrary sequence. For members δw = (δx, δu) and V of the
sequences {δw} and {V }, respectively, which have the same numbers,
we set
δu(t) if (t, u0 (t) + δu(t)) ∈ V ,
δuV (t) =
0 otherwise,
Proposition 2.47. The following relation holds for the above representation of the se-
quence {δw}:
δf = δ loc f + δ Vf + rfV ,
where
the function χ V is the characteristic function of the set M V = {t | δuV = 0}; and
def
Moreover,
rfV
∞ ≤
δ̄x f
∞ +
δx f
∞ = εf → 0 and
rfV
1 ≤ εf meas M V .
78 Chapter 2. Quadratic Conditions in the Calculus of Variations
Proof. We have
δf = δf (1 − χ V ) + δf χ V
= δ loc f (1 − χ V ) + (f (t, x 0 + δx, u0 + δu) − f (t, x 0 , u0 + δu)
+ f (t, x 0 , u0 + δu) − f (t, x 0 , u0 ))χ V
= δ loc f − δ loc f χ V + δ̄x f χ V + δ Vf
= δ loc f + δ Vf + (δ̄x f − δx f )χ V
= δ loc f + δ Vf + rfV .
Proof. Let there exist the sequences {δw}, {ε}, and {V } = {V (ε)} defined as above. Since
V → u0 , we have that, starting from a certain number, V ⊂ V, where V ⊂ Qtu is the
neighborhood of the compact set u0 from the definition of the function (t, u) given in
Section 2.3. Condition V ⊂ V implies (δuV )V = δuV (the meaning of this notation is
the same as above). Hence δuV = δuVV + δuV , where (t, u0 + δuVV ) ∈ V \ V on MVV =
{t | δuVV = 0} and (t, u0 + δuV ) ∈
/ V on M V = {t | δuV = 0}. The definitions of V, , and
V = V (ε) imply (t, u + δuV ) ≥ ε 2 on MV
0 V V . This implies γ V ≥ ε 2 meas M V , where γ V :=
V V V
tf V
0 + δuV ) dt. Since γ V ≤ γ V , we have meas M V ≤ γ . Further, it follows from the
(t, u V V V
t0 ε 2
t
definitions of and δuV that meas M V ≤ const γ V , where γ V := t0f (t, u0 + δuV ) dt, and
const depends on the entire sequence {δuV } but not on its specific member. Since γ V ≤ γ V
(because MV ⊂ MV ), we have meas M V ≤ const γ V . Therefore,
γ V 1
meas M V = meas MV V
+ meas MV ≤ 2 + const γ V = + const γ V . (2.114)
ε ε2
2.5. Estimation of the Basic Constant from Below 79
Taking into account that εf /ε 2 → 0, we obtain from (2.114) that εf meas MV = o(γ V ).
√
Moreover, since γ /ε 2 → 0 and γ V ≤ γ , (2.114) also implies
! ! !
γV
meas M ≤ V
+ const γ V γ V =o γV .
ε2
Finally, we obtain tf
δλ = δ loc λ + δ V H λ dt + ρ V λ . (2.116)
t0
Moreover, the conditions δf = δ loc f + δ Vf + rfV ,
δ Vf
1 ≤ const meas MV = o γV ,
and
rfV
1 = o(γ V ) imply
!
δ ẋ − δ loc f
1 ≤
δ ẋ − δf
1 + o γV . (2.117)
Proposition 2.49. Let {δw} ∈ be an arbitrary sequence, and let a sequence {ε} satisfy
the conditions of Proposition 2.48. Then conditions (2.115)–(2.117) hold for the corre-
sponding sequences {V (ε)}, {δwloc }, and {δwV }.
Proof. Let {δw} ∈ o(√γ ) be a sequence with γ > 0. Choose the sequence {ε} and the
corresponding sequence {V } as in Proposition 2.48. With the sequences {δw} and {V },
we associate the splitting {δw} = {δwloc } + {δwV }. Let us turn to formula (2.116). The
definition of the set M co (C) implies
for all sufficiently large numbers. Then (2.115) and (2.116) imply
where γ V = γ (δw V ). We set γ = γ (δw) and γ loc = γ (δwloc ). Then γ = γ loc + γ V . The
following two cases are possible:
loc
Case (a): lim γγ = 0.
loc
Case (b): lim γγ > 0.
Let us consider each of them.
loc
Case (a): Let lim γγ = 0. Choose subsequences such that γ loc /γ → 0 for them, and,
therefore,
γV
γ loc = o(γ ) and → 1. (2.120)
γ
We preserve the above notation for the subsequences. It follows from (2.119), (2.120), and
estimate (2.113) that δC ≥ Cγ + o(γ ). Hence
δC
lim ≥ C,
γ
where the lower limit is taken over the chosen subsequence.
Case (b): Now let limγ loc /γ > 0, and, therefore, γ ≤ O(γ loc ) and γ V ≤ O(γ
loc ).
√
Since
δ ẋ − δf
1 = o( γ ), it follows from (2.117) that
δ ẋ − δ loc f
1 = o( γ loc ).
Moreover,
! !
√ √
δ loc Fi = δFi ≤ o( γ ) = o1 γ loc , i ∈ I ; |δ loc K| = |δK| = o( γ ) = o γ loc .
γ loc γ V
+ = 1.
γ γ
Furthermore,
δ loc C
lim ≥ C γ C ; loc√
o( γ ) ,
γ loc
since {δwloc } ∈ loc
√
o( γ ) . Hence
δC
lim ≥ min Cγ C ; loc
√
o( γ ) , C .
γ
Therefore, we have shown that for every sequence {δw} ∈ o(√γ ) on which γ > 0, there
exists a subsequence such that lim δC /γ is not less than the right-hand side of inequality
(2.118). This implies inequality (2.118). The lemma is proved.
The method used in this section is very characteristic for decoding higher-order
conditions in optimal control in order to obtain sufficient conditions. Now this method is
related to the representation of the Pontryagin sequence as the sum {δw} = {δw loc } + {δw V }
and the use of the maximum principle of strictness C for {δwV }:
δ V H λ ≥ C(t, u0 + δuV ), λ ∈ M co (C).
In what follows, analogous consideration will be given to various consequences of the
minimum principle of strictness C, the Legendre conditions, and conditions related to
them. Lemmas 2.45 and 2.50 imply the following assertion.
2.5.4 Simplifications in the Definition of Cγ C ; loc
√
o( γ )
√ √
δFi ≤ o( γ ) ∀ i ∈ I , |δK| = o( γ ), (2.121)
√
δ ẋ − δf
1 = o( γ ). (2.122)
Obviously, conditions (2.121) are equivalent to the conditions
√ √
Fip δp ≤ o( γ ) ∀ i ∈ I , |Kp δp| = o( γ ).
Let us consider condition (2.122). Assume that the sequence {δw} is represented in the
canonical form (see Proposition 2.9)
By Proposition 2.10,
1
δf = fw δw 0 + fww δw0 , δw0 + δ ∗ f + δ ∗ fx δx + r̃,
2
t
where
r̃||1 = o(γ ) and γ :=
δx
2C + t0f |δu0 |2 dt. According to formula (2.50),
0 0
δ∗f = δk∗ f = ∗
[f ]k (χk− ∗
− χk+ ) + fuk+ δvk− + fuk− δvk+ + O(|δtk | + |δvk |2 ).
As was shown in Section 2.4 in proving the equivalence of conditions (2.66) and (2.69), we
have the estimates
tf
fww δw , δw
1 ≤ o(γ ),
0 0 0
δ ∗ fx δx dt ≤ o(γ ).
t0
Moreover,
fuk+ δvk−
1 ≤ |fuk+ | meas Mk−
δvk−
2 = o( γ ∗ ),
fuk− δvk+
1 ≤ |fuk− | meas Mk+
δvk+
2 = o( γ ∗ ),
t
t
where γ ∗ := k ( M∗ |δt| dt + t0f |δvk |2 dt) and
v
2 = ( t0f v(t), v(t) dt)1/2 is the norm
k
of the space L2 (, Rd(u) ) of Lebesgue square integrable functions v(t) : [t0 , tf ] → Rd(u) .
Therefore, condition (2.122) is equivalent to the condition
√
∗ ∗
δ ẋ − fw δw 0 − [f ]k (χk− − χk+ ) = o( γ ).
1
Finally, we obtain
√ √
loc
√ = {δw} ∈ loc | Fip δp ≤ o( γ ) ∀ i ∈ I ; |Kp δp| = o( γ );
o( γ )
√
∗ ∗
δ ẋ − fw δw0 − [f ]k (χk− − χk+ ) = o( γ ) .
1
In this relation, we use the canonical representation of the sequence {δw} = {δw 0 } + {δw ∗ },
where {δw0 } ∈ 0 , {δw∗ } ∈ ∗ , and |δu0 | · |δu∗ | = 0.
By Proposition 2.15, in calculating Cγ (C ; loc √
o( γ ) ), we can use the function
(see definition (2.61) for 1λ ) instead of the function C = maxM co (C) (λ, δw), and
since by Proposition 2.46the conditions [H λ ]k = 0 for all tk ∈ hold for any λ ∈ M co (C),
we can omit the terms k [H λ ]k (meas Mk− ∗ − meas M ∗ ) in the definition of 1λ , thus
k+
passing to the function
1
˜ 1λ (δw)
:= (λ, δw 0 )
2 s
tf
∗ ∗
+ k
D (H ) λ
|δtk | dt + [Hxλ ]k δx(χk− − χk+ ) dt
Mk∗ t0
k=1
1 tf
+ Huu
λk+
δvk− , δvk− + Huu
λk−
δvk+ , δvk+ dt ,
2 t0
2.5. Estimation of the Basic Constant from Below 83
where tf
(λ, δw) = lpp
λ
δp, δp + Hww
λ
δw, δw dt.
t0
Therefore, the constant Cγ C ; loc
√
o( γ ) does not change, if we replace the function C
˜ 1 , where
by the function C
Finally, we recall that the functional γ has the following form on local sequences represented
in the canonical form:
tf
tf
γ (δw) =
δx
2C + |δu0 |2 dt + 2 |δtk | dt + |δvk |2 dt
t0 k Mk∗ t0
where
˜1
˜ 1C ; loc√
Cγ o( γ ) = inf lim C .
loc
√
o( γ )
γ
˜ 1λ
We call attention to the fact that in the definition of this constant, the functions
∗
and γ are defined indeed on the set of triples (δw , M , δv) such that
0
Denote by a these triples, and denote by {a} an arbitrary sequences of triples a such that
{δw 0 } ∈ 0 ; Mk∗ → tk , k ∈ I ∗;
δv
∞ → 0.
To each sequence {δw} from loc represented in canonical form, we naturally associate the
sequence of triples {a} = {a(δw)} = {(δw 0 , M ∗ , δv)}. However, note that for an arbitrary
sequence {a}, we do not require the condition δu0 χ ∗ = 0, which holds for the sequences {a}
corresponding to the sequences {δw} ∈ loc in their canonical representations (as before,
χ ∗ is the characteristic function of the set M∗ ). Therefore, on the set of sequences {a}, we
84 Chapter 2. Quadratic Conditions in the Calculus of Variations
Indeed, the sequence {a} = {a(δw)} ∈ S1 corresponds to every sequence [δw} ∈ loc √
o( γ ) ,
˜ 1
and, moreover, the values of and γ are preserved under this correspondence. There is
C
no converse correspondence, since we omit the condition δu0 χ ∗ = 0 in the definition of S1 .
Lemma 2.51 and formulas (2.123) and (2.124) imply the following assertion.
˜ 1 ; S1 ), C}.
Lemma 2.52. Let the set M co (C) be nonempty. Then Cγ ≥ min{Cγ ( C
˜ 1 ; S1 ) from below.
In what follows, we estimate the constant Cγ ( C
Proof. Let λ ∈ M co (C), t ∈ [t0 , tf ]. Choose ε > 0 so small that the conditions ũ ∈ Rd(u)
and |ũ| < ε imply (t, x 0 (t), u0 (t) + ũ) ∈ Q and (t, u0 (t) + ũ) = |ũ|2 , and hence
H (t, x 0 (t), u0 (t) + ũ, ψ(t)) − H (t, x 0 (t), u0 (t), ψ(t)) ≥ C|ũ|2 .
2.5. Estimation of the Basic Constant from Below 85
In other words, the function ϕ(ũ) := H (t, x 0 (t), u0 (t) + ũ, ψ(t)) − C|ũ|2 defined on a neigh-
borhood of the origin of the space Rd(u) has a local minimum at zero. This implies
ϕ
(0) = 0 and ϕ
(0)ū, ū ≥ 0 for all ū ∈ Rd(u) . The first condition is equivalent to
Hu (t, x 0 (t), u0 (t), ψ(t)) = 0, and the second is equivalent to
This implies assertion (a) of the proposition. Assertion (b) is obtained from assertion (a) by
passing to the limit as t → tk + 0 and t → tk − 0, k ∈ I ∗ . The proposition is proved.
We now use assertion (b) only. Denote by b the pair (δw0 , M ∗ ) and by {b} the
sequence of pairs such that {δw0 } ∈ 0 , Mk∗ → tk for all k, M∗ = ∪Mk∗ . On each such
sequence, we define the functions
tf
1
∗ ∗
2λ (b) = (λ, δw ) +
0 k
D (H )λ
|δtk | dt + [Hx ]
λ k
δx(χk− − χk+ ) dt ,
2 Mk∗ t0
k
2C (b) = max 2λ (b),
M co (C)
tf
γ1 (b) =
δx
2C + |δu0 |2 dt + 2|δtk | dt.
t0 k Mk∗
We set
√ √
S2 = {b} = {(δw0 , M ∗ )} | Fip δp ≤ o( γ1 ), i ∈ I ; |Kp δp| = o( γ1 );
δ ẋ − fw δw 0 − [f ]k (χk− ∗ − χ ∗ )
= o(√γ ) .
k+ 1 1
˜ 1 , γ , and
We obtain the definitions of 2 , γ1 , and S2 from the corresponding definitions of
S1 setting δv = 0 everywhere. We set
2C
Cγ1 (2C ; S2 ) = inf lim .
S2 γ1
Proof. Let {a} ∈ S1 be an arbitrary sequence such that γ > 0 for all its members. For this
sequence, we set
tf
∗
{b} = {(δw , M )}, γ̂ (δv) =
0
|δv|2 dt.
t0
Then γ (a) = γ1 (b)+ γ̂ (δv), or, for short, γ = γ1 + γ̂ . Let λ ∈ M co (C). Proposition 2.53(b),
implies ˜ 1λ (δw0 ) ≥ 2λ (δw0 )+C γ̂ (δv). Consequently, ˜ 1 (δw0 ) ≥ 2 (δw 0 )+C γ̂ (δv),
C C
or, briefly, ˜ 1 ≥ 2 + C γ̂ . We consider the following two possible cases for the se-
C C
quence {a}.
86 Chapter 2. Quadratic Conditions in the Calculus of Variations
Case (a): Let lim(γ1 /γ ) = 0. Extract a subsequence such that γ1 = o(γ ) on it.
Let this condition hold for the sequence {a} itself. Then we obtain from the inequality
˜ 1 ≥ 2 + C γ̂ and the obvious estimate |2 | ≤ O(γ1 ) that
C C C
˜1
2 + C γ̂ o(γ ) + C γ̂ γ̂
lim C
≥ lim C = lim = C lim = C,
γ γ γ γ
since γ = γ1 + γ̂ = o(γ ) + γ̂ .
Case (b): Assume now that lim(γ1 /γ ) > 0, and hence γ ≤ const γ1 on the subse-
quence. The inequality ˜ 1 ≥ 2 + C γ̂ implies
C C
˜1
2C + C γ̂ γ1 2C γ̂ 2C
C
≥ = + C ≥ min ,C ,
γ γ γ γ1 γ γ1
But the conditions {a} ∈ S and γ ≤ const γ1 immediately imply {b} ∈ S2 . We obtain from
this that
˜1
lim C ≥ Cγ1 (2C , S2 ).
γ1
Consequently,
˜1
lim C
≥ min Cγ1 (2C ; S2 ), C .
γ
Therefore, we have shown that from any sequence {a} ∈ S1 at which γ > 0, it is possible to
˜ 1 /γ on it is not less than the right-
extract a subsequence such that the lower limit lim C
hand side of inequality (2.125). This obviously implies inequality (2.125). The lemma is
proved.
Lemma 2.55. Let the set M co (C) be nonempty. Then Cγ ≥ min Cγ1 (2C ; S2 ), C .
In what follows, we will estimate the constant Cγ1 (2C ; S2 ) from below.
Therefore, according to Proposition 2.24, there exists a constant 0 < q ≤ 1, q = 1/ (see
(2.73)), such that for any sequence {b} ∈ S2 , there exists a number starting from which we
have qγ1 (b) ≤ γ2 (b), or, briefly, qγ1 ≤ γ2 . Moreover, since |δx(t0 )| ≤
δx
C , we have
γ2 ≤ γ1 . Hence
γ2
q≤ ≤ 1.
γ1
This implies that the following relations hold for any sequence {b} ∈ S2 :
2C 2 γ2
lim = lim C · ≥ min Cγ2 (2C ; S2 ), qCγ2 (2C ; S2 ) .
γ1 γ2 γ1
Consequently,
Cγ1 (2C ; S2 ) ≥ min Cγ2 (2C ; S2 ), qCγ2 (2C ; S2 ) . (2.126)
Here,
2C
Cγ2 (2C ; S2 ) = inf lim .
S2 γ2
Inequality (2.126) and Lemma 2.55 imply the following assertion.
In what follows, we will estimate the constant Cγ2 (2C ; S2 ) from below.
Also, we set
1
3λ (b̄) =
(λ, w̄) + D k (H λ ) |δtk | dt + [Hxλ ]k x̄av
k
ξk ,
2 Mk∗
k
where ξk = meas Mk− ∗ − meas M ∗ . According to Lemma 2.26, for any sequence {b} ∈ S
k+ 2
for which γ2 > 0, there exists a sequence {b̄} such that
√
δp = p̄,
x̄˙ − fw w̄
1 = o( γ2 ), [x̄]k = [f ]k ξk ∀ tk ∈ ;
γ2 (b) = γ2 (b̄), 2C (b) = 3C (b̄) + o(γ2 ).
D k (H λ ) ≥ 2C, k ∈ I ∗.
Proof. Fix λ ∈ M co (C), k ∈ I ∗ . Take a small ε > 0 and construct a variation δu(t) in the
left neighborhood (tk − ε, tk ) of the point tk such that
u0k+ − u0 (t), t ∈ (tk − ε, tk ),
δu(t) =
0, t∈/ (tk − ε, tk ).
For a sufficiently small ε > 0, we have (t, x 0 (t), u0 (t) + δu(t)) ∈ Q. Consequently,
Denote by S4 the set of sequences {z̄} = {(ξ̄ , x̄, ū)} of elements of the space Z()
(defined by (2.89)) such that |ξ̄ | + ||w̄
∞ → 0 and, moreover,
√ √
Fip p̄ ≤ o( γ̄ ) (i ∈ I ), |Kp p̄| = o( γ̄ ),
√
x̄˙ − fw w̄
1 = o( γ̄ ), [x̄]k = [f ]k ξ̄k ∀ tk ∈ .
tf
Recall that γ̄ = γ̄ (z̄) := |x̄(t0 )|2 + t0 |ū|2 dt + |ξ̄ |2 . Also, recall that (see formula (2.13))
1 k λ 2 1
λ (z̄) := (D (H )ξ̄k + 2[Hxλ ]k x̄av
k
ξ̄k ) +
(λ, w̄),
2 2
k
tf
where
(λ, w̄) = lpp
λ p̄, p̄ +
t0 Hww
λ w̄, w̄ dt. We set
and
C
Cγ̄ (C ; S4 ) = inf lim .
S4 γ̄
Using Proposition 2.58, we prove the following estimate.
Proof. Let {b̄} = {(w̄, M ∗ )} ∈ S3 be a sequence on which γ2 > 0. The following inequalities
hold for each member of this sequence:
2 |δtk | dt ≥ ξ̄k2 , k ∈ I ∗ ,
Mk∗
∗ − meas M ∗ . We set
where ξ̄k = meas Mk− k+
μk = 2 |δtk | dt − ξ̄k2 , k ∈ I ∗, μ= μk .
Mk∗
90 Chapter 2. Quadratic Conditions in the Calculus of Variations
Then μk ≥ 0 for all k, and, therefore, μ ≥ 0. To the sequence {b̄} = {(w̄, M ∗ )}, we
naturally associate the sequence {z̄} = {ξ̄ , w̄)} with the same components w̄ and with
ξ̄k = meas Mk−∗ − meas M ∗ , k ∈ I ∗ . Then γ (b̄) = γ̄ (z̄) + μ, or, for short, γ = γ̄ + μ.
k+ 2 2
Moreover,
1
1
(b̄) =
3λ (λ, w̄) + D (H )(ξ̄k + μk ) + [Hx ] x̄av ξ̄k
k λ 2 λ k k
2 2
k
1 k λ
= λ (z̄) + D (H )μk .
2
k
or, briefly, 3C ≥ C + Cμ. Therefore, we have the following for the sequence {b̄}:
3C C + Cμ
≥ , (2.129)
γ2 γ2
where γ2 = γ̄ + μ. Consider the following two cases.
Case (a). Let lim γ̄ /γ2 = 0. Choose a subsequence such that γ̄ = o(γ2 ) on it. Let this
condition hold on the sequence {b̄} itself. Since γ2 = γ̄ + μ, we have μ/γ2 → 1. Let us
show that the following estimate holds in this case:
|C | = o(γ2 ). (2.130)
Indeed, the definition of the functional λ and the boundedness of the set M co (C) imply
the existence of a constant C > 0 such that
|C | ≤ C (
x̄
2∞ +
ū
22 + |ξ̄ |2 ). (2.131)
Lemma 2.61. The following relation holds: Cγ̄ (C ; S4 ) = Cγ̄ (C ; S5 ).
√
Conditions (2.132) easily imply
x̄
2∞ ≤ O(γ̄ ). Moreover,
δx
C = o( γ̄ ). This and con-
ditions (2.133) imply C (z̄) = C (z̄
) + o(γ̄ ). Consequently,
C (z̄) C (z̄
)
lim = lim ≥ Cγ̄ (C ; S5 ).
γ̄ (z̄) γ̄ (z̄
)
The inequality holds, since {z̄
} ∈ S5 . Since {z̄} is an arbitrary sequence from S4 on which
γ̄ > 0, we obtain from this that Cγ̄ (C ; S4 ) ≥ Cγ̄ (C ; S5 ). The inclusion S5 ⊂ S4 implies
the converse inequality. Therefore, we have an equality here. The lemma is proved.
Lemma 2.63. The following inequality holds: Cγ̄ (C ; S5 ) ≥ Cγ̄ (C ; K).
Proof. Let {z̄} be an arbitrary nonvanishing sequence from S5 . For this sequence, we have
Fip p̄ ≤ o( γ̄ ) (i ∈ I ), |Kp p̄| = o( γ̄ ),
w̄
∞ + |ξ̄ | → 0, (2.134)
˙x̄ = fw w̄, [x̄]k = [f ]k ξ̄k (k ∈ I ∗ ). (2.135)
Moreover, γ̄ (z̄) > 0 on the whole sequence, since it contains nonzero members. Let T
be the subspace in Z() defined by conditions (2.135). According to Lemma 2.41 (which
¯ =
follows from the Hoffman lemma), for the sequence {z̄}, there exists a sequence {z̄}
¯ ¯ ū)}
{(ξ̄ , x̄, ¯ in the subspace T such that
¯ ≤0
Fip (p̄ + p̄) (i ∈ I ), ¯ = 0,
Kp (p̄ + p̄)
√
and, moreover,
x̄
¯ ∞ + |ξ̄¯ | = o( γ̄ ). We set {z̄
} = {z̄ + z̄}.
¯ ∞ +
ū
¯ As in the proof of
Lemma 2.61, for the sequence {z̄
}, we have
Moreover,
γ̄ (z̄) = γ̄ (z̄
) + o(γ̄ ). (2.137)
¯ √
¯ ¯
√ follow from the estimate
x̄
∞ +
ū
∞ + |ξ̄ | = o( γ̄ ) and the estimate
These conditions
x̄
∞ ≤ O( γ̄ ), which, in turn, follows from (2.135). It follows from (2.137) and the
2.5. Estimation of the Basic Constant from Below 93
condition γ̄ (z̄) > 0 that γ̄ (z̄
) > 0 starting from a certain number. According to (2.136) and
(2.137), we have
C (z̄) C (z̄
)
lim = lim ≥ Cγ̄ (C ; K). (2.138)
γ̄ (z̄) γ̄ (z̄
)
The inequality holds, since z̄
∈ K \ {0}. Since this inequality holds for an arbitrary nonva-
nishing sequence {z̄} ∈ S5 , this implies Cγ̄ (C ; S5 ) ≥ Cγ̄ (C ; K).
Lemma 2.64. The following inequality holds for any real C such that the set M co (C) is
nonempty:
Cγ ≥ min{Cγ̄ (C ; K), qCγ̄ (C ; K), C, qC}. (2.139)
Let C be such that M co (C) is nonempty and
Then it follows from the definition of the constant Cγ̄ (C ; K) that Cγ̄ (C ; K) ≥ C.
This and (2.139) imply Cγ ≥ min{C, qC}. This inequality holds for all C such that the set
M co (C) is nonempty and condition (2.140) holds. Therefore, it also holds for the least
upper bound of these C. At the end of Section 2.4, we have denoted this upper bound by CK .
Therefore, we have proved the following theorem.
Recall that ≥ 1, 0 < q ≤ 1, and = 1/q. Bearing in mind inequalities (2.143) and (2.144),
we write
const
C γ = CK . (2.145)
This is the main result of decoding.
We now obtain the following important consequences of this result.
Case (a). Let CK ≥ 0. Then Cγ ≥ 0 by (2.143); according to Remark 2.44, this implies
the condition
M0co = ∅; max
co
λ (z̄) ≥ 0 ∀ z̄ ∈ K. (2.146)
M0
Conversely, if for a certain C > 0, condition (2.147) holds, then CK ≥ C, and hence
CK > 0. Therefore, the existence of C > 0 such that (2.147) holds is equivalent to the
inequality CK > 0. By Theorem 2.66, the latter is equivalent to the inequality Cγ > 0.
The following theorem summarizes what was said above.
Theorem 2.67. (a) The inequality Cγ ≥ 0 is equivalent to condition (2.146). (b) The
inequality Cγ > 0 is equivalent to the existence of C > 0 such that condition (2.147) holds.
Definition 2.68. Let (t, u) be an admissible function. We say that Condition Bco () holds
at the point w 0 if there exists C > 0 such that condition (2.147) holds.
As we already know, the condition Cγ > 0 is sufficient for the strict Pontryagin
minimum at the point w0 , and, according to Theorem 2.67, it is equivalent to Condition
Bco (). Therefore, Condition Bco () is also sufficient for the strict Pontryagin minimum.
Let us consider Condition Bco (). First of all, we show that it is equivalent to Condi-
tion B() in whose definition we have the set M(C) instead of the set M co (C). By defini-
tion, the set M(C) consists of tuples λ = (α0 , α, β, ψ) ∈ 0 such that ψ(t)(f (t, x 0 (t), u) −
f (t, x 0 (t), u0 (t))) ≥ C(t, u) if t ∈ [t0 , tf ]\ and (t, x 0 (t), u) ∈ Q. Therefore, the set M(C)
differs from the set M co (C) by that co 0 is replaced by 0 in the definition of the latter.
Condition B() means that the set M(C) is nonempty for a certain C > 0 and
max λ (z̄) ≥ C γ̄ (z̄) ∀ z̄ ∈ K. (2.148)
M(C)
Lemma 2.69. For any C > 0 such that M co (C) is nonempty, there exists 0 < ε < 1 such
that
C
M (C) ⊂ [ε, 1] ◦ M
co
,
ε
where [ε, 1] ◦ M is the set of tuples λ = ρ λ̃ such that ρ ∈ [ε, 1] and λ̃ ∈ M.
Proof. For an arbitrary λ = (α0 , α, β, ψ) ∈ co 0 , we set ν(λ) = α0 + |α| + |β|. Since the
function ν(λ) is convex and equals 1 on 0 , we have ν(λ) ≤ 1 for all λ ∈ co 0 . Further, let
C > 0 be such that M co (C) is nonempty. Then as is easily seen, the compact set M co (C)
does not contain zero. This implies
ε= min ν(λ) > 0,
λ∈M co (C)
Proof. Let Condition Bco () hold, i.e., there exists C > 0 such that M co (C) is non-
empty and
max
co
λ (z̄) ≥ C γ̄ (z̄) ∀ z̄ ∈ K.
M (C)
C (z̄) ≥ C γ̄ (z̄) ∀ z̄ ∈ K.
λ
max
[ε,1]◦M ε
C (z̄) ≥ C γ̄ (z̄) ∀ z̄ ∈ K.
λ
max
M ε
Therefore, Condition B() holds. We have shown that Condition Bco () implies Condition
B(). As mentioned above, the converse is also true. Therefore, Conditions Bco () and
B() are equivalent. The theorem is proved.
96 Chapter 2. Quadratic Conditions in the Calculus of Variations
Lemma 2.72. For any admissible function (t, u) and any C > 0, the set M(C) is a
compact set contained in the set Leg+ (M0+ ).
Lemma 2.73. For any nonempty compact set M ⊂ Leg+ (M0+ ), there exist an admissible
function (t, u) and a constant C > 0 such that M ⊂ M(C).
Lemma 2.74. Let M ⊂ Leg+ (0 ) be a nonempty compact set, and let (t, u) be an admis-
sible function. Then there exist a neighborhood U of the compact set u0 and a constant
C > 0 such that M ⊂ M U (C).
Proposition 2.75. Assume that there is a nonempty compact set M ⊂ 0 such that the
following conditions hold for each of its elements λ:
(a) for any t ∈ [t0 , tf ] \ ,
1
Huu (t, x 0 (t), u0 (t), ψ(t))ū, ū > 0 ∀ ū ∈ Rd(u) \ {0};
2
Then there exist C > 0 and ε > 0 such that for any λ ∈ M, the conditions
t ∈ [t0 , tf ] \ and |u − u0 (t)| < ε (2.151)
imply
H (t, x 0 (t), u, ψ(t)) − H (t, x 0 (t), u0 (t), ψ(t)) ≥ C|u − u0 (t)|2 . (2.152)
Proof. Assume the contrary. Let the compact set M ⊂ 0 be such that conditions (a) and
(b) of the proposition hold for each of its element, but there are no C > 0 and ε > 0 such that
conditions (2.151) imply inequality (2.152). Then there exist sequences {Cn }, {tn }, {λn },
and {ūn } such that
Cn → +0, tn ∈ [t0 , tf ] \ , λn ∈ M, ūn ∈ Rd(u) , |ūn | → 0,
(2.153)
H λn (tn , xn0 , u0n + ūn ) − H λn (tn , xn0 , u0n ) < Cn |ūn |2 ,
where xn0 = x 0 (tn ) and u0n = u0 (tn ). Without loss of generality, we assume that
ūn
tn → tˆ ∈ [t0 , tf ], λn → λ̂ ∈ M, → ū.
|un |
Then ūn = εn (ū + ũn ), where εn = |ūn | → 0, |ūn | → 0. In this case, we obtain from (2.153)
that
1 λn
H (tn , xn0 , u0n )ūn , ūn + o(εn2 ) < Cn εn2 . (2.154)
2 uu
Here, we have taken into account that Huλn (tn , xn0 , u0n ) = 0.
We first assume that tˆ ∈
/ . Dividing (2.154) by εn2 and passing to the limit, we obtain
1 λ̂
H (tˆ, x 0 (tˆ), u0 (tˆ))ū, ū ≤ 0.
2 uu
But this contradicts condition (a), since ū = 0. Analogously, in the case where tˆ ∈ , we
arrive at a contradiction to one of the conditions in (b). The proposition is proved.
In what follows, we need to use the assumption that each tk ∈ is an L-point of the
control u0 .
Proposition 2.76. Let M ⊂ 0 be a nonempty compact set such that the following conditions
hold for a fixed point tk ∈ and any λ ∈ M:
[H λ ]k = 0, D k (H λ ) > 0, (2.155)
1 λk−
H ū, ū > 0 ∀ ū ∈ Rd(u) \ {0}. (2.156)
2 uu
Then there exist C > 0 and ε > 0 such that for any λ ∈ M, the conditions
tk < t < tk + ε and |u − u0k− | < ε (2.157)
imply the inequality
1
H (t, x 0 (t), u, ψ(t)) − H (t, x 0 (t), u0 (t), ψ(t)) ≥ C |t − tk | + |u − u0k− |2 . (2.158)
2
98 Chapter 2. Quadratic Conditions in the Calculus of Variations
Proof. Let a compact set M ⊂ 0 satisfy the condition of the proposition, and let there be
no C > 0 and ε > 0 such that for any λ ∈ M, conditions (2.157) imply inequality (2.158).
Then there exist a sequence {C} and a sequence of triples {(t, u, λ)} such that
C → +0, t → tk + 0, u → u0k+ ,
1 (2.159)
H λ (t, x 0 (t), u) − H λ (t, x 0 (t), u0 (t)) < C |t − tk | + |u − u0k− |2
2
(we omit the serial numbers of members). We set t − tk = δt > 0, u − u0k− = δv, u0 (t) −
u0k+ = δu0 , x 0 (tk ) = x 0k , x 0 − x 0k = δx 0 , etc. Then we get
Subtracting the latter relation from the previous one, taking into account that Huλk+ =
Huλk− = 0 and [H λ ]k = 0, and also taking into account inequality (2.159), we obtain
1 λk−
− [Htλ ]k + [Hxλ ]k ẋ 0k+ + ψ̇ k+ [Hψλ ]k δt + Huu δv, δv + o(|δt| + |δv|2 )
2
1
< C |δt| + |δv| . 2
2
1 λk−
D k (H λ )δt + Huu δv, δv < o1 (|δt| + |δv|2 ). (2.160)
2
Proposition 2.77. Let M ⊂ 0 be a nonempty compact set such that the following condi-
tions hold for a fixed point tk ∈ and any λ ∈ M:
[H λ ]k = 0, D k (H λ ) > 0,
and
1 λk+
H ū, ū > 0 ∀ ū ∈ Rd(u) \ {0}.
2 uu
Then there exist C > 0 and ε > 0 such that for any λ ∈ M, the conditions tk − ε < t < tk ,
and |u − u0k+ | < ε imply
1
H (t, x (t), u, ψ(t)) − H (t, x (t), u (t), ψ(t)) ≥ C |t − tk | + |u − u | .
0 0 0 0k+ 2
2
Propositions 2.75, 2.76, and 2.77 directly imply Lemma 2.74.
Proof of Lemma 2.73. Assume that there exists a nonempty compact set M ⊂ Leg+ (M0+ ).
Let 1 (t, u) be a certain admissible function (as was shown in Section 2.3.7, there exists at
least one such function). According to Lemma 2.73, there exist a neighborhood U ⊂ Qtu
of the compact set u0 and a constant C > 0 such that M ⊂ M U (C1 ), i.e., M ⊂ 0 and the
conditions t ∈ [t0 , tf ] \ , (t, u) ∈ U, and (t, x 0 (t), u) ∈ Q imply
We set
1
h(t, u) = min {H λ (t, x 0 (t), u) − H λ (t, x 0 (t), u0 (t))},
C λ∈M (2.162)
(t, u) = min{h(t, u), 1 (t, u)}.
It is easy to see that the function (t, u) (defined by (2.162)) is admissible, and, moreover,
M ⊂ M(C). The lemma is proved.
We now recall the definition given in Section 2.1.8. We say that Condition B holds
for the point w0 if there exist a nonempty compact set M ⊂ Leg+ (M0+ ) and a constant C > 0
100 Chapter 2. Quadratic Conditions in the Calculus of Variations
such that
max λ (z̄) ≥ C γ̄ (z̄) ∀ z̄ ∈ K. (2.163)
M
The following assertion holds.
Proof. Let Condition B hold; i.e., there exist a nonempty compact set M ⊂ Leg+ (M0+ ) and
a constant C > 0 such that condition (2.163) holds. Then, according to Lemma 2.73, there
exist an admissible function (t, u) and a constant C1 > 0 such that M ⊂ M(C1 ). We set
C2 = min{C, C1 }. Then M(C1 ) ⊂ M(C2 ). Consequently, M ⊂ M(C2 ) and
max λ (z̄) ≥ max λ (z̄) ≥ C γ̄ (z̄) ≥ C2 γ̄ (z̄) ∀ z̄ ∈ K
M(C2 ) M
In Section 2.2, we have verified all the assumptions of the abstract scheme for the
canonical problem, the point w 0 , the sets P and Q (the absorbing set = W corresponds to
them in the space W ), and the set of Pontryagin sequences in the space W . In Section 2.3,
the corresponding assumptions of the abstract scheme were also verified for a higher order
γ on . Therefore, Theorem 1.7 is applicable. According to this theorem, the condition
Cγ > 0 is not only sufficient for the Pontryagin minimum at the point w 0 but is equivalent
to the γ -sufficiency on . The latter will be also called the Pontryagin γ -sufficiency. For
convenience, we define this concept here. Let be an admissible function, and let γ be the
higher order corresponding to it.
Definition 2.80. We say that the point w 0 yields the Pontryagin γ -sufficiency if there exists
ε > 0 such that for any sequence {δw} ∈ , there exists a number, starting from which the
condition σ (δw) ≥ εγ (δw) holds.
The equivalent condition for the Pontryagin γ -sufficiency consists of the following:
There is no sequence {δw} ∈ such that σ = o(γ ) on it.
The violation function σ was already defined in Section 2.2.4. In what follows, it is
convenient to use the following expression for σ :
tf
σ = (δJ )+ + Fi+ (p0 + δp) + |δK| + |δ ẋ − δf | dt, (2.164)
t0
2.5. Estimation of the Basic Constant from Below 101
Theorem 2.81. The condition Cγ > 0 is equivalent to the Pontryagin γ -sufficiency at the
point w 0 .
Since the Pontryagin γ -sufficiency implies the strict Pontryagin minimum, Condition
B is also sufficient for the latter. Therefore, Theorem 2.5 is proved.
However, Theorem 2.82 is a considerably stronger result than Theorem 2.5. It allows
us to proceed more efficiently in analyzing sufficient conditions. We will see this in what
follows.
where
M
Cγ (M ; o(√γ ) ) = inf√ lim
o( γ) γ
and M (δw) = maxλ∈M (λ, δw). We can further estimate Cγ (M ; o(√γ ) ) from below
exactly in the same way as was done for the constant Cγ (C ; o(√γ ) ) when M = M co (C).
All the arguments are repeated literally (see relations (2.118), (2.121), (2.124)–(2.128) and
Lemmas 2.61, 2.63). As a result, we arrive at the following estimate:
Cγ (M ; o(√γ ) ) ≥ min{Cγ̄ (M ; K), qCγ̄ (M ; K), C, qC)}, (2.166)
Now let M ⊂ Leg+ (M0+ ) be a nonempty compact set, and let there exist a constant
C > 0 such that
max λ (z̄) ≥ C γ̄ (z̄) ∀ z̄ ∈ K (2.167)
M
102 Chapter 2. Quadratic Conditions in the Calculus of Variations
(i.e., Condition B holds). Then by Lemma 2.73, there exist an admissible function
and a constant C1 such that M ⊂ M(C1 ). Condition (2.167) implies Cγ̄ (M ; K) ≥ C.
Then (2.166) implies Cγ (M ; o(√γ ) ) ≥ q min{C, C1 }. We set CM = 12 q min{C, C1 }. Then
Cγ (M ; o(√γ ) ) > CM . Therefore,
M ≥ CM · γ | o(√γ ) , (2.168)
i.e., for any sequence {δw} ∈ o(√γ ) , there exists a number starting from which we have
M ≥ CM γ . We have obtained the following result.
Lemma 2.83. Let M ⊂ Leg+ (M0+ ) be a nonempty compact set, and let there exist C > 0
such that condition (2.167) holds. Then there exists a constant CM > 0 such that condition
(2.168) holds.
As was already noted, it is slightly weaker than the necessary condition of Theorem 2.4,
since in Condition A, we have the set M0 , which is more narrow than the set M0co . However,
we will show that the obtained necessary condition remains valid under the replacement
of M0co by M0 , i.e., the necessary Condition A holds. We thus will complete the proof of
Theorem 2.4.
The passage to the auxiliary problem in [79, Part 1, Chapter 2, Section 7] and the
trajectory of this problem corresponding to the index ζ chosen in a special way allows us to
do this. For this trajectory, we write the necessary Condition Aco in the auxiliary problem
with the subsequent transform of this condition into the initial problem. Such a method
was already used in [79, Section 7, Part 1] in proving the maximum principle. We use the
notation, the concepts, and the results of [79, Section 7, Part 1], briefly mentioning the main
constructions. We stress that in contrast to [79, Section 7, Part 1], all the constructions here
refer to the problem on a fixed closed interval of time [t0 , tf ]. We write the condition that
the endpoints of the closed interval of time [t0 , tf ] are fixed as follows: t0 = t00 , tf = tf0 .
Therefore, let us consider the problem (2.1)–(2.4) in the form which corresponds to the
general problem considered in [79, Section 7, Part 1],
J (x0 , xf ) → min, F (x0 , xf ) ≤ 0, K(x0 , xf ) = 0,
(2.169)
(x0 , xf ) ∈ P , t0 − t00 = 0, tf − tf0 = 0;
dx
= f (t, x, u), (t, x, u) ∈ Q, (2.170)
dt
2.6. Completing the Proof of Theorem 2.4 103
such that
t i < t i+1 , i = 1, . . . , N − 1; (t i , x̂ 0 (t i ), uik ) ∈ Q, i, k = 1, . . . , N .
Here and below in this section, the fixed time interval is denoted by [t00 , tf0 ], while all t i ,
i = 1, . . . , N are internal points of this interval. Denote by D() the set of all ζ satisfying
the condition t i ∈ / , i = 1, . . . , N .
Further, recall the definition of the set ζ in [79, Section 7, Part 1]. For the problem
(2.169), (2.170), it consists of tuples μ = (α0 , α, β) such that
α0 ≥ 0, α ≥ 0, αF (x̂00 , x̂f0 ) = 0, α0 + αi + |β| = 1, (2.172)
and, moreover, there exist absolutely continuous functions ψ̂x (t) and ψ̂t (t) such that
ψ̂x (t00 ) = −lx
0 , ψ̂x (tf0 ) = lx
f , (2.173)
d ψ̂x
− = ψ̂x (t)fx
(t, x̂ 0 (t), û0 (t)),
dt (2.174)
d ψ̂t
− = ψ̂x (t)ft
(t, x̂ 0 (t), û0 (t)),
dt
t0
f
ψ̂x (t)f (t, x̂ 0 (t), û0 (t)) + ψ̂t (t) dt = 0, (2.175)
t00
where the intersection is taken over all subscripts ζ . At the end of Section 7 in [79, Part 1],
we have shown that elements μ ∈ satisfy the following minimum principle:
ψ̂x (t)f (t, x̂ 0 (t), u) + ψ̂t (t) ≥ 0 if t ∈ [t00 , tf0 ], u ∈ Rd(u) , (t, x 0 (t), u) ∈ Q; (2.178)
ψ̂x (t)f (t, x̂ 0 (t), û0 (t)) + ψ̂t (t) = 0 a.e. on [t00 , tf0 ]. (2.179)
By continuity, the latter condition extends to all points of the set [t00 , tf0 ] \ . This implies
⊂ N0 , where N0 is the projection of the set M0 under the injective mapping λ = (α0 , α,
β, ψx ) → μ = (α0 , α, β). We now consider the set
"
() = ζ .
ζ ∈D()
Clearly, elements μ ∈ () satisfy condition (2.178) at all points t ∈ [t00 , tf0 ] \ ; however,
by continuity, this condition extends to all points of the interval [t00 , tf0 ]. Consequently,
() = , and, therefore, "
ζ ⊂ N0
ζ ∈D()
(in fact, we have an equality here). In case (b) the following assertion holds.
Proposition 2.84. There exists a subscript ζ such that ζ ∩ (−ζ ) = ∅, and, moreover, the
following condition holds for all instants of time t i , i = 1, . . . , N , entering the definition of
the subscript ζ : t i ∈
/ , i = 1, . . . , N .
Proof. Assume that the proposition does not hold. Then each of the sets
Fζ := ζ ∩ (−ζ ), ζ ∈ D(),
is not empty. These sets compose a centered system of nonempty compact sets, and hence
their intersection "
F := Fζ
ζ ∈D()
is nonempty. Moreover, F ⊂ () ⊂ N0 , and the condition Fζ = −Fζ for all ζ ∈ D()
implies F = −F . Let μ = (α0 , α, β) ∈ F . Then μ ∈ N0 and (−μ) ∈ N0 ; therefore, we
have λ ∈ M0 and (−λ) ∈ M0 for the corresponding element λ. But this contradicts case (b)
considered. The proposition is proved.
dx dt dz
= (ϕ(η)z)f (t, x, u); = ϕ(η)z; = 0;
dτ dτ dτ (2.181)
(t, x, u) ∈ Q, η ∈ Q1 .
Here, z and η are of dimension N 2 + 1 and have the components
respectively.2 The open set Q1 is the union of disjoint neighborhoods Qθ and Qik of the
points eθ and eik , i, k = 1, . . . , N , respectively, which are the standard basis of RN +1 (eθ has
2
θ ik
the unit component eθ , and other components of e are zero, while e has the unit component
eik , and other components of eik are zero), and φ(η) : Q1 → Rd(η) is a function mapping
each of the mentioned neighborhoods into the element of the basis whose neighborhood
it is. Note that the functions u(τ ) and η(τ ) are controls, while the functions z(τ ), x(τ ),
and t(τ ) are state variables in Problem ZN .
Recall the definition of the point κ ζ = (z0 (τ ), x 0 (τ ), t 0 (τ ), u0 (τ ), η0 (τ )) (in Problem
ZN on the closed interval [τ0 , τf ]) corresponding to the subscript ζ and the trajectory
(x̂ 0 (t), û0 (t)), t ∈ [t00 , tf0 ].
We “insert” N closed intervals of unit length adjusting to each other into each point t i
of the closed interval [t00 , tf0 ]; thus, we enlarge the length of the closed interval by N 2 . Place
the left endpoint of the new closed interval at zero. We obtain the closed interval [τ0 , τf ]
(τ0 = 0) with N 2 closed intervals (denoted by ik , i, k = 1, . . . , N ) placed on it; moreover,
i1 , . . . , iN are closed intervals adjusted to each other, located in the same order, and
corresponding to the point t i , i = 1, . . . , N . We set
n
E = (τ0 , τf ) \ ij .
i,j =1
Let χE and χij be the characteristic functions of the sets E and ij , respectively. We set
zθ0 = 1, 0
zij = 0, i, j = 1, . . . , N ,
dt 0
= ϕ(η0 )z0 , t 0 (τ0 ) = t00 .
dτ
Then t 0 (τf ) = tf0 , since meas E = tf0 − t00 . We set x 0 (τ ) = x̂ 0 (t 0 (τ )), u0 (τ ) = û0 (t 0 (τ )). As
was shown in [79, Part 1, Section 7, Proposition 7.1], the point κ ζ defined in such a way
2 We preserve notation accepted in [79, Section 7, Part 1], where z and η were used to denote “zero-
θ θ
components” of vectors z and η, respectively.
106 Chapter 2. Quadratic Conditions in the Calculus of Variations
that it is a Pontryagin minimum point in Problem ZN on the fixed closed interval [τ0 , τf ].
In what follows, all the functions and sets related to N or ζ are endowed with the indices
N or ζ , respectively.
Since κ ζ yields the Pontryagin minimum in Problem ZN , the necessary Condition
A co ζ holds for it in this problem. We show that the necessary Condition Aζ also holds for
ζ ζ
the chosen index ζ . For this purpose, in Problem ZN , we consider the sets ζ , 0 , co 0 ,
ζ co ζ
M0 , and M0 for the trajectory κ ζ and find the relations between them. The definition
of the set ζ was given in [79, Section 7, Part 1], and the other sets were defined in [79,
Part 2].
2.6.4 Condition Aζ
The function l N has the form
lN = α0 J + αF + βK − αz z0 + βt0 (t0 − t00 ) + βtf (tf − tf0 )
= l − αz z0 + βt0 (t0 − t00 ) + βtf (tf − tf0 ).
such that
The gradients lx
0 and lx
f are taken at the point (x00 , xf0 , α0 , α, β).
In [79, Section 7, Part 1], we have shown that there is the following equivalent
ζ
normalization for the set 0 :
α0 + αi + |β| = 1 (2.192)
2.6. Completing the Proof of Theorem 2.4 107
ζ
Proposition 2.85. The convex hull co 0 does not contain zero.
ζ
Proof. Assume that this is not true. Then there exist an element λN ∈ 0 and a number
ζ
ρ > 0 such that −ρλN ∈ 0 . This implies that all nonnegative components α0 , α, and αz of
ζ ζ
the element λN (see (2.182)) vanish. But then the condition λN ∈ 0 implies −λN ∈ 0 ,
i.e., we may set ρ = 1.
Let μN = (α0 , α, β) = (0, 0, β) be the projection of the element λN . Then μN and
−μ belong to ζ . But the existence of such a μN contradicts the choice of the index ζ .
N
ζ ζ
Corollary 2.86. For any λN ∈ co 0 , there exists ρ > 0 such that ρλN ∈ 0 .
ζ ζ
Proof. Let λN ∈ co 0 . Then by Proposition 2.85, λN = 0. Obviously, co 0 is contained
ζ ζ ζ def
in the cone con 0 spanned by 0 . The conditions λN ∈ con 0 , λN = 0 imply ν(λN ) =
α0 + |α| + |β| > 0 (since ν = 1 is a normalization condition). We set
λN
λ̃N = .
ν(λN )
ζ ζ
Then λ̃N ∈ con 0 and ν(λ̃N ) = 1. Therefore, λ̃N ∈ 0 . It remains to set ρ = 1/ν(λN ).
The proposition is proved.
co ζ ζ
Corollary 2.86 and the definitions of the sets M0 and M0 imply the following
assertion.
co ζ ζ
Corollary 2.87. Let λN ∈ M0 . Then there exists ρ > 0 such that ρλN ∈ M0 .
The condition Aco ζ for the point κ ζ in Problem ZN has the form
max ζ (λN ; z̄N ) ≥ 0 ∀ z̄N ∈ K ζ .
co ζ
M0
Here, K ζ is the critical cone and ζ is the quadratic form of Problem ZN at the point κ ζ .
Let us show that this implies Condition Aζ :
max ζ (λN ; z̄N ) ≥ 0 ∀ z̄N ∈ K ζ .
ζ
M0
108 Chapter 2. Quadratic Conditions in the Calculus of Variations
co ζ
Indeed, let z̄N ∈ K ζ . Condition Aco ζ implies the existence of λN ∈ M0 such that
Lemma 2.88. Let ŵ0 be a Pontryagin minimum point in the problem (2.169), (2.170), and
let M0 ∩ (−M0 ) = ∅. Then there exists a superscript ζ ∈ D() such that Condition Aζ
holds.
In what follows, we fix a superscript ζ ∈ D() such that Condition Aζ holds. Now
our goal is to reveal which information about the trajectory ŵ 0 can be extracted from
Condition Aζ of superscript ζ . We show that Condition Aζ implies Condition A at the point
ŵ0 in the initial problem. For this purpose, consider in more detail the definitions of the
ζ
set M0 , cone K ζ , and quadratic form ζ at the point κ ζ in Problem ZN .
ζ
2.6.5 Relation Between the Sets M0 and M0
ζ ζ ζ
Consider the conditions defining the set M0 . By definition, M0 is the set of λN ∈ 0 such
that the following inequality holds for all τ in the closed interval [τ0 , τf ], except for a finite
set of discontinuity points of the controls u0 (τ ) and η0 (τ ):
(ϕ(η)z0 ) ψx (τ )f (t 0 (τ ), x 0 (τ ), u) + ψt (τ )
(2.194)
≥ (ϕ(η0 )z0 ) ψx (τ )f (t 0 (τ ), x 0 (τ ), u0 (τ )) + ψt (τ )
for all u ∈ Rd(u) such that (t 0 (τ ), x 0 (τ ), u) ∈ Q and all η ∈ Q1 . Let us analyze condition
(2.194). Choose a function η = η(τ ) ∈ Q1 so that the following condition holds:
Such a choice is possible, since the condition z0 = eθ and the definition of the function ϕ
imply
1, η ∈ Qθ ,
ϕ(η)z = ϕθ (η) =
0
0, η ∈ / Qθ ,
and, therefore, we may set η(τ ) = η∗ , where η∗ is an arbitrary point in Q1 \ Qθ , for example,
η∗ = e11 . Therefore, condition (2.195) holds for η(τ ) ≡ e11 . It follows from (2.194) and
(2.195) that the right-hand side of inequality (2.194) is nonpositive. But the integral of it
over the interval [τ0 , τf ] vanishes (this was shown in [79, Section 7, Part 1]; moreover, this
follows from the adjoint equation (2.190), conditions (2.183), and the transversality condi-
tions (2.187) considered for the component ψzθ only). But if the integral of a nonpositive
2.6. Completing the Proof of Theorem 2.4 109
function over a closed interval vanishes, then this function equals zero almost everywhere
on this closed interval. Hence
(ϕ(η0 (τ ))z0 ) ψx (τ )f (t 0 (τ ), x 0 (τ ), u0 (τ )) + ψt (τ ) = 0 (2.196)
a.e. on [τ0 , τf ]. Further, setting η = η0 (τ ) in (2.194) and taking into account (2.196), we
obtain
(ϕ(η0 (τ ))z0 ) ψx (τ )f (t 0 (τ ), x 0 (τ ), u) + ψt (τ ) ≥ 0 if (t 0 (τ ), x 0 (τ ), u) ∈ Q. (2.197)
We note that
x̂ 0 (t) = x 0 (τ 0 (t)), û0 (t) = u0 (τ 0 (t)). (2.199)
The first equation holds on [t00 , tf0 ], and the second holds at every continuity point of the
function û0 (t), i.e., on the set [t00 , tf0 ] \ . Also, recall that ϕ(η0 (τ ))z0 = χE (τ ), and hence
a.e. on [t00 , tf0 ]. Setting τ = τ 0 (t) in conditions (2.196) and (2.197) and taking into account
(2.198)–(2.200), for almost all t ∈ [t00 , tf0 ], we obtain
if (t, x̂ 0 (t), u) ∈ Q, u ∈ Rd(u) . Condition (2.201), which holds a.e. on [t00 , tf0 ], also holds at
every continuity point of the function û0 (t), i.e., on [t00 , tf0 ] \ ; condition (2.202) holds for
all t ∈ [t00 , tf0 ], since all functions entering this condition are continuous.
In [79, Section 7, Part 1], we have proved that equations (2.188) and (2.189) imply
the equations
d ψ̂x
− = ψ̂x (t)fx
(t, x̂ 0 (t), û0 (t)), (2.203)
dt
d ψ̂t
− = ψ̂t (t)ft
(t, x̂ 0 (t), û0 (t)). (2.204)
dt
110 Chapter 2. Quadratic Conditions in the Calculus of Variations
dt 0
= ϕ(η0 (τ ))z0 .
dτ
Finally, the transversality conditions (2.185) imply the following transversality conditions:
ψ̂x (t00 ) = −lx 0 (x̂00 , x̂10 ), ψ̂x (tf0 ) = lx f (x̂00 , x̂10 ), (2.205)
since t 0 (τ0 ) = t00 and t 0 (τf ) = tf0 . Conditions (2.201)–(2.205) and conditions (2.183)
and (2.192), which hold for a tuple (α0 , α, β, ψx (t), ψt (t)), imply that its projection
(α0 , α, β, ψx (t)) belongs to the set M0 of the problem (2.169), (2.170) at the point ŵ 0 (t).
Therefore, for the superscript ζ indicated in Lemma 2.88 and corresponding function τ 0 (t)
(defined above), we have proved the following assertion.
Lemma 2.89. Let a tuple λN = (α0 , α, β, αz , βt0 , βtf , ψx (τ ), ψt (τ ), ψz (τ )), belong to the
ζ
set M0 of Problem ZN at the point κ ζ . We set ψ̂x (t) = ψx (τ 0 (t)). Then the tuple λ =
(α0 , α, β, ψ̂x (t)) belongs to the set M0 of the problem (2.169), (2.170) at the point ŵ0 .
that ˜ ζ does not intersect the open set E . Therefore, the sets ζ and ˜ ζ are disjoint. We
denote their union by (ζ ).
By definition, the critical cone K ζ for the trajectory κ ζ in Problem ZN consists of
the tuples
z̄N = (ξ̄ , ξ̃ , t¯(τ ), x̄(τ ), z̄(τ ), ū(τ ), η̄(τ )) (2.206)
such that
where p̄ = (x̄(τ0 ), x̄(τf )) and the gradients Jp
, Fip
, and Kp
are taken at the point
(x̂00 , x̂10 ) = p̂ 0 ,
t¯(τ0 ) = 0, t¯(τf ) = 0, (2.209)
d x̄
= (ϕ(η0 (τ ))z0 ) ft
t¯(τ ) + fx
x̄(τ ) + fu
ū(τ )
dτ
+ ((ϕη
(η0 (τ ))η̄(τ ))z̄0 )f (t 0 (τ ), x 0 (τ ), u0 (τ )) (2.210)
+ (ϕ(η0 (τ ))z̄(τ ))f (t 0 (τ ), x 0 (τ ), u0 (τ )),
2.6. Completing the Proof of Theorem 2.4 111
Proof. Let z̄ˆ be an arbitrary element of the critical cone K of the problem (2.169), (2.170)
at the point ŵ0 having the form (2.219). Then by the definition of the cone K, we have
Let conditions (2.220) hold. We show that all conditions (2.214)–(2.218) defining the cone
ζ
K0 hold for the element z̄N (having form (2.221)).
Conditions (2.214) follow from (2.220). Conditions (2.215) follow from (2.222).
ˆ
Indeed, the function x̄(t) is piecewise absolutely continuous, and the function t 0 (τ ) is
ˆ 0 (τ )) is a piecewise absolutely continuous function
Lipschitz continuous. Hence x̄(τ ) = x̄(t
whose set of discontinuity points is contained in ζ . Further,
ˆ 0 (τ )) d x̄ˆ
d x̄(τ ) d x̄(t dt 0 (τ ) d x̄ˆ
= = · = χE (τ ) . (2.226)
dτ dτ dt 0 dτ dt 0
t=t (τ ) t=t (τ )
dt 0
Since χE2 = χE = dτ , we have
2 2 2
τf d x̄(τ ) τf d x̄ˆ 0 dt 0 (τ ) tf0 ˆ
d x̄(t)
dτ = (t (τ )) dτ = dt < +∞.
τ0 dτ τ0 dt dτ t00 dt
The function t 0 (τ ), and hence the function ū(t ˆ 0 (τ )) = ū(τ ), assumes finitely many values
on [τ0 , τf ] \ E ; hence the second integral in the sum is finite. For the first integral, we have
dt 0 (τ ) 0
ˆ 0 (τ ))2 dt (τ ) dτ
ū(τ )2 dτ = ū(τ )2 χE dτ = ū(τ )2 dτ = ū(t
E E E dτ E dτ
= ˆ 2 dt =
ū(t) ˆ 2 dt < +∞,
ū(t)
Ê [t00 ,tf0 ]
τ
since ūˆ is Lebesgue square integrable. Hence, τ0f ū(τ )2 dτ < +∞, i.e., ū(·) ∈
L2 ([τ0 , tf ], Rd(u) ). Further, condition (2.223) implies condition (2.216), since t 0 (τ0 ) = t00 ,
ˆ 0 ), x̄(τf ) = x̄(t
t 0 (τf ) = tf0 , and, therefore, x̄(τ0 ) = x̄(t ˆ 0 ). Consider the variational equation
0 f
(2.224). Making the change t = t 0 (τ ) in it, multiplying by χE (τ ), and taking into account
(2.226), we obtain
d x̄
= χE (τ )(fx
(t 0 (τ ), x 0 (τ ), u0 (τ ))x̄(τ ) + fu
(t 0 (τ ), x 0 (τ ), u0 (τ ))ū(τ )).
dτ
But χE (τ ) = ϕ(η0 (τ ))z0 . Therefore, the variational equation (2.217) holds for x̄ and ū.
Finally, we show that the jump conditions (2.218) hold. Note that
tk = t 0 (τk ), τk ∈ E , k = 1, . . . , s. (2.227)
Consequently, each τk is a continuity point of the function η0 (τ ) and
ϕ(η0 (τk ))z0 = 1, k = 1, . . . , s. (2.228)
2.6. Completing the Proof of Theorem 2.4 113
Analogously,
ˆ k ),
[x̄](τk ) = [x̄](t k = 1, . . . , s. (2.230)
Conditions (2.218) follow from (2.225), (2.229), and (2.230) and the relation ξ̄ = ξ̄ˆ . There-
ζ
fore, all the conditions defining the cone K0 hold for the tuple z̄N .
s
2ζ (λN , z̄N ) = (D k (H N )ξ̄k2 + 2[HxN ]k x̄av
k
ξ̄k )
k=1 τf (2.231)
+lpp p̄, p̄ + Hww
N
w̄(τ ), w̄(τ ) dτ .
τ0
Here,
p̄ = (x̄(τ0 ), x̄(τf )), lpp = lpp (x̂00 , x̂f0 ; α0 , α, β), (2.232)
N
Hww = (ϕ(η0 (τ ))z0 )Hww (t 0 (τ ), x 0 (τ ), u0 (τ ), ψx (τ ))
= χE (τ )Hww (t 0 (τ ), x 0 (τ ), u0 (τ ), ψx (τ )). (2.233)
Finally, by definition,
d N
D (H ) = − (k H )
k N
, k = 1, . . . , s.
dτ
τ =τk
114 Chapter 2. Quadratic Conditions in the Calculus of Variations
Since τk is a continuity point of the function η0 (τ ) and ϕ(η0 (τk ))z0 = 1, we have
(k H N )(τ )
= (ϕ(η0 (τk ))z
0 )ψ (τ ) f (t 0 (τ ), x 0 (τ ), u0 (τ +)) − f (t 0 (τ ), x 0 (τ ), u0 (τ −))
x k k
= ψ̂x (t 0 (τ )) f (t 0 (τ ), x̂ 0 (t 0 (τ )), û0 (tk +)) − f (t 0 (τ )), x̂ 0 (t 0 (τ )), û0 (tk −))
= (k H )(t 0 (τ )).
Consequently,
d N d dt 0
D (H ) = − (k H )
k N
=−
(k H ) ·
dτ τ =τk dt
t=tk dτ τ =τk
= D k (H )χE (τk ) = D k (H ). (2.235)
ζ
Let λ = (α0 , α, β, ψ̂x (t)) be the element corresponding to the tuple λN ∈ M0 , where
ψ̂x (t) = ψx (τ 0 (t)). Then λ ∈ M0 according to Lemma 2.89. Recall that by definition, the
quadratic form λ (z̄) ˆ for the problem (2.169), (2.170) at the point ŵ0 corresponding to the
tuple λ of Lagrange multipliers and calculated at the element z̄ˆ has the form
s tf0
ˆ =
2λ (z̄) k ˆ
(D k (H )ξ̄ˆk2 + 2[Hx ]k x̄ˆav ˆ p̄
ξ̄k ) + lpp p̄, ˆ + ˆ w̄
Hww w̄, ˆ dt. (2.237)
k=1 t00
ˆ ˆ
lpp p̄, p̄ = lpp p̄, p̄ , (2.238)
ˆ 0 ) = x̄(t
since x̄(t ˆ 0 (τ0 )) = x̄(τ0 ) and x̄(t
ˆ 0 ) = x̄(t
ˆ 0 (τf )) = x̄(τf ). Formulas (2.231) and
0 f
(2.234)–(2.238), imply
ˆ = ζ (λN ; z̄N ).
λ (z̄) (2.239)
Lemma 2.91. Let z̄ˆ be an arbitrary element (2.219) of the critical cone K, and let z̄N
ζ ζ
be element (2.221) of the cone K0 ⊂ K ζ obtained by formulas (2.220). Let λN ∈ M0 be
an arbitrary tuple (2.182), and let λ = (α0 , α, β, ψ̂x (t)) ∈ M0 be the tuple with the same
components α0 , α, β and with ψ̂x (t) = ψx (τ 0 (t)) corresponding to it by Lemma 2.89. Then
relation (2.239) holds for the quadratic forms.
Take an arbitrary element z̄ˆ ∈ K. According to Lemma 2.90, the element z̄N ∈ K0 ⊂ K ζ
ζ
ζ
corresponds to it by formulas (2.220). By (2.240), for z̄N , there exists λN ∈ M0 such that
Since z̄ˆ is an arbitrary element of K, this implies condition (2.241). Therefore, Condition
A also holds in case (b). The theorem is completely proved.
This fact, which had been previously absent in the classical calculus of variations, was
first discovered by Milyutin when studying problems of the calculus variations and optimal
control. We use this fact in this section.
We first define the concept of strong minimum, which will be considered here. It is
slightly different from the usual concept from the viewpoint of strengthening. The usual
concept used in the calculus of variations corresponds to the concept of minimum on the set
of sequences {δw} in the space W such that
δx
C → 0. It is not fully correct to extend it
to the canonical problem without any changes. Indeed, in the classical
calculus of variations,
t
it is customary to minimize an integral functional of the form J = t0f F (t, x, u) dt, where
u = ẋ. In passing to the canonical problem, we write the integral functional as the terminal
functional: J = y(tf ) − y(t0 ), but there arises a new state variable y such that ẏ = F (t, x, u).
Clearly, the requirement
δy
C → 0 must be absent in the canonical problem if we do not
want to distort the original concept of strong minimum in rewriting the problem.
How can this be taken into account if we have the canonical form in advance, and
it is not known from which problem it originates? It is easy to note that the new state
variables y arising in rewriting the integral functionals are characterized by the property
that they affinely enter the terminal functionals of the canonical form and are completely
absent in the control system of the canonical form. These variables are said to be unessential
and the other variables are said to be essential. In defining the strong minimum, we take
into account only the essential variables.
Let us give the precise definition. As before, we consider the canonical problem
(2.1)–(2.4) on a fixed closed interval of time [t0 , tf ].
Denote by x the vector composed of the essential components of the vector x. Simi-
larly, denote by δx the vector-valued function composed of the essential components of the
variation δx.
Denote by S the set of sequences {δw} in the space W such that |δx(t0 )|+
δx
C → 0.
Let us give the following definition for problem (2.1)–(2.4).
Definition 2.93. We say that w0 is a strong minimum point (with respect to the essential
state variables) if it is a minimum point on S .
In what follows, the strong minimum with respect to the essential variables will be
called the strong minimum, for brevity. By the strict strong minimum we mean the strict
minimum on S . Since ⊂ S , the strong minimum implies the Pontryagin minimum.
S
Denote by the set of sequences {δw} in W satisfying conditions (a) and (b) and
also the following additional conditions:
(c) starting from a certain number, (p0 + δp) ∈ P , (t, w 0 + δw) ∈ Q.
(d) σ (δw) → 0, where σ is the violation function (2.164).
Conditions (c) and (d) hold on every sequence “violating the minimum.” Therefore, we
S
may treat the bounded strong minimum as a minimum on . We will proceed in this way
in what follows, since we will need conditions (c) and (d). By the strict bounded strong
S S
minimum, we mean the strict minimum on . Since ⊂ ⊂ S , the strong minimum
implies the bounded strong minimum, and the latter implies the Pontryagin minimum.
A remarkable property is that the sufficient conditions obtained in Section 2.5 guaran-
tee not only the Pontryagin minimum but also the bounded strong minimum. This follows
from the theorem, which now will be proved. In what follows, w 0 is an admissible point
satisfying the standard assumptions of Section 2.1.
Theorem 2.95. For a point w 0 , let there exist an admissible function (t, u) and a constant
S
C > 0 such that the set M(C) is nonempty. Then = , and hence the Pontryagin
minimum is equivalent to the bounded strong minimum.
S
t
Proposition 2.96. Let λ ∈ 0 , {δw} ∈ . Then ( t0f δH λ dt)+ → 0, where a+ = max{a, 0}
and δH λ = H (t, x 0 + δx, u0 + δu, ψ) − H (t, x 0 , u0 , ψ).
Proof. Let λ ∈ 0 , and let δw be an admissible variation with respect to Q, i.e., (t, w 0 +
δw) ∈ Q. Then
tf
δl −
λ
ψ(δ ẋ − δf ) dt ≤ const σ (δw). (2.243)
t0
On the other hand, we have shown earlier that the conditions −ψ̇ = Hxλ , ψ(t0 ) = −lxλ0 , and
ψ(tf ) = lxλf imply
tf t tf tf
ψδ ẋ dt = ψδx tf − ψ̇δx dt = lp δp + Hxλ δx dt.
0
t0 t0 t0
Taking into account that ψδf = δH λ , we obtain from inequality (2.243) that
tf tf
δl λ − lp δp − Hxλ δx dt + δH λ dt ≤ const σ (δw). (2.244)
t0 t0
118 Chapter 2. Quadratic Conditions in the Calculus of Variations
S
tf
Let {δw} ∈ . The condition
δx
C → 0 implies Hxλ δx dt → 0 and (δl λ − lp δp) → 0.
t
t0
Moreover, σ (δw) → 0. Therefore, condition (2.244) implies ( t0f δH λ dt)+ → 0. The
proposition is proved.
Proposition 2.97. Let there exist an admissible function (t, u) and a constant C > 0 such
that the set M(C) is nonempty. Then the following condition holds for any sequence
S
t
{δw} ∈ : t0f (t, u0 + δu) dt → 0.
tf
Proof. Let C > 0, and let λ ∈ M(C). According to Proposition 2.96, ( t0 δH λ dt)+ → 0.
Represent δH λ as δH λ = δ̄x H λ + δu H λ , where
δ̄x H λ = H λ (t, x 0 + δx, u0 + δu) − H λ (t, x 0 , u0 + δu),
δu H λ = H λ (t, x 0 , u0 + δu) − H λ (t, x 0 , u0 ).
The conditions
δx
C →
0, (t, x 0 , u0 + δu) ∈ C, where C ⊂ Q is a compact set, imply
t
δ̄x H λ
∞ → 0. Hence ( t0f δu H λ dt)+ → 0. Further, the condition λ ∈ M(C) implies
t
t
δu H λ ≥ C(t, u0 + δu) ≥ 0. Consequently, t0f δu H λ dt ≥ C t0f (t, u0 + δu) dt ≥ 0. This
tf
t
and the condition ( t0 δu H λ dt)+ → 0 imply t0f (t, u0 + δu) dt → 0. The proposition is
proved.
Proposition 2.98. Let C ⊂ Q be a compact set, and let the variation δu ∈ L∞ (, Rd(u) ) be
such that (t, x 0 , u0 + δu) ∈ C a.e. on [t0 , tf ]. Then we have the estimate
δu
1 ≤
t
const( t0f (t, u0 + δu) dt)1/2 , where const depends only on C.
Proof. Let V be the neighborhood of the compact set u0 from the definition of the function
(t, u) in Section 2.3.7, and let V 0 and V ∗ be subsets of the neighborhood V defined in
Section 2.3.1. Represent δu as δu = δuv + δuv , where
δu if (t, u0 + δu) ∈ V,
δuv = δuv = δu − δuv .
0 otherwise,
Further, let the representation δuv = δu0 + δu∗ correspond to the partition V = V 0 ∪ V ∗ :
δuv if (t, u0 + δuv ) ∈ V 0 , ∗ δuv if (t, u0 + δuv ) ∈ V ∗ ,
δu =0
δu =
0 otherwise, 0 otherwise.
Then
δu
1 =
δuv
1 +
δu0
1 +
δu∗
1 . (2.245)
Let us estimate each of the summands separately.
(1) Since (t, x 0 , u0 + δuv ) ∈ C and (t, u0 + δuv ) ∈
/ V for δuv = 0, by the definition
of the function (t, u), there exists ε = ε(C) > 0 such that (t, u0 + δuv ) ≥ ε if δuv = 0.
2.7. Sufficient Conditions for the Strong Minimum 119
Proof of Theorem 2.95. Assume that the conditions of the theorem hold. Let us prove
the inclusion ¯ S ⊂ . Let {δw} ∈
¯ S . Then it follows from Propositions 2.97 and 2.98
that
δu
1 → 0. Further, the condition σ (δw) → 0 implies
δ ẋ − δf
1 → 0. But δf =
δ̄x f + δu f , where
δ̄x f = f (t, x 0 + δx, u0 + δu) − f (t, x 0 , u0 + δu),
δu f = f (t, x 0 , u0 + δu) − f (t, x 0 , u0 ).
Since
δx
C → 0 and there exists a compact set C ⊂ Q such that (t, x 0 , u0 + δu) ∈ C
starting from a certain number, we have
δ̄x f
∞ → 0. The conditions
δu
1 → 0 and
(t, x 0 , u0 + δu) ∈ C imply
δu f
1 → 0. Consequently,
δ ẋ
1 ≤
δ ẋ − δf
1 +
δf
1 ≤
δ ẋ − δf
1 +
δ̄x f
∞ (tf − t0 ) +
δu f
1 → 0.
The conditions
δ ẋ
1 → 0 and |δx(t0 )| → 0 imply
δx
1,1 → 0. Therefore, {δw} ∈ . The
¯ S ⊂ is proved. The converse inclusion always holds. Therefore,
inclusion ¯ S = .
The theorem is proved.
Theorem 2.99. Let the set Leg+ (M0+ ) be nonempty. Then = ¯ S , and hence the Pon-
tryagin minimum is equivalent to the bounded strong minimum.
Proof. Assume that Leg+ (M0+ ) is nonempty. Choose an arbitrary compact set M ⊂
Leg+ (M0+ ), e.g., a singleton. According to Lemma 2.73, there exist an admissible function
(t, u) and a constant C > 0 such that M ⊂ M(C). Therefore, M(C) is nonempty. Then
= ¯ S by Theorem 2.95. The theorem is proved.
Definition 2.100. We say that the point w 0 is a point of bounded strong γ -sufficiency if
¯ S such that σ = o(γ ) on it.
there is no sequence {δw} ∈
The condition that the set Leg+ (M0+ ) is nonempty is a counterpart of Condition B.
Therefore, Theorems 2.82 and 2.99 imply the following theorem.
The bounded strong γ -sufficiency implies the strict bounded strong minimum. There-
fore, Theorem 2.101 implies the following theorem.
Theorem 2.102. Condition B is sufficient for the strict bounded strong minimum at the
point w0 .
At this point, we complete the consideration of conditions for the bounded strong
minimum. Before passing to sufficient conditions for the strong minimum, we prove some
estimate for the function , which will be needed in what follows.
2.7. Sufficient Conditions for the Strong Minimum 121
(see Proposition 2.16). Let us show that the same estimate also holds on any Pontryagin
sequence but with a constant depending on the order γ and the sequence.
Lemma 2.103. Let the set 0 be nonempty. Let (t, u) be an admissible function, and let
γ be the higher order corresponding to it. Then for any sequence {δw} ∈ , there exists a
constant C > 0 such that
max |(λ, δw)| ≤ Cγ (δw).
co
0
Briefly, this property will be written as maxco || ≤ O(γ ) .
0
Proof. Proposition 2.16 implies the following assertion: there exist constants C > 0 and
ε > 0 and a neighborhood V of the compact set u0 such that the conditions δw ∈ W ,
δx
C ≤ ε, and (t, u0 + δu) ∈ V imply the estimate maxco |(λ, δw)| ≤ Cγ (δw). Let us
0
use this estimate. Let {δw} be an arbitrary sequence from . For each member δw = (δx, δu)
of the sequence {δw}, represent δu as δu = δuV + δuV , where
δu if (t, u0 + δu) ∈ V ,
δuV = δuV = δu − δuV .
0 otherwise,
We set δwV = (δx, δuV ) and δwV = (0, δuV ). Owing to a possible decrease of V , we
can assume that both sequences {δwV } and {δwV } are admissible with respect to Q; i.e.,
the conditions (t, x 0 + δx, u0 + δuV ) ∈ Q and (t, x 0 , u0 + δuV ) ∈ Q hold starting from a
certain number (such a possibility follows from the definition of Pontryagin sequence). We
assume that this condition holds for all numbers and
δx
C ≤ ε holds for all numbers.
We set MV = {t | δuV = 0}. The definitions of admissible function (t, u) and Pontryagin
sequence imply the existence of constants 0 < a < b such that a ≤ (t, u0 + δuV ) ≤ b | M V
for all
members of the sequence. This implies a· meas MV ≤ γ V ≤ b· meas M V , where
t
γ V = t0f (t, u0 + δuV ) dt = γ (δw V ). Therefore, γ V and meas MV are of the same order
of smallness. Moreover, the definitions of γ , δwV , and δwV imply γ (δw) = γ (δwV ) +
γ (δwV ), or, briefly, γ = γV + γ V . In what follows, we will need the formula
δf = δV f + δ̄ Vf , (2.249)
where δf = f (t, w 0 + δw) − f (t, w 0 ), δV f = f (t, w 0 + δwV ) − f (t, w 0 ), and δ̄ Vf =
f (t, x 0 + δx, u0 + δuV ) − f (t, x 0 + δx, u0 ). The fulfillment of this formula is proved by
the following calculation:
δf = f (t, x 0 + δx, u0 + δu) − f (t, x 0 + δx, u0 + δuV ) + δV f
= f (t, x 0 + δx, u0 + δu) − f (t, x 0 + δx, u0 + δuV ) χ V + δV f
= δ̄ Vf χ V + δV f = δ̄ Vf + δV f ,
122 Chapter 2. Quadratic Conditions in the Calculus of Variations
where χ V is the characteristic function of the set M V . Formula (2.249) implies the following
representation for (λ, δw) on the sequence {δw}:
tf tf
(λ, δw) := δl − λ
ψδ ẋ dt + ψδf dt
t0 t0
tf tf tf
= δl λ − ψδ ẋ dt + ψδV f dt + ψ δ̄ Vf dt
t0 t0 t0
tf
= (λ, δwV ) + δ̄ V H λ dt,
t0
According to the choice of V and ε, the first term of the sum on the right-hand side of the
inequality is estimated through γ (δwV ). The second term of this sum is estimated through
meas MV and hence through γ V . Since γV ≤ γ and γ V ≤ γ , the total sum is estimated
through γ with a certain positive constant as a multiplier. The lemma is proved.
Theorem 2.104. Let the following conditions hold for the point w0 :
(1) There exists a nonempty compact set M ⊂ Leg+ (M0+ ) such that
(a) for a certain C > 0, maxλ∈M λ (z̄) ≥ C γ̄ (z̄) for all z̄ ∈ K, i.e., Condition B
holds;
(b) for any ε > 0, there exist δ > 0 and a compact set C ⊂ Qtu such that for
all t ∈ [t0 , tf ] \ , the conditions (t, w) ∈ Q, |x − x 0 (t)| ≤ δ, (t, u) ∈
/ C imply
minλ∈M δ̄u H λ ≥ −ε|δ̄u f |;
(2) there exist δ0 > 0, ε0 > 0, a compact set C0 ⊂ Qtu , and an element λ0 ∈ M0 such
that for all t ∈ [t0 , tf ] \ the conditions (t, w) ∈ Q, |x − x 0 (t)| < δ0 , (t, u) ∈
/ C0 imply
δ̄u H λ0 ≥ ε0 |δ̄u f | > 0.
Then w0 is a strict strong minimum point.
Remark 2.105. For the point w 0 , let there exist δ0 , ε0 , C0 , and λ0 satisfying condition (2)
of Theorem 2.104. Moreover, let λ0 ∈ Leg+ (M0+ ), and for a certain C > 0, let λ0 (z̄) ≥
C γ̄ (z̄) for all z̄ ∈ K. Then, as is easily seen, all the conditions of Theorem 2.104 hold, and,
therefore, w0 is a strict strong minimum point.
2.7. Sufficient Conditions for the Strong Minimum 123
Proposition 2.106. The following conditions hold : (a) ϕ C → 0, (b) {δwC } ∈ , and hence
γC → 0.
Proof. By (2.252), we have the following for the sequence {δw} and the element λ = λ0 :
tf tf tf
δl − ψδ ẋ dt + δC H dt + δ̄ C H dt ≤ 0 (2.253)
t0 t0 t0
124 Chapter 2. Quadratic Conditions in the Calculus of Variations
where
δ̂Cx H = H (t, x 0 + δx, u0 + δuC ) − H (t, x 0 , u0 + δuC ),
δCu H = H (t, x 0 , u0 + δuC ) − H (t, x 0 , u0 ).
Then we obtain from (2.253) that
tf tf tf tf
δl − ψδ ẋ dt + δ̂Cx H dt + δCu H dt + δ̄ C H dt ≤ 0. (2.254)
t0 t0 t0 t0
The conditions
δx
C → 0 and (t, u0 + δuC ) ∈ C imply
tf
δ̂Cx H dt → 0. (2.255)
t0
We continue the proof of the theorem. Consider the following two possible cases for
the sequence {δw}.
Case (a). Assume that there exist a compact set C satisfying conditions (2.251) and a sub-
sequence of the sequence {δw} such that the following conditions hold on this subsequence:
Assume that these conditions hold for the sequence {δw} itself. Inequality (2.252) and the
conditions λ0 ∈ M0 ⊂ 0 imply that the following inequality holds on the sequence {δw}:
tf
δ̄ C H λ0 dt ≤ −λ0 (δwC ). (2.262)
t0
As was already mentioned in the proof of Proposition 2.106, the following inequalities hold
for all members of the sequence {δw} having sufficiently large numbers:
tf
ε0 ϕ C ≤ δ̄ C H λ0 dt. (2.263)
t0
On the other hand, according to Lemma 2.103, the conditions {δwC } ∈ , λ0 ∈ M0 ⊂ 0
imply the estimate
−λ0 (δwC ) ≤ O(γC ). (2.264)
We obtain from (2.261)–(2.264) that 0 < ϕ C ≤ O(γC ) = o(ϕ C ). This is a contradiction.
Case (b). Consider the second possibility. Assume that for any compact set C satisfying
conditions (2.251), there exists a constant N > 0 such that the following estimate holds on
the sequence {δw}:
ϕ C ≤ N γC . (2.265)
We show that this also leads to a contradiction. We will thus prove the theorem. First
of all, we note that the constant N in (2.265) can be chosen common for all compact sets
C satisfying conditions (2.251). Indeed, let N0 correspond to a compact set C0 , i.e., the
estimate ϕ C0 ≤ N0 γC0 holds on the sequence {δw}. Let C be an arbitrary compact set such
that (2.251) hold. Then we have the following on the sequence {δw} for all sufficiently large
numbers: ϕ C ≤ ϕ C0 ≤ N0 γC0 ≤ N0 γC . Therefore, N = N0 is also appropriate for C. Also,
we note that for any C satisfying (2.251), there exists a (serial) number of the sequence
starting from which γC > 0. Indeed, otherwise, there exist a compact set C, satisfying
(2.251) and a subsequence of the sequence {δw} such that γC = 0 on the subsequence, and
then ϕ C = 0 by (2.265). By assumption (2) of the theorem, this implies that all members of
the subsequence vanish. The latter is impossible, since the sequence {δw} contains nonzero
members by assumption.
126 Chapter 2. Quadratic Conditions in the Calculus of Variations
Now let the compact set C satisfy conditions (2.251). Inequality (2.252) and the
inclusion M ⊂ 0 imply that the following inequality holds on the sequence {δw}:
tf
max (λ, δwC ) ≤ − min δ̄ C H λ dt. (2.266)
M M t0
Obviously,
tf tf
min δ̄ C H λ dt ≥ min δ̄ C H λ dt. (2.267)
M t0 t0 M
Condition 1(b) of the theorem implies that, for any ε > 0, there exists a compact set C
satisfying (2.251) such that
tf
min δ̄ C H λ dt ≥ −ε· ϕ C (2.268)
t0 M
for all sufficiently large numbers of the sequence. We obtain from (2.265)–(2.268) that
for any ε > 0, there exists a compact set C satisfying (2.251) such that following estimate
holds starting from a certain number:
where N = N0 is independent of C.
We now estimate the left-hand side of inequality (2.269) from below. For this pur-
pose, we show that for any compact set C satisfying (2.251), the sequence {δwC } belongs
to o(√γ ) . Let C be an arbitrary compact set satisfying (2.251). According to Proposi-
tion 2.106, {δwC } ∈ . Since σ (δw) = 0, we have (δJ )+ + Fi+ (p 0 + δp) + |δK| = 0
on the whole sequence. Moreover, the conditions δ ẋ = δf , δf = δC f + δ̄ Cf ,
δ¯C f
1 =
ϕ C ≤ O(γC ) imply
δ ẋ − δC f
1 ≤ O(γC ). Therefore, {δwC } ∈ σ γ ⊂ o(√γ ) . But then,
by Lemma 2.83, condition 1(a) of the theorem implies that there exists a constant CM > 0
such that, starting from a certain number,
Moreover, the constant CM is independent of the sequence from o(√γ ) and hence is
independent of C. Comparing estimates (2.269) and (2.270), we obtain the following
result: For any ε > 0, there exists a compact set C satisfying (2.251) such that, starting from
a certain number, 0 < CM γC ≤ εN γC . Choosing 0 < ε < CM /N , we obtain a contradiction.
The theorem is proved.
Chapter 3
In Sections 3.1 and 3.2 of this chapter, following [92], we extend the quadratic conditions
obtained in Chapter 2 to the general problem with the local relation g(t, x, u) = 0 using
a special method of projection contained in [79]. In Section 3.3, we extend these condi-
tions to the problem on a variable interval of time using a simple change of time variable.
In Section 3.4, we formulate (without proofs) quadratic conditions in an optimal control
problem with the local relations g(t, x, u) = 0 and ϕ(t, x, u) ≤ 0.
127
128 Chapter 3. Quadratic Conditions for Optimal Control Problems
where p0 = (x 0 (t0 ), x 0 (tf )), δw = (δx, δu), and δp = (δx(t0 ), δx(tf )). A (strict) minimum
on (see Section 2.1.3) is said to be a (strict) Pontryagin minimum.
Our goal is to obtain quadratic conditions in problem (3.1), (3.2) using the quadratic
conditions obtained in problem (2.1)–(2.4). In this case, we will use the same method for
passing to a problem without local constraints, which was already used in [79, Section 17,
Part 1] for obtaining first-order conditions. Recall that in [79, Section 17, Part 1], we have
introduced the set
G = {(t, x, u) ∈ Q | g(t, x, u) = 0}.
We have shown that there exist a neighborhood Q1 ⊂ Q of the set G and a continuously
differentiable function U (t, x, u) : Q1 → Rd(x) such that
Proof. The property that w0 is a Pontryagin minimum point in the problem (3.1), (3.2)
implies that w0 is an admissible point in the problem (3.1), (3.2), and then by the second
property (3.4) of the projection, w0 is also admissible in the problem (3.1), (3.5). Suppose
that w0 is not a Pontryagin minimum point in the problem (3.1), (3.5). Then there exist
C ∈ Q1 and a sequence {δw} = {(δx, δu)} such that
δx
1,1 → 0,
δu
1 → 0, (3.6)
and the following conditions hold for all members of the sequence:
where p0 = (x 0 (t0 ), x 0 (tf )) and δp = (δx(t0 ), δx(tf )). Therefore, {δw} is a Pontryagin
sequence “violating the Pontryagin minimum” at the point w 0 in the problem (3.1), (3.5).
We set {δw1 } = {(δx, δu1 )}, where
δx
C → 0 and (t, x 0 + δx, u0 + δu) ∈ C ⊂ Q1 (3.11)
imply (t, x 0 , u0 + δu) ∈ Q1 for all sufficiently large serial numbers (here, we use the com-
pactness of C and the openness of Q1 ). Representation (3.10) implies
δu1
1 ≤
U (t, x 0 + δx, u0 + δu) − U (t, x 0 , u0 + δu)
∞ (tf − t0 )
+
U (t, x 0 , u0 + δu) − U (t, x 0 , u0 )
1 .
This implies g(t, x 0 + δx, u0 + δu1 ) = 0. Finally, we note that the sequence {w 0 + δw1 }
satisfies the differential equation of system (3.2):
transforms the Pontryagin sequence {δw} of the system (3.1), (3.5) at the point w 0 into
the Pontryagin sequence {δw1 } of the system (3.1), (3.2) at the same point. Since the
sequence {δw} “violates” the Pontryagin minimum in the problem (3.1), (3.5) at the point
w 0 , conditions (3.8) hold. But these conditions can be also referred to as the sequence
{δw1 }, since the members δx of these two sequence coincide. Therefore, {δw 1 } “violates”
the Pontryagin minimum at the point w0 in the problem (3.1), (3.2). Therefore, the absence
130 Chapter 3. Quadratic Conditions for Optimal Control Problems
of the Pontryagin minimum at the point w 0 in the problem (3.1), (3.5) implies the same case
in the problem (3.1), (3.2). This implies what was required.
3.1.2 Set M0
For the problem (3.1), (3.2), we set l = α0 J + αK, H = ψf , and H̄ = H + νg, where
ν ∈ (Rd(g) )∗ . Therefore, H̄ = H̄ (t, x, u, ψ, ν). We also set
H (t, x, ψ) = min ψf (t, x, u).
{u|(t,x,u)∈G}
For problem (3.1), (3.2) and the point w 0 , we introduce the set M0 consisting of tuples
λ = (α0 , α, β, ψ(t), ν(t)) such that
α0 ≥ 0, α ≥ 0, αF (p 0 ) = 0, α0 + |α| + |β| = 1, (3.12)
ψ(t0 ) = −lx0 , ψ(tf ) = lxf , (3.13)
− ψ̇ = H̄x , H̄u = 0, (3.14)
H (t, x 0 , u0 , ψ) = H (t, x 0 , ψ). (3.15)
Here, ψ(t) is an absolutely continuous function and ν(t) is a bounded measurable function.
All the derivatives are taken for p = p 0 and w = w 0 (t). The results of [79, Section 17,
Part 1] imply that the set M0 of problem (3.1), (3.5) at the point w0 can be represented
in this form. More precisely, the linear projection (α0 , α, β, ψ, ν) → (α0 , α, β, ψ) yields a
one-to-one correspondence between the elements of the set M0 of the problem (3.1), (3.2)
at the point w 0 and the elements of the set M0 of the problem (3.1), (3.5) at the same point.
To differentiate these two sets from one another, we denote the latter set by M0U . We will
equip all objects referring to the problem (3.1), (3.5) with the superscript U .
In what follows, we will assume that all assumptions of Section 2.1 hold for the
point w 0 , i.e., u0 (t) is a piecewise continuous function whose set of discontinuity points is
= {t1 , . . . , ts } ⊂ (t0 , tf ), and each point of the set is an L-point. The condition
−fu∗ (t, x 0 (t), u0 (t))ψ ∗ (t) = gu∗ (t, x 0 (t), u0 (t))ν ∗ (t), (3.16)
which is equivalent to the condition H̄u = 0, and also the full-rank condition (3.3) imply
that ν(t) has the same properties as u0 (t): the function ν(t) is piecewise continuous and
each of its point of discontinuity is an L-point which belongs to . To verify this, it suffices
to premultiply the above relation by the matrix gu (t, x 0 (t), u0 (t)),
−gu (t, x 0 (t), u0 (t))fu∗ (t, x 0 (t), u0 (t))ψ ∗ (t) = gu (t, x 0 (t), u0 (t))gu∗ (t, x 0 , (t), u0 (t))ν ∗ (t),
and use the properties of the functions g, f , x 0 , and u0 ; in particular, the property
| det gu (t, x 0 , u0 )gu∗ (t, x 0 , u0 )| ≥ const > 0,
3.1. Quadratic Necessary Conditions on a Fixed Time Interval 131
which is implied by the full-rank condition (3.3). Therefore, the basic properties of the
function ν(t) are proved exactly in the same way as in [79, Section 17, Part 1] for the
bounded measurable control u0 (t).
All the derivatives are taken for w = w0 (t). Setting ũ = Ux x̄ + Uu ū, we obtain
This is the usual variational equation, but for the pair w̃ = (x̄, ũ). Let us show that the pair
(x̄, ũ) also satisfies the condition
gx x̄ + gu ũ = 0. (3.19)
By the first condition in (3.4), we have
Differentiating this relation in x, u, and t, as in [79, Section 17, Part 1], we obtain
gx + gu Ux = 0, (3.21)
gu Uu = 0, (3.22)
gt + gu Ut = 0. (3.23)
These relations hold on Q1 , but it suffices to consider them only on the trajectory
(t, x 0 (t), u0 (t)), and, moreover, we now need only the first two conditions. By (3.21) and
(3.22), we have
Proposition 3.2. Let a pair of functions (x̄, ū) satisfy the variational equation (3.17) of
system (3.5). We set ũ = Ux x̄ + Uu ū. Then conditions (3.18) and (3.19) hold for (x̄, ũ).
In this proposition, x̄ ∈ P W 1,2 (, Rd(x) ), ū ∈ L2 (, Rd(u) ), and ũ ∈ L2 (, Rd(u) ),
where = [t0 , tf ]. Also, we are interested in the possibility of the converse passage from
conditions (3.18) and (3.19) to condition (3.17). For this purpose, we prove the following
proposition.
Proposition 3.3. Let a pair of functions (x̄, ũ) be such that gx x̄ + gu ũ = 0. Then set-
ting ū = ũ − Ux x̄, we obtain Uu ū = ū, and hence ũ = Ux x̄ + Uu ū. Here, as above, x̄ ∈
P W 1,2 (, Rd(x) ), ū ∈ L2 (, Rd(u) ), and ũ ∈ L2 (, Rd(u) ); all the derivatives are taken for
x = x 0 (t) and u = u0 (t).
132 Chapter 3. Quadratic Conditions for Optimal Control Problems
Proof. First of all, we note that properties (3.4) of the function U (t, x, u) imply the follow-
ing assertion: at each point (t, x, u) ∈ G, the finite-dimensional linear operator
Indeed, let (t, x, u) ∈ G, and then condition (3.22) implies that the image of operator (3.24)
is contained in subspace (3.25). The condition
is easily proved by using the Lyusternik theorem [28]. Indeed, let ū ∈ Lg (t, x, u), i.e.,
gu (t, x, u)ū = 0. Let ε → +0. Then
where |rg (ε)| = o(ε). By the Lyusternik theorem [28], there exists a “sequence” {ug (ε)}
such that ug (ε) ∈ Rd(u) , |ug (ε)| = o(ε), and, moreover, g(t, x, u + ε ū + ug (ε)) = 0, i.e.,
(t, x, u + εū + ug (ε)) ∈ G. Hence U (t, x, u + εū + ug (ε)) = u + ε ū + ug (ε). But U (t, x, u +
ε ū + ug (ε)) = U (t, x, u) + Uu (t, x, u)(ε ū + ug (ε)) + rU (ε), where |rU (ε)| = o(ε). We ob-
tain from the latter two conditions and the condition U (t, x, u) = u that ε ū + ug (ε) =
Uu (t, x, u)(ε ū + ug (ε)) + rU (ε). Dividing this relation by ε and passing to the limit as
ε → +0, we obtain ū = Uu (t, x, u)ū. Condition (3.26) is proved. Condition (3.26) holds at
each point (t, x, u) ∈ G, but we use it only at the trajectory (t, x 0 (t), u0 (t)) | t ∈ [t0 , tf ].
If a pair of functions x̄(t), ũ(t) satisfies the condition gx x̄ + gu ũ = 0, then by (3.21),
we have −gu Ux x̄ + gu ũ = 0, i.e., −Ux x̄ + ũ ∈ Lg (t, x 0 , u0 ). Then the condition Uu ū = ū
also holds for ū = −Ux x̄ + ũ, and hence ũ = Ux x̄ + Uu ū. The proposition is proved.
Proposition 3.4. Let a pair of functions (x̄, ũ) be such that conditions (3.18) and (3.19) hold :
x̄˙ = fx x̄ +fu ũ and gx x̄ +gu ũ = 0. We set ū = −Ux x̄ + ũ. Then Uu ū = ū, and the variational
equation (3.17) of system (3.5) holds for the pair of functions (x̄, ū) at the point w 0 .
Proof. Indeed,
Definition 3.5. The critical cone K of problem (3.1), (3.2) at the point w 0 is the set of
triples z̄ = (ξ̄ , x̄, ū) satisfying the following conditions:
3.1. Quadratic Necessary Conditions on a Fixed Time Interval 133
Let us compare this definition with the definition of the critical cone K U of the
problem (3.1), (3.5) at the point w0 . According to Section 2.1, the latter is defined by the
same conditions (3.27) and (3.28), the variational equation (3.17), and the jump condition
(3.30). Thus, all the conditions referring to the components ξ̄ and x̄ in the definitions of
two critical cones coincide. This and Proposition 3.2 imply the following assertion.
Lemma 3.6. Let z̄ = (ξ̄ , x̄, ū) be an arbitrary element of the critical cone K U of the problem
(3.1), (3.5) at the point w0 . We set ũ = Ux x̄ + Uu ū. Then z̃ = (ξ̄ , x̄, ũ) is an element of the
critical cone K of the problem (3.1), (3.2) at the point w0 .
Lemma 3.7. Let z̃ = (ξ̄ , x̄, ũ) be an arbitrary element of the critical cone K of the problem
(3.1), (3.2) at the point w0 . We set ū = ũ − Ux x̄. Then z̄ = (ξ̄ , x̄, ū) is an element of the
critical cone K U of the problem (3.1), (3.5) at the point w0 , and, moreover, Uu ū = ū, which
implies ũ = Ux x̄ + Uu ū.
This is the connection between the critical cones at the point w 0 in the problems
(3.1), (3.2) and (3.1), (3.5). We will need Lemma 3.6 later in deducing quadratic sufficient
conditions in the problem with a local equality; now, in deducing necessary conditions, we
use Lemma 3.7. Preparatorily, we find the connection between the corresponding quadratic
forms.
We omit the superscripts λ and λU in the notation. For each t ∈ [t0 , tf ], let us calculate the
quadratic form
Hww
U
w̄, w̄ = Hxx
U
x̄, x̄ + 2Hxu
U
ū, x̄ + Huu
U
ū, ū , (3.34)
134 Chapter 3. Quadratic Conditions for Optimal Control Problems
where x̄ ∈ Rd(x) , ū ∈ Rd(u) , and all the second derivatives are taken at the point (t, x, u) =
(t, x 0 (t), u0 (t)).
Let us calculate Hxx
U x̄, x̄ . It follows from (3.33) that
HxU = Hx + Hu Ux . (3.35)
Hxx
U x̄, x̄ = Hxx x̄, x̄ + 2Hxu (Ux x̄), x̄ + Huu (Ux x̄), (Ux x̄)
(3.36)
+ (Hu Uxx )x̄, x̄ ,
d(u)
where (Hu Uxx )x̄, x̄ = ( i=1 Hui Uixx )x̄, x̄ by definition. Further, let us calculate
Hxu
U ū, x̄ . Differentiating (3.35) in u and multiplying it by ū and x̄, we obtain
Hxu
U
ū, x̄ = Hxu (Uu ū), x̄ + Huu (Uu ū), (Ux x̄) + (Hu Uxu )ū, x̄ , (3.37)
d(u)
where (Hu Uxu )ū, x̄ = ( i=1 Hui Uixu )ū, x̄ . Finally, let us calculate Huu U ū, ū . It follows
by ū, we obtain
HuuU
ū, ū = Huu (Uu ū), (Uu ū) + (Hu Uuu )ū, ū , (3.38)
d(u)
where (Hu Uuu )ū, ū = ( i=1 Hui Uiuu )ū, ū . Formulas (3.34), (3.36)–(3.38) imply
Hww
U w̄, w̄ = Hxx x̄, x̄ + 2Hxu (Ux x̄), x̄ + Huu (Ux x̄), (Ux x̄)
+ (Hu Uxx )x̄, x̄ + 2Hxu (Uu ū), x̄ + 2Huu (Uu ū), (Ux x̄)
+ 2(Hu Uxu )ū, x̄ + Huu (Uu ū), (Uu ū) + (Hu Uuu )ū, ū (3.39)
= Hxx x̄, x̄ + 2Hxu ũ, x̄ + Huu ũ, ũ
+ (Hu Uxx x̄, x̄ + 2(Hu Uxu )ū, x̄ + (Hu Uuu )ū, ū ,
gixx x̄, x̄ + 2gixu (Ux x̄), x̄ + giuu (Ux x̄), (Ux x̄) + giu Uxx x̄, x̄ = 0,
i = 1, . . . , d(g). Multiplying each of these relations by the ith component νi of the vector-
valued function ν(t), summing with respect to i, and using the relation Hu + νgu = 0, we
obtain
νgxx x̄, x̄ + 2νgxu (Ux x̄), x̄ + νguu (Ux x̄), (Ux x̄) − Hu Uxx x̄, x̄ = 0, (3.41)
where νgxx = νi gixx , νgxu = νi gixu , and νguu = νi giuu . Differentiating the same
relations (3.40) in u and multiplying by ū and x̄, we obtain
gixu (Uu ū), x̄ + giuu (Uu ū), (Ux x̄) + (giu Uxu )ū, x̄ = 0.
3.1. Quadratic Necessary Conditions on a Fixed Time Interval 135
Multiplying each of these relations by 2νi , summing with respect to i, and using the property
that Hu + νgu = 0, we obtain
2νgxu (Uu ū), x̄ + 2νguu (Uu ū), (Ux x̄) − 2(Hu Uxu )ū, x̄ = 0, (3.42)
where νgxu = νi gixu and νguu = νi giuu . Finally, differentiating in u the relations
giu Uu = 0, i = 1, . . . , d(g), which hold on Q1 , and twice multiplying the result by ū, we
obtain giuu (Uu ū), (Uu ū) + (giu Uuu )ū, ū = 0. Multiplying each of these equations by νi
and using the property that Hu + νgu = 0, we obtain
Huu
U
ū, ū = Huu (Uu ū), (Uu ū) + νguu (Uu ū), (Uu ū) . (3.44)
Huu
U
ū, ū = H̄uu (Uu ū), (Uu ū) . (3.45)
We will use this relation later in Section 3.2. Summing relations (3.41)–(3.43), we obtain
Similarly to what was done for (k H )(t) in Section 2.3 (see Lemma 2.12), we can show
that the function (k H̄ )(t) is continuously differentiable at the point tk ∈ , and its deriva-
tive at this point coincides with −D k (H̄ ). Therefore, we can obtain the value of D k (H̄ )
calculating the left or right limit of the derivatives of the function (k H̄ )(t) defined by
formula (3.49):
d
D k (H̄ ) = − (k H̄ )(tk ).
dt
We now show that
D k (H̄ ) = D k (H U ), k = 1, . . . , s. (3.50)
Indeed, by definition,
Furthermore,
HxU = Hx + Hu Ux = Hx − νgu Ux = Hx + νgx = H̄x . (3.52)
Finally,
HtU = Ht + Hu Ut = Ht − νgu Ut = Ht + νgt = H̄t , (3.54)
since gt + gu Ut = 0. The formulas (3.48), (3.51)–(3.54) imply relation (3.50). Further, note
that relations (3.52) imply
where
[H̄x ]k = H̄x (tk , x 0 (tk ), u0k+ , ψ(tk ), ν k+ ) − H̄x (tk , x 0 (tk ), u0k− , ψ(tk ), ν k− ) (3.56)
is the jump of the function H̄x (t, x 0 (t), u0 (t), ψ(t), ν(t)) at the point tk ∈ .
For the problem (3.1), (3.2) and the point w0 , we define the following quadratic form
in z̄ = (ξ̄ , x̄, ū) for each λ = (α0 , α, β, ψ, ν) ∈ M0 :
s
tf
2λ (z̄) = D k (H̄ λ )ξ̄k2 + 2[H̄xλ ]k x̄av
k
ξ̄k + lpp
λ
p̄, p̄ + H̄ww
λ
w̄, w̄ dt. (3.57)
k=1 t0
Therefore, the quadratic form in the problem with local equality-type constraints is defined
in the same way as in the problem without local constraints; the only difference is that
instead of the function H = ψf in the definition of the new quadratic form, we must use
the function H̄ = H + νg.
3.1. Quadratic Necessary Conditions on a Fixed Time Interval 137
According to Section 2.1, the quadratic form takes the following form for the problem
(3.1), (3.5) and the point w 0 :
s
% &
U U U
2U λ (z̄) = D k H U λ ξ̄k2 + 2 HxU λ x̄av k
ξ̄k
k=1 tf (3.58)
λU U
+ lpp p̄, p̄ + H U λ w̄, w̄ dt.
t0
Here, as above, we have used the superscript U in the notation U of the quadratic form
in order to stress that this quadratic form corresponds to the problem (3.1), (3.5) being
considered for a given projection U (t, x, u). We have denoted by λU the tuple (α0 , α, β, ψ),
which uniquely defines the tuple λ = (α0 , α, β, ψ, ν) by the condition ψfu + νgu = 0. In
what follows, these tuples correspond to one another and belong to the sets M0U and M0
of the problems (3.1), (3.5) and (3.1), (3.2), respectively. Formulas (3.47), (3.50), (3.55),
(3.57), and (3.58) imply the following assertion.
Lemma 3.8. Let z̄ = (ξ̄ , x̄, ū) be an arbitrary element of the space Z2 (), and let z̃ =
(ξ̄ , x̄, ũ) = (ξ̄ , w̃), where ũ = Ux x̄ + Uu ū. Let λU be an arbitrary element of M0U , and let λ
U
be the corresponding element of M0 . Then λ (z̃) = U λ (z̄).
Theorem 3.9. If w 0 is a Pontryagin minimum point in the problem (3.1), (3.2), then the
following Condition A holds: the set M0 is nonempty and
where K is the critical cone at the point w 0 defined by conditions (3.27)–(3.31), λ (z̄)
is the quadratic form at the same point defined by (3.57), and M0 is the set of tuples of
Lagrange multipliers satisfying the minimum principle defined by (3.12)–(3.15).
Proof. Let w0 be a Pontryagin minimum point in the problem (3.1), (3.2). Then according
to Proposition 3.1, w0 is a Pontryagin minimum point in the problem (3.1), (3.5). Hence, by
Theorem 2.4, the following necessary Condition AU holds at the point w 0 in the problem
(3.1), (3.5): the set M0U is nonempty and
U
max U λ (z̄) ≥ 0 ∀ z̄ ∈ K U . (3.60)
M0U
Let us show that the necessary Condition A holds at the point w 0 in the problem (3.1),
(3.2). Let z̃ = (ξ̄ , x̄, ũ) be an arbitrary element of the critical cone K at the point w0 in
the problem (3.1), (3.2). According to Lemma 3.7, there exists a function ū(t) such that
ũ = Ux x̄ + Uu ū, and, moreover, z̄ = (ξ̄ , x̄, ū) is an element of the critical cone K U at
the point w0 in the problem (3.1), (3.5) (we can set ū = ũ − Ux x̄). Since the necessary
138 Chapter 3. Quadratic Conditions for Optimal Control Problems
U
Condition AU holds, there exists an element λU ∈ M0U such that U λ (z̄) ≥ 0. According
to Section 3.1.2, an element λ ∈ M0 corresponds to the element λU ∈ M0U . Moreover, by
U
Lemma 3.8, λ (z̃) = U λ (z̄). Therefore, λ (z̃) ≥ 0. Since z̃ is an arbitrary element of K,
Condition A holds. The theorem is proved.
Therefore, we have obtained the final form of the quadratic necessary condition for
the Pontryagin minimum in the problem with local equality-type constraints, Condition A,
in which there is no projection U . This condition is a natural generalization of the quadratic
necessary Condition A in the problem without local constraints.
y0 = 0, yf ≤ 0, (3.62)
1
ẏ = (u − U (t, x, u))2 , (3.63)
2
ẋ = f (t, x, U (t, x, u)), (t, x, u) ∈ Q1 , (3.64)
where x0 = x(t0 ), xf = x(tf ), y0 = y(t0 ), and yf = y(tf ).
Problem (3.61)–(3.64) is called the auxiliary problem V . By the superscript V we
denote all objects referring to this problem. Therefore, the auxiliary problem V differs from
the auxiliary problem (3.1), (3.5) or problem U by the existence of the additional constraints
(3.62) and (3.63).
If (y, x, u) is an admissible triple in problem (3.61)–(3.64), then y = 0, and the pair
w = (x, u) satisfies the constraints of problem (3.1), (3.2). Indeed, (3.62) and (3.63) imply
and then
ẋ(t) − f (t, x(t), u(t)) = ẋ(t) − f (t, x(t), U (t, x(t), u(t))) = 0,
g(t, x(t), u(t)) = g(t, x(t), U (t, x(t), u(t))) = 0,
since (t, x(t), u(t)) ∈ Q1 . The converse is also true: if w = (x, u) is an admissible pair in
problem (3.1), (3.2) and y = 0, then the triple (y, x, u) is admissible in Problem V , since
conditions (3.65) hold for it.
We will use the concept of a bounded strong minimum and also that of a bounded
strong γ V -sufficiency at the point (0, w 0 ) in the auxiliary problem V . Let us introduce
analogous concepts for problem (3.1), (3.2) at the point w 0 .
As in Section 2.7 (see Definition 2.92), we denote by δx the tuple of essential com-
ponents of the variation δx. (The definition of unessential components in problem (3.1),
(3.2) is the same as in problem (2.1)–(2.4), but now neither functions f nor g depends on
these components.)
Therefore, the violation function σ (δw) in problem (3.1), (3.2) contains no term
related to the local constraint g(t, x, u) = 0; however, in the definition of the bounded
strong γ -sufficiency of this problem, it is required that the sequence {w0 + δw} satisfies
this constraint. The local constraint does not have the same rights as the other constraints.
Obviously, the following assertion holds.
Proposition 3.12. The bounded strong γ -sufficiency at the point w0 in problem (3.1), (3.2)
implies the strict bounded strong minimum.
Our goal is to obtain a sufficient condition for the bounded strong γ -sufficiency in
problem (3.1), (3.2) using the sufficient condition for the bounded strong γ V -sufficiency in
the auxiliary problem V without local constraints. For this purpose, we prove the following
assertion.
Proof. Suppose that w0 is not a point of the bounded strong γ -sufficiency in problem
(3.1), (3.2). Then there exist a sequence {δw} containing nonzero members and a compact
set C ⊂ Q such that conditions (3.69) and (3.70) hold. We show that in this case, (0, w0 )
is not a point of bounded strong γ V -sufficiency in Problem V . Consider the sequence
{(0, δw)} with δy = 0. Condition (3.70) implies
U (t, x 0 + δx, u0 + δu) = u0 + δu, (3.72)
and then
δf = f (t, x 0 + δx, u0 + δu) − f (t, x 0 , u0 )
= f (t, x 0 + δx, U (t, x 0 + δx, u0 + δu)) − f (t, x 0 , u0 ) = δf U . (3.73)
It follows from (3.66)–(3.73) that
σ V (0, δw) = σ (δw) = o(γ (δw)) = o(γ V (0, δw)).
Therefore, (0, w 0 ) is not a point of the bounded strong γ V -sufficiency in Problem V . The
proposition is proved.
Next, we formulate the main result of this section: the quadratic sufficient conditions
for the bounded strong γ -sufficiency at the point w 0 in problem (3.1), (3.2). Then we
3.2. Quadratic Sufficient Conditions on a Fixed Time Interval 141
show that these conditions guarantee the bounded strong γ V -sufficiency in Problem V , and
hence, by Proposition 3.13, this implies the bounded strong γ -sufficiency at the point w 0
in problem (3.1), (3.2). This is our program. We now formulate the main result.
(3) the following condition C k− holds for each point tk ∈ : the form
(4) the following condition C k+ holds for each point tk ∈ : the quadratic form
Further, denote by M0+ the set of λ ∈ M0 such that the following conditions hold:
H (tk , x 0 (tk ), u, ψ(tk )) > H (tk , x 0 (tk ), u0k− , ψ(tk )) = H (tk , x 0 (tk ), u0k+ , ψ(tk )) (3.82)
if tk ∈ , u ∈ U(tk , x 0 (tk )), u ∈ / {u0k− , u0k+ }. (3.83)
Denote by Leg+ (M0+ ) the set of all strictly Legendrian elements λ ∈ M0+ .
142 Chapter 3. Quadratic Conditions for Optimal Control Problems
Definition 3.15. We say that Condition B holds at the point w 0 in problem (3.1), (3.2) if
the set Leg+ (M0+ ) is nonempty and there exist a compact set M ⊂ Leg+ (M0+ ) and a constant
C > 0 such that
max λ ≥ C γ̄ (z̄) ∀ z̄ ∈ K, (3.84)
λ∈M
where the quadratic form λ (z̄), the critical cone K for problem (3.1), (3.2), and the point
w0 were defined by relations (3.57) and (3.27)–(3.31), respectively, and
tf
γ̄ (z̄) = ξ̄ , ξ̄ + x̄(t0 ), x̄(t0 ) + ū(t), ū(t) dt (3.85)
t0
Theorem 3.16. If Condition B holds for the point w 0 in the problem (3.1), (3.2), then we
have the bounded strong γ -sufficiency at this point.
We now prove this theorem. As was said, by Proposition 3.13, it suffices to show
that Condition B guarantees the bounded strong γ V -sufficiency at the point (0, w0 ) in
the problem without local constraints. For this purpose, we write the quadratic sufficient
condition of Section 2.1 for the auxiliary Problem V at the point (0, w0 ).
if t ∈ [t0 , tf ] \ and (t, x 0 (t), u) ∈ Q1 . Let us analyze these conditions. Since yf0 = y 0 (tf ) =
0, condition (3.90) holds automatically, and we can exclude it from consideration. Further,
since HxV = HxU − ψy (u0 − U (t, w0 ))Ux (t, w0 ) = HxU , condition (3.93) is equivalent to the
condition
ψ̇x = −HxU . (3.96)
We now turn to the minimum condition (3.95). It follows from (3.86) and (3.97) that it is
equivalent to the condition
1
H U (t, x 0 (t), u, ψx (t)) + αy (u − U (t, x 0 (t), u))2
2 (3.99)
≥ H U (t, x 0 (t), u0 (t), ψx (t))
whenever t ∈ [t0 , tf ] \ and (t, x 0 (t), u) ∈ Q1 . Therefore, we can identify the set M0V with
the set of tuples λV = (α0 , α, β, ψx , αy ) such that conditions (3.88), (3.89), (3.98), (3.96),
(3.94), and (3.99) hold.
Let there exist an element λ = (α0 , α, β, ψ, ν) of the set M0 of problem (3.1), (3.2) at
the point w 0 . Then its projection λU = (α0 , α, β, ψ) is an element of the set M0U of problem
(3.1), (3.5) (of problem U ) at the same point. Let 0 ≤ αy ≤ 1. We set
i.e., the normalization condition (3.98) holds for λV . Also conditions (3.88), (3.89), (3.96),
and (3.94) hold.
Let us verify the minimum condition (3.99). Since λU ∈ M0U, the conditions
imply
H U (t, x 0 (t), u, ψx (t)) ≥ H U (t, x 0 (t), u0 (t), ψx (t)). (3.102)
144 Chapter 3. Quadratic Conditions for Optimal Control Problems
1
αy (u − U (t, x 0 (t), u))2 ≥ 0. (3.103)
2
Adding inequalities (3.102) and (3.103), we obtain the minimum condition (3.99). There-
fore, we have proved the following assertion.
Then according to Proposition 3.17, λV ∈ M0V . Let us show that λV ∈ M0V + , i.e., the strict
minimum principle at the point (0, w0 ) in Problem V holds for λV . First, let t ∈ [t0 , tf ] \ ,
and let
u ∈ Rd(u) , (t, x 0 (t), u) ∈ Q1 , u = u0 (t). (3.105)
If g(t, x 0 (t), u) = 0 in this case, then
and the strict inequality (3.80) holds. It follows from (3.106) and (3.80) that
1
H U (t, x 0 (t), u, ψx (t)) + αy (u − U (t, x 0 (t), u))2
2 (3.108)
> H U (t, x 0 (t), u0 (t), ψx (t)),
i.e., the strict minimum condition holds for the function H V defined by relation (3.86). If,
along with conditions (3.105), the condition g(t, x 0 (t), u) = 0 holds, then this implies
and then
1
αy (u − U (t, x 0 (t), u))2 > 0, (3.110)
2
since αy > 0. Since λU ∈ M0U, (3.105) implies the nonstrict inequality
Again, inequalities (3.110) and (3.111) imply the strict inequality (3.108). Therefore, for
t∈/ , we have the strict minimum in u at the point u0 (t) for H V (t, 0, x 0 (t), u, ψy , ψx (t)).
The case t = tk ∈ is considered analogously. Therefore, we have proved the following
assertion.
Proposition 3.18. The conditions λ = (α0 , α, β, ψ, ν) ∈ M0+ , 0 < αy < 1, λU = (α0 , α, β, ψ),
and λV = ((1 − αy )λU , αy ) imply λV ∈ M0V + .
Let Leg+ (M0 ) be the set of all strictly Legendre elements λ ∈ M0 in the problem (3.1),
(3.2) at the point w0 . The definition of these elements was given in Section 3.2.3. Also,
denote by Leg+ (M0V ) the set of all strictly Legendre elements λV ∈ M0V of Problem V at
the point (0, w0 ). The definition of these elements was given in Section 2.1. Let us prove
the following assertion.
1
H V = (1 − αy )H U + αy (u − U )2 , (3.112)
2
HyV = 0. (3.114)
Also, (3.112) implies that HtV = (1 − αy )HtU on the trajectory (t, w0 (t)). But, according to
(3.54), HtU = H̄t . Hence
Finally, the definitions of the functions H V and H̄ imply that the following relations hold
on the trajectory (t, w 0 (t)):
We obtain from the definitions of D k (H V ) and D k (H̄ ) and also from conditions (3.113)–
(3.117) that
Since λ ∈ Leg+ (M0 ) by condition, D k (H̄ ) > 0 for all tk ∈ . This and (3.118) together with
the inequality 1 − αy > 0 imply
D k (H V ) > 0 ∀ tk ∈ . (3.119)
Let us verify the conditions for the strict Legendre property of the element λV . For
this purpose, we calculate the quadratic form Huu V ū, ū , where ū ∈ Rd(u) , for this element
0
on the trajectory (t, w (t)). Differentiating relation (3.112) in u and multiplying it by ū, we
obtain
HuV ū = (1 − αy )HuU ū + αy (u − U ), (ū − Uu ū) .
The repeated differentiation in u and the multiplication by ū yield
Huu
V
ū, ū = (1 − αy )Huu
U
ū, ū + αy (ū − Uu ū)2 − αy (u − U )Uuu ū, ū .
Substituting (t, w) = (t, w 0 (t)) and, moreover, taking into account that u0 = U (t, w0 ),
we obtain Huu V ū, ū = (1 − α )H U ū, ū + α (ū − U ū)2 . Finally, according to (3.45),
y uu y u
Huu ū, ū = H̄uu (Uu ū), (Uu ū) , where H̄uu = H̄uu
U λ corresponds to the element λ. Hence
Huu
V
ū, ū = (1 − αy )H̄uu (Uu ū), (Uu ū) + αy (ū − Uu ū)2 . (3.120)
The values of the derivatives are taken for (t, w) = (t, w 0 (t)).
It is easy to verify that for each t ∈ [t0 , tf ], form (3.120) quadratic in ū is positive-
definite on Rd(u) . Indeed, suppose first that t ∈ [t0 , tf ] \ . Recall that the mapping ū ∈
Rd(u) → Uu (t, w 0 (t))ū ∈ Rd(u) is the projection on the subspace
This and the condition λ ∈ Leg+ (M0 ) imply that the quadratic form H̄uu (Uu ū), (Uu ū)
is positive semidefinite on Rd(u) and positive definite on subspace (3.121). Furthermore,
the quadratic form (ū − Uu ū)2 is positive semidefinite on Rd(u) and positive outside sub-
space (3.121). This and the conditions 0 < αy < 1 imply the positivity of the quadratic
form HuuV ū, ū outside the origin of the space Rd(u) and, therefore, the positive definiteness
ered similarly. Therefore, all the conditions needed for element λV to belong to the set
Leg+ (M0V ) hold. The proposition is proved.
3.2. Quadratic Sufficient Conditions on a Fixed Time Interval 147
Lemma 3.20. The conditions λ = (α0 , α, β, ψ, ν) ∈ Leg+ (M0+ ), 0 < αy < 1, λU = (α0 , α,
β, ψ), λV = ((1 − αy )λU , αy ) imply λV ∈ Leg+ (M0V + ).
Fix a number αy such that 0 < αy < 1, e.g., αy = 1/2, and consider the linear operator
λ = (α0 , α, β, ψ, ν) → λV = ((1 − αy )λU , αy ), (3.122)
where λU = (α0 , α, β, ψ) is the projection of the element λ. Lemma 3.20 implies the
following assertion.
Now let us consider the critical cone K V of Problem V at the point (0, w0 ). Accord-
ing to the definition given in Section 2.1, it consists of elements z̄V = (ξ̄ , ȳ, w̄) = (ξ̄ , ȳ, x̄, ū)
such that the following conditions hold:
z̄ = (ξ̄ , x̄, ū) = (ξ̄ , w̄) ∈ Z2 (), ȳ ∈ P W 1,2 (, R),
Jp p̄ ≤ 0, Fip p̄ ≤ 0 ∀ i ∈ I , Kp p̄ = 0,
ȳ0 = 0, ȳf ≤ 0,
ȳ˙ = 0, [ȳ]k = 0 ∀ tk ∈ ,
x̄˙ = fx x̄ + fu (Uw w̄),
[x̄]k = [f ]k ξ̄k ∀ tk ∈ .
These conditions imply ȳ = 0, z̄ = (ξ̄ , x̄, ū) ∈ K U , where K U is the critical cone of prob-
lem U , i.e., the problem (3.1), (3.5), at the point w 0 . Then, according to Lemma 3.6,
z̃ = (ξ̄ , x̄, Uw w̄) is an element of the critical cone K of the problem (3.1), (3.2) at the
point w 0 . Therefore, we have proved the following assertion.
Lemma 3.22. Let z̄V = (ξ̄ , ȳ, x̄, ū) = (ξ̄ , ȳ, w̄) ∈ K V . Then ȳ = 0, z̃ = (ξ̄ , x̄, Uw w̄) ∈ K.
Therefore, the linear operator
(ξ̄ , ȳ, x̄, ū) → (ξ̄ , x̄, Ux x̄ + Uu ū) (3.123)
transforms the critical cone K V into the critical cone K.
We now consider the quadratic forms. Let λ ∈ M0 be an arbitrary element, and let
λV be its image under mapping (3.122). According to Proposition 3.17, λV ∈ M0V . For
V
Problem V and the point (0, w0 ), let us consider the quadratic form V λ (corresponding
to the element λV ), which was defined in Section 2.1, and let us study its relation with the
quadratic form λ (which corresponds to the element λ) at the point w0 of the problem
(3.1), (3.2). We have already shown that
D k (H V ) = (1 − αy )D k (H̄ ), tk ∈ (3.124)
(we omit the superscripts λV and λ of H V and H̄ , respectively). It follows from (3.113) that
[HxV ]k = (1 − αy )[H̄x ]k , tk ∈ , (3.125)
148 Chapter 3. Quadratic Conditions for Optimal Control Problems
where w̄ = (x̄, ū), x̄ ∈ P W 1,2 (, Rd(x) ), ū ∈ L2 (, Rd(u) ). According to (3.47),
Hww
U
w̄, w̄ = H̄ww w̃, w̃ , (3.128)
where w̃ = (x̄, ũ) = (x̄, Uw w̄). Let us calculate the second summand in formula (3.127).
We have
1 ∂
(u − U )2 w̄ = (u − U )(ū − Uw w̄).
2 ∂w
Therefore, for w = w 0 (t), we have
' ( ) *
1 ∂2
(u − U )2
w̄, w̄ = ( ū − Uw w̄) 2
− u0
− U (w 0
, t) Uww w̄, w̄
2 ∂w2
= (ū − Uw w̄)2 = (ū − ũ)2 . (3.129)
Hww
V
w̄, w̄ = (1 − αy )H̄ww w̃, w̃ + αy (ū − ũ)2 , (3.130)
where ũ = Ux x̄ + Uu ū and z̃ = (ξ̄ , x̄, ũ). Therefore, we have proved the following assertion.
Lemma 3.23. Let an element λ ∈ M0 and a tuple z̄V = (ξ̄ , ȳ, x̄, ū) satisfying conditions
(3.132) be given. Let λV be the image of λ under mapping (3.122). Then formula (3.133)
V
holds for the quadratic form V λ calculated for Problem V, the point (0, w 0 ), and the
element λV and also holds for the quadratic form λ calculated for the problem (3.1),
(3.2), the point w0 , and the element λ.
3.2. Quadratic Sufficient Conditions on a Fixed Time Interval 149
We now assume that the following sufficient Condition B holds at the point w0 in
problem (3.1), (3.2): there exist a nonempty compact set M ⊂ Leg+ (M0+ ) and a constant
C > 0 such that
max λ (z̄) ≥ C γ̄ (z̄) ∀ z̄ ∈ K. (3.134)
M
Let us show that in this case, the sufficient condition of Section 2.1 (denoted by BV ) holds
at the point (0, w 0 ) in Problem V .
Let z̄V = (ξ̄ , ȳ, x̄, ū) be an arbitrary element of the critical cone K V of Problem V at
the point (0, w0 ). Then ȳ = 0, and, by Lemma 3.22, z̃ = (ξ̄ , x̄, ũ) ∈ K, where ũ = Uw w̄.
Condition (3.134) implies the existence of λ ∈ M such that
λ (z̃) ≥ C γ̄ (z̃). (3.135)
Let λVbe the image of λ under the mapping defined by operator (3.122). Then, by
Lemma 3.23, formula (3.133) holds. It follows from (3.133) and (3.135) that
tf
V λV V
(z̄ ) ≥ (1 − αy )C γ̄ (z̃) + αy (ū − ũ)2 dt. (3.136)
t0
Therefore,
V
tf
max V λ (z̄V ) ≥ (1 − αy )C γ̄ (z̃) + αy (ū − ũ)2 dt, (3.137)
λV ∈M V t0
where M V is the image of the compact set M under the mapping defined by operator (3.122).
By Lemma 3.21,
M V ⊂ Leg+ (M0V + ). (3.138)
Conditions (3.137) and (3.138) imply that the sufficient Condition BV holds at the point
(0, w 0 ) in Problem V . To verify this, it suffices to show that the right-hand side of inequality
(3.137) on the cone K V is estimated from below by the functional
tf
γ̄ (z̄ ) := ξ̄ + ȳ0 + x̄0 +
V V 2 2 2
ū2 dt = γ̄ (z̄) + ȳ02
t0
with a small coefficient ε > 0. Since ȳ =
0 for all z̄ ∈ K V , it suffices to prove that there
exists ε > 0 such that (1 − αy )C γ̄ (z̃) + αy (ū − ũ)2 dt ≥ εγ̄ (z̄) or
tf tf tf
(1 − αy )C ξ̄ 2 + x̄02 + ũ2 dt + αy (ū − ũ)2 dt ≥ ε ξ̄ 2 + x̄02 + ū2 dt . (3.139)
t0 t0 t0
We set ū − ũ = û. Then ū2 = û2 + 2ûũ + ũ2 ≤ 2û2 + 2ũ2 . This obviously implies the
estimate required. Therefore, we have proved the following assertion.
Lemma 3.24. Let the sufficient Condition B hold at the point w 0 in problem (3.1), (3.2).
Then the sufficient Condition BV of Section 2.1 holds at the point (0, w 0 ) in Problem V .
where P and Q are open sets, and x, u, F , K, f , and g are vector functions.
We assume that the functions J , F , and K are defined and twice continuously differen-
tiable on P , and that the functions f and g are defined and twice continuously differentiable
on Q. It is also assumed that the gradients with respect to the control giu (t, x, u), i =
1, . . . , d(g) are linearly independent at each point (t, x, u) ∈ Q such that g(t, x, u) = 0. Here
d(g) is a dimension of the vector g.
where ψ and ν are row vectors of the dimensions d(x) and d(g), respectively. Let us define
the endpoint Lagrange function
such that ψ(·) : → (Rd(x) )∗ and ψ0 (·) : → R1 are piecewise smooth functions, continu-
ously differentiable on each interval of the set \ , and ν(·) : → (Rd(g) )∗ is a piecewise
continuous function, Lipschitz continuous on each interval of the set \ .
Denote by M0 the set of the normed tuples λ satisfying the conditions of the minimum
principle for the trajectory T :
α0 ≥ 0, α ≥ 0, αF (p) = 0, α0 + αi + |βj | = 1,
ψ̇ = −H̄x , ψ̇0 = −H̄t , H̄u = 0, t ∈ \ ,
ψ(t0 ) = −lx0 , ψ(tf ) = lxf , ψ0 (t0 ) = −lt0 , ψ0 (tf ) = ltf , (3.149)
min H (t, x(t), u, ψ(t)) = H (t, x(t), u(t), ψ(t)), t ∈ \ ,
u∈U(t,x(t))
H (t, x(t), u(t), ψ(t)) + ψ0 (t) = 0, t ∈ \ ,
where U(t, x) = {u ∈ Rd(u) | g(t, x, u) = 0, (t, x, u) ∈ Q}. The derivatives lx0 and lxf are
at (p, α0 , α, β), where p = (t0 , x(t0 ), tf , x(tf )), and the derivatives H̄x , H̄u , and H̄t are at
(t, x(t), u(t), ψ(t), ν(t)), where t ∈ \ . (Condition H̄u = 0 follows from the others con-
ditions in this definition, and therefore could be excluded; yet we need to use it later.)
Let us give the definition of Pontryagin minimum in problem (3.140)–(3.144) on a
variable interval [t0 , tf ].
Definition 3.26. The trajectory T affords a Pontryagin minimum if for each compact
set C ⊂ Q there exists ε > 0 such that J(T˜ ) ≥ J(T ) for all admissible trajectories T˜ =
(x̃(t), ũ(t) | t ∈ [t˜0 , t˜f ]) satisfying the following conditions:
(a) |t˜0 − t0 | < ε, |t˜f − tf | < ε;
(b)
max∩˜ |x̃(t) − x(t)| < ε, where ˜ = [t˜0 , t˜f ];
(c) ∩˜ | ũ(t) − u(t)| dt < ε;
(d) (t, x̃(t), ũ(t)) ∈ C a.e. on . ˜
Theorem 3.27. If the trajectory T affords a Pontryagin minimum, then the set M0 is
nonempty.
Assume that M0 is nonempty. Using the definition of the set M0 and the full rank
condition of the matrix gu on the surface g = 0, one can easily prove the following statement.
Proposition 3.28. The set M0 is a finite-dimensional compact set, and the mapping λ →
(α0 , α, β) is injective on M0 .
where H̄xk− = H̄x (tk , x(tk ), u(tk −), ψ(tk ), ν(tk −)), H̄xk+ = H̄x (tk , x(tk ), u(tk +), ψ(tk ), ν(tk +)),
[H̄t ]k = H̄tk+ − H̄tk− , etc.
D k (H̄ ) ≥ 0, k = 1, . . . , s. (3.151)
Thus, conditions (3.151) follows from the minimum principle conditions (3.149). The
following is an alternative method for calculating D k (H̄ ): For λ ∈ M0 , tk ∈ , consider the
function
(k H̄ )(t) = H̄ (tk , x(t), u(tk +), ψ(t), ν(tk +)) − H̄ (tk , x(t), u(tk −), ψ(t), ν(tk −)).
d d
(k H̄ )t=t − = (k H̄ )t=t + = −D k (H̄ ), k = 1, . . . , s. (3.152)
dt k dt k
Hence, for λ ∈ M0 the function (k H̄ )(t) has a derivative at the point tk ∈ equal
to −D k (H̄ ), k = 1, . . . , s. Let us formulate a quadratic necessary condition of a Pontryagin
minimum for the trajectory T . First, for this trajectory, we introduce a Hilbert space
Z2 () and the critical cone K ⊂ Z2 (). We denote by P W 1,2 (, Rd(x) ) the Hilbert
space of piecewise continuous functions x̄(·) : → Rd(x) , absolutely continuous on each
interval of the set \ and such that their first derivative is square integrable. For each
x̄ ∈ P W 1,2 (, Rd(x) ), tk ∈ , we set
Thus,
z̄ ∈ Z2 () := R2 × Rs × P W 1,2 (, Rd(x) ) × L2 (, Rd(u) ).
Moreover, for given z̄ we set
By IF (p) = {i ∈ {1, . . . , d(F )} | Fi (p) = 0}, we denote the set of active indices of the
constraints Fi (p) ≤ 0.
Let K be the set of all z̄ ∈ Z2 () satisfying the following conditions:
where p = (t0 , x(t0 ), tf , x(tf )), w = (x, u). It is obvious that K is a convex cone in the
Hilbert space Z2 (), and we call it the critical cone. If the interval is fixed, then we
set p := (x0 , xf ) = (x(t0 ), x(tf )), and in the definition of K we have t¯0 = t¯f = 0, x̄¯0 = x̄0 ,
x̄¯f = x̄f , and p̄¯ = p̄ := (x̄0 , x̄f ).
Let us introduce a quadratic form on Z2 (). For λ ∈ M0 and z̄ ∈ K, we set
ωe (λ, z̄) = lpp p̄,¯ p̄
¯ − 2ψ̇(tf )x̄(tf )t¯f − ψ̇(tf )ẋ(tf ) + ψ̇0 (tf ) t¯f2
+ 2ψ̇(t0 )x̄(t0 )t¯0 + ψ̇(t0 )ẋ(t0 ) + ψ̇0 (t0 ) t¯02 , (3.156)
where lpp = lpp (p, α0 , α, β), p = (t0 , x(t0 ), tf , x(tf )). We also set
tf
ω(λ, z̄) = ωe (λ, z̄) + H̄ww w̄(t), w̄(t) dt, (3.157)
t0
where H̄ww = H̄ww (t, x(t), u(t), ψ(t), ν(t)). Finally, we set
s
2(λ, z̄) = ω(λ, z̄) + D k (H̄ )ξ̄k2 − 2[ψ̇]k x̄av
k
ξ̄k , (3.158)
k=1
k = 1 (x̄ k− + x̄ k+ ), [ψ̇]k = ψ̇ k+ − ψ̇ k− .
where x̄av 2
Now, we formulate the main necessary quadratic condition of Pontryagin minimum
in the problem on a variable time interval.
Theorem 3.31. If the trajectory T yields a Pontryagin minimum, then the following Con-
dition A holds: the set M0 is nonempty and
Definition 3.32. We say that the trajectory T affords a bounded strong minimum if there
is no sequence of admissible trajectories T n = (x n (t), un (t) | t ∈ [t0n , tfn ]), n = 1, 2, . . . such
that
(a) J(T n ) < J(T );
(b) t0n → t0 , tfn → tf , x n (t0n ) → x(t0 ) (n → ∞);
(c) maxn ∩ |x n (t) − x(t)| → 0 (n → ∞), where n = [t0n , tfn ];
(d) there exists a compact set C ⊂ Q such that (t, x n (t), un (t)) ∈ C a.e. on n for all n.
Definition 3.33. The trajectory T affords a bounded strong minimum if for each compact
set C ⊂ Q there exists ε > 0 such that J(T˜ ) ≥ J(T ) for all admissible trajectories T˜ =
(x̃(t), ũ(t) | t ∈ [t˜0 , t˜f ]) satisfying the following conditions:
(a) |t˜0 − t0 | < ε, |t˜f − tf | < ε, |x̃(t˜0 ) − x(t0 )| < ε;
(b) max∩ ˜ |x̃(t) − x(t)| < ε, where ˜ = [t˜0 , t˜f ];
(c) (t, x̃(t), ũ(t)) ∈ C a.e. on . ˜
The strict bounded strong minimum is defined in a similar way, with the nonstrict
inequality J(T˜ ) ≥ J(T ) replaced by the strict one and the trajectory T˜ required to be
different from T . Finally, we define a (strict) strong minimum in the same way but omit
condition (c) in the last definition. The following statement is quite obvious.
Proposition 3.34. If there exists a compact set C ⊂ Q such that {(t, x, u) ∈ Q | g(t, x, u) =
0} ⊂ C, then a (strict) strong minimum is equivalent to a (strict) bounded strong minimum.
of Hilbert space Z2 (). Recall that the critical cone K is contained in the subspace (3.161).
Theorem 3.35. For the trajectory T , assume that the following Condition B holds: The
set Leg+ (M0+ ) is nonempty and there exist a nonempty compact set M ⊂ Leg+ (M0+ ) and a
number C > 0 such that
max (λ, z̄) ≥ C γ̄ (z̄) (3.162)
λ∈M
for all z̄ ∈ K. Then the trajectory T affords a strict bounded strong minimum.
156 Chapter 3. Quadratic Conditions for Optimal Control Problems
3.3.4 Proofs
The proofs are based on the quadratic optimality conditions, obtained in this chapter for
problems on a fixed interval of time. We will give the proofs but omit some details. In
order to extend the proofs to the case of a variable interval [t0 , tf ] we use a simple change
of the time variable. Namely, we associate the fixed admissible trajectory T = (x(t), u(t) |
t ∈ [t0 , tf ]) in the problem on a variable time interval (3.140)–(3.144) with a trajectory
T τ = (v(τ ), t(τ ), x(τ ), u(τ ) | τ ∈ [τ0 , τf ]), considered on a fixed interval [τ0 , τf ], where
τ0 = t0 , τf = tf , t(τ ) ≡ τ , v(τ ) ≡ 1. This is an admissible trajectory in the following
problem on a fixed interval [τ0 , τf ]: Minimize the cost function
In this problem, x(τ ), t(τ ), and v(τ ) are state variables, and u(τ ) is a control variable.
For brevity, we refer to problem (3.140)–(3.144) as problem P (on a variable interval
= [t0 , tf ]), and to problem (3.163)–(3.167) as problem P τ (on a fixed interval [τ0 , τf ]).
We denote by Aτ the necessary quadratic Condition A for problem P τ on a fixed interval
[τ0 , τf ]. Similarly, we denote by Bτ the sufficient quadratic Condition B for problem P τ
on a fixed interval [τ0 , τf ].
Recall that the control u(·) is a piecewise Lipschitz continuous function on the interval
= [t0 , tf ] with the set of discontinuity points = {t1 , . . . , ts }, where t0 < t1 < · · · < ts < tf .
Hence, for each λ ∈ M0 , the function ν(t) is also piecewise Lipschitz continuous on the
interval , and, moreover, all discontinuity points of ν belong to . This easily follows
from the equation H̄u = 0 and the full-rank condition for matrix gu . Consequently, u̇ and
ν̇ are bounded measurable functions on . The proof of Theorem 3.31 is composed of the
following chain of implications:
(i) A Pontryagin minimum is attained on the trajectory T in problem P =⇒
(ii) A Pontryagin minimum is attained on the trajectory T τ in problem P τ =⇒
(iii) Condition Aτ holds for the trajectory T τ in problem P τ =⇒
(iv) Condition A holds for the trajectory T in problem P .
The first implication is readily verified, the second follows from Theorem 3.9. The
verification of the third implication (iii) ⇒ (iv) is not short and rather technical: we have
to compare the sets of Lagrange multipliers, the critical cones, and the quadratic forms in
both problems. This will be done below.
In order to prove the sufficient conditions in problem P , given by Theorem 3.35, we
have to check the following chain of implications:
(v) Condition B holds for the trajectory T in problem P =⇒
(vi) Condition Bτ holds for the trajectory T τ in problem P τ =⇒
3.3. Quadratic Conditions on a Variable Time Interval 157
l τ = α0 J + αF + βK = l,
H τ = ψvf + ψ0 v + ψv · 0 = v(ψf + ψ0 ), H̄ τ = H τ + νg.
The set M0τ in problem P τ for the trajectory T τ consists of all tuples of Lagrange multipliers
λτ = (α0 , α, β, ψ, ψ0 , ψv , ν) such that the following conditions hold:
α0 + |α| + |β| = 1,
dψ dψ0 dψv
− = vψfx + νgx , − = vψft + νgt , − = ψf + ψ0
dτ dτ dτ
ψ(τ0 ) = −lx0 , ψ(τf ) = lxf , ψ0 (τ0 ) = −lt0 , ψ0 (τf ) = ltf , (3.168)
ψv (τ0 ) = ψv (τf ) = 0, vψfu + νgu = 0,
v(τ ) ψ(τ )f (t(τ ), x(τ ), u) + ψ0 (τ )
≥ v(τ ) ψ(τ )f (t(τ ), x(τ ), u(τ )) + ψ0 (τ ) .
The last inequality holds for all u ∈ Rd(u) such that g(t(τ ), x(τ ), u) = 0, (t(τ ), x(τ ), u) ∈ Q.
Recall that v(τ ) ≡ 1, t(τ ) ≡ τ , τ0 = t0 , and τf = tf . In (3.168), the function f and its deriva-
tives fx , fu , ft , gx gu , and gt are taken at (t(τ ), x(τ ), u(τ )), τ ∈ [τ0 , τf ] \ , while the deriva-
tives lt0 , lx0 , ltf , and lxf are calculated at (t(τ0 ), x(τ0 ), t(τf ), x(τf )) = (t0 , x(t0 ), tf , x(tf )).
τ
Conditions −dψv /dτ = ψf + ψ0 and ψv (τ0 ) = ψv (τf ) = 0 imply that τ0f (ψf +
ψ0 ) dτ = 0. As is well known, conditions (3.168) of the minimum principle also imply that
ψf + ψ0 = const, whence ψf + ψ0 = 0 and ψv = 0. Taking this fact into account and
comparing the definitions of the sets M0τ (3.168) and M0 (3.149), we see that the projector
α0 , α, β, ψ, ψ0 , ψv , ν → α0 , α, β, ψ, ψ0 , ν (3.169)
realizes a one-to-one correspondence between these two sets. (Moreover, in the definition of
the set M0τ one could replace the relations −dψv /dτ = ψf + ψ0 and ψv (τ0 ) = ψv (τf ) = 0
with ψf + ψ0 = 0, and thus identify M0τ with M0 .)
We say that an element λτ ∈ M0τ corresponds to an element λ ∈ M0 if λ is the projec-
tion of λτ under the mapping (3.169).
Comparison of the critical cones. For brevity, we set ! = (v, t, x, u) = (v, t, w).
Let us define the critical cone K τ in problem P τ for the trajectory T τ . It consists of all
158 Chapter 3. Quadratic Conditions for Optimal Control Problems
where the derivatives Jt0 , Jx0 , Jtf Jxf , etc. are calculated at (t(τ0 ), x(τ0 ), t(τf ), x(τf )) =
(t0 , x(t0 ), tf , x(tf )), while f , ft , fx , fu gt , gx , and gu are taken at (t(τ ), x(τ ), u(τ )), τ ∈
[τ0 , τf ] \ . Let (ξ̄ , v̄, t¯, x̄, ū) be an element of the critical cone K τ . We can use the following
change of variables:
x̃ = x̄ − t¯ẋ, ũ = ū − t¯u̇, (3.176)
or, briefly,
w̃ = w̄ − t¯ẇ. (3.177)
Since v = 1, ẋ = f , and t = τ , equation (3.173) is equivalent to
d x̄
= v̄ ẋ + ft t¯ + fw w̄. (3.178)
dt
Using the relation x̄ = x̃ + t¯ẋ in (3.178) along with t˙¯ = v̄, we get
ẍ = ft + fw ẇ. (3.180)
The relations
[x̄]k = [ẋ]k ξ̄k , x̄ = x̃ + t¯ẋ
imply
[x̃]k = [ẋ]k ξ̃k , (3.182)
where
ξ̃k = ξ̄k − t¯k , t¯k = t¯(tk ), k = 1, . . . , s. (3.183)
Further, relation (3.175) may be written as gt t¯ + gw w̄ = 0. Differentiating the relation
g(t, w(t)) = 0, we obtain
gt + gw ẇ = 0. (3.184)
3.3. Quadratic Conditions on a Variable Time Interval 159
gw w̃ = 0. (3.185)
where t¯0 = t¯(t0 ) and t¯f = t¯(tf ). The vector in the right-hand side of the last equality has the
same form as the vector p̄¯ in definition (3.154). Consequently, all relations in definition
(3.155) of the critical cone K in problem P are satisfied for the element z̃ = (t¯0 , t¯f , ξ̃ , w̃).
We have proved that the obtained element z̃ belongs to the critical cone K in problem P .
Conversely, if (t¯0 , t¯f , ξ̃ , w̃) is an element of the critical cone in problem P , then by setting
t¯f − t¯0
v̄ = , t¯ = v̄(τ − τ0 ) + t¯0 , w̄ = w̃ + t¯ẇ, ξ̄k = ξ̃k + t¯(τk ), k = 1, . . . , s,
tf − t0
we obtain an element (ξ̄ , v̄, t¯, w̄) of the critical cone (3.170)–(3.175) in problem P τ . Thus,
we have proved the following lemma.
Lemma 3.36. If (ξ̄ , v̄, t¯, w̄) is an element of the critical cone (3.170)–(3.175) in problem P τ
for the trajectory T τ and
then (t¯0 , t¯f , ξ̃ , w̃) is an element of the critical cone (3.155) in problem P for the trajectory
T . Moreover, relations (3.187) define a one-to-one correspondence between elements of
the critical cones in problems P τ and P .
We say that an element (ξ̄ , v̄, t¯, w̄) of the critical cone in problem P τ corresponds to
an element (t¯0 , t¯f , ξ̃ , w̃) of the critical cone in problem P if relations (3.187) hold.
Comparison of the quadratic forms. Assume that the element λτ ∈ M0τ corre-
sponds to the element λ ∈ M0 . Let us show that the quadratic form τ (λτ , ·), calculated on
the element (ξ̄ , v̄, t¯, w̄) of the critical cone in problem P τ for the trajectory T τ , can be trans-
formed into the quadratic form (λ, ·) calculated on the corresponding element (t¯0 , t¯f , ξ̃ , w̃)
of the critical cone in problem P for the trajectory T .
(i) The relations H̄ τ = v(H + ψ0 ) + νg, H̄ = H + νg, v = 1 imply
H̄!!
τ
!, ¯ = H̄ww w̄, w̄ + 2H̄tw w̄t¯ + H̄tt t¯2 + 2v̄(Hw w̄ + Ht t¯),
¯ ! (3.188)
H̄ww w̄, w̄ = H̄ww w̃, w̃ + 2H̄ww ẇ, w̄ t¯ − H̄ww ẇ, ẇ t¯2 . (3.189)
we obtain
Hw w̄ + Ht t¯ = H̄w w̄ + H̄t t¯ − ν(gw w̄ + gt t¯)
(3.190)
= H̄w w̄ + H̄t t¯ = H̄x x̄ + H̄t t¯ = −ψ̇ x̄ − ψ̇0 t¯.
H̄!! ¯ !
τ !, ¯ = H̄ww w̃, w̃ + 2H̄ww ẇ, w̄ t¯ + 2H̄tw w̄t¯
(3.191)
− H̄ww ẇ, ẇ t¯2 + H̄tt t¯2 − 2v̄ ψ̇ x̄ + ψ̇0 t¯ .
(ii) Let us transform the terms 2H̄ww ẇ, w̄ t¯ + 2H̄tw w̄t¯ in (3.191). By differentiating −ψ̇ =
H̄x with respect to t, we obtain
Since (ξ̄ , v̄, t¯, w̄) is an element of the critical cone in problem P τ , from (3.173) and (3.175)
we get fw w̄ = x̄˙ − v̄ ẋ − ft t¯, gw w̄ = −gt t¯. Therefore, equation (3.194) can be represented
in the form
d
H̄tw w̄ + H̄ww ẇ, w̄ = v̄(ψ̇ ẋ) − (ψ̇ x̄) + ψ̇ft + ν̇gt t¯, (3.195)
dt
which implies
d
2H̄ww ẇ, w̄ t¯ + 2H̄tw w̄ t¯ = 2t¯v̄(ψ̇ ẋ) − 2t¯ (ψ̇ x̄) + 2 ψ̇ft + ν̇gt t¯2 . (3.196)
dt
(iii) Let us transform the term −H̄ww ẇ, ẇ t¯2 in (3.191). Multiplying (3.192) by ẋ and
(3.193) by u̇ and summing the results, we obtain
(iv) Finally, let us transform the term H̄tt t¯2 in (3.191). Differentiating −ψ̇0 = H̄t with
respect to t and using the relations H̄ψt = ft and H̄νt = gt , we get
−ψ̈0 = H̄tt + H̄tw ẇ + ψ̇ft + ν̇gt . (3.200)
Consequently,
H̄tt t¯2 = −ψ̈0 t¯2 − H̄tw ẇt¯2 − ψ̇ft + ν̇gt t¯2 . (3.201)
(v) Summing (3.199) and (3.201), we obtain
d
−H̄ww ẇ, ẇ t¯2 + H̄tt t¯2 = −ψ̈0 t¯2 − 2 ψ̇ft + ν̇gt t¯2 + t¯2 (ψ̇ ẋ). (3.202)
dt
Using relations (3.196) and (3.202) in (3.191), we get
d
H̄!!
τ
!, ¯ = H̄ww w̃, w̃ + 2t¯v̄(ψ̇ ẋ) − 2t¯
¯ ! (ψ̇ x̄)
dt
d
− ψ̈0 t¯2 + t¯2 (ψ̇ ẋ) − 2v̄ ψ̇ x̄ + ψ̇0 t¯ . (3.203)
dt
But
d d d
ψ̈0 t¯2 + 2v̄ t¯ψ̇0 = ψ̇0 t¯2 , t¯ (ψ̇ x̄) + v̄(ψ̇ x̄) = t¯ψ̇ x̄ ,
dt dt dt
d d
2t¯v̄(ψ̇ ẋ) + t¯2 (ψ̇ ẋ) = (ψ̇ ẋ t¯2 ).
dt dt
Therefore,
d
H̄!!
τ
¯ !
!, ¯ = H̄ww w̃, w̃ + (ψ̇ ẋ)t¯2 − ψ̇0 t¯2 − 2ψ̇ x̄ t¯ . (3.204)
dt
Finally, using the change of the variable x̄ = x̃ + t¯ẋ in the right-hand side of this relation,
we obtain
d
H̄!!
τ
¯ !
!, ¯ = H̄ww w̃, w̃ − (ψ̇0 + ψ̇ ẋ)t¯2 + 2ψ̇ x̃ t¯ . (3.205)
dt
We have proved the following lemma.
(vi) Recall that λτ is an arbitrary element of the set M0τ (consequently ψv = 0) and λ is
the corresponding element of the set M0 , i.e., λ is the projection of λτ under the mapping
(3.169). The quadratic form τ (λτ , ·) in problem P τ for the trajectory T τ has the following
representation:
s
¯ =
τ (λτ ; ξ̄ , !) D k (H̄ τ )ξ̄k2 − 2[ψ̇]k x̄av
k
ξ̄k − 2[ψ̇0 ]k t¯av
k
ξ̄k
k=1 τf (3.206)
+ lpp p̄, p̄ + H̄!!
τ
¯ !
!, ¯ dτ .
τ0
162 Chapter 3. Quadratic Conditions for Optimal Control Problems
Comparing the definitions of D k (H̄ τ ) and D k (H̄ ) (see (3.152)) and taking into account that
H̄ τ = v(ψf + ψ0 ) + νg and v = 1, we get
Let z̄τ = (ξ̄ , !)¯ = (ξ̄ , v̄, t¯, x̄, ū) be an element of the critical cone K τ in the problem P τ
for the trajectory T τ , and let z̃ = (t¯0 , t¯f , ξ̃ , x̃, ũ) be the corresponding element of the critical
cone K in the problem P for the trajectory T ; i.e., relations (3.187) hold. Since [t¯]k = 0,
k = 1, . . . , s, we have
t¯av
k
= t¯k , k = 1, . . . , s (3.208)
where t¯k = t¯(tk ), k = 1, . . . , s. Also recall that τ0 = t0 , τf = tf , t(τ ) = τ , dt = dτ . Since
the functions ψ̇0 , ψ̇, ẋ, and x̃ may have discontinuities only at the points of the set , the
following formula holds:
tf d tf
(ψ̇0 + ψ̇ ẋ)t¯2 + 2ψ̇ x̃ t¯ dt = (ψ̇0 + ψ̇ ẋ)t¯2 + 2ψ̇ x̃ t¯
t0 dt t0
s
(3.209)
− [ψ̇0 + ψ̇ ẋ]k t¯(tk )2 + 2[ψ̇ x̃]k t¯(tk ) .
k=1
Let us transform the terms related to the discontinuity points tk of the control u(·), k =
1, . . . , s. For any λ ∈ M0 , the following lemma holds.
Proof. In this proof, we omit the subscript and superscript k. We also write t¯ instead of
t¯(tk ). Set a = D(H̄ ). Using the relations
we obtain
a ξ̄ 2 − 2[ψ̇]x̄av ξ̄ − 2[ψ̇0 ]t¯ξ̄ + [ψ̇0 + ψ̇ ẋ]t¯2 + 2[ψ̇ x̃]t¯
= a ξ̃ 2 + 2a ξ̃ t¯ + a t¯2 − 2[ψ̇]x̃av ξ̄ − 2[ψ̇]ẋav t¯ξ̄ − 2[ψ̇0 ]t¯ξ̄ + [ψ̇0 + ψ̇ ẋ]t¯2 + 2[ψ̇ x̃]t¯
= a ξ̃ 2 − 2[ψ̇]x̃av ξ̃ + r,
(3.214)
where
r = 2a ξ̃ t¯ + a t¯2 − 2[ψ̇]x̃av t¯ − 2[ψ̇]ẋav t¯ξ̄ − 2[ψ̇0 ]t¯ξ̄ + [ψ̇0 + ψ̇ ẋ]t¯2 + 2[ψ̇ x̃]t¯. (3.215)
It suffices to show that r = 0. Using the relations (3.213) in formula (3.215), we get
r = 2a(ξ̄ − t¯)t¯ + a t¯2 − 2[ψ̇](x̄av − t¯ẋav )t¯ − 2[ψ̇]ẋav t¯ξ̄ − 2[ψ̇0 ]t¯ξ̄
+ [ψ̇0 + ψ̇ ẋ]t¯2 + 2[ψ̇(x̄ − t¯ẋ)]t¯
= t¯2 − a + 2[ψ̇]ẋav + [ψ̇0 ] − [ψ̇ ẋ] + 2t¯ξ̄ a − [ψ̇]ẋav − [ψ̇0 ]
+ 2t¯ − [ψ̇]x̄av + [ψ̇ x̄] .
The coefficient of t¯2 in the right-hand side of the last equality vanishes:
−a + 2[ψ̇]ẋav + [ψ̇0 ] − [ψ̇ ẋ] = − ψ̇ + ẋ − − ψ̇ − ẋ + + [ψ̇0 ] + (ψ̇ + − ψ̇ − )(ẋ + + ẋ − )
+ [ψ̇0 ] − ψ̇ + ẋ + + ψ̇ − ẋ − = 0.
The coefficient of 2t¯ξ̄ is equal to
1
a − [ψ̇]ẋav − [ψ̇0 ] = ψ̇ + ẋ − − ψ̇ − ẋ + + [ψ̇0 ] − (ψ̇ + − ψ̇ − )(ẋ − + ẋ + ) − [ψ̇0 ]
2
1 + − 1
= ψ̇ ẋ − ψ̇ − ẋ + − [ψ̇ ẋ].
2 2
The coefficient of 2t¯ is equal to
1
−[ψ̇]x̄av + [ψ̇ x̄] = − (ψ̇ + − ψ̇ − )(x̄ − + x̄ + ) + (ψ̇ + x̄ + − ψ̇ − x̄ − )
2
1 + 1
= ψ̇ [x̄] + ψ̇ − [x̄] = ψ̇av [ẋ]ξ̄ ,
2 2
since [x̄] = [ẋ]ξ̄ . Consequently,
1 + − 1
r = 2t¯ξ̄ ψ̇ ẋ − ψ̇ − ẋ + − [ψ̇ ẋ] + ψ̇av [ẋ]
2 2
+ −
= t¯ξ̄ (ψ̇ ẋ − ψ̇ ẋ ) − (ψ̇ ẋ + − ψ̇ − ẋ − )
− + +
+ (ψ̇ − + ψ̇ + )(ẋ + − ẋ − ) = 0.
In view of (3.214) the equality r = 0 proves the lemma.
Relation (3.210) along with equality (3.212) gives the following transformation of
quadratic form τ (see (3.206)) on the element z̄τ = (ξ̄ , !)
¯ of the critical cone K τ ,
s
¯
τ (λτ ; ξ̄ , !) = D k (H̄ )ξ̃k2 − 2[ψ̇]k x̃av
k
ξ̃k
tf
k=1
tf
+ lpp p̄, p̄ − (ψ̇0 + ψ̇ ẋ)t¯2 + 2ψ̇ x̃ t¯ + H̄ww w̃, w̃ dτ .
t0 t0
(3.216)
164 Chapter 3. Quadratic Conditions for Optimal Control Problems
Taking into account (3.186) and definitions (3.156)–(3.158) of quadratic forms ωe , ω, and ,
we see that the right-hand side of (3.216) is the quadratic form (λ, z̃) (see (3.158)) in
problem P for the trajectory T , where z̃ = (t¯0 , t¯f , ξ̃ , w̃) is the corresponding element of the
critical cone K. Thus we have proved the following theorem.
Theorem 3.39. Let z̄τ = (ξ̄ , v̄, t¯, w̄) be an element of the critical cone K τ in problem P τ for
the trajectory T τ . Let z̃ = (t¯0 , t¯f , ξ̃ , w̃) be the corresponding element of the critical cone K
in problem P for the trajectory T , i.e., relations (3.187) hold. Then for any λτ ∈ M0τ and
the corresponding projection λ ∈ M0 (under the mapping (3.169)) the following equality
holds: τ (λτ , z̄τ ) = (λ, z̃).
This theorem proves the implications (iii) ⇒ (iv) and (v) ⇒ (vi) (see the beginning
of this section), and thus completes the proofs of Theorems 3.31 and 3.35.
Assumption 3.40. (a) The functions J (p), F (p), and K(p) are defined and twice contin-
uously differentiable on the open set P ⊂ R2d(x) , and the functions f (t, w), g(t, w), and
ϕ(t, w) are defined and twice continuously differentiable on the open set Q ⊂ Rd(x)+d(u)+1 .
(b) The gradients with respect to the control giu (t, w), i = 1, . . . , d(g), ϕj u (t, w), j ∈ Iϕ (t, w)
3.4. Quadratic Conditions for Mixed Control-State Constrained Problems 165
are linearly independent at all points (t, w) ∈ Q such that g(t, w) = 0 and ϕ(t, w) ≤ 0. Here
gi and ϕj are the components of the vector functions g and ϕ, respectively, and
Iϕ (t, w) = {j ∈ {1, . . . , d(ϕ)} | ϕj (t, w) = 0} (3.220)
is the set of indices of active inequality constraints ϕj (w, t) ≤ 0 at (t, w) ∈ Q.
We refer to (b) as the linear independence assumption for the gradients of the active
mixed constraints with respect to the control. Let a pair w0 (·) = (x 0 (·), u0 (·)) ∈ W satisfying
constraints (3.218)–(3.219) of the problem be the point tested for optimality.
Assumption 3.41. The control u0 (·) is a piecewise continuous function such that all its
discontinuity points are L-points (see Definition 2.1). Let = {t1 , . . . , ts }, t0 < t1 < · · · <
ts < tf be the set of all discontinuity points of u0 (·). It is also assumed that (tk , x 0 (tk ), u0k− ) ∈
Q, (tk , x 0 (tk ), u0k+ ) ∈ Q, k = 1, . . . , s, where u0k− = u0 (tk − 0), u0k+ = u0 (tk + 0).
In what follows, we assume for definiteness that the set of discontinuity points of
u0 is nonempty. Whenever this set is empty, all statements admit obvious simplifications.
Definition 3.43. We say that w0 is a point of Pontryagin minimum if this point is a mini-
mum on the set of sequences {δwn } in W satisfying
t the following two conditions:
(a)
δxn
1,1 +
δun
1 → 0, where
δun
1 = t0f |δun | dt;
(b) there exists a compact set C ⊂ Q (which depends on the choice of the sequence) such
that for all sufficiently large n, we have (t, w 0 (t) + δwn (t)) ∈ C a.e. on [t0 , tf ].
Any sequence satisfying conditions (a) and (b) will be referred to as a Pontryagin
sequence on Q.
α0 ≥ 0, α ≥ 0, αF (p 0 ) = 0, (3.227)
)
d(F
d(K)
α0 + αi + |βj | = 1, (3.228)
i=1 j =1
where αi and βj are components of the row vectors α and β, respectively, and H̄x , H̄u , lx0 ,
and lxf are gradients with respect to the corresponding variables.
It is well known that if w0 is a weak minimum, then 0 is nonempty (see, e.g., [30]).
The latter condition is just the local minimum principle. Note that 0 can consist of more
than one element. The following result pertain to this possibility.
Proposition 3.44. The set 0 is a finite-dimensional compact set, and the projection λ =
(α0 , α, β, ψ, ν, μ) → (α0 , α, β) is injective on 0 .
This property of 0 follows from the linear independence assumption for the gra-
dients of the active mixed constraints with respect to the control. This assumption also
guarantees the following property.
Proposition 3.45. Let λ ∈ 0 be an arbitrary tuple. Then its components ν(t) and μ(t) are
continuous at each point of continuity of the control u0 (t). Consequently, ν(t) and μ(t) are
piecewise continuous functions such that all their discontinuity points belong to the set .
The adjoint variable ψ(t) is a piecewise smooth function such that all its break points belong
to the set .
3.4. Quadratic Conditions for Mixed Control-State Constrained Problems 167
where [H̄tλ ]k = H̄tλk+ − H̄tλk− is the jump of H̄t (t, x 0 (t), u0 (t), ψ(t), ν(t), μ(t)) at tk . Note that
D k (H̄ λ ) is linear in λ.
Since conditions D k (H̄ λ ) ≥ 0 for all tk ∈ follow from the minimum principle, they
are necessary conditions for the Pontryagin minimum at the point w 0 .
As in previous problems, there is another way, convenient for practical use, to calculate
the quantities D k (H̄ λ ). Given any λ ∈ 0 and tk ∈ , we set
The function (k H̄ λ )(t) is continuously differentiable at each point of the set [t0 , tf ] \ ,
since this property hold for x 0 (t) and ψ(t). The latter follows from the equation
Proposition 3.49. For any λ ∈ 0 and any tk ∈ , the following equalities hold:
d d
D k (H̄ λ ) = − (k H̄ λ )(tk − 0) = − (k H̄ λ )(tk + 0). (3.245)
dt dt
Relations (3.244), (3.230), (3.235), (3.236), and formula (3.242) can be used for obtaining
one more representation of the value D k (H̄ λ ).
3.4. Quadratic Conditions for Mixed Control-State Constrained Problems 169
where the function ψt (t) is defined by ψt (t) = −H (t, x 0 (t), u0 (t), ψ(t)), the value [ψ̇t ]k =
ψ̇tk+ − ψ̇tk− is the jump of the derivative ψ̇t (t) at the point tk , and the vectors ẋ 0k− , ψ̇ k− , ψ̇tk−
and ẋ 0k+ , ψ̇ k+ , ψ̇tk+ are the left and the right limit values of the derivatives ẋ 0 (t), ψ̇(t) and
ψ̇t (t) at tk , respectively.
of the variable ū is positive semidefinite on the cone formed by the vectors ū ∈ Rd(u) such
that
gu (t, x 0 (t), u0 (t))ū = 0,
ϕj u (t, x 0 (t), u0 (t))ū ≤ 0 ∀ j ∈ Iϕ (t, x 0 (t), u0 (t)), (3.249)
μj (t)ϕj u (t, x 0 (t), u0 (t))ū = 0 ∀ j ∈ Iϕ (t, x 0 (t), u0 (t)),
where H̄uu is the matrix of second derivatives with respect to u of the function H̄ , and
Iϕ (t, x, u) is the set of indices of active inequality constraints ϕj (t, x, u) ≤ 0 at (t, x, u),
defined by (3.220).
(LC−
) For any tk ∈ , the quadratic form
of the variable ū is positive semidefinite on the cone formed by the vectors ū ∈ Rd(u) such
that
gu (tk , x 0 (tk ), u0k− )ū = 0,
ϕj u (tk , x 0 (tk ), u0k− )ū ≤ 0 ∀ j ∈ Iϕ (tk , x 0 (tk ), u0k− ), (3.251)
μj ϕj u (tk , x (tk ), u )ū = 0 ∀ j ∈ Iϕ (tk , x (tk ), u )
k− 0 0k− 0 0k−
(LC+
) For any tk ∈ , the quadratic form
of the variable ū is positive semidefinite on the cone formed by the vectors ū ∈ Rd(u) such
that
gu (tk , x 0 (tk ), u0k+ )ū = 0,
ϕj u (tk , x 0 (tk ), u0k+ )ū ≤ 0 ∀ j ∈ Iϕ (tk , x 0 (tk ), u0k+ ), (3.253)
μk+
j ϕ ju k(t , x 0 (t ), u0k+ )ū = 0 ∀ j ∈ I (t , x 0 (t ), u0k+ ).
k ϕ k k
Let M be an arbitrary subset of the compact set 0 . Denote by Leg(M) the subset of all
Legendrian elements λ ∈ M. It follows from Theorems 3.48 and 3.51 and Proposition 3.46
that
Leg(M0 ) = M0 . (3.255)
Now we introduce the critical cone K and the quadratic form λ (·) which will be used for
the statement of the quadratic optimality condition.
Critical cone. As above, we denote by P W 1,2 ([t0 , tf ], Rd(x) ) the space of piecewise
continuous functions x̄(·) : [t0 , tf ] → Rd(x) that are absolutely continuous on each interval
of the set [t0 , tf ] \ and have a square integrable first derivative. Given tk ∈ and x̄(·) ∈
P W 1,2 ([t0 , tf ], Rd(x) ), we use the notation x̄ k− = x̄(tk − 0), x̄ k+ = x̄(tk + 0), [x̄]k = x̄ k+ −
x̄ k− . Denote by Z2 () the space of triples z̄ = (ξ̄ , x̄, ū) such that
the set of indices of active inequality constraints Fi (p) ≤ 0 at the point p 0 , where Fi are the
components of the vector function F .
Let K denote the set of z̄ = (ξ̄ , x̄, ū) ∈ Z2 () such that
where M0 (ϕj0 ) = {t ∈ [t0 , tf ] | ϕj (t, w0 (t)) = 0}, p̄ = (x̄(t0 ), x̄(tf )), w̄ = (x̄, ū), and [ẋ 0 ]k is
the jump of the function ẋ 0 (t) at the point tk ∈ , i.e., [ẋ 0 ]k = ẋ 0k+ − ẋ 0k− = ẋ 0 (tk + 0) −
ẋ 0 (tk − 0), tk ∈ . Obviously, K is a closed convex cone in the space Z2 (). We call it the
critical cone of problem (3.217)–(3.219) at the point w 0 .
The following question is of interest: Which inequalities in the definition of K can
be replaced by equalities without affecting K? This question is answered below.
3.4. Quadratic Conditions for Mixed Control-State Constrained Problems 171
Note that conditions (3.260) and (3.261) can be written in brief as α0 J
(p 0 )p̄ = 0,
αF
(p0 )p̄= 0, and μ(t)ϕw (t, w0 (t))w̄(t) = 0. Proposition 3.52 gives an answer to the
question posed above. According to this proposition, for any λ ∈ 0 , conditions (3.256)–
(3.261) also define K. It follows that if, for some λ = (α0 , α, β, ψ, ν, μ) ∈ 0 , the condition
α0 > 0 holds, then, in the definition of K, the inequality J
(p 0 )p̄ ≤ 0 can be replaced by
the equality J
(p0 )p̄ = 0. If, for some λ ∈ 0 and i0 ∈ {1, . . . , d(F )}, the condition αi0 > 0
holds, then the inequality Fi
0 (p0 )p̄ ≤ 0 can be replaced by the equality Fi
0 (p0 )p̄ = 0.
Finally, for any j ∈ {1, . . . , d(ϕ)} and λ ∈
0 , the inequality ϕj w (t, w (t))w̄(t) ≤ 0 can be
0
replaced by the equality ϕj w (t, w (t))w̄(t) = 0 a.e. on the set {t ∈ [t0 , tf ] | μj (t) > 0} ⊂
0
M0 (ϕj0 ). Every such change gives an equivalent system of conditions still defining K.
The following question is also of interest: Under what conditions can one of the
endpoint inequalities in the definition of K be omitted without affecting K? In particular,
when can the inequality J
(p 0 )p̄ ≤ 0 be omitted?
combined with (3.257)–(3.259) and (3.261) imply that J
(p0 )p̄ = 0; i.e., K can be defined
by conditions (3.257)–(3.259), (3.261), and (3.262)) as well.
λ = (α0 , α, β, ψ, ν, μ) ∈ 0 ,
we set
[H̄xλ ]k = H̄xλk+ − H̄xλk− , k = 1, . . . , s, (3.263)
where H̄xλk− and H̄xλk+ are defined by (3.239). Thus, [H̄xλ ]k denotes a jump of the function
H̄x (t, x 0 (t), u0 (t), ψ(t), ν(t), μ(t)) at t = tk ∈ . It follows from the adjoint equation (3.230)
that
[H̄xλ ]k = −[ψ̇]k , k = 1, . . . , s, (3.264)
172 Chapter 3. Quadratic Conditions for Optimal Control Problems
∂ 2l 0 ∂ 2 H̄
λ
lpp (p0 ) = (p , α0 , α, β), H̄ λ
ww (w 0
) = (t, x 0 (t), u0 (t), ψ(t), ν(t), μ(t)). (3.265)
∂p 2 ∂w2
1
k
x̄av = (x̄ k− + x̄ k+ ), k = 1, . . . , s. (3.266)
2
k is an average value of the function x̄ at t ∈ .
Here x̄av k
We are now ready to introduce the quadratic form, which takes into account the
discontinuities of the control u0 . For any λ ∈ 0 and z̄ = (ξ̄ , x̄, ū) ∈ Z2 (), we set
1 k λ 2
s
λ (z̄) = D (H̄ )ξ̄k + 2[H̄xλ ]k x̄av
k
ξ̄k
2
k=1 (3.267)
1 λ 0 1 tf λ
+ lpp (p )p̄, p̄ + H̄ww (t, w0 )w̄, w̄ dt,
2 2 t0
where w̄ = (x̄, ū), p̄ = (x̄(t0 ), x̄(tf )). Recall that the value D k (H̄ λ ) is defined by (3.242),
and it is nonpositive for any λ ∈ M0 . Obviously, λ is quadratic in z̄ and linear in λ. Set
1 k λ 2
s
λ
ω (ξ̄ , x̄) = D (H̄ )ξ̄k + 2[H̄xλ ]k x̄av
k
ξ̄k . (3.268)
2
k=1
This quadratic form is related to the discontinuities of control u0 , and we call it the internal
form. According to (3.247) and (3.264) it can be written as follows:
1 k+ 0k−
s
λ
ω (ξ̄ , x̄) = ψ̇ ẋ − ψ̇ k− ẋ 0k+ + [ψ̇t ]k ξ̄k2 − 2[ψ̇]k x̄av
k
ξ̄k . (3.269)
2
k=1
Furthermore, we set
tf
1 λ 0 1
ωλ (w̄) = lpp (p )p̄, p̄ + H̄ww
λ
(t, w 0 )w̄, w̄ dt. (3.270)
2 2 t0
This quadratic form is the second variation of the Lagrangian of problem (3.217)–(3.219)
at the point w 0 . We call it the external form. Thus, the quadratic form λ (z̄) is a sum of the
internal and external forms:
λ (z̄) = ω
λ
(ξ̄ , x̄) + ωλ (w̄). (3.271)
3.4. Quadratic Conditions for Mixed Control-State Constrained Problems 173
Theorem 3.54. If w0 is a Pontryagin minimum, then the following Condition A holds: The
set M0 is nonempty and
max λ (z̄) ≥ 0 ∀ z̄ ∈ K.
λ∈M0
The proof of this theorem was given in [86] and published in [95].
Next we formulate the basic sufficient condition in problem (3.217)–(3.219). We call
it briefly Condition B(). It is sufficient not only for a Pontryagin minimum, but also for
a bounded strong minimum defined below. We give a preliminary definition of a strong
minimum which is slightly different from the commonly used definition.
Set M(C ). Let be an order function (see Definition 2.17). For any C > 0, we
denote by M(C) the set of all λ ∈ M0 such that the following condition holds:
Condition (3.272) strengthens the minimum condition (3.234), and we call (3.272) the
minimum condition of strictness C (or C-growth condition for H ). For any C > 0,
M(C) is a closed subset in M0 and, therefore, a finite dimensional compact set.
174 Chapter 3. Quadratic Conditions for Optimal Control Problems
On the subspace (3.257)–(3.258) of the space Z 2 (), the value γ̄ (z̄) is a norm, which is
equivalent to the norm of the space Z 2 (). Let be an order function.
Definition 3.55. We say that a point w0 satisfies condition B() if there exists C > 0 such
that the set M(C) is nonempty and
Theorem 3.56. If there exists an order function (t, u) such that Condition B() holds,
then w0 is a strict bounded strong minimum.
Condition B() obviously holds if for some C > 0 the set M(C) is nonempty, and
if the cone K consists only of zero. Therefore, Theorem 3.56 implies the following.
Corollary 3.57. If for some C > 0 the set M(C) is nonempty, and if K = {0}, then w 0 is
a strict bounded strong minimum.
Corollary 3.57 states the first-order sufficient condition of a bounded strong minimum.
γ -sufficiency. Quadratic Condition B() implies not only a bounded strong mini-
mum, but also a certain strengthening of this concept which is called the γ -sufficiency on
the set of bounded strong sequences. Below, we introduce the (higher) order γ and formu-
late two concepts: γ -sufficiency for Pontryagin minimum and γ -sufficiency for bounded
strong minimum. Regarding the point w0 = (x 0 , u0 ), tested for optimality, we again use
Assumption 3.41. Let (t, u) be an order function. Set
tf
γ (δw) =
δx
2C + (t, u0 + δu) dt.
t0
The functional γ is defined on the set of all variations δw = (δx, δu) ∈ W such that
The functional γ is the order associated with the order function (t, u). Following the
general theory [55], we also call γ the higher order. Thus, with the point w0 we associate
the family of order functions and the family of corresponding orders γ . Let us denote the
latter family by Ord(w0 ).
Let us introduce the violation function of problem (3.217)–(3.219) at the point w0 :
where
σJFK (δp) = max J (p 0 + δp) − J (p0 ), max Fi (p0 + δp), |K(p 0 + δp)| ,
tf i=1,...,d(F )
Definition 3.58. We say that w0 is a point of γ -sufficiency on S if there exists an ε > 0 such
that, for any sequence {δwn } ∈ S, we have σ (δwn ) ≥ εγ (δwn ) for all sufficiently large n.
Let us now introduce a set of sequences related to a bounded strong minimum. Denote
by Sbs the set of all sequences {δwn } which are bounded strong on Q and satisfy the following
conditions:
(a) (p0 + δpn ) ∈ P for all sufficiently large n,
(b) σ (δwn ) → 0 as n → ∞.
Conditions (a) and (b) hold for every sequence {δwn } that violates minimality, so a bounded
strong minimum can be treated as a minimum on Sbs (the subscript bs means “bounded
strong”).
The proof of this theorem was given in [86] and published in [94, 95]. Theorem 3.59
particularly shows a nontrivial character of minimum guaranteed by condition B(). Let
us explain this in more detail. A sequence {δwn } is said to be admissible if the sequence
{w0 + δwn } satisfies all constraints of the canonical problem. We say that w0 is a point of
γ -minimum on Sbs (or the γ -growth condition for the cost function holds on Sbs ) if there
exists ε > 0 such that, for any admissible sequence {δwn } ∈ Sbs , we have J (p0 + δpn ) −
J (p0 ) ≥ εγ (δwn ) for all sufficiently large n. Clearly, γ -sufficiency on Sbs implies γ -
minimum on Sbs . In fact, it is the sufficient Condition B() that ensures γ -minimum on
Sbs . A nontrivial character of γ -minimum on Sbs is caused by a nontrivial definition of the
order function which specifies the higher-order γ .
Now, let us discuss an important question concerning characterization of Condition
λ ∈ M(C).
δH [t, v] := H (t, x 0 (t), u0 (t) + v, ψ(t)) − H (t, x 0 (t), u0 (t), ψ(t)). (3.274)
176 Chapter 3. Quadratic Conditions for Optimal Control Problems
Definition 3.60. We say that, at the point w 0 , the Hamiltonian satisfies a local quadratic
growth condition if there exist ε > 0 and α > 0 such that for all t ∈ [t0 , tf ] \ the following
inequality holds:
δH [t, v] ≥ α|v|2 if v ∈ Rd(u) , g(t, x 0 (t), u0 (t) + v) = 0,
(3.275)
ϕ(t, x (t), u (t) + v) ≤ 0, |v| < ε.
0 0
the first and second derivative with respect to u of the augmented Hamiltonian, and adopt a
similar notation for the Hamiltonian function H . Similarly, we denote gu (t) :=
gu (t, x 0 (t), u0 (t)), ϕi (t) := ϕi (t, x 0 (t), u0 (t)), ϕiu (t) := ϕiu (t, x 0 (t), u0 (t)), i = 1, . . . , d(ϕ).
We shall formulate a generalization of the strengthened Legendre condition using the quad-
ratic form H̄uu (t)v, v complemented by some special nonnegative term ρ(t, v) which will
be homogeneous (not quadratic) of the second degree with respect to v. Let us define this
additional term.
For any number a, we set a + = max{a, 0} and a − = max{−a, 0}, so that a + ≥ 0,
a ≥ 0, and a = a + − a − . Denote by
−
the characteristic function of the set {τ | ϕi (τ ) < 0}, i = 1, . . . , d(ϕ). If d(ϕ) > 1, then, for
any t ∈ [t0 , tf ] and any v ∈ Rd(u) , we set
d(ϕ)
μj (t) − +
ρ(t, v) = max χi (t) ϕj u (t)v ϕiu (t)v . (3.277)
1≤i≤d(ϕ) |ϕi (t)|
j =1
Here, by definition,
μj (t)
χi (t) = 0 if ϕi (t) = 0, i, j = 1, . . . , d(ϕ).
|ϕi (t)|
μ1 (t) − +
ρ(t, v) = χ2 (t) ϕ1u (t)v ϕ2u (t)v
|ϕ2 (t)|
− + (3.278)
μ2 (t)
+ χ1 (t) ϕ2u (t)v ϕ1u (t)v .
|ϕ1 (t)|
Definition 3.61. We say that the Hamiltonian satisfies the generalized strengthened Legen-
dre condition if
∃ α > 0, > 0 such that ∀ t ∈ [tf , tf ] \ :
(3.280)
2 H̄uu (t)v, v + ρ(t, v) ≥ α|v|
1 2 ∀ v ∈ C ().
t
Theorem 3.62. A local quadratic growth condition for the Hamiltonian is equivalent to the
generalized strengthened Legendre condition.
This theorem was proved in [8] for the control constrained problem (without mixed
constraints).
We note that Ct () is in general a larger set than the local cone Ct of critical direc-
tions for the Hamiltonian, i.e., the directions v ∈ Rd(u) , such that
gu (t)v = 0, ϕj u (t)v ≤ 0 if ϕj (t) = 0,
(3.281)
ϕj u (t)v = 0 if μj (t) > 0, j = 1, . . . , d(ϕ).
Let us note that this inequality is stronger than (3.280), since the function ρ(t, v) is nonneg-
ative.
Theorem 3.63. Condition (3.282) implies a local quadratic growth of the Hamiltonian.
Denote by M0+ the set of λ ∈ M0 such that the following conditions hold:
(a) H (t, x 0 (t), u, ψ(t)) > H (t, x 0 (t), u0 (t), ψ(t))
if t ∈ [t0 , tf ]\, u ∈ U(t, x 0 (t)), u = u0 (t), where
U(t, x) = {u ∈ Rd(u) | (t, x, u) ∈ Q, g(t, x, u) = 0, ϕ(t, x, u) ≤ 0};
(b) H (tk , x 0 (tk ), u, ψ(tk )) > H k
if tk ∈ , u ∈ U(tk , x 0 (tk )), u ∈
/ {u0k− , u0k+ }, where
H := H (tk , x (tk ), u , ψ(tk )) = H (tk , x 0 (tk ), u0k+ , ψ(tk )).
k 0 0k−
Denote by Leg+ (M0+ ) the set of all strictly Legendrian elements λ ∈ M0+ .
Theorem 3.64. An element λ ∈ Leg+ (M0+ ) iff there exists C > 0 such that λ ∈ M(C).
where P and Q are open sets, and x, u, F , K, f , g, and ϕ are vector functions.
We assume that the functions J , F , and K are defined and twice continuously
differentiable on P , and the functions f , g, and ϕ are defined and twice contin-
uously differentiable on Q; moreover, g and ϕ satisfy the linear independence assump-
tion (see Assumption 3.40).
where ψ, ν, and μ are row vectors of the dimensions d(x), d(g), and d(ϕ), respectively.
Let us define the endpoint Lagrange function
such that ψ(·) : → (Rd(x) )∗ and ψ0 (·) : → R1 are piecewise smooth functions, con-
tinuously differentiable on each interval of the set \ , and ν(·) : → (Rd(g) )∗ and
3.4. Quadratic Conditions for Mixed Control-State Constrained Problems 179
μ(·) : → (Rd(ϕ) )∗ are piecewise continuous functions, Lipschitz continuous on each in-
terval of the set \ .
Denote by M0 the set of the normed tuples λ satisfying the conditions of the minimum
principle for the trajectory T :
)
d(F
d(K)
α0 ≥ 0, α ≥ 0, αF (p) = 0, α0 + αi + |βj | = 1,
i=1 j =1
ψ̇ = −H̄x , ψ̇0 = −H̄t , H̄u = 0, t ∈ \ ,
(3.293)
ψ(t0 ) = −lx0 , ψ(tf ) = lxf , ψ0 (t0 ) = −lt0 , ψ0 (tf ) = ltf ,
min H (t, x(t), u, ψ(t)) = H (t, x(t), u(t), ψ(t)), t ∈ \ ,
u∈U(t,x(t))
H (t, x(t), u(t), ψ(t)) + ψ0 (t) = 0, t ∈ \ ,
where U(t, x) = {u ∈ Rd(u) | g(t, x, u) = 0, ϕ(t, x, u) ≤ 0, (t, x, u) ∈ Q}. The derivatives lx0
and lxf are at (p, α0 , α, β), where p = (t0 , x(t0 ), tf , x(tf )), and the derivatives H̄x , H̄u , and
H̄t are at (t, x(t), u(t), ψ(t), ν(t), μ(t)), where t ∈ \ . (Condition H̄u = 0 follows from
the others conditions in this definition, and therefore could be excluded; yet we need to use
it later.)
We define the Pontryagin minimum in problem (3.284)–(3.288) on a variable interval
[t0 , tf ] as in Section 3.3.2 (see Definition 3.25). The condition M0 = ∅ is equivalent to
the Pontryagin’s minimum principle. It is a first-order necessary condition of Pontryagin
minimum for the trajectory T . Thus, the following theorem holds (see, e.g., [76]).
Theorem 3.65. If the trajectory T affords a Pontryagin minimum, then the set M0 is
nonempty.
Assume that M0 is nonempty. Using the definition of the set M0 and the linear
independence assumption for g and ϕ one can easily prove the following statement.
Proposition 3.66. The set M0 is a finite-dimensional compact set, and the mapping λ →
(α0 , α, β) is injective on M0 .
where H̄xk− = H̄x (tk , x(tk ), u(tk −), ψ(tk ), ν(tk −), μ(tk −)), H̄xk+ = H̄x (tk , x(tk ), u(tk +), ψ(tk ),
ν(tk +), μ(tk +)), [H̄t ]k = H̄tk+ − H̄tk− , etc.
and such that their first derivative is square integrable. For each x̄ ∈ P W 1,2 (, Rd(x) ),
tk ∈ , we set x̄ k− = x̄(tk −), x̄ k+ = x̄(tk +), [x̄]k = x̄ k+ − x̄ k− . Further, we let z̄ = (t¯0 , t¯f ,
ξ̄ , x̄, ū), where
Thus,
z̄ ∈ Z2 () := R2 × Rs × P W 1,2 (, Rd(x) ) × L2 (, Rd(u) ).
Moreover, for given z̄, we set
By IF (p) = {i ∈ {1, . . . , d(F )} | Fi (p) = 0} we denote the set of active indices of the con-
straints Fi (p) ≤ 0.
Let K be the set of all z̄ ∈ Z2 () satisfying the following conditions:
where M0 (ϕj0 ) = {t ∈ [t0 , tf ] | ϕj (t, w 0 (t)) = 0}, p = (t0 , x(t0 ), tf , x(tf )), w = (x, u). It is
obvious that K is a convex cone in the Hilbert space Z2 (), and we call it the critical cone.
If the interval is fixed, then we set p := (x0 , xf ) = (x(t0 ), x(tf )), and in the definition of K
we have t¯0 = t¯f = 0, x̄¯0 = x̄0 , x̄¯f = x̄f , and p̄¯ = p̄ := (x̄0 , x̄f ). Define quadratic forms ωe ,
ω, and by formulas (3.156), (3.157), and (3.158), respectively. Now, we formulate the
main necessary quadratic condition of a Pontryagin minimum in the problem on a variable
time interval.
Theorem 3.68. If the trajectory T yields a Pontryagin minimum, then the following Con-
dition A holds: The set M0 is nonempty and
of Hilbert space Z2 (). Recall that the critical cone K is contained in the subspace (3.300).
For any C > 0, we denote by M(C) the set of all λ ∈ M0 such that condition (3.272) holds.
Theorem 3.69. For the trajectory T , assume that the following Condition B() holds:
There exists C > 0 such that the set M(C) is nonempty and
for all z̄ ∈ K. Then the trajectory T affords a strict bounded strong minimum.
Again we can use the characterization of the condition λ ∈ M(C) formulated in the
previous section.
Chapter 4
Here we derive tests for the positive semidefiniteness, respectively, positive definiteness of
the quadratic form on the critical cone K (introduced in Chapter 2 for extremals with
jumps of the control). In Section 4.1, we derive such tests for the simplest problem of
the calculus of variations and for an extremal with only one corner point. We come to a
generalization of the concept of conjugate point which allows us to formulate both necessary
and sufficient second-order optimality conditions for broken extremals. Three numerical
examples illustrate this generalization. Further, we concentrate on sufficient conditions for
positive definiteness of the quadratic form in the auxiliary problem. We show that if there
exists a solution to the Riccati matrix equation satisfying a certain jump condition, then
the quadratic form can be transformed into a perfect square. This gives a possibility of
proving a sufficient condition for positive definiteness of the quadratic form in the auxiliary
problem and thus to obtain one more sufficient condition for optimality of broken extremals.
At the end of Section 4.1, we obtain such condition for the simplest problem of the calculus
of variations, and then, in Section 4.2, we prove it for the general problem (without constraint
g(t, x, u) = 0).
Here, x(·) : [t0 , tf ] → Rm is absolutely continuous, and u(·) : [t0 , tf ] → Rm is bounded and
measurable. Set w(·) = (x(·), u(·)). Then w(·) is an element of the space
We say that w(·) = (x(·), u(·)) is an admissible pair if w(·) ∈ W and the constraints (4.2),
(4.3) hold for it. Let an admissible pair w(·) = (x(·), u(·)) be given. Assume that u(·) is a
piecewise continuous function with a unique point of discontinuity t∗ ∈ (t0 , tf ). Denote by
the singleton {t∗ }. For t∗ , we set u− = u(t∗ −), u+ = u(t∗ +), and [u] = u+ − u− . Thus,
[u] is a jump of the function u(·) at the point t∗ .
In correspondence with (4.3), assume that (t, x(t), u(t)) ∈ Q for all t ∈ [t0 , tf ] \ , and
(t∗ , x(t∗ ), u− ) ∈ Q, (t∗ , x(t∗ ), u+ ) ∈ Q. Also assume that there exist a constant C > 0 and a
small number ε > 0 such that
Here ψ ∈ (Rm )∗ is a row vector while x, u ∈ R m are column vectors. For the Pontryagin
function
H (t, x, u, ψ) = ψu + F (t, x, u), (4.6)
we set H − = H (t∗ , x(t∗ ), u− , ψ(t∗ )), H + = H (t∗ , x(t∗ ), u+ , ψ(t∗ )), and [H ] = H + − H − .
The equalities
[H ] = 0, [ψ] = 0 (4.7)
are the Weierstrass–Erdmann conditions. They are known as necessary conditions for the
strong minimum. However, they are also necessary for the Pontryagin minimum introduced
in Chapter 2. For convenience, let us recall the definitions of Pontryagin and bounded strong
minima in the simplest problem.
We say that the pair of functions w = (x, u) is a point of Pontryagin minimum in
problem (4.1)–(4.3) if for each compact set C ⊂ Q there exists ε > 0 such that J (w̃) ≥ J (w)
for all admissible pairs w̃ = (x̃, ũ) such that
(a) max[t0 ,tf ] |x̃(t) − x(t)| < ε,
t
(b) t0f |ũ(t) − u(t)| dt < ε,
(c) (t, x̃(t), ũ(t)) ∈ C.
4.1. Jacobi-Type Conditions and Riccati Equation for Broken Extremals 185
We say that the pair w = (x, u) is a point of bounded strong minimum in problem (4.1)–(4.3),
if for each compact set C ⊂ Q there exists ε > 0 such that J (w̃) ≥ J (w) for all admissible
pairs w̃ = (x̃, ũ), such that
(a) max[t0 ,tf ] |x̃(t) − x(t)| < ε,
(b) (t, x̃(t), ũ(t)) ∈ C.
Clearly, the following implications hold: strong minimum =⇒ bounded strong minimum =⇒
Pontryagin minimum =⇒ weak minimum.
As already mentioned, the Weierstrass–Erdmann conditions are necessary for the
Pontryagin minimum. As shown in Chapter 2, they can be supplemented by additional
condition of the same type. We set
where ψ̇ − = ψ̇(t∗ −), ψ̇ + = ψ̇(t∗ +), ẋ − = ẋ(t∗ −), ẋ + = ẋ(t∗ +), and [Ft ] = Ft (t∗ , x(t∗ ), u+ )−
Ft (t∗ , x(t∗ ), u− ). Then a ≥ 0 is a necessary condition for the Pontryagin minimum.
As we know, the value D(H ) can be computed in a different way. Consider the
function
(H )(t) := H (t, x(t), u+ , ψ(t)) − H (t, x(t), u− , ψ(t))
(4.9)
= ψ(t)[ẋ] + F (t, x(t), u+ ) − F (t, x(t), u− ) .
d
(H )|t∗ +0 = ψ̇ + [ẋ] − [ψ̇]ẋ + + [Ft ], (4.10)
dt
d
(H )|t∗ −0 = ψ̇ − [ẋ] − [ψ̇]ẋ − + [Ft ]. (4.11)
dt
d d
(H )|t∗ −0 = (H )|t∗ +0 = −D(H ). (4.12)
dt dt
Note that the inequality a ≥ 0 and the Weierstrass–Erdmann conditions are implied
by the conditions of the minimum principle, which is equivalent to the Weierstrass con-
dition in this problem. Here, the minimum principle has the form: H (t, x(t), u, ψ(t)) ≥
H (t, x(t), u(t), ψ(t)) if t ∈ [t0 , tf ] \ , u ∈ Rm , (t, x(t), u) ∈ Q. Let us also formulate the
strict minimum principle:
(a) H (t, x(t), u, ψ(t)) > H (t, x(t), u(t), ψ(t))
for all t ∈ [t0 , tf ] \ , u ∈ Rm , (t, x(t), u) ∈ Q, u = u(t),
(b) H (t∗ , x(t∗ ), u, ψ(t∗ )) > H (t∗ , x(t∗ ), u− , ψ(t∗ )) = H (t∗ , x(t∗ ), u+ , ψ(t∗ ))
for all u ∈ Rm , (t, x(t), u) ∈ Q, u = u− , u = u+ .
Now we define a quadratic form that corresponds to an extremal w(·) with a cor-
ner point. As in Section 2.1.5, denote by P W 1,2 [t0 , tf ], Rm the space of all piecewise
continuous functions x̄(·) : [t0 , tf ] −→ Rm that are absolutely continuous on each of the
intervals in [t0 , tf ]\ whose derivatives are square Lebesgue integrable. For t∗ , we set
186 Chapter 4. Jacobi-Type Conditions and Riccati Equation for Broken Extremals
x̄ − = x̄(t∗ −), x̄ + = x̄(t∗ +), [x̄] = x̄ + − x̄ − . Recall that the space P W 1,2 ([t0 , tf ], Rm )
t
˙ ȳ(t)
with the inner product (x̄, ȳ) = x̄(0), ȳ(0) + [x̄], [ȳ] + t0f x̄(t), ˙ dt is a Hilbert space.
We set
Z2 () = R1 × P W 1,2 [t0 , tf ], Rm × L2 [t0 , tf ], Rm
and denote by z̄ = (ξ̄ , x̄, ū) an element of the space Z2 (), where
ξ̄ ∈ R1 , x̄(·) ∈ P W 1,2 [t0 , tf ], Rm , ū(·) ∈ L2 [t0 , tf ], Rm .
In the Hilbert space Z2 (), we define a subspace and a quadratic form by setting
and
1 2 1 tf
(z̄) = a ξ̄ − 2[ψ̇]x̄av ξ̄ + Fww w̄(t), w̄(t) dt, (4.14)
2 2 t0
respectively, where
1 −
x̄av = x̄ + x̄ + , Fww = Fww (t, x(t), u(t)), w = (x, u),
2
Fxx Fxu
Fww = , w̄(·) = (x̄(·), ū(·)).
Fux Fuu
The auxiliary minimization problem for an extremal w(·) with a corner point is for-
mulated as follows:
Theorem 4.1. If w(·) is a Pontryagin minimum, then the following conditions hold:
(a) the Euler equation,
(b) the minimum principle (the Weierstrass condition),
4.1. Jacobi-Type Conditions and Riccati Equation for Broken Extremals 187
Theorem 4.2. If the following conditions hold, then w(·) is a point of a strict bounded
strong minimum:
(a) the Euler equation,
(b) the strict minimum principle (the strict Weierstrass condition),
(c) the strengthened Legendre condition (SL),
(d) a > 0,
t
(e) there exists ε > 0 such that (z̄) ≥ ε(ξ̄ 2 + t0f ū(t), ū(t) dt) for all z̄ ∈ K.
Theorems 4.1 and 4.2 follow from Theorems 2.4 and 2.102, respectively. Note that
the functional tf
γ̄ (z̄) = ξ̄ 2 + ū(t), ū(t) dt (4.19)
t0
on the subspace K is equivalent to the squared norm: γ̄ ∼ (z̄, z̄).
Now our goal is to derive the tests for positive semidefiniteness and positive definite-
ness of the quadratic form on the subspace K in the case of an extremal with a single
corner point. We give such tests in the form of the Jacobi-type conditions and in terms of
solutions to the Riccati equations.
A −1 F ,
= −Fuu −1 ,
B = −Fuu
ux
−1 ∗ −1 . (4.20)
C = −Fxu Fuu Fux + Fxx , A = −Fxu Fuu
All derivatives are computed along the trajectory (t, w(t)). Note that
B ∗ = B, C ∗ = C,
(4.21)
|det B| ≥ const > 0 on [t0 , tf ],
and A, B, and C are matrices with piecewise continuous entries on [t0 , tf ] that are continu-
ous on each of the intervals of the set (t0 , tf ) \ .
Further, we formulate Jacobi-type conditions for an extremal w(·) with a single corner
point t∗ . Denote by X(t) and "(t) two square matrices of order m, where t ∈ [t0 , tf ]. For
X(t) and "(t), we consider the set of differential equations
Ẋ = AX + B",
˙ (4.22)
−" = CX + A∗ "
with the initial conditions
X(t0 ) = O, "(t0 ) = −I , (4.23)
where O and I are the zero matrix and the identity matrix, respectively.
188 Chapter 4. Jacobi-Type Conditions and Riccati Equation for Broken Extremals
Recall that a continuous (and hence a piecewise smooth) solution X(t), "(t) to the
Cauchy problem (4.22), (4.23) allows one to formulate the classical concept of the conju-
gate point. Namely, a point τ ∈ (t0 , tf ] is called conjugate (to the point t0 ) if det X(τ ) = 0.
The absence of a conjugate point in (t0 , tf ) is equivalent to the positive semidefiniteness of
the quadratic form tf
ω= Fww w̄, w̄ dt
t0
on the subspace K0 consisting of pairs w̄ = (x̄, ū) such that
The latter condition is necessary for the weak minimum. The absence of a conjugate point in
(t0 , tf ] is equivalent to the condition of positive definiteness of ω on K0 , which is a sufficient
condition for the strict weak minimum. This is the classical Jacobi condition.
We note that and K pass into ω and K0 , respectively, if we set ξ̄ = 0 in the
definition of the first pair. In our tests of positive semidefiniteness and positive definiteness
of the quadratic form on the subspace K, we use a discontinuous solution X, " to the
Cauchy problem (4.22), (4.23) with certain jump conditions at the point t∗ . Namely, let a
pair X(t), "(t) be a continuous solution to the problem (4.22), (4.23) on the half-interval
[t0 , t∗ ). We set X− = X(t∗ −), and " − = "(t∗ −). The jumps [X] and ["] of the matrix-
valued functions X(t) and "(t) at the point t∗ are uniquely defined by using the relations
a[X] = [ẋ] −[ẋ]∗ " − + [ψ̇]X− , (4.24)
∗
∗ − −
a["] = [ψ̇] −[ẋ] " + [ψ̇]X , (4.25)
where [ẋ] and [ψ̇]∗ are column matrices, while [ẋ]∗ and [ψ̇] are row matrices. Let us define
the right limits X + and " + of the functions X and " at t∗ in the following way:
Then we continue the process of solution of system (4.22) on (t∗ , T ] by using the initial
conditions for X and " at t∗ given by the conditions
Thus, on [t0 , tf ], we obtain a piecewise continuous solution X(t), "(t) to system (4.22) with
the initial conditions (4.23) at t0 and the jump conditions (4.24) and (4.25) at t∗ . On each
of the intervals of the set [t0 , tf ] \ , the matrix-valued functions X(t) and "(t) are smooth.
Briefly, this pair of functions will be called a solution to the problem (4.22)–(4.25) on [t0 , tf ].
Theorem 4.3. The form is positive semidefinite on K iff the solution X(t), "(t) to the
problem (4.22)–(4.25) on [t0 , tf ] satisfies the conditions
where Q(t) = "(t)X−1 (t), Q− = Q(t∗ −), and X −1 (t) is the inverse matrix to X(t).
4.1. Jacobi-Type Conditions and Riccati Equation for Broken Extremals 189
Theorem 4.4. The form is positive definite on K iff the solution X(t), "(t) to the
problem (4.22)–(4.25) on [t0 , tf ] satisfies conditions (4.26)–(4.28), together with the addi-
tional condition
det X(tf ) = 0. (4.29)
The conditions for positive semidefiniteness, and those for positive definiteness of
on K, which are given by these two theorems can easily be reformulated in terms of a
solution to the corresponding matrix Riccati equation. Indeed, if X(t), "(t) is a solution
to system (4.22) on a certain interval ⊂ [t0 , tf ] with det X(t) = 0 on , then, as is well
known, the matrix-valued function Q(t) = "(t)X −1 (t) satisfies the Riccati equation
Q̇ + QA + A∗ Q + QBQ + C = 0 (4.30)
on . Let us prove this assertion. Differentiating the equality " = QX and using (4.22),
we obtain
−CX − A∗ " = " ˙ = Q̇X + QẊ = Q̇X + Q(AX + B").
Consequently,
CX + A∗ " + Q̇X + QAX + QB" = 0.
Multiplying this equation by X−1 from the right, we obtain (4.30). Using (4.20), we can
also represent (4.30) as
−1
Q̇ − (Q + Fxu )Fuu (Q + Fux ) + Fxx = 0. (4.31)
The solution Q = "X −1 has a singularity at the zero point, since X(t0 ) = O. The
question is: How do we correctly assign the initial condition for Q? We can do this in
the following way. In a small half-neighborhood [t0 , t0 + ε), ε > 0, we find a solution to the
Riccati equation for R = Q−1 = X" −1 with the initial condition
R(t0 ) = O, (4.32)
which is implied by (4.23). This Riccati equation for R can easily be obtained. Namely,
differentiating the equality
X = R" (4.33)
and using (4.22), we obtain, for small ε > 0,
Using (4.20), we can transform this Riccati equation into the form
−1
Ṙ + (RF xu + I )Fuu (Fux R + I ) − RF xx R = 0, t ∈ [t0 , t0 + ε]. (4.35)
Thus, we solve the Riccati equation (4.34) or (4.35) with initial condition (4.32) in a certain
half-neighborhood [t0 , t0 +ε) of t0 . Recall that the matrices B and C are symmetric on [t0 , tf ].
Consequently, R is also symmetric, and therefore,
is symmetric.
190 Chapter 4. Jacobi-Type Conditions and Riccati Equation for Broken Extremals
Let Q(t) be the continuous solution of the Riccati equation (4.30) with initial condi-
tion (4.36) on the interval (t0 , t∗ ). The existence of such a solution is a necessary condition
for positive semidefiniteness of on K. Since B(t) and C(t) are symmetric on [t0 , tf ]
and Q(ε) is also symmetric, we have that Q(t) is symmetric on (t0 , t∗ ). Consequently,
Q− = Q(t∗ −) is symmetric. Further, we define a jump condition for Q at t∗ that corre-
sponds to the jump conditions (4.24) and (4.25). This condition has the form
a − (q− )[ẋ] [Q] = (q− )∗ (q− ), (4.37)
where
q− = [ẋ]∗ Q− − [ψ̇]. (4.38)
Note that q− and [ẋ]∗ )∗
are row vectors while (q− and [ẋ] are column vectors. The jump
[Q] of the matrix Q at the point t∗ is uniquely defined by using (4.37) since, according to
Theorem 4.3, the condition
a − (q− )[ẋ] > 0
is necessary for positive semidefiniteness of on K. Note that (q− )∗ (q− ) is a symmetric
positively semidefinite matrix. Hence, [Q] is a symmetric negatively semidefinite matrix.
The right limit Q+ = Q(t∗ +) is defined by the relation
Q+ = Q− + [Q]. (4.39)
It follows that Q+ is symmetric. Using Q+ as the initial condition for Q at the point t∗ ,
we continue the solution of the Riccati equation (4.30) for t > t∗ . The matrix Q(t) is also
symmetric for t > t∗ . Assume that this symmetric solution Q(t) is extended to a certain
interval (t0 , τ ) or half-interval (t0 , τ ], where t∗ < τ ≤ tf . It will be called a solution to
problem (4.30), (4.36), (4.37) on (t0 , τ ) or on (t0 , τ ], respectively.
Theorem 4.5. The form is positive semidefinite on K iff there exists a solution Q to the
problem (4.30), (4.36), (4.37) on (t0 , tf ) that satisfies
Theorem 4.6. The form is positive definite on K iff there exists a solution Q to the
problem (4.30), (4.36), (4.37) on (t0 , tf ] that satisfies inequality (4.40).
We note that the inequality (4.40) is equivalent to condition (4.28). Moreover, set
b− = a − (q− )[ẋ]. Then the inequality (4.40) and the jump condition (4.37) obtain the form
respectively.
use some ideas from [25, 26]. Everywhere below, we assume that condition (SL) holds. It
follows from condition (SL) that is a Legendre form on K (cf., e.g., [42]) in the abstract
sense; i.e., it is a weakly lower semicontinuos functional on K, and the conditions
z̄n ∈ K ∀ n, z̄ ∈ K, (4.41)
ξ̄ n → ξ̄ , ūn → ū weakly in L2 , (4.42)
(z̄n ) → (z̄) (4.43)
imply
ūn − ū
2 → 0, (4.44)
tf
and hence z̄ → z̄ strongly in Z2 (). Here
v
2 = ( t0 v(t), v(t) dt) is the norm in L2 .
n 1/2
Theorem 4.7. If is not positive definite on K, then there exists τ0 ∈ (t0 , tf ] such that
(a) (·) is positive definite on K(τ ) for all τ ∈ (t0 , τ0 ) and
(b) (·) ≥ 0 on K(τ0 ) and there exists z̄ ∈ K(τ0 )\{0} such that (z̄) = 0.
192 Chapter 4. Jacobi-Type Conditions and Riccati Equation for Broken Extremals
Using this theorem, we can define τ0 as the minimum value among all τ ∈ (t0 , tf ]
such that the quadratic form has a nontrivial zero z̄ on the subspace K(τ ). To prove this
theorem, we need two auxiliary assertions.
Proposition 4.8. There exists τ ∈ (t0 , tf ] such that is positive definite on K(τ ).
Choose a certain τ ∈ (t0 , t∗ ) and let z̄ = (ξ̄ , x̄, ū) ∈ K(τ )\{0} be an arbitrary element. Then
ū = 0 and
ūχ[t0 ,τ ] = ū, x̄˙ = ū, x̄(t0 ) = 0,
x̄χ[t0 ,t∗ ] = x̄, x̄(τ ) = x̄ − , x̄ − + [ẋ]ξ̄ = 0,
where χM is the characteristic function of a set M. Consequently,
√
√ |x̄ − | τ − t0
x̄
∞ ≤
ū
1 ≤ τ − t0
ū
2 , |ξ̄ | = ≤
ū
2 ,
√ |[ ẋ]| |[ ẋ]|
|2x̄av | = |x̄ − | ≤ τ − t0
ū
2 .
Therefore,
t
f
Fxx x̄, x̄ dt ≤
Fxx
∞
x̄
2∞ (t∗ − t0 ) ≤
Fxx
∞ (t∗ − t0 )(τ − t0 )
ū
22 ,
t0
t
f
2F ū, x̄ dt ≤ 2
Fxu
∞
x̄
∞
ū
1 ≤ 2
Fxu
∞ (τ − t0 )
ū
2 ,
xu 2
t0
|a| |[Fx ]|
|aξ 2 | ≤ (τ − t0 )
ū
22 , |2[Fx ]x̄av ξ̄ | ≤ (τ − t0 )
ū
22 .
|[ẋ]|2 |[ẋ]|
Consequently,
tf
(z̄) ≥ Fuu ū, ū dt − M(τ − t0 )
ū
22 ≥ #L
ū
22 − M(τ − t0 )
ū
22 , (4.49)
t0
where
|a| |[Fx ]|
M =
Fxx
∞ (t∗ − t0 ) + 2
Fxu
∞ + + .
|[ẋ]|2 |[ẋ]|
Let τ be such that t0 < τ < t0 + #L /M. Then (4.49) implies that is positive definite on
K(τ ).
Proposition 4.9. Assume that is not positive definite on K. Then is positive semi-
definite on K(τ0 ), where τ0 is defined as in (4.47).
We omit the simple proof of this proposition. Now we are ready to prove the theorem.
Proof of Theorem 4.7. We have already proved that is positive definite on K(τ ) for all
τ > t0 sufficiently close to t0 (Proposition 4.8). Consequently, τ0 > t0 . Further, we consider
4.1. Jacobi-Type Conditions and Riccati Equation for Broken Extremals 193
only the nontrivial case where τ0 < tf . We know that (·) ≥ 0 on K(τ0 ) (Proposition 4.9).
We have to show that is not positive on K(τ0 ), i.e., there exists z̄ ∈ K(τ0 )\{0} such that
(z̄) = 0 (the passage through zero). Now we follow [26]. For any τ > τ0 (τ ≤ tf ), is
not positive on K(τ ) . Therefore, for each
1
τn = τ0 + < tf (4.50)
n
there exists
z̄n ∈ K(τn ) (4.51)
such that
(z̄n ) ≤ 0, (4.52)
γ̄ (z̄n ) = 1. (4.53)
The sequence {z̄n } is bounded in K. Therefore, without loss of generality, we assume that
(z̄0 ) ≥ 0. (4.57)
and then
z̄n −→ z̄0 strongly (4.60)
since is a Legendre form. Conditions (4.53) and (4.60) imply
γ̄ (z̄0 ) = 1. (4.61)
yields the minimum in this problem. Then the following first-order necessary optimality
condition holds:
(z̄), z̃ = 0 ∀ z̃ ∈ K(τ ), (4.63)
where z̃ = (ξ̃ , x̃, ũ) = (ξ̃ , w̃),
(z̄) is the Fréchet derivative of the functional at the point
z̄, and (·, ·) is the inner product in Z2 (); in more detail,
tf
(
(z̄), z̃) = a ξ̄ ξ̃ − [ψ̇]x̄av ξ̃ − [ψ̇]x̃av ξ̄ + Fww w̄, w̃ dt. (4.64)
t0
Corollary 4.10. Assume that is not positive definite on K. Then, for τ = τ0 ∈ (t0 , tf ]
(given by (4.47)), there exists a nonzero element z̄ that satisfies (4.62) and (4.63).
1
( (z̄), z̄) = (z̄). (4.65)
2
Proposition 4.11. If, for certain τ ∈ [t0 , tf ], there exists a nonzero element z̄ satisfying
(4.62) and (4.63), then (z̄) = 0, and hence, is not positive definite on K(τ ).
Theorem 4.12. Assume that is not positive definite on K. Then τ0 , given by (4.47), is
minimal among all τ ∈ (t0 , tf ] such that there exists a nonzero element z̄ satisfying condi-
tions (4.62) and (4.63).
4.1.4
-Conjugate Point
In this section, we obtain a dual test for condition (4.63), and then we use it to obtain an
analogue of a conjugate point for a broken extremal. The most important role is played by
the following lemma.
Lemma 4.13. Let τ ∈ (t0 , tf ]. A triple z̄ = (ξ̄ , x̄, ū) satisfies conditions (4.62) and (4.63)
iff there exists a function ψ̄(·) : [t0 , tf ] −→ (Rm )∗ such that for the tuple ξ̄ , x̄, ū, ψ̄ ,
4.1. Jacobi-Type Conditions and Riccati Equation for Broken Extremals 195
Proof. Let z̄ = (ξ̄ , x̄, ū) satisfy conditions (4.62) and (4.63). Consider condition (4.63).
Define the following subspace:
L̃2 ([τ , tf ], Rm ) := {ṽ ∈ L2 ([t0 , tf ], Rm ) | ṽ = 0 a.e. on [t0 , τ ]}.
The operator z̃ −→ (x̃˙ − ũ, ũχ[τ ,tf ] ) maps the space Z2 () onto the space L2 ([t0 , tf ], Rm ) ×
L̃2 ([τ , tf ], Rm ). The operator z̃ −→ ([x̃] − [ẋ]ξ̃ , x̃(t0 ), x̃(tf )) is finite dimensional. Conse-
quently, the image of the operator
z̃ → [x̃] − [ẋ]ξ̃ , x̃˙ − ũ, ũχ[τ ,tf ] , x̃(t0 ), x̃(tf ) ,
which maps from Z2 () into
Rn × L2 ([t0 , tf ], Rm ) × L̃2 ([τ , tf ], Rm ) × Rm × Rm
is closed. The kernel of this operator is equal to K(τ ). Consequently, an arbitrary linear
functional z$ that vanishes on the kernel of this operator admits the following representation:
tf tf
(z , z̃) = ζ̄ ([x̃] − [ẋ]ξ̃ ) −
$ ˙
ψ̄(x̃ − ũ) dt + ν̄ ũ dt + c̄0 x̃(t0 ) + c̄f x̃(tf ), (4.74)
t0 t0
where
ζ̄ ∈ (Rn )∗ , ψ̄ ∈ L2 ([t0 , tf ], (Rm )∗ ), ν̄ ∈ L̃2 ([τ , tf ], (Rm )∗ ), c̄0 , c̄f ∈ (Rm )∗ . (4.75)
Consequently, the condition (4.63) is equivalent to the existence of ζ̄ , ψ̄, ν̄, c̄0 , and c̄f that
satisfy (4.75) and are such that for z$ defined by formula (4.74), we have
(
(z̄), z̃) + (z$ , z̃) = 0 ∀ z̃ ∈ Z2 ().
The exact representation of the latter condition has the form
a ξ̄ ξ̃ − [ψ̇]x̄av ξ̃ − [ψ̇]x̃av ξ̄
tf
+ (Fxx x̄, x̃ + Fxu ū, x̃ + Fux x̄, ũ + Fuu ū, ũ ) dt
t0 tf tf (4.76)
+ ζ̄ ([x̃] − [ẋ]ξ̃ ) − ψ̄(x̃˙ − ũ) dt + ν̄ ũ dt + c̄0 x̃(t0 ) + c̄f x̃(tf )
t0 t0
=0 ∀ z̃ = (ξ̃ , x̃, ũ) ∈ Z2 ().
196 Chapter 4. Jacobi-Type Conditions and Riccati Equation for Broken Extremals
Consequently,
x̄ ∗ Fxu + ū∗ Fuu + ψ̄ = −ν̄. (4.78)
The latter equation is equivalent to condition (4.69).
(b) We set ξ̃ = 0 and ũ = 0 in (4.76). Then
tf tf
−[ψ̇]x̃av ξ̄ + (Fxx x̄, x̃ ) + Fxu ū, x̃ ) dt + ζ̄ [x̃] − ψ̄ x̃˙ dt
t0 t0
+ c̄0 x̃(t0 ) + c̄f x̃(tf ) = 0 ∀ x̃ ∈ P W 1,2 . (4.79)
Using (4.80) and the definitions x̃av = 12 (x̃ − + x̃ + ), [x̃] = x̃ + − x̃ − in (4.79), we obtain
tf
1 − +
− [ψ̇](x̃ + x̃ )ξ̄ + Fxx x̄ + Fxu ū, x̃ dt
2 t0
+ ζ̄ (x̃ + − x̃ − ) − ψ̄(tf )x̃(tf ) + ψ̄(t0 )x̃(t0 ) + ψ̄ + x̃ + − ψ̄ − x̃ −
tf
+ ψ̄˙ x̃ dt + c̄0 x̃(t0 ) + c̄f x̃(tf ) = 0 ∀ x̃ ∈ P W 1,2 . (4.81)
t0
Equation (4.81) implies that the coefficients of x̃ − , x̃ + , x̃(t0 ), x̃(tf ) and the coefficient of x̃
in the integral vanish:
1
− [ψ̇]ξ̄ − ψ̄ − − ζ̄ = 0, (4.82)
2
1
− [ψ̇]ξ̄ + ψ̄ + + ζ̄ = 0, (4.83)
2
ψ̄(t0 ) + c̄0 = 0, (4.84)
− ψ̄(tf ) + c̄f = 0, (4.85)
x̄ ∗ F + ū∗ F + ψ̄˙ = 0.
xx ux (4.86)
Thus, (4.70) and (4.72) hold. Subtracting (4.82) from (4.83) and dividing the result by two,
we obtain
ζ̄ = −ψ̄av . (4.88)
4.1. Jacobi-Type Conditions and Riccati Equation for Broken Extremals 197
Using (4.88) in (4.90), we obtain (4.73). Conditions (4.67), (4.68), and (4.71) and the first
three conditions in (4.66) are implied by (4.62). Thus, all conditions (4.66)–(4.73) hold.
Conversely, if a tuple (ξ̄ , x̄, ū, ψ̄) satisfies conditions (4.66)–(4.73), then one easily verifies
that z̄ = (ξ̄ , x̄, ū) satisfies (4.62) and (4.63).
Let z̄ = (ξ̄ , x̄, ū) ∈ K. Obviously, the condition z̄ = 0 is equivalent to x̄(·) = 0. Thus,
we obtain the following theorem from Theorem 4.12 and Lemma 4.13 under the above
condition (SL).
Theorem 4.14. Assume that is not positive definite on K. Then τ0 (given by equa-
tion (4.47)) is a minimal among all τ ∈ (t0 , tf ] such that there exists a tuple (ξ̄ , x̄, ū, ψ̄) that
satisfies (4.66)–(4.73) and the condition x̄(·) = 0.
In what follows, we will assume that a > 0 and, as above, condition (SL) holds.
Assume that τ0 < tf . We know that ≥ 0 on K(τ0 ). Theoretically, it is possible that
≥ 0 on K(τ1 ) for a certain τ1 > τ0 . In this case, the closed interval [τ0 , τ1 ] is called a
table [25]. Tables occur in optimal control problems [25], but, for a smooth extremal, they
never arise in the calculus of variations. We now show that for the simplest problem of
the calculus of variations, the closed interval [τ0 , tf ] cannot serve as a table in the case of a
broken extremal. To this end, we complete Lemma 4.13 by the following two propositions.
Proposition 4.15. If the functions x̄, ψ̄ ∈ P W 1,2 and ū ∈ L2 satisfy the system
x̄˙ = ū, −ψ̄˙ = x̄ ∗ Fxx + ū∗ Fux , −ψ̄ = x̄ ∗ Fxu + ū∗ Fuu (4.91)
on a certain closed interval ⊂ [t0 , tf ], then the functions x̄ and ψ̄ satisfy the system
on the same closed interval , where A, B, and C are the same as in (4.20).
Substituting these expressions for ū and ū∗ into the first and second equations in (4.91),
respectively, we obtain (4.92).
Proof. Using the expressions ψ̄av = ψ̄ + − 0.5[ψ̄] and x̄av = x̄ + − 0.5[x̄], together with
(4.71) and (4.72), in (4.73), we obtain (4.93).
Definition 4.18. A point τ ∈ (t0 , tf ] is called -conjugate (to t0 ) if there exists a triple
(ξ̄ , x̄, ψ̄) that satisfies conditions (4.94)–(4.100).
Theorem 4.19. The form is positive semidefinite on K iff there is no point that is
-conjugate to t0 on the interval (t0 , tf ). The form is positive definite on K iff there is no
point that is -conjugate to t0 on the half-interval (t0 , tf ].
Now let us examine the condition for positive definiteness of on K(t∗ ). This
condition implies the positive definiteness of ω on K0 (t∗ ) which is defined as a subspace of
4.1. Jacobi-Type Conditions and Riccati Equation for Broken Extremals 199
pairs w̄ = (x̄, ū) ∈ K0 (see Section 4.1.2) such that ū(t) = 0 on [t∗ , tf ]. Let X(t), "(t) be a
matrix-valued solution of the Cauchy problem (4.22)–(4.23) on [t0 , t∗ ]. As is well known,
the positive definiteness of ω on K0 (t∗ ) is equivalent to the condition
Assume that this condition holds. We set Q(t) = "(t)X−1 (t), t ∈ (t0 , t∗ ]. Consider condi-
tions (4.94)–(4.100) for τ ≤ t∗ . These conditions imply that
Let (ξ̄ , x̄, ψ̄) be a solution of this system on the closed interval [t0 , τ ], where τ ∈ (t0 , t∗ ].
Then there exists c̄ ∈ Rm such that
Consequently, ψ̄(t)∗ = "(t)c̄ = "(t)X−1 (t)(X(t)c̄) = Q(t)x̄(t). Using this relation, together
with (4.105), in (4.106), we obtain
If ξ̄ = 0, then from (4.105) and (4.107), we obtain x̄(·) = 0, ψ̄(·) = 0; this contradicts
(4.103). Therefore ξ̄ = 0, and then (4.108) implies
We have obtained this relation from the system (4.102)–(4.106). Conversely, if (4.109)
holds, then, setting ξ̄ = −1 and c̄ = X −1 (τ )[ẋ] and defining x̄, ψ̄ by formulas (4.107), we
obtain a solution of the system (4.102)–(4.106). We have proved the following lemma.
Lemma 4.20. Assume that condition (4.101) holds. Then, τ ∈ (t0 , t∗ ] is a point that is
-conjugate to t0 iff condition (4.109) holds.
We set
μ(t) = a − [ẋ]∗ (Q(t) − (t))[ẋ] + [ψ̇][ẋ]. (4.110)
Then, by Lemma 4.20, the absence of a -conjugate point in (t0 , t∗ ] is equivalent to the
condition
μ(t) = 0 ∀ t ∈ (t0 , t∗ ]. (4.111)
200 Chapter 4. Jacobi-Type Conditions and Riccati Equation for Broken Extremals
We show further that the function μ(t) does not increase on (t0 , t∗ ] and μ(t0 + 0) = +∞.
Consequently, condition (4.111) is equivalent to μ(t∗ ) > 0, which is another form of condi-
tion (4.28) or condition (4.40).
Proposition 4.21. Assume that a symmetric matrix Q(t) satisfies the Riccati equation (4.31)
on (t0 , τ ), where τ ≤ t∗ . Then
Consequently,
−1
X(t) = (t − t0 )Fuu (t0 , w(t0 )) + o(t), "(t) = −I + o(1) as t → t0 + 0. (4.117)
Thus,
1
Q(t) = − (Fuu (t0 , w(t0 )) + o(1)) as t → t0 + 0; (4.118)
t − t0
this implies
1
−Q(t)[ẋ], [ẋ] = (Fuu (t0 , w(t0 ))[ẋ], [ẋ] + o(1)) → +∞ as t → t0 + 0. (4.119)
t − t0
Now (4.110) and (4.119) imply (4.113).
Theorem 4.22. Assume that condition (4.101) holds. Then the absence of a point τ that is
-conjugate to t0 on (t0 , t∗ ] is equivalent to condition (4.28).
all conditions on [t0 , τ ] only. As a result, we arrive at the following equivalent system on
[0, τ ]:
x̄, ψ̄ ∈ P W 1,2 [t0 , τ ], (4.120)
x̄(0) = x̄(τ ) = 0, ψ̄(0) = 0, (4.121)
x̄˙ = Ax̄ + B ψ̄ ∗ , −ψ̄˙ = x̄ ∗ C + ψ̄A on [t0 , τ ], (4.122)
a[x̄] = [ẋ] − ψ̄ − [ẋ] + [ψ̇]x̄ − , (4.123)
a[ψ̄] = [ψ̇] − ψ̄ − [ẋ] + [ψ̇]x̄ − . (4.124)
We have proved the following lemma.
Lemma 4.23. A point τ ∈ (t∗ , tf ] is -conjugate to t0 iff there exists a pair of functions x̄, ψ̄
that satisfies conditions (4.120)–(4.124).
Proof. Assume that condition (4.125) holds, in which X, " is a solution to the problem
(4.22)–(4.25). Then there exists c̄ ∈ Rm such that
X(τ )c̄ = 0, c̄ = 0. (4.126)
We set
x̄ = X c̄, ψ̄ = c̄∗ " ∗ . (4.127)
Then x̄ and ψ̄ satisfy (4.120)–(4.124). Conversely, let x̄ and ψ̄ satisfy (4.120)–(4.124).
We set c̄ = −ψ̄(t0 )∗ . Then conditions (4.126) and (4.127) hold. Conditions (4.126) imply
condition (4.125).
Note that Theorems 4.19, 4.22, and 4.24 imply Theorems 4.3 and 4.4.
To complete the proof of the results of Section 4.1.2, we have to consider a jump
condition for a solution Q to the Riccati equation (4.30) with initial condition (4.36). In
what follows, we assume that is nonnegative on K and that the pair X, " is a solution
to the problem (4.22)–(4.25). Then conditions (4.26) and (4.27) hold. We set Q = "X−1 .
Then " = QX. Using the relations q− = [ẋ]∗ Q− − [ψ̇] and " − = Q− X− in the jump
conditions (4.24) and (4.25) for X and ", we obtain
a[X] = −[ẋ](q− )X − , (4.128)
a["] = −[ψ̇]∗ (q− )X − . (4.129)
Now (4.130) follows from this relation considered together with (4.27) and the inequality
a > 0.
This relation and the formula (q− )∗ = Q− [ẋ] − [ψ̇]∗ imply the equality
[Q] aI − [ẋ](q− ) = (q− )∗ (q− ).
L I
-
IC ID IR
# ? ? ? #
AA D 6
V
t
V0 C R -
"!
? "!
The strict Legendre–Clebsch condition holds in view of Huu (t) ≡ 2 > 0 . The adjoint
equation ψ̇ = −Hx yields
Since the final state is specified, there are no boundary conditions for the adjoint
variable. The boundary value problem (4.133), (4.134) and (4.138) with control u = −2ψ2
was solved using the multiple shooting code BNDSCO developed by Oberle and Grimm
[82]. The optimal state, control, and adjoint variables are shown in Figure 4.2. We get the
following initial and final values for the adjoint variables:
Nearly identical numerical results can be obtained by solving the discretized control problem
with a high number of gridpoints.
The optimality of this extremal solution may be checked by producing a finite solu-
tion of the Riccati equation (4.30). Since the Rayleigh problem has dimension n = 2 , we
(c) adjoint variables and (d) solutions Q11, Q12, Q22 of Riccati equation
1 2
2 2.5
0 2
-2 1.5
-4 1
-6 0.5
-8 0
-10 -0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Figure 4.2. Rayleigh problem with regular control. (a) State variables. (b) Con-
trol. (c) Adjoint variables. (d) Solutions of the Riccati equation (4.140).
4.1. Jacobi-Type Conditions and Riccati Equation for Broken Extremals 205
Since the final state is fixed, no boundary condition is prescribed for Q(t1 ) . Choosing for
convenience the boundary condition Q(t1 ) = 0 , we obtain the bounded solution of the
Riccati equation shown in Figure 4.2. The initial values are computed as
Thus the unconstrained solution shown in Figure 4.2 provides a local minimum.
To conclude the discussion of the Rayleigh problem, we consider the control problem
with free final state x(tf ). The solution is quite similar to that shown in Figure 4.2, with
the only difference being that the boundary inequality Q(tf ) > 0 should hold. However,
it suffices to find a solution Q(t) satisfying the boundary condition Q(tf ) = 0 which was
imposed earlier. Due to the continuous dependence of solutions on initial or terminal
conditions, equation (4.140) then has a solution with Q(tf ) = # · I2 > 0 for # > 0 small. We
obtain
x1 (4.5) = −0.0957105, x2 (4.5) = −0.204377,
ψ1 (0) = −9.00126, ψ2 (0) = −2.67259,
(4.141)
Q11 (0) = 2.00607, Q22 (0) = 0.470491,
Q22 (0) = −0.351606.
The unknown initial value ẋ(0) can be determined by shooting methods; cf. Stoer and
Bulirsch [106]. The boundary value problem (4.143) has the explicit solution x (1) (t) =
4/(1 + t)2 and a second solution x (2) (t) with initial values
Both solutions x (k) (t), k = 1, 2, may be tested for optimality by the classical Jacobi condition.
The variational system (4.22), (4.23), along the two extremals, yields for k = 1, 2
ẍ (k) (t) = 3
2 x (k) (t), x (k) (0) = 4, ẋ (k) (0) as in (4.144),
(4.145)
ÿ (k) (t) = 3 x (k) (t) y (k) (t), y (k) (0) = 0, ẏ (k) (0) = 1.
For k = 1, 2 the extremals x (k) (k = 1, 2) and variational solutions y (k) (k = 1, 2) are displayed
in Figure 4.3. The extremal x (1) (t) = 4/(1 + t)2 is optimal in view of
whereas the second extremal x (2) is not optimal, since it exhibits the conjugate point tc =
0.674437 with y (2) (tc ) = 0. The conjugate point tc has the property that the envelope of
Figure 4.3. Top left: Extremals x (1) , x (2) (lower graph). Top right: Variational
solutions y (1) and y (2) (lower graph) to (4.145). Bottom: Envelope of neighboring extremals
illustrating the conjugate point tc = 0.674437.
4.1. Jacobi-Type Conditions and Riccati Equation for Broken Extremals 207
extremals corresponding to neighboring initial slopes ẋ(0) touches the extremal x (2) (t) at
tc ; cf. Figure 4.3.
and then
[Q] = (b− )−1 (q− )2 = (b− )−1 4 cot 2 ϕ,
4 cot 2 ϕ
Q+ = Q− + [Q] = − cot ϕ + = −2(b− )−1 < 0.
b−
Now, for t ∈ (ϕ, tf ], we have to solve the Cauchy problem
Since Q(ϕ+) < 0, the solution has the form Q(t) = − cot(t + τ ) with certain τ > 0. It is
clear that this solution is defined at least on the closed interval [ϕ, ϕ + 0.5π ] which contains
the closed interval [ϕ, tf ]. By Theorem 4.6 the quadratic form is positive definite on the
subspace K. It means that the extremal with single switching point satisfies the sufficient
second-order optimality condition. By Theorem 4.2, such an extremal is a point of a strict
bounded strong minimum in the problem.
where q− is a row vector and (q− )∗ is a column vector. (Recall that (q− )∗ (q− ) is a sym-
metric positive semidefinite matrix.) In this case, we will say that the matrix Q(t) is the
symmetric solution of the problem (4.150)–(4.152) on [t0 , tf ].
For the quadratic form on the subspace K, defined by relations (4.14) and (4.13),
respectively, the following theorem holds.
Theorem 4.28. If there exists a symmetric solution Q(t) of the problem (4.150)–(4.152)
on [t0 , tf ], then the quadratic form has the following transformation into a perfect
square on K:
tf
−1 − 2
2(z̄) = (b− ) (a ξ̄ + q− x̄ ) + B(t)v̄(t), v̄(t) dt, (4.153)
t0
where
v̄(t) = Q(t) + Fux (t, w(t)) x̄(t) + Fuu (t, w(t))ū(t). (4.154)
Proposition 4.29. Let Q(t) be a symmetric matrix on [t0 , tf ] with piecewise continuous
entries, which are absolutely continuous on each interval of the set [t0 , tf ]\. Then
2(z̄) = b− ξ̄ 2 + 2q− x̄ + ξ̄ + [Q]x̄ + , x̄ +
tf
(4.155)
+ (Fxx + Q̇)x̄, x̄ + 2(Fxu + Q)ū, x̄ + Fuu ū, ū dt
t0
we obtain
connected to the jump point t∗ ∈ of the control u(·), and let us represent it as a function
of x̄ + , ξ̄ ,
ω∗ (ξ̄ , x̄) = (a − Q[ẋ] − [ψ̇]∗ , [ẋ] )ξ̄ 2 + 2Q− [ẋ] − [ψ̇]∗ , x̄ + ξ̄ + [Q]x̄ + , x̄ + .
Proposition 4.30. If Q satisfies the jump condition b− [Q] = (q− )∗ (q− ), then
Proof. Using the jump conditions for [Q] and [x̄] and the definition (4.148) of b− , we
obtain
b− ξ̄ 2 + 2q− x̄ + ξ̄ + [Q]x̄ + , x̄ + = (b− )−1 (b− )2 ξ̄ 2 + 2q− x̄ + b− ξ̄ + (q− x̄ + )2
2
= (b− )−1 b− ξ̄ + q− x̄ +
2
= (b− )−1 (a − q− [ẋ])ξ̄ + q− x̄ +
2
= (b− )−1 a ξ̄ − q− [x̄] + q− x̄ +
2
= (b− )−1 a ξ̄ + q− x̄ − .
QA + A∗ Q + QBQ + C
−1 −1 −1 −1
= −QFuu Fux − Fxu Fuu Q − QFuu Q − Fxu Fuu Fux + Fxx
−1 −1
= −QFuu (Q + Fux ) − Fxu Fuu (Q + Fux ) + Fxx
−1
= −(Q + Fxu )Fuu (Q + Fux ) + Fxx .
Proof. By Theorem 4.28, formula (4.153) holds for on K. Hence, the form is
nonnegative on K. Let us show that is positive on K. Let z̄ ∈ K and (z̄) = 0. Then
by (4.153) and (4.154), we have
x̄˙ = −Fuu
−1
(Q + Fux )x̄ on [t0 , tf ]\. (4.166)
Remark. Theorem 4.33 was proved for a single point of discontinuity of the control u(·).
A similar result could be proved for finitely many points of discontinuity of the control u(·).
Now, using Theorem 4.33, we can prove, for s = 1, that the existence of a symmetric
solution Q(t) of the problem (4.150)–(4.152) on [t0 , tf ] is not only sufficient (as it was
stated by Theorem 4.28) but also necessary for the positive definiteness of on K. This can
212 Chapter 4. Jacobi-Type Conditions and Riccati Equation for Broken Extremals
be done in different ways. We choose the following one. Let be positive definite on K,
take a small # > 0, and put u(t) = u(t0 ) for t ∈ [t0 − #, t0 ].
Thus, u(t) is continued to the left-hand side of the point t0 by constant value equal to
the value at t0 . Now we have w(t) defined on [t0 − #, tf ] with the same single discontinuity
point t∗ , and therefore the “continued” quadratic form
1 2 1 tf
# (z̄) = a ξ̄ − 2[ψ̇]x̄av + Fww (t, w(t))w̄(t), w̄(t) dt
2 2 t0 −#
Using the same technique as in [96], one can easily prove that there is # > 0 such that # (z̄)
is positive definite on K# (note that condition (SL) is satisfied for # on [t0 − #, tf ], which
is important for this proof). Then by Theorem 4.33 applied for [t0 − #, tf ] there exists a
solution Q of the Riccati equation (4.150) on (t0 − #, tf ] satisfying inequality (4.151) and
jump condition (4.152), and hence we have this solution on the segment [t0 , tf ]. Thus, for
s = 1 the following theorem holds.
The formulation of the jump condition (4.152) allows us to solve problem (4.150)–
(4.152) in a forward time direction. Certainly it is possible to move in opposite direction.
Let us prove the following.
where
q+ = [ẋ]∗ Q+ − [ψ̇]. (4.169)
Hence
a[Q] = (q+ )∗ (q− ). (4.171)
Using again (4.170) in (4.171), we obtain a[Q] = (q+ )∗ (q+ − [ẋ]∗ [Q]). It follows from this
formula that a[Q] = (q+ )∗ (q+ ) − (q+ )∗ [ẋ]∗ [Q]. Consequently
Let us transform the left-hand side of this equation, using (4.171). Since
1 1
(q+ )∗ [ẋ]∗ [Q] = (q+ )∗ [ẋ]∗ (q+ )∗ (q− ) = (q+ )∗ [ẋ]∗ (q+ )∗ (q− )
a a
1 1
= (q+ )∗ [ẋ]∗ (q+ )∗ (q− ) = (q+ )∗ (q− ) q+ [ẋ]
a a
= [Q] q+ [ẋ] ,
we have
(aI + (q+ )∗ [ẋ]∗ )[Q] = a[Q] + (q+ )∗ [ẋ]∗ [Q] = a[Q] + q+ [ẋ] [Q] = (a + q+ [ẋ])[Q].
q+ [Q](q+ )∗ ≥ 0. (4.175)
1
ω∗ (ξ̄ , x̄) := a ξ̄ 2 − [ψ̇](x̄ − + x̄ + )ξ̄ − Q− x̄ − , x̄ − + Q+ x̄ + , x̄ k+ = (a ξ̄ + q+ x̄ + )2 .
b+
Proof. We have
Theorem 4.37. If the assumptions of Theorem 4.28 are satisfied, then the quadratic form
has the following transformation to a perfect square on K:
tf
2(z̄) = (b+ )−1 (a ξ̄ + q+ x̄ + )2 + B(t)v̄(t), v̄(t) dt, (4.177)
t0
We denote by 0 the set of the normed collections λ satisfying the conditions of the
local minimum principle for an admissible trajectory w(·) = (x(·), u(·)),
α0 ≥ 0, α0 + |βj | = 1, ψ̇ = −Hx , ψ(t0 ) = −lx0 , ψ(tf ) = lxf , Hu = 0,
(4.182)
where all derivatives are calculated on the trajectory w(·), respectively, at the endpoints
(x(t0 ), x(tf )) of this trajectory. The condition 0 = ∅ is necessary for a weak minimum at
the point w(·).
We set U(t, x) = {u ∈ Rd(u) | (t, x, u) ∈ Q}. Denote by M0 the set of tuples λ ∈ 0
such that for all t ∈ [t0 , tf ]\, the condition u ∈ U(t, x 0 (t)) implies the inequality
|x(t0 ) − x 0 (t0 )| < ε, max |x(t) − x 0 (t)| ≤ ε, (t, w(t)) ∈ C a.e. on [t0 , tf ],
t∈[t0 ,tf ]
where x is the vector composed of the essential components of the vector x. Recall that
the component xi of a vector x = (x1 , . . . , xd(x) ) is said to be unessential if the function f is
independent of xi and the functions J and K affinely depend on xi0 = xi (t0 ), xif = xi (tf ).
Now we shall formulate quadratic sufficient conditions for a strict bounded strong
minimum at the point w(·), which follow from the corresponding conditions in Section
2.7.2; see Theorem 2.102. For λ ∈ 0 and tk ∈ , we put (k H )(t) = H (t, x(t), uk+ , ψ(t))−
H (t, x(t), uk− , ψ(t)), where uk− = u(tk − 0), uk+ = u(tk + 0), and
d d
D k (H ) := − (k H )t −0 = − (k H )t +0 . (4.184)
dt k dt k
The second equality is a consequence of the minimum principle. As we know, the following
formula holds:
D k (H ) = −Hxk+ Hψk− + Hxk− Hψk+ − [Ht ]k . (4.185)
Let K denote the subspace of all z̄ = (ξ̄ , x̄(·), ū(·)) ∈ Z2 () such that
Kpp̄ = 0, x̄˙ = fw w̄, [x̄]k = [ẋ]k ξ̄k , k = 1, . . . , s, (4.186)
where fw = fw (t, w(t)), Kp = Kp (x(t0 ), x(tf )), [ẋ]k is the jump of x(·) at tk , and [x̄]k is the
jump of x̄(·) at tk . We call K the critical subspace.
As in (2.13), we define the quadratic form, which corresponds to the element λ ∈ 0 , as
1 k
s
1 tf 1
(z̄) =
λ
D (H )ξ̄k2 − 2[ψ̇]k x̄av
k
ξ̄k + Hww w̄, w̄ dt + lpp p̄, p̄ , (4.187)
2 2 t0 2
k=1
where lpp = lpp (x(t0 ), x(tf ), α0 , β), Hww = Hww (t, w(t), ψ(t)). We set
tf
γ̄ (z̄) = ξ̄ , ξ̄ + x̄(t0 ), x̄(t0 ) + ū, ū dt.
t0
Theorem 4.38. Assume that there exists a strictly Legendrian element λ ∈ M0+ such that
α0 > 0 and the quadratic form λ (·) is positive definite on K, i.e., there exists # > 0 such
that
λ (z̄) ≥ # γ̄ (z̄) ∀ z̄ ∈ K. (4.188)
Then, w(·) is a strict bounded strong minimum.
Now, let us show that the quadratic form could be transformed into a perfect square
if the corresponding Riccati equation has a solution Q(t) defined on [t0 , tf ], satisfying certain
jump conditions at each point of the set . Define the Riccati equation along x(t), u(t), and
ψ(t) by
Q̇ + Qfx + fx∗ Q + Hxx − (Hxu + Qfu )Huu
−1
(Hux + fu∗ Q) = 0, t ∈ [t0 , tf ]\, (4.189)
for a piecewise continuous function Q(t) which
∗ is continuously differentiable on each
interval of the set [t0 , tf ]\. Set qk− = [ẋ]k Qk− − [ψ̇]k and define the conditions
Theorem 4.39. Assume that there exists a symmetric solution Q(t) (piecewise continuous
on [t0 , tf ] and continuously differentiable on each interval of the set [t0 , tf ]\) of Riccati
equation (4.189) which satisfies at each point tk ∈ conditions (4.190) and jump condi-
tions (4.191). Then the quadratic form λ (z̄) (see (4.187)) has the following transformation
into a perfect square on the subspace K (see (4.186)),
s 2 tf
−1 −1
2 (z̄) =
λ
(bk− ) ak ξ̄k + (qk− )x̄ k−
+ Huu v̄, v̄ dt + M p̄, p̄ , (4.192)
k=1 t0
where
v̄ = (Hux + fu∗ Q)x̄ + Huu ū, (4.193)
lx0 x0 + Q(t0 ) lx0 xf
M= . (4.194)
lxf x0 lxf xf − Q(tf )
Proof. In perfect analogy with Theorem 4.28 we have to prove statements similar to
Propositions 4.29–4.32. They need only small changes. Below we point out these changes.
Take a symmetric matrix Q(t) on [t0 , tf ] with piecewise continuous coefficients, which are
absolutely continuous on each interval of the set [t0 , tf ]\. Using, for z̄ ∈ K, formula
(4.156) and the equalities
˙ x̄ = 2Qfx x̄, x̄ + 2Qfu ū, x̄ = (Qfx + fx∗ Q)x̄, x̄ + Qfu ū, x̄ + (fu∗ Qx̄, ū ,
2Qx̄,
we obtain (similar to (4.157)) the following zero form on K:
s
0= (Qk+ x̄ k+ , x̄ k+ − Qk− x̄ k− , x̄ k− )
k=1
tf
Adding this zero form (4.195) to the form 2λ (z̄) (see (4.187)) considered for arbitrary
z̄ ∈ K, we obtain
tf
2λ (z̄) = M p̄, p̄ + (Hxx + Q̇ + Qfx + fx∗ Q)x̄, x̄
t0
Using the Riccati equation (4.189) and definition (4.193) for v̄, we obtain for the integral
part of (4.196),
tf
(Hxx + Q̇ + Qfx + fx∗ Q)x̄, x̄ + (Hxu + Qfu )ū, x̄
t0
namely,
tf
−1
tf −1
Huu v̄, v̄ dt = Huu (Hux + fu∗ Q)x̄, (Hux + fu∗ Q)x̄
t0 t0
−1
Theorem 4.40. Assume that (a) there exists a symmetric solution Q(t), piecewise continu-
ous on [t0 , tf ] and continuously differentiable on each interval of the set [t0 , tf ]\, of Riccati
equation (4.189) which satisfies at each point tk ∈ conditions (4.190) and jump conditions
(4.191). In addition, assume that (b) M p̄, p̄ ≥ 0 for all p̄ ∈ R2n such that Kpp̄ = 0. Also,
assume that (c) the conditions Kpp̄ = 0 and M p̄, p̄ = 0 imply that x̄0 = 0 or x̄f = 0. Then
the quadratic form λ (z̄) (see (4.187)) is positive definite on the subspace K (see (4.186));
i.e., condition (4.188) holds with some # > 0.
Proof. From Theorem 4.39 it follows that λ (·) is nonnegative on K. Let us show that
λ (·) is positive on K. Assume that (z̄) = 0 for some z̄ ∈ K. Then by formula (4.192),
we have
M p̄, p̄ = 0, (4.201)
v̄ = 0, (4.202)
ak ξ̄k = −qk− x̄ k− , k = 1, . . . , s. (4.203)
4.2. Riccati Equation in the General Problem 219
Together with the initial condition x̄(t0 ) = 0 this implies that x̄(t) = 0 for all t ∈ [t0 , t1 ).
Hence x̄ 1− = 0. Then by (4.203), we obtain ξ̄1 = 0. Then [x̄]1 = 0 by the condition
[x̄]1 = [ẋ]1 ξ̄1 . Hence x̄ 1+ = x̄ 1− + [x̄]1 = 0 and then again by (4.205) x̄(t) = 0 for all
t ∈ (t1 , t2 ), etc. By induction, we get x̄(t) = 0 on [t0 , tf ], ξ̄ = 0, and then by (4.204), we get
ū(t) = 0 a.e. on [t0 , tf ]. Thus z̄ = (ξ̄ , x̄, ū) = 0. Consequently, λ (·) is positive on K.
(ii) Consider the case x̄f = 0. Then equation (4.205) and condition x̄(tf ) = 0 imply that
x̄(t) = 0 for all t ∈ (ts , tf ]. Hence x̄ s+ = 0. Then [x̄]s = −x s− . Using this condition in
(4.203), we obtain as ξ̄s − qs− [x̄]s = 0, or
because [x̄]s = [ẋ]s ξ̄s . Since bs− = as − qs− [ẋ]s > 0, condition (4.206) implies that ξ̄s = 0.
Hence [x̄]s = 0 and then x̄ s− = 0. Then, by virtue of (4.205), x̄(t) = 0 on (ts−1 , ts ), etc.
By induction we get x̄(·) = 0, ξ̄ = 0, and then, by (4.204), ū = 0, whence z̄ = 0. Thus, we
have proved that is positive on K. It means that is positive definite on K, since is
a Legendre form.
Notes on SSC, Riccati equations, and sensitivity analysis. For regular controls, sev-
eral authors have used the Riccati equation approach to verify SSC. Maurer and Pickenhain
[73] considered optimal control problems with mixed control-state inequality constraints
and derived SSC and the associated Riccati matrix equation on the basis of Klötzler’s
duality theory. Similar results were obtained by Zeidan [116, 117]. Extensions of these
results to control problems with free final time are to be found in Maurer and Oberle [68].
It is well known that SSC are fundamental for the stability and sensitivity analysis
of parametric optimal control problems; cf., Malanowski and Maurer [58, 59], Augustin
and Maurer [3], Maurer and Augustin [64, 65], and Maurer and Pesch [71, 72]. SSC also
lay firm theoretical grounds to the method of determining neighboring extremals (Bryson
and Ho [12] and Pesch [101]) and to real-time control techniques (see Büskens [13] and
Büskens and Maurer [14, 15, 16]).
Chapter 5
Second-Order Optimality
Conditions in Optimal Control
Problems Linear in a Part of
Controls
In this chapter, we derive quadratic optimality conditions for optimal control problems
with a vector control variable having two components: a continuous unconstrained control
appearing nonlinearly in the control system and a control appearing linearly and belonging
to a convex polyhedron. It is assumed that the control components appearing linearly
are of bang-bang type. In Section 5.1, we obtain quadratic conditions in the problem with
continuous and bang-bang control components on a fixed time interval (that we call the main
problem). The case of a nonfixed time interval is considered in Section 5.2. In Section 5.3,
we show that, also for the mixed continuous-bang case, there exists a technique to check the
positive definiteness of the quadratic form on the critical cone via a discontinuous solution of
an associated Riccati equation with appropriate jump conditions at the discontinuity points
of the bang-bang control [98]. In Section 5.4, this techniques is applied to an economic
control problem in optimal production and maintenance. We show that the numerical
solution obtained by Maurer, Kim, and Vossen [67] satisfies the second-order test derived
in this chapter, while existing sufficiency results fail to hold.
223
224 Chapter 5. Second-Order Optimality Conditions in Optimal Control
Here, F , K, a are column vector functions, B is a d(x) × d(u) matrix function, P ⊂ R2d(x) ,
Q ⊂ R1+d(x)+d(v) are open sets, and U ⊂ Rd(u) is a convex polyhedron. The functions
J , F , K are assumed to be twice continuously differentiable on P , and the functions a, B
are twice continuously differentiable on Q. The dimensions of F , K are denoted by d(F ),
d(K). By = [t0 , tf ] we denote the interval of control and use the abbreviations x0 = x(t0 ),
xf = x(tf ), p = (x0 , xf ).
A process T = {(x(t), u(t), v(t)) | t ∈ [t0 , tf ] } is said to be admissible if x(·) is ab-
solutely continuous, u(·), v(·) are measurable bounded on , and the triple of functions
(x(t), u(t), v(t)), together with the endpoints p = (x(t0 ), x(tf )), satisfies the constraints (5.2)
and (5.3). Thus, the main problem is considered in the space
Definition 5.1. An admissible process T affords a Pontryagin minimum if for each com-
pact set C ⊂ Q there exists ε > 0 such that J(T˜ ) ≥ J(T ) for all admissible
processes
T˜ = {(x̃(t), ũ(t), ṽ(t)) | t ∈ [t0 , tf ] } such that (a) max |x̃(t) − x(t)| < ε, (b) |ũ(t) −
u(t)| dt < ε, (c) (t, x̃(t), ṽ(t)) ∈ C a.e. on .
where ψ is a row vector of dimension d(ψ) = d(x), while x, u, f , F , and K are column
vectors. The row vector of dimension d(u),
will be called the switching function for the u-component of the control. Denote by l the
endpoint Lagrange function
where α and β are row vectors with d(α) = d(F ) and d(β) = d(K), and α0 is a number. We
introduce a tuple of Lagrange multipliers λ = (α0 , α, β, ψ(·)) such that ψ(·) : → Rd(x) is
continuous on and continuously differentiable on each interval of the set \ . In what
follows, we will denote first- or second-order partial derivatives by subscripts referring to
the variables.
Denote by M0 the set of the normalized tuples λ satisfying the minimum principle
conditions for the process T :
)
d(F
d(K)
α0 ≥ 0, α ≥ 0, αF (p) = 0, α0 + αi + |βj | = 1, (5.7)
i=1 j =1
ψ̇ = −Hx ∀ t ∈ \ , (5.8)
ψ(t0 ) = −lx0 , ψ(tf ) = lxf , (5.9)
H (t, x(t), u, v, ψ(t)) ≥ H (t, x(t), u(t), v(t), ψ(t))
∀ t ∈ \ , u ∈ U , v ∈ Rd(v) such that (t, x(t), v) ∈ Q. (5.10)
The derivatives lx0 and lxf are taken at the point (p, α0 , α, β), where p = (x(t0 ), x(tf )), and
the derivative Hx is evaluated along the trajectory (t, x(t), u(t), v(t), ψ(t)), t ∈ \ . The
condition M0 = ∅ constitutes the first-order necessary condition of a Pontryagin minimum
for the process T which is called the Pontryagin minimum principle, cf. Pontryagin et al.
[103], Hestenes [40], and Milyutin and Osmolovskii [79]. The set M0 is a finite-dimensional
compact set and the projector λ → (α0 , α, β) is injective on M0 .
In the following, it will be convenient to use the simple abbreviation (t) for indicating
all arguments (t, x(t), u(t), v(t), ψ(t)), e.g.,
H (t) = H (t, x(t), u(t), v(t), ψ(t)), φ(t) = φ(t, x(t), v(t), ψ(t)).
(k H )(t) = H (t, x(t), uk+ , v(tk ), ψ(t)) − H (t, x(t), uk− , v(tk ), ψ(t))
= φ(t, x(t), v(tk ), ψ(t)) [u]k . (5.11)
Consequently, for each λ ∈ M0 , the function (k H )(t) has a derivative at the point
tk ∈ . In what follows, we will consider the quantities
d
D k (H ) = − (k H )(tk ) = −φ̇(tk ±)[u]k , k = 1, . . . , s. (5.12)
dt
Then the minimum condition (5.10) implies the following property.
where Hxk− and Hxk+ are the left-hand and right-hand values of the function Hx (t) at tk ,
respectively, [Ht ]k is the jump of the function Ht (t) := Ht (t, x(t), u(t), v(t), ψ(t)) at tk , etc.,
and ψ0 (t) = −H (t).
Definition 5.4. The process T affords a bounded strong minimum if for each compact set
C ⊂ Q there exists ε > 0 such that J(T+) ≥ J(T ) for all admissible processes T+ =
{(x̃(t), ũ(t), ṽ(t)) | t ∈ [t0 , tf ] } such that (a) |x̃(t0 ) − x(t0 )| < ε, (b) max |x̃(t) − x(t)| < ε,
(c) (t, x̃(t), ṽ(t)) ∈ C a.e. on .
The strict bounded strong minimum is defined in a similar way, with the nonstrict
inequality J(T+) ≥ J(T ) replaced by the strict one and the process T+ required to be different
from T .
the set of indices of all active endpoint inequalities Fi (p) ≤ 0 at the point p = (x(t0 ), x(tf )).
Denote by K the set of all z̄ ∈ Z2 () satisfying the following conditions:
where
etc. Note that the functional (λ, z̄) is linear in λ and quadratic in z̄. The following theorem
gives the main second-order necessary condition of optimality.
Theorem 5.5. If the process T affords a Pontryagin minimum, then the following Condition
A holds: The set M0 is nonempty and maxλ∈M0 (λ, z̄) ≥ 0 for all z̄ ∈ K.
Definition 5.6. For a given admissible process T with a piecewise constant control u(t)
and continuous control v(t), we say that u(t) is a strict bang-bang control if the set M0 is
nonempty and there exists λ ∈ M0 such that
where [u(t−), u(t+)] denotes the line segment spanned by the vectors u(t−), u(t+).
φ(t) = 0 ∀ t ∈ \ .
It is easy to show that if the set M0+ is nonempty, then u(t) is a strict bang-bang control.
Denote by Leg+ (M0+ ) the set of all strictly Legendrian elements λ ∈ M0+ , and set
tf
γ̄ (z̄) = ξ̄ , ξ̄ + x̄(t0 ), x̄(t0 ) + v̄(t), v̄(t) dt.
t0
Theorem 5.8. Let the following Condition B be fulfilled for the process T :
(a) The set Leg+ (M0+ ) is nonempty;
5.1. Quadratic Optimality Conditions in the Problem on a Fixed Time Interval 229
(b) there exists a nonempty compact set M ⊂ Leg+ (M0+ ) and a number C > 0
such that maxλ∈M (λ, z̄) ≥ C γ̄ (z̄) for all z̄ ∈ K.
Then T is a strict bounded strong minimum.
Remark 5.9. If the set Leg+ (M0+ ) is nonempty and K={0}, then Condition (b) is fulfilled
automatically. This case can be considered as a first-order sufficient optimality condition
for a strict bounded strong minimum.
As mentioned in the introduction, the proof of Theorem 5.8 is very similar to the
proof of the sufficient quadratic optimality condition for the pure bang-bang case given in
Milyutin and Osmolovskii [79, Theorem 12.4, p. 302], and based on the sufficient quadratic
optimality condition for broken extremals in the general problem of calculus of variations;
see Part I of the present book. The proofs of Theorems 5.5 and 5.8 will be given below.
where P ⊂ R2d(x) , Q ⊂ R1+d(x)+d(v) are open sets, U ⊂ Rd(u) is a compact set, and
= [t0 , tf ] is a fixed time interval. The functions J , F , and K are assumed to be twice
continuously differentiable on P , and the functions a and B are twice continuously differ-
entiable on Q. The compact set U is specified by
for all u ∈ Qg such that g(u) = 0. It follows from (5.22) that d(g) ≤ d(u), but it is possible,
in particular, that d(g) = d(u). In this latter case for U we can take any finite set of points
in Rd(u) , for example, the set of vertices of a convex polyhedron.
For brevity, we will refer to problem (5.18)–(5.20) as the problem Z. Thus, the
only difference between problem Z and the main problem (5.1)–(5.4) is that a convex
polyhedron U is replaced by an arbitrary compact set U specified by (5.21). The definitions
of the Pontryagin minimum, the bounded strong minimum, and the strict bounded strong
minimum in the problem Z are the same as in the main problem.
230 Chapter 5. Second-Order Optimality Conditions in Optimal Control
We will consider the problem (5.18), (5.19) not only for control system (5.20), but
also for its convexification with respect to u:
ẋ = a(t, x, v) + B(t, x, v)u, u ∈ co U, (t, x, v) ∈ Q, (5.23)
where co U is the convex hull of the compact set U. We will refer to the problem (5.18),
(5.19), (5.23) as the problem co Z.
We will be interested in the relationships between conditions for a minimum in the
problems Z and co Z. Naturally, these conditions concern a process satisfying the con-
straints of the problem Z. Let w 0 (·) = (x 0 (·), u0 (·), v 0 (·)) be such a process. Then it
satisfies the constraints of the problem co Z as well. We assume that w0 satisfy the follow-
ing conditions: The function u0 (t) is piecewise continuous with the set = {t1 , . . . , ts } of
discontinuity points, control v 0 (t) is continuous, and each point tk ∈ is an L-point of the
controls u0 (t) and v 0 (t). The latter means that there exist constants C > 0 and ε > 0 such
that for each point tk ∈ , we have
|u0 (t) − u0k− | ≤ C|t − tk | for t ∈ (tk − ε, tk ),
|u0 (t) − u0k+ | ≤ C|t − tk | for t ∈ (tk , tk + ε),
|v 0 (t) − v 0 (tk )| ≤ C|t − tk | for t ∈ (tk − ε, tk + ε).
We can formulate quadratic conditions for the point w0 in the problem Z. To what extent
can they be carried over to the problem co Z?
Regarding the necessary quadratic conditions (see Condition A in Theorem 3.9) this is
a simple question. If a point w 0 yields a Pontryagin minimum in the problem co Z, this point,
a fortiori, affords a Pontryagin minimum in the problem Z. Hence any necessary condition
for a Pontryagin minimum at the point w0 in the problem Z is a necessary condition for a
Pontryagin minimum at this point in the problem co Z as well. Thus we have the following
theorem.
Theorem 5.10. Condition A (given by Theorem 3.9) for the point w 0 in the problem Z is a
necessary condition for a Pontryagin minimum at this point in the problem co Z.
where f (t, x, u, v) = a(t, x, v) + B(t, x, v)u and a + = max{a, 0} for a ∈ R1 . For an arbitrary
variation δw = (δx, δu, δv) ∈ W , let
γ1 (δw) = (
δu
1 )2 , (5.25)
tf
where
δu
1 = t0 |δu(t)| dt.
Let {w n } = {(x n , un , v n )} be a bounded sequence in W . In what follows, the notation
σ (δw ) = o(γ1 (w n − w 0 )) means that there exists a sequence of numbers εn → 0 such that
n
Definition 5.11. We say that the bounded strong γ1 -sufficiency holds at the point w0 in
the problem Z if there are no compact set C ⊂ Q and sequence {wn } = {(x n , un , v n )} in W
such that
σ (wn ) = o(γ1 (wn − w 0 )), max |x n (t) − x 0 (t)| → 0, |x n (t0 ) − x 0 (t0 )| → 0,
t∈
and for all n, w n = w 0 , (t, x n (t), v n (t)) ∈ C, un (t) ∈ U a.e. on , where x n is composed
of the essential components of vector x n .
Proposition 5.12. A bounded strong γ1 -sufficiency at the point w0 implies a strict bounded
strong minimum at this point.
Proof. Let w 0 be an admissible point in the problem Z. Assume that w0 is not a point of
a strict bounded strong minimum. Then there exist a compact set C ⊂ Q and a sequence
of admissible points {wn } = {(x n , un , v n )} such that maxt∈ |x n (t) − x 0 (t)| → 0, |x n (t0 ) −
x 0 (t0 )| → 0 and, for all n,
J (p n ) − J (p 0 ) < 0, (t, x n (t), v n (t)) ∈ C, un (t) ∈ U a.e. on , wn = w0 .
It follows that σ (w n ) = 0 for all n. Hence w0 is not a point of a bounded strong γ1 -sufficiency
in the problem Z.
Theorem 5.13. Suppose that for an admissible point w 0 in the problem Z the bounded
strong γ1 -sufficiency holds. Then w 0 is a point of the strict bounded strong minimum in the
problem co Z.
Lemma 5.14. Let U ⊂ Rd(u) be a bounded set, and let u(t) be a measurable function on
such that u(t) ∈ co U a.e. on . Then there exists a sequence un (t) of measurable functions
on such that for every n we have un (t) ∈ U , t ∈ , and
t t
un (τ ) dτ → u(τ ) dτ uniformly in t ∈ . (5.26)
t0 t0
232 Chapter 5. Second-Order Optimality Conditions in Optimal Control
This lemma is a consequence of Theorem 16.1 in [79, Appendix, p. 361]. In the proof
of Theorem 5.13 we will also use the following theorem, which is similar to Theorem 16.2
in [79, Appendix, p. 366].
where Q, a, and B are the same as in the problem Z, and c0 ∈ Rd(x) . Suppose that there is
a sequence un ∈ L∞ ([t0 , tf ], Rd(u) ) such that
sup
un
∞ < +∞ (5.29)
n
where
δx = x n − x ∗ , δu = un − u∗ ,
δa = a(t, x n , v ∗ ) − a(t, x ∗ , v ∗ ) = a(t, x ∗ + δx, v ∗ ) − a(t, x ∗ , v ∗ ),
δB = B(t, x n , v ∗ ) − B(t, x ∗ , v ∗ ) = B(t, x ∗ + δx, v ∗ ) − B(t, x ∗ , v ∗ ).
We have here
δx(t0 ) = 0. (5.34)
5.1. Quadratic Optimality Conditions in the Problem on a Fixed Time Interval 233
These properties mean that the sequence {δu} converges to zero ∗-weakly (L1 -weakly) in
L∞ . Therefore the functions
t
δy(t) := B(τ , x ∗ (τ ), v ∗ (τ ))δu(τ ) dτ
t0
converge to zero pointwise on [t0 , tf ], and hence uniformly on [t0 , tf ] because they possess
the common Lipschitz constant. Thus
δy
C → 0. (5.36)
δ ż = δa + (δB)un ,
where
Thus
δ ż = a(t, x ∗ + δy + δz, v ∗ ) − a(t, x ∗ , v ∗ )
+ (B(t, x ∗ + δy + δz, v ∗ ) − B(t, x ∗ , v ∗ )) un , (5.39)
δz(t0 ) = 0, (t, x ∗ + δy + δz, v ∗ ) ∈ Q.
This system is equivalent to the system (5.31) in the following sense: x n solves (5.31) iff
x n = x ∗ + δy + δz, (5.40)
δz
C → 0. Therefore (5.32) holds.
234 Chapter 5. Second-Order Optimality Conditions in Optimal Control
Proof of Theorem 5.13. Suppose that w 0 is not a strict bounded strong minimum point in the
problem co Z. Then there exist a compact set C ⊂ Q and a sequence {w n } = {(x n , un , v n )}
such that for all n one has wn = w0 and
We will show that w0 is not a point of the bounded strong γ1 -sufficiency in the problem
Z. If un = u0 for infinitely many terms, w0 is not even a strict bounded strong minimum
point in the problem Z. Hence, we assume that un = u0 for all n. Apply Lemma 5.14 to
each function un . By virtue of this lemma there exists a sequence of measurable functions
{unk }∞
k=1 such that, for all k,
According to Theorem 5.15 this system has a solution for all sufficiently large k and
x nk − x n
C → 0 as k → ∞. (5.50)
where p nk = (x nk (t0 ), x nk (tf )) = (x n (t0 ), x nk (tf )). Combined with (5.49) this implies that
σ (wnk ) → 0 as k → ∞, (5.52)
where w nk = (x nk , unk , v n ). It follows from (5.48) and (5.52) that there is a number k = k(n)
such that tf 2
1
σ (wnk ) ≤ |unk (t) − u0 (t)| dt ∀ k ≥ k(n). (5.53)
n t0
By (5.50), k(n) also can be chosen so that
1
x nk(n) − x n
C ≤ . (5.54)
n
5.1. Quadratic Optimality Conditions in the Problem on a Fixed Time Interval 235
Finally, from (5.42) and (5.54) it follows that there exists a compact set C1 such that
C ⊂ C1 ⊂ Q, and for all sufficiently large n we have
(x nk(n) (t0 ), x nk(n) (tf )) ∈ P , (t, x nk(n) (t), v n (t)) ∈ C1 a.e. on . (5.59)
The existence of a compact set C1 ⊂ Q and a sequence {w nk(n) } in the space W satisfying
(5.55)–(5.59) means that the bounded strong γ1 -sufficiency fails at the point w0 in the
problem Z.
Theorem 5.16. Condition B (given in Definition 3.15) for the point w0 in the problem Z is
a sufficient condition for a strict bounded strong minimum at this point in the problem co Z.
Proof. Assume that Condition B for the point w0 in the problem Z is satisfied. Then
by Theorem 3.16 a bounded strong γ -sufficiency holds in problem Z at the point w 0 .
The latter means (see Definition 3.11) that there are no compact set C ⊂ Q and sequence
{wn } = {(x n , un , v n )} in W such that
and for all n we have wn = w0 , (t, x n (t), v n (t)) ∈ C, g(un (t)) = 0, un (t) ∈ Qg a.e. on ,
where γ is the higher order defined in Definition 2.17.
Note that for each n the conditions g(un (t)) = 0, un (t) ∈ Qg mean that un (t) ∈ U,
where U is a compact set. It follows from Proposition 2.98 that for any compact set C ⊂ Q
there exists a constant C > 0 such that for any w = (x, u, v) ∈ W satisfying the conditions
u(t) ∈ U and (t, x(t), v(t)) ∈ C a.e. on , we have γ1 (w − w0 ) ≤ Cγ (w − w0 ). By virtue
of this inequality, a bounded strong γ -sufficiency at the point w 0 in problem Z implies a
bounded strong γ1 -sufficiency at w 0 in the same problem. Then by Theorem 5.13, w0 is a
point of a strict bounded strong minimum in problem co Z.
Proofs of Theorems 5.5 and 5.8. Consider the main problem again:
The last inequality implies the conditions of a local minimum principle with respect to u
and v:
ψfu (t, x 0 (t), u0 (t), v 0 (t)) + ν(t)g (u0 (t)) = 0, ψfv (t, x 0 (t), u0 (t), v 0 (t)) = 0.
Since g
(u) = I , the first equality uniquely determines the multiplier ν(t).
Note that H̄x = Hx . Therefore conditions (5.67) are equivalent to the minimum
principle conditions (5.7)–(5.10). More precisely, the linear projection
(α0 , α, β, ψ, ν) → (α0 , α, β, ψ)
yields a one-to-one correspondence between the elements of the set (5.67) and the elements
of the set (5.7)–(5.10).
Consider the critical cone K for the point w0 (see Definition 3.5). To the constraint
g(u) = 0, u ∈ Qg there corresponds the condition g
(u0 )ū = 0. But g
= I . Hence ū = 0.
This condition implies that the critical cone K can be identified with the set of triples
z̄ = (ξ , x̄, v̄) ∈ Z2 () such that conditions (5.14)–(5.16) are fulfilled.
The definition of the set M0+ in Section 5.1.6 corresponds to the definition of this set
in Section 3.2.3. The same is true for the set Leg+ (M0+ ) (again, due to the equality g
= I ).
Further, for the point w 0 we write the quadratic form (see formula (3.57)), where
we set ū = 0. Since H̄x = Hx and H̄ψ = Hψ , we have [H̄x ]k = [Hx ]k , D k (H̄ ) = D k (H ),
k = 1, . . . , s. Moreover, H̄xx = Hxx , H̄xv = Hxv , H̄vv = Hvv . Combined with condition ū = 0
this implies
H̄ww w̄, w̄ = Hxx x̄, x̄ + 2Hxv v̄, x̄ + Hvv v̄, v̄ . (5.68)
Thus, in view of condition ū = 0 the quadratic form becomes defined by formula (5.17).
Now, Theorem 5.5 easily follows from Theorem 5.10, and similarly Theorem 5.8 becomes
a simple consequence of Theorem 5.16.
ẋ(t) = f (t, x(t), u(t), v(t)), u(t) ∈ U , (t, x(t), v(t)) ∈ Q, (5.70)
F (t0 , x(t0 ), tf , x(tf )) ≤ 0, K(t0 , x(t0 ), tf , x(tf )) = 0,
(5.71)
(t0 , x(t0 ), tf , x(tf )) ∈ P .
238 Chapter 5. Second-Order Optimality Conditions in Optimal Control
whereas the control variable v appears nonlinearly in the dynamics. Here, F , K, and a
are column vector functions, B is a d(x) × d(u) matrix function, P ⊂ R2+2d(x) and Q ⊂
R1+d(x)+d(v) are open sets, and U ⊂ Rd(u) is a convex polyhedron. The functions J , F , and
K are assumed to be twice continuously differentiable on P and the functions a and B are
twice continuously differentiable on Q. The dimensions of F and K are denoted by d(F )
and d(K). By = [t0 , tf ] we denote the interval of control. We shall use the abbreviations
x0 = x(t0 ), xf = x(tf ), and p = (t0 , x0 , tf , xf ). A process T = {(x(t), u(t), v(t)) | t ∈ [t0 , tf ] }
is said to be admissible, if x(·) is absolutely continuous, u(·), v(·) are measurable bounded
on = [t0 , tf ], and the triple of functions (x(t), u(t), v(t)) together with the endpoints p =
(t0 , x(t0 ), tf , x(tf )) satisfies the constraints (5.70) and (5.71).
(d) there exists a compact set C ⊂ Q (which depends on the choice of the sequence)
such that for all sufficiently large n, we have (t, x n (t), v n (t)) ∈ C a.e. on n .
Definition 5.18. The process T affords a Pontryagin minimum if for each compact set
C ⊂ Q there exists ε > 0 such that J(T˜ ) ≥ J(T ) for all admissible processes T˜ =
{(x̃(t), ũ(t), ṽ(t)) | t ∈ [t˜0 , t˜f ] } such that
(a) |t˜0 − t0 | < ε, |t˜f − tf | < ε;
(b) max∩ |x̃(t) − x(t)| < ε, where ˜ = [t˜0 , t˜f ];
˜
(c) ˜
∩ |ũ(t) − u(t)| dt < ε; ∩ ˜ |ṽ(t) − v(t)| dt < ε;
(d) ˜
(t, x̃(t), ṽ(t)) ∈ C a.e. on .
)
d(F
d(K)
α0 ≥ 0, α ≥ 0, αF (p) = 0, α0 + αi + |βj | = 1, (5.75)
i=1 j =1
Theorem 5.19. If the process T affords a Pontryagin minimum, then the set M0 is
nonempty. The set M0 is a finite-dimensional compact set and the projector λ → (α0 , α, β)
is injective on M0 .
Again we use the simple abbreviation (t) for indicating all arguments
(t, x(t), u(t), v(t), ψ(t)).
Let λ = (α0 , α, β, ψ(·), ψ0 (·)) ∈ M0 . From condition (5.79), it follows that H (t) is a contin-
uous function. In particular, we have [H ]k = H k+ − H k− = 0 for each tk ∈ , where
H k− := H (tk , x(tk ), u(tk −), v(tk ), ψ(tk )), H k+ := H (tk , x(tk ), u(tk +), v(tk ), ψ(tk )).
240 Chapter 5. Second-Order Optimality Conditions in Optimal Control
where Hxk− and Hxk+ are the left- and right-hand values of the function Hx (t) at tk , respec-
tively, [Ht ]k is the jump of the function Ht (t) at tk , etc.
Definition 5.20. We say that the process T affords a bounded strong minimum if there is
no sequence of admissible processes T n = {(x n (t), un (t), v n (t)) | t ∈ [t0n , tfn ] }, n = 1, 2, . . .,
such that
(a) J(T n ) < J(T );
(b) t0n → t0 , tfn → tf , x n (t0 ) → x(t0 ) (n → ∞);
(c) maxn ∩ |x n (t) − x(t)| → 0 (n → ∞), where n = [t0n , tfn ];
(d) there exists a compact set C ⊂ Q (which depends on the choice of the sequence)
such that for all sufficiently large n we have (t, x n (t), v n (t)) ∈ C a.e. on n .
Definition 5.21. The process T affords a bounded strong minimum if for each compact
set C ⊂ Q there exists ε > 0 such that J(T˜ ) ≥ J(T ) for all admissible processes T˜ =
{(x̃(t), ũ(t), ṽ(t)) | t ∈ [t˜0 , t˜f ] } such that
(a) |t˜0 − t0 | < ε, |t˜f − tf | < ε, |x̃(t0 ) − x(t0 )| < ε;
(b) max∩ ˜ |x̃(t) − x(t)| < ε, where ˜ = [t˜0 , t˜f ];
(c) (t, x̃(t), ṽ(t)) ∈ C a.e. on . ˜
The strict bounded strong minimum is defined in a similar way, with the nonstrict
inequality J(T˜ ) ≥ J(T ) replaced by the strict one and the process T˜ required to be dif-
ferent from T . Below, we shall formulate a quadratic necessary optimality condition of a
Pontryagin minimum (Definition 5.17) for given control process T . A strengthening of this
quadratic condition yields a quadratic sufficient condition of a bounded strong minimum
(Definition 5.20).
5.2. Quadratic Optimality Conditions on a Variable Time Interval 241
x̄¯0 = x̄(t0 ) + t¯0 ẋ(t0 ), x̄¯f = x̄(tf ) + t¯f ẋ(tf ), p̄¯ = (t¯0 , x̄¯0 , t¯f , x̄¯f ). (5.81)
The vector p̄¯ is considered a column vector. Note that t¯0 = 0, respectively, t¯f = 0, holds
for a fixed initial time t0 , respectively, final time tf . Denote by IF (p) = {i ∈ {1, . . . , d(F )} |
Fi (p) = 0} the set of indices of all active endpoint inequalities Fi (p) ≤ 0 at the point
p = (t0 , x(t0 ), tf , x(tf )). Denote by K the set of all z̄ ∈ Z2 () satisfying the following
conditions:
where p = (t0 , x(t0 ), tf , x(tf )), [ẋ] = ẋ(tk +) − ẋ(tk −). It is obvious that K is a convex cone
with finitely many faces in the space Z2 (). The cone K is called the critical cone.
where
¯ p̄
ωe (λ, z̄) = lpp p̄, ¯ − 2ψ̇(tf )x̄(tf )t¯f − ψ̇(tf )ẋ(tf ) + ψ̇0 (tf ) t¯f2
+ 2ψ̇(t0 )x̄(t0 )t¯0 + ψ̇(t0 )ẋ(t0 ) + ψ̇0 (t0 ) t¯02 , (5.86)
1
lpp = lpp (p, α0 , α, β, p), p = (t0 , x(t0 ), tf , x(tf )),
k
x̄av = (x̄ k− + x̄ k+ ),
2
Hxx (t) = Hxx (t, x(t), u(t), v(t), ψ(t)), etc.
Note that for a problem on a fixed time interval [t0 , tf ] we have t¯0 = t¯f = 0 and, hence,
¯ p̄ .
the quadratic form (5.86) reduces to lpp p̄, ¯ The following theorem gives the main
second-order necessary condition of optimality.
242 Chapter 5. Second-Order Optimality Conditions in Optimal Control
Theorem 5.22. If the process T affords a Pontryagin minimum, then the following Condi-
tion A holds: The set M0 is nonempty and
Denote by Leg+ (M0+ ) the set of all strictly Legendrian elements λ ∈ M0+ and set
tf
γ̄ (z̄) = t¯02 + t¯f2 + ξ̄ , ξ̄ + x̄(t0 ), x̄(t0 ) + v̄(t), v̄(t) dt.
t0
Theorem 5.24. Let the following Condition B be fulfilled for the process :
(a) The set Leg+ (M0+ ) is nonempty;
(b) there exists a nonempty compact set M ⊂ Leg+ (M0+ ) and a number C > 0
such that maxλ∈M (λ, z̄) ≥ C γ̄ (z̄) for all z̄ ∈ K.
Then T is a strict bounded strong minimum.
If the set Leg+ (M0+ ) is nonempty and K = {0}, then (b) is fulfilled automatically.
This is a first-order sufficient optimality condition of a strict bounded strong minimum. Let
us emphasize that there is no gap between the necessary Condition A and the sufficient
Condition B.
5.2. Quadratic Optimality Conditions on a Variable Time Interval 243
5.2.7 Proofs
Let us consider the following problem on a variable time interval which is similar to the
general problem (5.69)–(5.71):
ẋ(t) = f (t, x(t), u(t), v(t)), u(t) ∈ U, (t, x(t), v(t)) ∈ Q, (5.88)
F (t0 , x(t0 ), tf , x(tf )) ≤ 0, K(t0 , x(t0 ), tf , x(tf )) = 0,
(5.89)
(t0 , x(t0 ), tf , x(tf )) ∈ P .
whereas the control variable v appears nonlinearly. Here, F , K, and a are column vector
functions, B is a d(x) × d(u) matrix function, P ⊂ R2+2d(x) and Q ⊂ R1+d(x)+d(v) are
open sets, and U ⊂ Rd(u) is a compact set. The functions J , F , and K are assumed to be
twice continuously differentiable on P and the functions a and B are twice continuously
differentiable on Q. The dimensions of F and K are denoted by d(F ) and d(K). By
= [t0 , tf ] we denote the interval of control. The compact set U is specified by
Consider now the same question for sufficient quadratic conditions. It can be solved
with the aid of Theorem 5.16 obtained for the problem Z on a fixed time interval, but first
we have to make a simple time change. Namely, with the admissible trajectory T in the
problem A we associate the trajectory
T τ = z(τ ), t(τ ), x(τ ), u(τ ), v(τ ) | τ ∈ [τ0 , τf ] ,
Now, recall that the representation of the set of vertices of polyhedron U in the form of
equality-type constraint g(u) = 0, u ∈ Qg allowed us to consider the main problem (5.1)–
(5.4) as a special case of the problem co Z (5.18), (5.19), (5.23) and thus to obtain both
necessary Condition A and sufficient Condition B in the main problem as the consequences
of these conditions in problem co Z; more precisely, we have shown that Theorems 5.5 and
4 Note that the function z(τ ) in problem Aτ corresponds to the function v(τ ) in problem P τ .
5.3. Riccati Approach 245
5.8 follow from Theorem 5.10 and 5.16, respectively (see the proofs of Theorems 5.5 and
5.8). Similarly, this representation allows us to consider the general problem on a variable
time interval (5.69)–(5.72) as a special case of the problem co A and thus to obtain Theorems
5.22 and 5.24 as the consequences of Theorems 5.25 and 5.26, respectively.
These relations imply that J
(x(tf ))x̄(tf ) = 0 since α0 > 0. The quadratic form is given by
s
(λ, z̄) = lxf xf (x(tf ))x̄f , x̄f + (D k (H )ξ̄k2 − 2[ψ̇]k x̄av
k
ξ̄k )
k=1
tf
+ Hxx (t)x̄(t), x̄(t) + 2Hxv (t)v̄(t), x̄(t) + Hvv (t)v̄(t), v̄(t) dt,
t0
where, by definition, x̄f = x̄(tf ). We assume that D k (H ) > 0, k = 1, . . . , s, and the strength-
ened Legendre (SL) condition with respect to v is satisfied:
Hvv (t)v̄, v̄ ≥ cv̄, v̄ ∀ v̄ ∈ Rd(v) , ∀ t ∈ [t0 , tf ] \ (c > 0).
246 Chapter 5. Second-Order Optimality Conditions in Optimal Control
5.3.2 Q-Transformation of on K
Let Q(t) be a symmetric n × n matrix on [t0 , tf ] with piecewise continuous entries which
are absolutely continuous on each interval of the set [t0 , tf ] \ . For each z̄ ∈ K we obvi-
ously have
tf tf
d
s
Qx̄, x̄ dt = Qx̄, x̄ − [Qx̄, x̄ ]k , (5.100)
dt t0 t0 k=1
where [Qx̄, x̄ ] is the jump of the function Qx̄, x̄ at the point tk ∈ . Using the equation
k
where [ẋ]k is a column vector, while qk− , ([ẋ]k )∗ and [ψ̇]k are row vectors, and bk− is a
number. We shall assume that
where (qk− ) is a row vector, (qk− )∗ is a column vector, and hence (qk− )∗ (qk− ) is a symmetric
n × n matrix. Then, as it was shown in Section 4.2,
ωk = (bk− )−1 ((bk− )ξ̄k + (qk− )(x̄ k+ ))2 = (bk− )−1 (D k (H )ξ̄k + (qk− )(x̄ k− ))2 . (5.105)
Thus, we obtain the following transformation of the quadratic form = (λ, z̄) to perfect
squares on the critical cone K:
s
2 tf
= lxf xf − Q(tf ) x̄f , x̄f + (bk− )−1 (D k (H )ξ̄k + (qk− )(x̄ k− ) + −1
Hvv h̄, h̄ dt,
k=1 t0
where h̄ = (Hvx + fv∗ Q)x̄ + Hvv v̄. In addition, let us assume that
lxf xf − Q(tf ) x̄f , x̄f ≥ 0
for all x̄f ∈ Rd(x) \ {0} such that Kxf (x(tf ))x̄f = 0. Then, obviously, (λ, z̄) ≥ 0 on K. Now
let us show that (λ, z̄) > 0 for each nonzero element z̄ ∈ K. This will imply that (λ, z̄)
is positive definite on the critical cone K since (λ, z̄) is a Legendre quadratic form.
Assume that (λ, z̄) = 0 for some element z̄ ∈ K. Then, for this element, the following
equations hold:
x̄(t0 ) = 0, (5.106)
D k (H )ξ̄k + (qk− )(x̄ k− ) = 0, k = 1, . . . , s, (5.107)
h̄(t) = 0 a.e. in . (5.108)
Using this formula in the equation x̄˙ = fx x̄ + fv v̄, we see that x̄ is a solution to the linear
equation
x̄˙ = (fx − fv Hvv−1
(Hvx + fv∗ Q))x̄. (5.110)
This equation together with initial condition x̄(t0 ) = 0 implies that x̄(t) = 0 for all t ∈ [t0 , t1 ).
Consequently, x̄ 1− = 0, and then, by virtue of (5.107), ξ̄1 = 0. This equality, together
with the jump condition [x̄]1 = [ẋ]1 ξ̄1 , implies that [x̄]1 = 0, i.e., x̄ is continuous at t1 .
Consequently, x̄ 1+ = 0. From the last condition and equation (5.110) it follows that x̄(t) = 0
for all t ∈ (t1 , t2 ). Repeating this argument, we obtain ξ̄1 = ξ̄2 = · · · = ξ̄s = 0, x̄(t) = 0 for
all t ∈ [t0 , tf ]. Then from (5.109) it follows that v̄ = 0. Consequently, we have z̄ = 0 and
thus have proved the following theorem; cf. [98].
248 Chapter 5. Second-Order Optimality Conditions in Optimal Control
Theorem 5.27. Assume that there exists a symmetric matrix Q(t), defined on [t0 , tf ], such
that
(a) Q(t) is piecewise continuous on [t0 , tf ] and continuously differentiable on each
interval of the set [t0 , tf ] \ ;
(b) Q(t) satisfies the Riccati equation
Q̇ + Qfx + fx∗ Q + Hxx − (Hxv + Qfv )Hvv
−1
(Hvx + fv∗ Q) = 0 (5.111)
on each interval of the set [t0 , tf ] \ ;
(c) at each point tk ∈ matrix Q(t) satisfies the jump condition
[Q]k = (bk− )−1 (qk− )∗ (qk− ),
where qk− = ([ẋ]k )∗ Qk− − [ψ̇]k , bk− = D k (H ) − (qk− )[ẋ]k > 0;
(d) lxf xf − Q(tf ) x̄f , x̄f ≥ 0 for all x̄f ∈ Rd(x) \ {0} such that Kxf (x(tf ))x̄f = 0.
Then (λ, z̄) is positive definite on the subspace K.
according to
⎧ ⎫
⎨ mmax if φm (t) > 0 ⎬
m(t) = 0 if φm (t) < 0 . (5.122)
⎩ singular if φ (t) ≡ 0 for t ∈ Ising ⊂ [0, tf ] ⎭
m
For the final time tf = 1 which was considered in [22] and [67], the maintenance control
contains a singular arc. However, the computations in [67] show that for final times tf ∈
[0.15, 0.98] the maintenance control is bang-bang and has only one switching time:
0 for 0 ≤ t < t1
m(t) = . (5.123)
mmax = 4 for t1 ≤ t ≤ tf
Let us study the control problem with final time tf = 0.9 in more detail. To compute a
solution candidate, we apply nonlinear programming methods to the discretized control
problem with a large number N of grid points τi = i · tf /N , i = 0, 1, . . . , N ; cf. [5, 14]. Both
the method of Euler and the method of Heun are employed for integrating the differential
equation. We use the programming language AMPL developed by Fourer et al. [33] and
the interior point optimization code IPOPT of Wächter and Biegler [114]. For N = 5000
grid points, the computed state, control and adjoint functions are displayed in Figure 5.1.
The following values for the switching time, functional value, and selected state and adjoint
(a) inventory x1 and good items x2 (b) production v and bang-bang maintenance m
3
4
2.5 3.5
3
2
2.5
1.5 2
1.5
1
1
0.5 0.5
0
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
(c) adjoint variables 1 and 2 (d) maintenance rate m and switching function m
-3
4
-4
3
-5
2
-6
1
-7
-8 0
-9 -1
-10 -2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Figure 5.1. Optimal production and maintenance, final time tf = 0.9. (a) State
variables x1 , x2 . (b) Regular production control v and bang-bang maintenance control m.
(c) Adjoint variables ψ1 , ψ2 . (d) Maintenance control m with switching function φm .
5.4. Numerical Example: Optimal Control of Production and Maintenance 251
q1+ = ([ẋ 1 ]T Q1+ − [ψ̇]1 = (0, M(1 − x2 (t1 ))Q22 (t1 +) − Mψ2 (t1 )
= (0, 8.2439),
b1+ = D 1 (H ) + (q1+ )[ẋ 1 ]
= D 1 (H ) + M 2 ((1 − x2 (t1 ))Q22 (t1 +) − ψ2 (t1 ))ψ2 (t1 )
= 27.028 + 133.55 = 165.58 > 0.
reduces to a jump condition for Q22 at t1 . However, we do not need to evaluate this jump
condition explicitly because the linear equation (5.129) has a solution regardless of the value
Q22 (t1 −). Hence, we conclude from Theorem 5.27 that the numerical solution characterized
by (5.124) and displayed in Figure 5.1 provides a strict bounded strong minimum.
We may hope to improve on the benefit by choosing a larger time horizon. For the
final time tf = 1.1, we get a bang-singular-bang maintenance control m(t) as shown in
Figure 5.2.
-10 0
-11
-1
-12
-13 -2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Figure 5.2. Optimal production and maintenance, final time tf = 1.1. (a) State
variables x1 , x2 . (b) Regular production control v and bang-singular-bang maintenance
control m. (c) Adjoint variables ψ1 , ψ2 . (d) Maintenance control m with switching func-
tion φm .
5.4. Numerical Example: Optimal Control of Production and Maintenance 253
J = 23.3567,
x1 (tf ) = 0, x2 (tf ) = 0.64926,
(5.131)
ψ1 (0) = −11.93, ψ2 (0) = −9.332,
ψ1 (tf ) = −12.97, ψ2 (tf ) = −8.956.
Apparently, the larger time horizon tf = 1.1 results in a smaller gain J = 23.357 compared to
J = 26.705 for the final time tf = 0.9. We are not aware of any type of sufficient optimality
conditions that would apply to the extremal for tf = 1.1, where one control component has
a bang-singular-bang structure. Thus one is lead to ask: What is the optimal lifetime of
the machine to give maximal gain? This amounts to solving the control problem (5.113)–
(5.116) with free final time tf . The solution is very similar to that shown in Figure 5.1. The
maintenance control m(t) is bang-bang with one switching time t1 = 0.6523. The optimal
final time tf = 0.8633 gives the gain J = 26.833 which slightly improves on J = 26.705
for the final time tf = 0.9.
Chapter 6
Second-Order Optimality
Conditions for Bang-Bang
Control
In this chapter, we investigate the pure bang-bang case, where the second-order necessary
or sufficient optimality conditions amount to testing the positive (semi)definiteness of a
quadratic form on a finite-dimensional critical cone. In Section 6.2, we deduce these con-
ditions from the results obtained in the previous chapter. Although the quadratic conditions
turned out to be finite-dimensional, the direct numerical test works only in some special
cases. Therefore, in Section 6.3, we study various transformations of the quadratic form
and the critical cone which are tailored to different types of control problems in practice.
In particular, by a solution to a linear matrix differential equation, the quadratic form can be
converted to perfect squares. In Section 6.5, we study second-order optimality conditions
for time-optimal control problems with control appearing linearly. In Section 6.6, we show
that an approach similar to the above mentioned Riccati equation approach is applicable for
such problems. Again, the test requires us to find a solution of a linear matrix differential
equation which satisfies certain jump conditions at the switching points. In Section 6.7, we
discuss two numerical examples that illustrate the numerical procedure of verifying positive
definiteness of the corresponding quadratic forms. In Section 6.8, following [79], we study
second-order optimality conditions in a simple, but important class of time-optimal control
problems for linear autonomous systems.
255
256 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
Here, F , K, and a are vector functions, B is a d(x) × d(u) matrix function, P ⊂ R2+2d(x) ,
Q ⊂ R1+d(x) are open sets, and U ⊂ Rd(u) is a convex polyhedron. The functions J , F ,
and K are assumed to be twice continuously differentiable on P , and the functions a, B are
twice continuously differentiable on Q. The dimensions of F and K are denoted by d(F )
and d(K). We shall use the abbreviations x0 = x(t0 ), xf = x(tf ), p = (t0 , x0 , tf , xf ). A tra-
jectory T = (x(t), u(t) | t ∈ [t0 , tf ]) is said to be admissible if x(·) is absolutely continuous,
u(·) is measurable bounded, and the pair of functions (x(t), u(t)) on the interval = [t0 , tf ]
with the endpoints p = (t0 , x(t0 ), tf , x(tf )) satisfies the constraints (6.2) and (6.3). We set
J(T ) := J (t0 , x(t0 ), tf , x(tf )).
Obviously, the basic problem (6.1)–(6.3) is a special case of the general problem
(5.69)–(5.72) studied in the previous chapter. This special case corresponds to the assump-
tion that the function f does not depend on the variable v, i.e., f = f (t, x, u). Under this
assumption it turned out to be possible to obtain certain deeper results than in the general
problem. More precisely, we formulate the necessary quadratic Condition A in the prob-
lem (6.1)–(6.3), which is a simple consequence of the Condition A in the general problem,
whereas the correspondent sufficient quadratic Condition B will be slightly simplified.
Let us give the definition of Pontryagin minimum for the basic problem. The tra-
jectory T affords a Pontryagin minimum if there is no sequence of admissible trajectories
T n = (x n (t), un (t) | t ∈ [t0n , tfn ]), n = 1, 2, . . . , such that the following properties hold with
n = [t0n , tfn ]:
(a) J(T n ) < J(T ) for all n and t0n → t0 , tfn → tf for n → ∞;
(b) maxn ∩ |x n (t) − x(t)| → 0 for n → ∞;
n ∩ |u (t) − u(t)| dt → 0 for n → ∞.
(c) n
Note that for a fixed time interval , a Pontryagin minimum corresponds to an L1 -local
minimum with respect to the control variable.
)
d(F
d(K)
α0 ≥ 0, α ≥ 0, αF (p) = 0, α0 + αi + |βj | = 1, (6.7)
i=1 j =1
Theorem 6.1. If the trajectory T affords a Pontryagin minimum, then the set M0 is
nonempty. The set M0 is a finite-dimensional compact set, and the projector λ → (α0 , α, β)
is injective on M0 .
Consequently, for each λ ∈ M0 the function (k H )(t) has a derivative at the point
tk ∈ . Set
d
D k (H ) = − (k H )(tk ).
dt
258 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
Then, for each λ ∈ M0 , the minimum condition (6.10) implies the inequality
D k (H ) ≥ 0, k = 1, . . . , s. (6.13)
where Hxk− and Hxk+ are the left- and right-hand values of the function
at tk , respectively, [Ht ]k is a jump of the function Ht (t) at tk , etc. It also follows from the
above representation that we have
where the values on the right-hand side agree for the derivative φ̇(tk +) from the right and
the derivative φ̇(tk −) from the left. In the case of a scalar control u, the total derivative
φt + φx ẋ + φψ ψ̇ does not contain the control variable explicitly and hence the derivative
φ̇(t) is continuous at tk .
Proof. The equalities (6.15) follow from the equality ψ(t)ẋ(t) + ψ0 (t) = 0 evaluated for
t = t0 and t = tf together with the transversality conditions
The component y was called an unessential component in the augmented problem. The
general definition was given in Section 5.2.3: the state variable xi is called unessential if
the function f does not depend on xi and if the functions F , J , and K are affine in xi0 = xi (t0 )
and xi1 = xi (tf ). Let x denote the vector of all essential components of state vector x. Now
we can define a strong minimum in the basic problem.
We say that the trajectory T affords a strong minimum if there is no sequence of
admissible trajectories T n = (x n (t), un (t) | t ∈ [t0n , tfn ]), n = 1, 2, . . . , such that
(a) J(T n ) < J(T );
(b) t0n → t0 , tfn → tf , x n (t0 ) → x(t0 ) (n → ∞);
(c) maxn ∩ |x n (t) − x(t)| → 0 (n → ∞), where n = [t0n , tfn ].
The strict strong minimum is defined in a similar way, with the strict inequality (a) replaced
by the nonstrict one and the trajectory T n required to be different from T for each n.
where [u(t−), u(t+)] denotes the line segment spanned by the vectors u(t−) and u(t+) in
Rd(u) and φ(t) := φ(t, x(t), ψ(t)) = ψ(t)B(t, x(t)). Note that [u(t−), u(t+)] is a singleton
{u(t)} at each continuity point of the control u(t) with u(t) being a vertex of the polyhe-
dron U . Only at the points tk ∈ does the line segment [uk− , uk+ ] coincide with an edge
of the polyhedron.
As it was already mentioned in Section 5.1.7, if the control is scalar, d(u) = 1,
and U = [umin , umax ], then the strict bang-bang property is equivalent to φ(t) = 0 for all
t ∈ \ which yields the control law
umin if φ(t) > 0
u(t) = ∀ t ∈ \ . (6.20)
umax if φ(t) < 0
For vector-valued control inputs, condition (6.19) imposes further restrictions. For exam-
ple, if U is the unit cube in Rd(u) , condition (6.19) precludes simultaneous switching of
the control components; the case of simultaneous switching was studied in Felgenhauer
[31]. This property holds in many examples. The condition (6.19) is indispensable in the
sensitivity analysis of optimal bang-bang controls.
The vector p̄¯ is considered as a column vector. Note that t¯0 = 0, respectively, t¯f = 0, for
fixed initial time t0 , respectively, final time tf . Let IF (p) = {i ∈ {1, . . . , d(F )} | Fi (p) = 0}
be the set of indices of all active endpoint inequalities Fi (p) ≤ 0 at the point p = (t0 , x(t0 ),
tf , x(tf )). Denote by K the set of all z̄ ∈ Z() satisfying the following conditions:
where p = (t0 , x(t0 ), tf , x(tf )). It is obvious that K is a convex finite-dimensional and finite-
faced cone in the space Z(). We call it the critical cone. Each element z̄ ∈ K is uniquely
defined by numbers t¯0 , t¯f , a vector ξ̄ , and the initial value x̄(t0 ) of the function x̄(t).
Proof. Integrating the equality ψ(x̄˙ − fx x̄) = 0 on [t0 , tf ] and using the adjoint equation
t t
ψ̇ = −ψfx we obtain t0f dtd (ψ x̄) dt = 0, whence (ψ x̄)|tf0 − sk=1 [ψ x̄]k = 0. From the
jump conditions [x̄]k = [ẋ]k ξ̄k and the equality ψ(t)ẋ(t) + ψ0 (t) = 0 it follows that [ψ x̄]k =
t
ψ(tk )[ẋ]k ξ̄k = [ψ ẋ]k ξ̄k = −[ψ0 ]k ξ̄k = 0 for all k. Then the equation (ψ x̄)|tf0 = 0, together
with the transversality conditions ψ(t0 ) = −lx0 and ψ(tf ) = lxf , implies (6.24).
s
α0 J
(p)p̄¯ + αi Fi
(p)p̄¯ + βK
(p)p̄¯ = 0. (6.25)
i=1
t¯0 (lx0 ẋ(t0 ) + lt0 ) + t¯f (lxf ẋ(tf ) + ltf ) + lx0 x̄(t0 ) + lxf x̄(tf ) = 0.
Now using the equalities x̄¯0 = x̄(t0 ) + t¯0 ẋ(t0 ), x̄¯f = x̄(tf ) + t¯f ẋ(tf ), and p̄¯ = t¯0 , x̄¯0 , t¯f , x̄¯f ,
we get lp p̄¯ = 0 which is equivalent to condition (6.25).
6.2. Quadratic Necessary and Sufficient Optimality Conditions 261
Two important properties of the critical cone follow from Proposition 6.4.
Proposition 6.5. For any λ ∈ M0 and z̄ ∈ K, we have α0 J
(p)p̄¯ = 0 and αi Fi
(p)p̄¯ = 0 for
all i ∈ IF (p).
Proposition 6.6. Suppose that there exist λ ∈ M0 with α0 > 0. Then adding the equali-
ties αi Fi
(p)p̄¯ = 0 for all i ∈ IF (p) to the system (6.22), (6.23) defining K, one can omit
the inequality J
(p)p̄¯ ≤ 0 in that system without affecting K.
Thus, K is defined by condition (6.23) and by the condition p̄¯ ∈ K0 , where K0 is the
cone in R2d(x)+2 given by (6.22). But if there exists λ ∈ M0 with α0 > 0, then we can put
K0 = {p̄¯ ∈ Rd(x)+2 | Fi
(p)p̄¯ ≤ 0, αi Fi
(p)p̄¯ = 0 ∀ i ∈ IF (p), K
(p)p̄¯ = 0}. (6.26)
If, in addition, αi > 0 holds for all i ∈ IF (p), then K0 is a subspace in Rd(x)+2 .
An explicit representation of the variations x̄(t) in (6.23) is obtained as follows. For
each k = 1, . . . , s, define the vector functions y k (t) as the solutions to the system
ẏ = fx (t)y, y(tk ) = [ẋ]k , t ∈ [tk , tf ].
For t < tk we put y k (t) = 0 which yields the jump [y k ]k = [ẋ]k . Moreover, define y 0 (t) as
the solution to the system
ẏ = fx (t)y, y(t0 ) = x̄(t0 ) =: x̄0 .
By the superposition principle for linear ODEs it is obvious that we have
s
x̄(t) = y k (t)ξ̄k + y 0 (t)
k=1
Furthermore, denote by x(t; t1 , . . . , ts ) the solution of the state equation (6.2) using the
values of the optimal bang-bang control with switching points t1 , . . . , ts . It easily follows
from elementary properties of ODEs that the partial derivatives of state trajectories with
respect to the switching points are given by
∂x
(t; t1 , . . . , ts ) = −y k (t) for t ≥ tk , k = 1, . . . , s. (6.28)
∂tk
This gives the following expression for x̄(t):
s
∂x
x̄(t) = − (t)ξ̄k + y 0 (t). (6.29)
∂tk
k=1
In a special case that frequently arises in practice, we can use these formulas to show that
K = {0}. This property then yields a first-order sufficient condition in view of Theorem 6.10.
262 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
Namely, consider the following problem with an integral cost functional, where the initial
time t0 = tˆ0 is fixed, while the final time tf is free and where the initial and final values of
the state variables are given: Minimize
tf
J= f0 (t, x, u)dt (6.30)
t0
subject to
ẋ = f (t, x, u), x(t0 ) = x̂0 , x(tf ) = x̂f , u(t) ∈ U , (6.31)
where f is defined by (6.4), and f0 is defined by (6.16). In the definition of K we then
have t¯0 = 0, x̄(t0 ) = 0, x̄¯f = 0. The condition x̄(t0 ) = 0 implies that y 0 (t) ≡ 0, whereas the
condition x̄¯f = 0 yields in view of the representation (6.27),
s
y k (tf )ξ̄k + ẋ(tf )t¯f = 0.
k=1
We conclude this subsection with a special property of the critical cone for time-
optimal control problems with fixed initial time and state,
tf → min, ẋ = f (t, x, u), u ∈ U , t0 = tˆ0 , x(t0 ) = x̂0 , K(x(tf )) = 0, (6.32)
where f is defined by (6.4). The following result will be used in Section 6.2.3; cf. Propo-
sition 6.11.
Proposition 6.8. Suppose that there exists (ψ0 , ψ) ∈ M0 such that α0 > 0. Then t¯f = 0
holds for each z̄ = (t¯f , ξ̄ , x̄) ∈ K.
Proof. For arbitrary λ ∈ M0 and z̄ = (t¯f , ξ̄ , x̄) ∈ K we infer from the proof of Proposition
6.3 that ψ(t)x̄(t) is a constant function on [t0 , tf ]. In view of the relations
ψ(tf ) = βKxf (x(tf )), Kxf (x(tf ))x̄¯f = 0, x̄¯f = x̄(tf ) + ẋ(tf )t¯f ,
we get
0 = (ψ x̄)(t0 ) = (ψ x̄)(tf ) = ψ(tf )(x̄¯f − ẋ(tf )t¯f ) = −ψ(tf )ẋ(tf )t¯f = ψ0 (tf )t¯f .
Since ψ0 (tf ) = α0 > 0, this relation yields t¯f = 0.
In the case α0 > 0, we note as a consequence that the critical cone is a subspace defined by
the conditions
x̄˙ = fx (t)x̄, [x̄]k = [ẋ]k ξ̄k (k = 1, . . . , s),
(6.33)
t¯0 = t¯f = 0, x̄(t0 ) = 0, Kxf (x(tf ))x̄(tf ) = 0.
6.2. Quadratic Necessary and Sufficient Optimality Conditions 263
where
¯ p̄
Ap̄, ¯ p̄
¯ = lpp p̄, ¯ + 2ψ̇(t0 )x̄¯0 t¯0 + (ψ̇0 (t0 ) − ψ̇(t0 )ẋ(t0 ))t¯02
− 2ψ̇(tf )x̄¯f t¯f − (ψ̇0 (tf ) − ψ̇(tf )ẋ(tf ))t¯f2 , (6.35)
lpp = lpp (α0 , α, β, p), p = (t0 , x(t0 ), tf , x(tf )), Hxx = Hxx (t, x(t), u(t), ψ(t)),
1
k
x̄av = (x̄ k− + x̄ k+ ).
2
Note that for a problem on a fixed time interval [t0 , tf ] we have t¯0 = t¯f = 0 and, hence,
¯ p̄
the quadratic form (6.35) reduces to Ap̄, ¯ = lpp p̄, p̄ . The following theorem gives the
main second-order necessary condition of optimality.
Theorem 6.9. If the trajectory T affords a Pontryagin minimum, then the following Con-
dition A holds: The set M0 is nonempty and maxλ∈M0 (λ, z̄) ≥ 0 for all z̄ ∈ K.
Theorem 6.10. Let the following Condition B be fulfilled for the trajectory T :
(a) there exists λ ∈ M0 such that D k (H ) > 0, k = 1, . . . , s, and condition (6.19) holds
(i.e., u(t) is a strict bang-bang control),
(b) maxλ∈M0 (λ, z̄) > 0 for all z̄ ∈ K \ {0}.
Then T is a strict strong minimum.
Note that the condition (b) is automatically fulfilled, if K = {0}, which gives a first-
order sufficient condition for a strong minimum in the problem. A specific situation where
K = {0} holds was described in Proposition 6.7. Also note that the condition (b) is auto-
matically fulfilled if there exists λ ∈ M0 such that
Example: Resource allocation problem. Let x(t) be the stock of a resource and
let the control u(t) be the investment rate. The control problem is to maximize the overall
consumption tf
x(t)(1 − u(t)) dt
0
264 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
We conclude this subsection with the case of a time-optimal control problem (6.32)
with a single switching point, i.e., s = 1. Assume that α0 > 0 for a given λ ∈ M0 . Then by
Proposition 6.8, we have t¯f = 0, and thus the critical cone is the subspace defined by (6.33).
In this case, the quadratic form can be computed explicitly as follows. Denote by y(t),
t ∈ [t1 , tf ], the solution to the Cauchy problem
ẏ = fx y, y(t1 ) = [ẋ]1 .
The following assertion is obvious: If (ξ̄ , x̄) ∈ K, then x̄(t) = 0 for t ∈ [t0 , t1 ) and x̄(t) =
y(t)ξ̄ for t ∈ (t1 , tf ]. Therefore, the inequality Kxf (x(tf ))y(tf ) = 0 would imply K = {0}.
Consider now the case Kxf (x(tf ))y(tf ) = 0. This condition always holds for time-optimal
problems with a scalar function K and α0 > 0. Indeed, the condition dtd (ψy) = 0 implies
(ψy)(t) = const in [t1 , tf ], whence
(ψy)(tf ) = (ψy)(t1 ) = ψ(t1 )[ẋ]1 = φ(t1 )[u]1 = 0.
Using the transversality condition ψ(tf ) = βKxf (x(tf )) and the inequality β = 0 (if β = 0,
then ψ(tf ) = 0, and hence ψ(t) = 0 and ψ0 (t) = 0 in [t0 , tf ]), we see that the equality
(ψy)(tf ) = 0 implies the equality Kxf (x(tf ))y(tf ) = 0.
Observe now that the cone K is a one-dimensional subspace, on which the quadratic
form has the representation = ρ ξ̄ 2 , where
tf
ρ := D 1 (H ) − [ψ̇]1 [ẋ]1 + (y(t))∗ Hxx (t)y(t) dt + (y(tf ))∗ (βK)xf xf y(tf ). (6.37)
t1
6.2. Quadratic Necessary and Sufficient Optimality Conditions 265
Proposition 6.11. Suppose that we have found an extremal for the time-optimal control
problem (6.32) that has one switching point and satisfies α0 > 0 and Kxf (x(tf ))y(tf ) = 0.
Then the inequality ρ > 0 with ρ defined in (6.37) is equivalent to the positive definiteness
of on K.
Thus, Theorem 6.9, which gives necessary quadratic conditions in the problem (6.1)–(6.3),
is a consequence of Theorem 5.22.
Now let us proceed to sufficient quadratic conditions in the same problem. Here the
set M0+ consists of all those elements λ ∈ M0 for which condition (6.19) is fulfilled, and
the set Leg+ (M0+ ) consists of all those elements λ ∈ M0+ for which
D k (H ) > 0, k = 1, . . . , s. (6.38)
Denote for brevity Leg+ (M0+ ) = M.Thus the set M consists of all those elements λ ∈ M0
for which (6.19) and (6.38) are fulfilled. Let us also note that the strict bounded strong
minimum in the problem (6.1)–(6.3) is equivalent to the strict strong minimum, since U is
a compact set. Thus Theorem 5.24 implies the following result.
Theorem 6.12. For a trajectory T in the problem (6.1)–(6.3) let the following Condition B
be fulfilled: The set M is nonempty and there exist a nonempty compact M ⊂ M and ε > 0
such that
max (λ, z̄) ≥ ε γ̄ (z̄) ∀ z̄ ∈ K, (6.39)
λ∈M
where γ̄ (z̄) = t¯02 + t¯f2 + ξ̄ , ξ̄ + x̄(t0 ), x̄(t0 ) . Then the trajectory T affords a strict strong
minimum in this problem.
Remarkably, the fact that the critical cone K in the problem (6.1)–(6.3) is finite-
dimensional (since each element z̄ = (t¯0 , t¯f , ξ̄ , x̄) is uniquely defined by the parameters
266 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
t¯0 , t¯f , ξ̄ , x̄(t0 )) implies that condition B in Theorem 6.12 is equivalent to the following,
generally weaker condition.
Condition B0 . The set M is nonempty and
Proof. As we pointed out, Condition B always implies Condition B0 . We will show that,
in our case, the inverse assertion also holds. Let Condition B0 be fulfilled. Let S1 (K) be
the set of elements z̄ ∈ K satisfying the Condition γ̄ (z̄) = 1. Then
Recall that M0 is a finite-dimensional compact set. It is readily verified that the relative
interior of the cone con M0 generated by M0 is contained in the cone con M generated
by M, i.e., reint(con M0 ) ⊂ con M. Combined with (6.41) this implies that for any element
z̄ ∈ S1 (K) there exist λ ∈ M and a neighborhood Oz̄ ⊂ S1 (K) of element z̄ such that
(λ, ·) > 0 on Oz̄ . The family of neighborhoods {Oz̄ } forms an open covering of the compact
set S1 (K). Select a finite subcovering. To this subcovering there corresponds a finite
compact set M = {λ1 , . . . , λr } such that
Theorem 6.12 and Lemma 6.13 imply Theorem 6.10, where B = B0 . In what fol-
lows, for the problems of the type of basic problem (6.1)–(6.3), by Condition B we will
mean Condition B0 .
Now our aim is to find sufficient conditions for the positive definiteness of the quadratic
form on the cone K. In what follows, we shall use the ideas and results presented in
Chapter 4 (see also [69]).
6.3.1 Q-Transformation of on K
Let Q(t) be a symmetric matrix on [t0 , tf ] with piecewise continuous entries which are
absolutely continuous on each interval of the set [t0 , tf ] \ . Therefore, Q may have a jump
at each point tk ∈ . For z̄ ∈ K, formula (5.100) holds:
tf
tf d s
Qx̄, x̄ dt = Qx̄, x̄ − [Qx̄, x̄ ]k ,
t0 dt t0 k=1
where [Qx̄, x̄ ]k is the jump of the function Qx̄, x̄ at the point tk ∈ . Using the equation
x̄˙ = fx x̄ with fx = fx (t, x(t), u(t)), we obtain
s tf
[Qx̄, x̄ ]k + (Q̇ + fx∗ Q + Qfx )x̄, x̄ dt − Qx̄, x̄ (tf ) + Qx̄, x̄ (t0 ) = 0,
k=1 t0
where the asterisk denotes transposition. Adding this zero form to and using the equality
[Hx ]k = −[ψ̇]k , we get
¯ p̄
= Ap̄, ¯ − Qx̄, x̄ (tf ) + Qx̄, x̄ (t0 )
s
+ D k (H )ξ̄k2 − 2[ψ̇]k x̄av
k
ξ̄k + [Qx̄, x̄ ]k (6.42)
k=1
tf
+ (Hxx + Q̇ + fx∗ Q + Qfx )x̄, x̄ dt.
t0
It is interesting to note that the same equation is obtained from the modified Riccati equa-
tion in Maurer and Pickenhain [73, Equation (47)], when all control variables are on the
boundary of the control constraints. Using (6.43) the quadratic form (6.42) reduces to
s
= ω0 + ωk , (6.44)
k=1
ωk := D k (H )ξ̄k2 − 2[ψ̇]k x̄av
k
ξ̄k + [Qx̄, x̄ ]k , k = 1, . . . , s, (6.45)
¯ p̄
ω0 := Ap̄, ¯ − Qx̄, x̄ (tf ) + Qx̄, x̄ (t0 ). (6.46)
268 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
Proposition 6.14. Let Q(t) satisfy the linear differential equation (6.43) on [t0 , tf ] \ .
Then for each z̄ ∈ K the representation (6.44) holds.
Now our goal is to derive conditions such that ωk > 0, k = 0, . . . , s, holds on K \ {0}.
To this end, as in Section 4.1.6, we shall express ωk via the vector (ξ̄k , x̄ k− ). We use the
formula
x̄ k+ = x̄ k− + [ẋ]k ξ̄k , (6.47)
which implies
Consequently,
Here [ẋ]k and x̄ k− are column vectors while ([ẋ]k )∗ , (x̄ k− )∗ , and [ψ̇]k are row vectors. By
putting
qk+ = ([ẋ]k )∗ Qk+ − [ψ̇]k , bk+ = D k (H ) + (qk+ )[ẋ]k , (6.49)
we get
ωk = (bk+ )ξ̄k2 + 2(qk+ )x̄ k− ξ̄k + (x̄ k− )∗ [Q]k x̄ k− . (6.50)
Note that ωk is a quadratic form in the variables (ξ̄k , x̄ k− ) with the matrix
bk+ qk+
Mk+ = , (6.51)
(qk+ )∗ [Q]k
where qk+ is a row vector and (qk+ )∗ is a column vector. Similarly, using the relation
x̄ k− = x̄ k+ − [ẋ]k ξ̄k , we obtain
where
qk− = ([ẋ]k )∗ Qk− − [ψ̇]k , bk− = D k (H ) − (qk− )[ẋ]k . (6.53)
6.3. Sufficient Conditions for Positive Definiteness of the Quadratic Form 269
We consider (6.52) as a quadratic form in the variables (ξ̄k , x̄ k+ ) with the matrix
bk− qk−
Mk− = . (6.54)
(qk− )∗ [Q]k
Since the right-hand sides of equalities (6.50) and (6.52) are connected by relation (6.47),
the following statement obviously holds.
Proposition 6.15. For each k = 1, . . . , s, the positive (semi)definiteness of the matrix Mk−
is equivalent to the positive (semi)definiteness of the matrix Mk+ .
Theorem 6.16. Assume that s = 1. Let Q(t) be a solution to the linear differential equation
(6.43) on [t0 , tf ] \ which satisfies two conditions:
(i) The matrix M1+ is positive semidefinite;
(ii) the quadratic form ω0 is positive on the cone K0 \ {0}.
Then is positive on K \ {0}.
Proof. Take an arbitrary element z̄ ∈ K. Conditions (i) and (ii) imply that ωk ≥ 0 for
k = 0, 1, and hence = ω0 + ω1 ≥ 0 for this element. Assume now that = 0. Then
ωk = 0 for k = 0, 1. By virtue of (ii) the equality ω0 = 0 implies that t¯0 = t¯f = 0 and
x̄(t0 ) = x̄(tf ) = 0. The last two equalities and the equation x̄˙ = fx x̄ show that x̄(t) = 0 in
[t0 , t1 ) ∪ (t1 , tf ]. Now using formula (6.45) for ω1 = 0, as well as the conditions D 1 (H ) > 0
and x̄ 1− = x̄ 1+ = 0, we obtain that ξ̄1 = 0. Consequently, we have z̄ = 0 which means that
is positive on K \ {0}.
Theorem 6.17. Assume that s ≥ 2. Let Q(t) be a solution to the linear differential equation
(6.43) on [t0 , tf ] \ which satisfies the following conditions:
(a) The matrix Mk+ is positive semidefinite for each k = 1, . . . , s;
(b) bk+ := D k (H ) + (qk+ )[ẋ]k > 0 for each k = 1, . . . , s − 1;
(c) the quadratic form ω0 is positive on the cone K0 \ {0}.
Then is positive on K \ {0}.
Proof. Take an arbitrary element z̄ ∈ K. Conditions (a) and (c) imply that ωk ≥ 0 for
k = 0, 1, . . . , s, and then ≥ 0 for this element. Assume that = 0. Then ωk = 0 for k =
0, 1, . . . , s. By virtue of (c) the equality ω0 = 0 implies that t¯0 = t¯f = 0 and x̄(t0 ) = x̄(tf ) = 0.
The last two equalities and the equation x̄˙ = fx x̄ show that x̄(t) = 0 in [t0 , t1 ) ∪ (ts , tf ] and
hence x̄ 1− = x̄ s+ = 0. The conditions ω1 = 0, x̄ 1− = 0, and b1+ > 0 by formula (6.50)
(with k = 1) yield ξ̄1 = 0. Then [x̄]1 = 0 and hence x̄ 1+ = 0. The last equality and
the equation x̄˙ = fx x̄ show that x̄(t) = 0 in (t1 , t2 ) and hence x̄ 2− = 0. Similarly, the
conditions ω2 = 0, x̄ 2− = 0 and b2+ > 0 by formula (6.50) (with k = 2) imply that ξ̄2 = 0
and x̄(t) = 0 in (t2 , t3 ). Therefore, x̄ 3− = 0, etc. Continuing this process we get that x̄ ≡ 0
and ξ̄k = 0 for k = 1, . . . , s −1. Now using formula (6.45) for ωs = 0, as well as the conditions
D s (H ) > 0 and x̄ ≡ 0, we obtain that ξ̄s = 0. Consequently, z̄ = 0, and hence is positive
on K \ {0}.
270 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
Similarly, using representation (6.52) for ωk we can prove the following statement.
Theorem 6.18. Let Q(t) be a solution to the linear differential equation (6.43) on [t0 , tf ] \
which satisfies the following conditions:
(a
) The matrix Mk− is positive semidefinite for each k = 1, . . . , s;
(b
) bk− := D k (H ) − (qk− )[ẋ]k > 0 for each k = 2, . . . , s
(if s = 1, then this condition is not required);
(c) the quadratic form ω0 is positive on the cone K0 \ {0}.
Then is positive on K \ {0}.
Remark. Noble and Schättler [80] and Ledzewicz and Schättler [53] use also the linear
ODE (6.43) for deriving sufficient conditions. It would be of interest to relate their approach
to the results in Theorem 6.18.
Theorem 6.19. Assume that the initial values t0 = tˆ0 and x(t0 ) = x̂0 are fixed in the problem
(6.1)–(6.3), and let s = 1. Let Q(t) be a continuous solution of the linear differential
equation (6.43) on [t1 , tf] which satisfies twoconditions:
(i) b1 := D 1 (H ) + ([ẋ]1 )∗ Q(t1 ) − [ψ̇]1 [ẋ]1 ≥ 0;
(ii) the quadratic form ω0 is positive on the cone K0 \ {0}.
Then is positive on K \ {0}.
6.3. Sufficient Conditions for Positive Definiteness of the Quadratic Form 271
Proof. Continue Q(t) arbitrarily as a solution of differential equation (6.43) to the whole
interval [t0 , tf ] with possible jump at the point t1 . Note that the value b1 in condition (i) is
the same as the value b1+ for the continued solution, and hence b1+ ≥ 0. Let z̄ ∈ K, and
hence x̄ 1− = 0. Then by (6.50) with k = 1 the condition b1+ ≥ 0 implies the inequality
ω1 ≥ 0. Condition (ii) implies the inequality ω0 ≥ 0. Consequently = ω0 + ω1 ≥ 0.
Further arguments are the same as in the proof of Theorem 6.16.
Theorem 6.20. Assume that the initial values t0 = tˆ0 and x(t0 ) = x̂0 are fixed in the problem
(6.1)–(6.3) and s ≥ 2. Let Q(t) be a solution of the linear differential equation (6.43) on
(t1 , tf ] \ which satisfies the following conditions:
(a) The matrix Mk+ is positive semidefinite for each k = 2, . . . , s;
(b) bk+ := D k (H ) + (qk+ )[ẋ]k > 0 for each k = 1, . . . , s − 1;
(c) the quadratic form ω0 is positive on the cone K0 \ {0}.
Then is positive on K \ {0}.
Proof. Again, without loss of generality, we can consider Q(t) as a discontinuous solution
of equation (6.43) on the whole interval [t0 , tf ]. Let z̄ ∈ K. Then by (6.50) with k = 1 the
conditions b1+ > 0 and x̄ 1− = 0 imply the inequality ω1 ≥ 0. Further arguments are the
same as in the proof of Theorem 6.17.
where (qk+ )∗ is a column vector while qk+ is a row vector (defined as in (6.49)). Then ωk
can be written as the perfect square
2
ωk = (bk+ )−1 (bk+ )ξ̄k + (qk+ )(x̄ k− )
2 (6.60)
= (bk+ )−1 D k (H )ξ̄k + (qk+ )(x̄ k+ ) .
Theorem 6.22. Let Q(t) satisfy the linear differential equation (6.43) on [t0 , tf ] \ , and
let conditions (6.58) and (6.59) hold for each k = 1, . . . , s. Let ω0 be positive on K0 \ {0}.
Then is positive on K \ {0}.
272 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
Proof. By Proposition 6.21 and formulae (6.50), (6.51) the matrix Mk+ is positive semi-
definite for each k = 1, . . . , n. Now using Theorem 6.16 for s = 1 and Theorem 6.17 for
s ≥ 2, we obtain that is positive on K \ {0}.
Similar assertions hold for the jump conditions that use left-hand values of Q at each
point tk ∈ . Suppose that
Theorem 6.23. Let Q(t) satisfy the linear differential equation (6.43) on [t0 , tf ] \ , and
let conditions (6.61) and (6.62) hold for each k = 1, . . . , s. Let ω0 be positive on K0 \ {0}.
Then is positive on K \ {0}.
The final time tf is unspecified. The following values of the constants will be used in
computations:
In view of the integral cost criterion (6.64), we consider the Pontryagin function
(Hamiltonian) (cf. (6.17)) in normalized form taking α0 = 1,
u c
H (x1 , x2 , ψ1 , ψ2 , u) = u + ψ1 x2 + ψ2 − αg − x22 . (6.68)
mx2 m
The adjoint equations ψ̇ = −Hx are
u 2c
ψ̇1 = 0, ψ̇2 = −ψ1 + ψ2 + x2 . (6.69)
mx22 m
The transversality condition (6.11) evaluated at the free final time tf yields the additional
boundary condition
u(tf ) 9c
u(tf ) + 3ψ1 (tf ) + ψ2 (tf ) − αg − = 0. (6.70)
3m m
The switching function
ψ2
φ(x, ψ) = Du H = 1 + , φ(t) = φ(x(t), ψ(t)),
mx2
determines the control law
umin if φ(t) > 0
u(t) = .
umax if φ(t) < 0
Computations give evidence to the fact that the optimal control is bang-bang with one
switching time t1 ,
umin , 0 ≤ t < t1
u(t) = .
umax , t1 ≤ t ≤ tf
We compute an extremal using the code BNDSCO of Oberle and Grimm [82] or the code
NUDOCCCS of Büskens [13]. The solution is displayed in Figure 6.1. Results for the
switching time t1 , final time tf , and adjoint variables ψ(t) are
t1 = 3.924284, tf = 4.254074,
ψ1 (0) = −42.24170, ψ2 (0) = −3.876396,
x1 (t1 ) = 9.086464, x2 (t1 ) = 2.367329,
ψ1 (tf ) = −42.24170, ψ2 (tf ) = −17.31509.
We will show that this trajectory satisfies the assumptions of Proposition 6.7. The critical
cone is K = {0}, since the computed vectors
∂x
(tf ) = (−0.6326710, −0.7666666)∗ , ẋ(tf ) = (3.0, 0.7666666)∗
∂t1
are linearly independent. Moreover, we find in view of (6.14) that
D 1 (H ) = −φ̇(t1 ) [u]1 = 0.472397 · 40 > 0.
Theorem 6.10 then asserts that the computed bang-bang control provides a strict strong
minimum.
274 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
10
140
8
130
6
120
4
110
2
100
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
(c) adjoint variable 2
(d) switching function
-2
0.1
-4
-6 0
-8 -0.1
-10 -0.2
-12
-0.3
-14
-16 -0.4
-18 -0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Figure 6.1. Minimal fuel consumption of a car. (a) State variables x1 , x2 . (b) Bang-
bang control u. (c) Adjoint variable ψ2 . (d) Switching function φ.
Here, a0 and a1 are given points in Rd(x) , Q ⊂ R1+d(x) is an open set, and U ⊂ Rd(u) is a con-
vex polyhedron. The functions a, B are twice continuously differentiable on Q, with B being
a d(x) × d(u) matrix function. A trajectory or control process T = { (x(t), u(t)) | t ∈ [0, tf ] }
is said to be admissible if x(·) is absolutely continuous, u(·) is measurable and essen-
tially bounded, and the pair of functions (x(t), u(t)) satisfies the constraints (6.72)–(6.74)
6.5. Quadratic Conditions in Time-Optimal Bang-Bang Control 275
on the interval = [0, tf ]. Let us define the Pontryagin and the strong minimum in the
problem.
An admissible trajectory T is said to be a Pontryagin minimum if there is no sequence
of admissible trajectories T n = {(x n (t), un (t)) | t ∈ [0, tfn ]}, n = 1, 2, . . . , with
(a) tfn < tf for n = 1, 2, . . . ;
(b) tfn → tf for n → ∞;
n|x (t) − x(t)| → 0 for n → ∞, where = [0, tf ];
(c) max n n n n
values ψ(0) are assumed to belong to the unit ball of Rd(x) . The case that there exists
a multiplier (ψ0 , ψ) ∈ M0 with ψ0 (tf ) > 0 will be called the nondegenerate or normal
case. Again we use the simple abbreviation (t) for all arguments (t, x(t), u(t), ψ(t)), e.g.,
φ(t) = φ(t, x(t), ψ(t)).
Let us introduce the quantity D k (H ). For (ψ0 , ψ) ∈ M0 and tk ∈ consider the
function
(k H )(t) = H (t, x(t), uk+ , ψ(t)) − H (t, x(t), uk− , ψ(t)) = φ(t, x(t), ψ(t))[u]k .
D k (H ) ≥ 0 for k = 1, . . . , s. (6.81)
We need the definition of a strict bang-bang control (see Section 6.1.4) to obtain the suffi-
cient conditions in Theorem 6.27. For a given extremal trajectory T = { (x(t), u(t)) | t ∈ }
with piecewise constant control u(t) we say that u(t) is a strict bang-bang control if there
exists (ψ0 , ψ) ∈ M0 such that
where
[u(t−), u(t+)] = {αu(t−) + (1 − α)u(t+) | 0 ≤ α ≤ 1 }
denotes the line segment in Rd(u) . As it was mentioned already in Section 6.1.4, if U is
the unit cube in Rd(u) , condition (6.82) precludes simultaneous switching of the control
components. However, this property holds for all numerical examples in Chapter 8.
In order to formulate quadratic optimality condition for a given extremal T with bang-
bang control u(·) we shall introduce the space Z(), the critical subspace K ⊂ Z(), and
the quadratic form defined in Z().
Then K is a subspace of the space Z() which we call the critical subspace. Each element
z̄ ∈ K is uniquely defined by the number t¯f and the vector ξ̄ . Consequently, the subspace
K is finite-dimensional.
An explicit representation of the variations x̄(t) in (6.83) is obtained as in Section
6.2.1. For each k = 1, . . . , s, define the vector functions y k (t) as the solutions to the system
ẏ = fx (t)y, y(tk ) = [ẋ]k , t ∈ [tk , tf ]. (6.85)
For t < tk we put y k (t) = 0 which yields the jump [y k ]k = [ẋ]k . Then
s
x̄(t) = y k (t)ξ̄k (6.86)
k=1
Furthermore, denote by x(t; t1 , . . . , ts ) the solution of the state equation (6.72) using the
optimal bang-bang control with switching points t1 , . . . , ts . Then the partial derivatives of
state trajectories with respect to the switching points are given by
∂x
(t; t1 , . . . , ts ) = −y k (t) for t ≥ tk , k = 1, . . . , s. (6.88)
∂tk
This relation holds for all t ∈ [0, tf ]\{tk }, because for t < tk we have ∂t
∂x
k
(t) = 0 and y k (t) = 0.
Hence, (6.86) yields
s
∂x
x̄(t) = − (t)ξ̄k . (6.89)
∂tk
k=1
In the nondegenerate case ψ0 (tf ) > 0, the critical subspace is simplified as follows.
Proposition 6.24. If there exists (ψ0 , ψ) ∈ M0 such that ψ0 (tf ) > 0, then t¯f = 0 holds for
each z̄ = (t¯f , ξ̄ , x̄) ∈ K.
Now we discuss the case of two switching points, i.e., s = 2, to prepare the numerical
example in Section 6.5.4. Let us assume that ψ0 (tf ) > 0 (for some (ψ0 , ψ) ∈ M0 ) and
[ẋ]1 = 0, [ẋ]2 = 0. By virtue of Proposition 6.24, we have t¯f = 0 and hence x̄(tf ) = 0 for
each element z̄ ∈ K. Then the relations (6.84) and (6.86) yield
0 = x̄(tf ) = y 1 (tf )ξ̄1 + y 2 (tf )ξ̄2 . (6.90)
The conditions [ẋ]1 = 0 and [ẋ]2 = 0 imply that y 1 (tf ) = 0 and y 2 (tf ) = 0, respectively.
Furthermore, assume that K = {0}. Then (6.90) shows that the nonzero vectors y 1 (tf ) and
y 2 (tf ) are collinear, i.e.,
y 2 (tf ) = αy 1 (tf ) (6.91)
with some factor α = 0. As a consequence, the relation y 2 (t) = αy 1 (t) is valid for all
t ∈ (t2 , tf ]. In particular, we have y 2 (t2 +) = αy 1 (t2 ) and thus
Theorem 6.26. Let a trajectory T affords a Pontryagin minimum. Then the following
Condition A holds for the trajectory T : The set M0 is nonempty and
max (ψ0 , ψ, z̄) ≥ 0 ∀ z̄ ∈ K \ {0}.
(ψ0 ,ψ)∈M0
Theorem 6.27. Let the following Condition B be fulfilled for the trajectory T :
(a) there exists λ ∈ M0 such that D k (H ) > 0, k = 1, . . . , s, and condition (6.82) holds
(i.e., u(t) is a strict bang-bang control),
(b) max(ψ0 ,ψ)∈M0 (ψ0 , ψ, z̄) > 0 for all z̄ ∈ K \ {0}.
Then T is a strict strong minimum.
6.5. Quadratic Conditions in Time-Optimal Bang-Bang Control 279
Remarks.
1. The sufficient Condition B is a natural strengthening of the necessary Condition A.
2. Condition (b) is automatically fulfilled if K = {0} holds (cf. Proposition 6.25) which
gives a first-order sufficient condition for a strong minimum.
3. If there exists (ψ0 , ψ) ∈ M0 such that (ψ0 , ψ, z̄) > 0 for all z̄ ∈ K \ {0}, then condi-
tion (b) is obviously fulfilled.
4. For boxes U = {u = (u1 , . . . , ud(u) ) ∈ Rd(u) : umini ≤ ui ≤ umax
i , i = 1, . . . , d(u)}, the
condition D (H ) > 0, k = 1, . . . , s, is equivalent to the property φ̇i (tk ) = 0 if tk is a switch-
k
ing point of the ith control component ui (t). Note again that condition (6.82) precludes the
simultaneous switching of two or more control components.
5. A further remark concerns the case that the set M0 of Pontryagin multipliers is not a
singleton. This case was illustrated in [89] by the following time-optimal control problem
for a linear system:
where x = (x1 , x2 , x3 , x4 ). It was shown in [89] that for some a and b there exists an extremal
in this problem with two switching points of the control such that, under appropriate nor-
malization, the set M0 is a segment. For this extremal, the maximum of the quadratic forms
is positive on each nonzero element of the critical subspace, and hence the sufficient
conditions of Theorem 6.27 are satisfied. But this is not true for any single quadratic form
(corresponding to an element of the set M0 ).
˙ = fx (t)x̄(t),
x̄(t) [x̄]k = [ẋ]k ξ̄k (k = 1, . . . , s), x̄(0) = 0, x̄(tf ) = 0. (6.95)
In particular, these conditions imply x̄(t) ≡ 0 on [0, t1 ) and (ts , tf ]. Hence we have x̄ 1− =
x̄ s+ = 0 for all z̄ ∈ K. Then the quadratic form (6.94) is equal to
s
tf
(ψ, z̄) = D k (H )ξk2 + 2[Hx ]k x̄av
k
ξk + Hxx (t)x̄(t), x̄(t) dt. (6.96)
k=1 0
This case of a time-optimal (autonomous) control problem was studied by Sarychev [104].
He used a special transformation of the problem and obtained sufficient optimality condition
for the transformed problem. It is not easy, but it is possible, to reformulate his results in
terms of the original problem. The comparison of both types of conditions reveals that
Sarychev used the same critical subspace, but his quadratic form is a lower bound for .
Namely, in his quadratic form the positive term D k (H )ξ̄k2 has the factor 14 instead of the
factor 1 for the same term in . Therefore, the sufficient Condition B is always fulfilled
whenever Sarychev’s condition is fulfilled. However, there is an example of a control
problem where the optimal solution satisfies Condition B but does not satisfy Sarychev’s
280 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
condition. Finally, Sarychev proved that his condition is sufficient for an L1 -minimum with
respect to the control (which is a Pontryagin minimum in this problem). In fact, it could be
proved that his condition is sufficient for a strong minimum.
Theorem 6.28. Let the following conditions be fulfilled for the trajectory T :
(a) u(t) is a bang-bang control with one switching point, i.e., s = 1,
(b) there exists (ψ0 , ψ) ∈ M0 such that D 1 (H ) > 0 and condition (6.82) holds
(i.e., u(t) is a strict bang-bang control),
(c) there exists (ψ0 , ψ) ∈ M0 with ψ0 (tf ) > 0.
Then T is a strict strong minimum.
Now we turn our attention to the case of two switching points where s = 2. Assume
the nondegenerate case ψ0 (tf ) > 0 and suppose that [ẋ]1 = 0, [ẋ]1 = 0 and y 2 (tf ) = αy 1 (tf )
as in (6.91). Otherwise, K = {0} holds and, hence, the first-order sufficient condition for
a strong minimum is satisfied. For any element z̄ ∈ K, we have t¯f = 0, x̄ 1− = 0, x̄ 2+ = 0.
Consequently,
1 1 1 1 1
1
x̄av = [x̄]1 = [ẋ]1 ξ̄1 , 2
x̄av = x̄ 2− = y 1 (t2 )ξ̄1 = [ẋ]2 ξ̄1
2 2 2 2 2α
in view of x̄(t) = y 1 (t)ξ̄1 + y 2 (t)ξ̄2 , y 2 (t2 −) = 0 and (6.92). Using these relations in the
quadratic form (6.96) together with (6.93) and the conditions y 2 (t) = 0 for all t < t2 ,
[Hx ]k = −[ψ̇]k , k = 1, 2, we compute the quadratic form for the element of the critical
subspace as
t2
= D 1 (H )ξ̄12 + D 2 (H )ξ̄22 − 2[ψ̇]1 x̄av 1
ξ̄1 − 2[ψ̇]2 x̄av
2
ξ̄2 + Hxx x̄, x̄ dt
t1 t2
1 2 1
= D (H )ξ̄1 + 2 D (H )ξ̄1 − [ψ̇] [ẋ] ξ̄1 + 2 [ψ̇] [ẋ] ξ̄1 +
1 2 2 1 1 2 2 2 2
Hxx y , y dt ξ̄12
1 1
α α t1
= ρ ξ̄12 ,
where
1 t2
ρ = D (H ) − [ψ̇] [ẋ]
1 1 1
+ 2 D 2 (H ) + [ψ̇]2 [ẋ]2 + Hxx y 1 , y 1 dt. (6.97)
α t1
Proposition 6.29. Assume that ψ0 (tf ) > 0, s = 2, [ẋ]1 = 0, [ẋ]2 = 0, and y 2 (tf ) = αy 1 (tf )
(which is equivalent to (6.91)) with some factor α. Then the condition of the positive
definiteness of on K is equivalent to the inequality ρ > 0, where ρ is defined as in (6.97).
6.6. Sufficient Conditions for Time-Optimal Control Problems 281
6.6.1 Q-Transformation of on K
Here we shall use the same arguments as in Sections 5.3.2 and 6.3.1. Let Q(t) be a sym-
metric matrix on [t1 , ts ] with piecewise continuous entries which are absolutely continuous
on each interval of the set [t1 , ts ] \ . Therefore, Q may have a jump at each point tk ∈
including t1 , ts , and thus the symmetric matrices Q1− and Qs+ are also defined. For z̄ ∈ K,
we obviously have
ts t s +
d s
Qx̄, x̄ dt = Qx̄, x̄ − [Qx̄, x̄ ]k ,
t1 dt t1 − k=1
where [Qx̄, x̄ ]k is the jump of the function Qx̄, x̄ at the point tk ∈ . Using the conditions
x̄˙ = fx (t)x̄ and x̄ 1− = x̄ s+ = 0, we obtain
s ts
[Qx̄, x̄ ]k + (Q̇ + fx∗ Q + Qfx )x̄, x̄ dt = 0, (6.98)
k=1 t1
where the asterisk denotes transposition. Adding this zero form to , we get
s
= D k (H )ξ̄k2 − 2[ψ̇]k x̄av
k
ξ̄k + [Qx̄, x̄ ]k
k=1
ts
(6.99)
+ (Hxx + Q̇ + fx∗ Q + Qfx )x̄, x̄ dt.
t1
We call this formula the Q-transformation of on K.
To eliminate the integral term in , we assume that Q(t) satisfies the following linear
matrix differential equation:
Q̇ + fx∗ Q + Qfx + Hxx = 0 on [t1 , ts ] \ . (6.100)
Using (6.100), the quadratic form (6.99) reduces to
s
= ωk , ωk := D k (H )ξk2 − 2[ψ̇]k x̄av
k
ξk + [Qx̄, x̄ ]k . (6.101)
k=1
282 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
Proposition 6.30. Let Q(t) satisfy the linear differential equation (6.100) on [t1 , ts ] \ .
Then for each z̄ ∈ K the representation (6.101) holds.
Now our goal is to derive conditions such that ωk > 0 holds on K \ {0} for k = 1, . . . , s.
We shall use the representations of ωk given in Section 6.3.1. According to (6.50),
ωk = D k (H ) + (qk+ )[ẋ]k ξ̄k2 + 2(qk+ )x̄ k− ξ̄k + (x̄ k− )∗ [Q]k x̄ k− , (6.102)
where qk+ = ([ẋ]k )∗ Qk+ − [ψ̇]k . We immediately see from this representation that one
way to enforce ωk > 0 is to impose the following conditions:
In practice, however, it might be difficult to check this condition since it is necessary to satisfy
the d(x) equality constraints qk+ = ([ẋ]k )∗ Qk+ − [ψ̇]k = 0 and the inequality constraints
[Q]k ≥ 0. It is more convenient to express ωk as a quadratic form in the variables (ξ̄k , x̄ k− )
with the matrix
D k (H ) + (qk+ )[ẋ]k qk+
Mk+ = , (6.104)
(qk+ )∗ [Q]k
where qk+ is a row vector and (qk+ )∗ is a column vector.
Similarly, according to (6.52), the following representation holds:
ωk = D k (H ) − (qk− )[ẋ]k ξ̄k2 + 2(qk− )x̄ k+ ξ̄k + (x̄ k+ )∗ [Q]k x̄ k+ , (6.105)
where qk− = ([ẋ]k )∗ Qk− − [ψ̇]k . Again, we see that ωk > 0 holds if we require the condi-
tions
D k (H ) > 0, qk− = ([ẋ]k )∗ Qk− − [ψ̇]k = 0, [Q]k ≥ 0. (6.106)
To find a more general condition for ωk > 0, we consider (6.105) as a quadratic form in the
variables (ξ̄k , x̄ k+ ) with the matrix
D k (H ) − (qk− )[ẋ]k qk−
Mk− = . (6.107)
(qk− )∗ [Q]k
Since the right-hand sides of equalities (6.102) and (6.105) are connected by the relation
x̄ k+ = x̄ k− + [ẋ]k ξ̄k , the following statement obviously holds.
Proposition 6.31. For each k = 1, . . . , s, the positive (semi)definiteness of the matrix Mk−
is equivalent to the positive (semi)definiteness of the matrix Mk+ .
Theorem 6.32. Let Q(t) be a solution of the linear differential equation (6.100) on [t1 , ts ]\
which satisfies the following conditions:
6.6. Sufficient Conditions for Time-Optimal Control Problems 283
Proof. Take an arbitrary element z̄ = (ξ̄ , x̄) ∈ K. Let us show that ≥ 0 for this element.
Condition (a) implies that ωk ≥ 0 for k = 2, . . . , s. Condition (b) for k = 1 together with
condition x̄ 1− = 0 implies that ω1 ≥ 0. Consequently, ≥ 0. Assume that = 0. Then
ωk = 0, k = 1, . . . , s. The conditions ω1 = 0, x̄ 1− = 0, and b1+ > 0 by formula (6.102)
(with k = 1) yield ξ̄1 = 0. Then [x̄]1 = 0 and hence x̄ 1+ = 0. The last equality and
the equation x̄˙ = fx (t)x̄ show that x̄(t) = 0 in (t1 , t2 ) and hence x̄ 2− = 0. Similarly, the
conditions ω2 = 0, x̄ 2− = 0, and b2+ > 0 by formula (6.102) (with k = 2) imply that
ξ̄2 = 0 and x̄(t) = 0 in (t2 , t3 ). Therefore, x̄ 3− = 0 etc. Continuing this process we get
x̄ ≡ 0 and ξ̄k = 0 for k = 1, . . . , s − 1. Now using formula (6.101) for ωs = 0, as well
as the conditions D s (H ) > 0 and x̄ ≡ 0 we obtain ξ̄s = 0. Consequently, we have z̄ = 0
which means that is positive on K \ {0}.
Similarly, using representation (6.105) for ωk we can prove the following statement.
Theorem 6.33. Let Q(t) be a solution of the linear differential equation (6.100) on [t1 , ts ]\
which satisfies the following conditions:
(a) The matrix Mk− is positive semidefinite for each k = 1, . . . , s − 1;
(b) bk− := D k (H ) − (qk− )[ẋ]k > 0 for each k = 2, . . . , s.
Then is positive on K \ {0}.
where (qk− )∗ is a column vector while qk− is a row vector. Then according to (6.63),
2 2
ωk = (bk− )−1 (bk− )ξ̄k + (qk− )(x̄ k+ ) = (bk− )−1 D k (H )ξ̄k + (qk− )(x̄ k− ) . (6.110)
Theorem 6.34. Let Q(t) satisfy the linear differential equation (6.100) on [t1 , ts ] \ .
Let condition (6.108) hold for each k = 1, . . . , s and condition (6.109) hold for each k =
1, . . . , s − 1. Then is positive on K \ {0}.
Proof. According to (6.110), the matrix Mk− is positive semidefinite for each k = 1, . . . ,
s − 1 (cf. (6.105) and (6.107)), and hence both conditions (a) and (b) of Theorem 6.33 are
fulfilled. Then by Theorem 6.33, is positive on K \ {0}.
284 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
Similar assertions hold for the jump conditions that use right-hand values of Q at each
point tk ∈ . Suppose that
Then
2 2
ωk = (bk+ )−1 (bk+ )ξ̄k + (qk+ )(x̄ k− ) = (bk+ )−1 D k (H )ξ̄k + (qk+ )(x̄ k+ ) . (6.113)
Theorem 6.35. Let Q(t) satisfy the linear differential equation (6.100) on [t1 , ts ] \ . Let
condition (6.111) hold for each k = 1, . . . , s and condition (6.112) hold for each k = 2, . . . , s.
Then is positive on K \ {0}.
Theorem 6.36. Let Q(t) satisfy the linear differential equation (6.100) on (t1 , t2 ) and the
following inequalities hold at t1 , t2 :
and
b1+ := D 1 (H ) + q1 [ẋ]1 > 0, b2− := D 2 (H ) − q2 [ẋ]2 > 0. (6.115)
Define the jumps [Q]1 and [Q]2 by the conditions
Then Q1− and Q2+ are also symmetric matrices. Thus, we obtain a symmetric matrix
Q(t) satisfying (6.100) on (t1 , t2 ), the inequalities (6.115), and the jump conditions (6.116).
By formulas (6.110) and (6.113) the terms ω1 and ω2 are nonnegative. In view of (6.101),
6.6. Sufficient Conditions for Time-Optimal Control Problems 285
ψ̇ = −ψax ,
where
D k (H ) = ψ̇(tk )B[u]k , k = 1, . . . , s.
In the case of two switching points with s = 2, the conditions (6.114) take the form
(a) state variables x1 and x2 (b) time-optimal control u and switching function
1
1
0.8
0.6 0.5
0.4
0
0.2
0 -0.5
-0.2
-1
-0.4
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
(d) adjoint variables 1 and 2
(c) phaseportrait (x1, x2)
1 0.8
0.6
0.9 0.4
0.2
0.8
0
0.7 -0.2
-0.4
0.6 -0.6
-0.8
0.5
-1
0.4 -1.2
-0.4 -0.2 0 0.2 0.4 0.6 0 0.2 0.4 0.6 0.8 1 1.2
Figure 6.2. Time-optimal solution of the van der Pol oscillator, fixed terminal state
(6.120). (a) State variables x1 and x2 (dashed line). (b) Control u and switching function
ψ2 (dashed line). (c) Phase portrait (x1 , x2 ). (d) Adjoint variables ψ1 and ψ2 (dashed line).
Imposing this boundary condition, we aim at steering the van der Pol oscillator to a small
neighborhood of the origin. The adjoint equation (6.124) remains valid. The transversality
condition for the adjoint variable gives
The boundary condition (6.11) associated with the free final time tf yields
ψ1 (tf )x2 (tf ) + ψ2 (tf ) (−x1 (tf ) + x2 (tf )(1 − x1 (tf )2 ) + u(tf )) + 1 = 0. (6.131)
Again, the switching function (6.18) is given by φ(t) = Hu (t) = ψ2 (t). The structure of the
bang-bang control agrees with that in (6.132),
1 for 0 ≤ t < t1
u(t) = , (6.132)
−1 for t1 ≤ t ≤ tf
Using either the boundary value solver BNDSCO of Oberle and Grimm [82] or the direct
optimization routine NUDOCCCS of Büskens [13, 14], we obtain the extremal solution
depicted in Figure 6.3 and the following values for the switching, final time, state, and
adjoint variables:
t1 = 0.7139356, tf = 2.864192,
ψ1 (0) = 0.9890682, ψ2 (0) = 0.9945782,
x1 (t1 ) = 1.143759, x2 (t1 ) = −0.5687884,
ψ1 (t1 ) = 1.758128, ψ2 (t1 ) = 0.0, (6.134)
x1 (tf ) = 0.06985245, x2 (tf ) = −0.1874050,
ψ1 (tf ) = 0.4581826, ψ2 (tf ) = −1.229244,
β = 3.279646.
There are two alternative ways to check sufficient conditions. We may either use
Theorem 6.19 and solve the linear equation (6.43) or evaluate directly the quadratic form
in Proposition 6.11. We begin by testing the assumptions of Theorem 6.19 and consider the
symmetric 2 × 2 matrix
Q11 (t) Q12 (t)
Q(t) = .
Q12 (t) Q22 (t)
The linear equations Q̇ = −Qfx − fx∗ Q − Hxx in (6.100) yield the following ODEs:
In view of Theorem 6.19 we must find a solution Q(t) only in the interval [t1 , tf ] such that
Figure 6.3. Time-optimal solution of the van der Pol oscillator, nonlinear boundary
condition (6.129). (a) State variables x1 and x2 (dashed line). (b) Control u and switching
function ψ2 (dashed line). (c) Phase portrait (x1 , x2 ). (d) Adjoint variables ψ1 and ψ2
(dashed line).
holds and the quadratic form ω0 in (6.55)–(6.57) is positive definite on the cone K0 defined
in (6.26). Since ψ2 (t1 ) = 0 we get from (6.130),
which is satisfied by any initial value Q22 (t1 ) > −0.879064. By Proposition 6.8, we have
t¯f = 0 for every element z̄ = (t¯f , ξ , x̄) ∈ K. Therefore, by (6.57) we must check that the
matrix B22 = 2β I2 − Q(tf ) is positive definite on the critical cone K0 defined in (6.26),
i.e., on the cone
Thus the variations (v1 , v2 ) are related by v2 = −v1 x1 (tf )/x2 (tf ). Evaluating the quadratic
form (2β I2 − Q(tf ))(v1 , v2 ), (v1 , v2 ) with v2 = −v1 x1 (tf )/x2 (tf ), we arrive at the test
2 2
x1 x1 x1
c = 2β 1 + − Q11 − 2 Q12 + Q22 (tf ) > 0.
x2 x2 x2
290 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
A straightforward integration of the ODEs (6.135) using the solution data (6.134) and the
initial values Q11 (t1 ) = Q12 (t1 ) = Q22 (t1 ) = 0 gives the numerical results
which yield the positive value c = 7.593456 > 0. Thus Theorem 6.19 asserts that the bang-
bang control characterized by (6.134) provides a strict strong minimum.
The alternative test for second-order sufficient conditions (SSC) is based on Proposi-
tion 6.11. The variational system ẏ(t) = fx (t)y(t), y(t1 ) = [ẋ]1 , for the variation y = (y1 , y2 )
leads to the variational system
ẏ1 = y2 , y1 (t1 ) = 0,
ẏ2 = −(1 + 2x1 2x2 )y1 + (1 − x12 )y2 , y2 (t1 ) = 2,
Note that the relation Kxf (x(tf ))y(tf ) = 2(x1 (tf )y1 (tf ) + x2 (tf )y2 (tf ) = 0 holds. By Propo-
sition 6.11 we have to show that the quantity ρ in (6.37) is positive,
tf
ρ = D 1 (H ) − [ψ̇]1 [ẋ]1 + (y(t))∗ Hxx (t)y(t) dt + (y(tf ))∗ (βK)xf xf y(tf ) > 0.
t1
Using [ψ̇]1 = 0 and (y(tf ))∗ (βK)xf xf y(tf ) = 2β(y1 (tf )2 + y2 (tf )2 ), we finally obtain
ρ = D 1 (H ) + 184.550 > 0.
ẋ1 (t) = x2 (t), ẋ2 (t) = −x1 (t) + x2 (t)(1.4 − 0.14x2 (t)2 ) + u(t), (6.136)
x1 (0) = x2 (0) = −5, x1 (tf ) = x2 (tf ) = 0, (6.137)
| u(t) | ≤ 4 for t ∈ [0, tf ]. (6.138)
Note that we have shifted the factor 4 to the control variable in the dynamics (4.134) to
the control constraint (6.138). The Pontryagin function (Hamiltonian) (see (6.75)) for this
problem is
H (x, u, ψ) = ψ1 x2 + ψ2 (−x1 + x2 (1.4 − 0.14x22 ) + u). (6.139)
6.7. Numerical Examples of Time-Optimal Control Problems 291
As for the van der Pol oscillator, it is easy to show that there are no singular arcs with
ψ2 (t) ≡ 0 holding on a time interval Is ⊂ [0, tf ]. Hence, the optimal control is bang-bang.
Applying nonlinear programming methods to the suitably discretized Rayleigh problem,
one realizes that the optimal control comprises the following three bang-bang arcs with
two switching times t1 , t2 :
⎧ ⎫
⎨ 4 for 0 ≤ t < t1 ⎪
⎪ ⎬
u(t) = −4 for t1 ≤ t < t2 . (6.142)
⎩ 4 for t ≤ t ≤ t ⎪
⎪ ⎭
2 f
Thus, we have to solve the multipoint boundary value problem comprising equations
(6.136)–(6.143). The codes BNDSCO [82] and NUDOCCS [13, 14] yield the extremal
depicted in Figure 6.4. The final time, the switching points, and some values for the adjoint
variables are computed as
tf = 3.66817339,
t1 = 1.12050658, t2 = 3.31004698,
ψ1 (0) = −0.12234128, ψ2 (0) = −0.08265161, (6.144)
ψ1 (t1 ) = −0.21521225, ψ1 (t2 ) = 0.89199176,
ψ1 (tf ) = 0.84276186, ψ2 (tf ) = −0.25.
Now we are going to show that the computed control satisfies the assumptions of Theorem
6.19 and thus provides a strict strong minimum. Consider the symmetric 2 × 2 matrix
Q11 (t) Q12 (t)
Q(t) = ,
Q12 (t) Q22 (t)
The linear equation (6.100), dQ/dt = −Q fx − fx∗ Q − Hxx , leads to the following three
equations:
Q̇11 = 2Q12 ,
Q̇12 = −Q11 − Q12 (1.4 − 0.42x22 ) + Q22 , (6.145)
Q̇22 = −Q12 − Q22 (1.4 − 0.42x22 ) + 0.84ψ2 x2 .
We must find a solution of these equations satisfying the estimates (6.114) at the switching
times t1 and t2 . From
(a) state variables x1 and x2 (b) time-optimal control u and switching function (×4)
6
4
4 3
2 2
1
0
0
-2
-1
-4 -2
-6 -3
-4
-8
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
Figure 6.4. Time-optimal control of the Rayleigh equation. (a) State variables
x1 and x2 (dashed line). (b) Control u and switching function φ (dashed line). (c) Phase
portrait (x1 , x2 ). (d) Adjoint variables ψ1 and ψ2 (dashed line).
we get
Furthermore, we have [ẋ]k = [u]k (0, 1)∗ = (0, 8)∗ and thus obtain Qk [ẋ]k , [ẋ]k =
64 Q22 (tk ), for k = 1, 2. The next step is to find a solution for the equations (6.145) in
the interval [t1 , t2 ] that satisfies the inequalities
These conditions can be satisfied by choosing, e.g., the following initial values at the switch-
ing time t1 :
Q11 (t1 ) = 0, Q12 (t1 ) = 0.25, Q22 (t1 ) = 0.
Integration yields the value Q22 (t2 ) = −0.1677185 which shows that the estimates (6.146)
hold. Note that these estimates do not hold for the choice Q(t1 ) = 0, since this initial
value would give Q22 (t2 ) = 0.70592. In summary, Theorem 6.36 asserts that the computed
solution provides a strict strong minimum.
6.8. Time-Optimal Control Problems for Linear Systems with Constant Entries 293
Definition 6.37. We say that the admissible triple (tf , x, u) affords an almost global mini-
mum if there is no sequence of admissible triples (tf n , x n , un ), n = 1, 2, . . . , such that tfn < tf
for n = 1, 2, . . . , and tfn → tf .
Proposition 6.38. Suppose that, in the problem (6.147), there exists an u∗ ∈ U such that
Aa + Bu∗ = 0 or Ab + Bu∗ = 0
(for example, let b = 0, 0 ∈ U ). Then the almost global minimum is equivalent to the
global one.
Proof. If, for example, Ab + Bu∗ = 0, u∗ ∈ U , then any pair (x, u) admissible on [0, tf ] can
be extended to the right of tf by putting x(t) = b, u(t) = u∗ .
Let (tf , x, u) be an admissible triple for which the conditions of the minimum principle
are fulfilled: There exists a smooth function ψ : [0, tf ] → Rd(x) such that
Theorem 6.39 (Milyutin). Suppose there exists ψ ∈ M0 such that α0 := −ψ ẋ > 0. Then
(tf , x, u) affords an almost global minimum.
294 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
In what follows, we assume that (tf , x, u) is an extremal triple such that u(t) is a
piecewise constant function taking values in the vertices of the polyhedron U . We denote
by = {t1 , . . . , ts } the set of discontinuity points of the control u(t). Let ψ ∈ M0 . An
important role in formulations of optimality conditions in problem (6.147) is played by the
product (ψ̇ ẋ)(t).
Proposition 6.40. The product (ψ̇ ẋ)(t) is a monotone nonincreasing step function with
discontinuities only at the discontinuity points of the control u(t).
We now formulate yet another simple sufficient condition for the almost global min-
imum, obtained in [79, Part 2].
Theorem 6.41. Suppose that there exists ψ ∈ M0 such that the product (ψ̇ ẋ) fulfills one of
the following two conditions: (ψ̇ ẋ)(0) < 0, (ψ̇ ẋ)(tf ) > 0; i.e., (ψ̇ ẋ) strictly retains its sign
on [0, tf ]. Then (tf , x, u) affords an almost global minimum.
Corollary 6.42. Let (tf , x, u) be an extremal triple such that for any point tk ∈ the vectors
are different from zero and equally directed (so that ẋ k− = ck ẋ k+ for some ck > 0). Suppose
that there exists ψ ∈ M0 such that (ψ̇ ẋ) is not identically zero. Then (tf , x, u) affords an
almost global minimum.
It follows from the condition Ht = 0 that ψ̇0 = 0. Further, the condition that (ψ̇ ẋ) is
constant on (ts , tf ) implies (ψ̇ ẋ)(tf ) = (ψ̇ ẋ)(ts +). Thus according to (6.94),
s
(ψ, z̄) = − [(ψ̇ ẋ)]k ξ̄k2 + t¯f2 (ψ̇ ẋ)s+ . (6.149)
k=1
This formula holds for any ψ ∈ M0 and any z̄ = (t¯f , ξ̄ , x̄) which belongs to the subspace K
as in (6.83) and (6.84).
6.8. Time-Optimal Control Problems for Linear Systems with Constant Entries 295
Let us see what is the form of K in the present case. Since x̄(t) satisfies the linear
system x̄˙ = Ax̄ on each interval \ , condition x̄(0) = 0 in (6.84) can be replaced by
x̄(t1 −) = 0. Since ẋ satisfies the same system, ẍ = Aẋ, condition x̄(tf ) + ẋ(tf )t¯f = 0 in
(6.84) can be replaced by x̄(ts +) + ẋ(ts +)t¯f = 0. For brevity, put
For ψ ∈ M0 , ξ̄ ∈ Rs , let
s
Q(ψ, ξ̄ ) = − [ψ̇ ẋ]k ξ̄k2 + (ψ̇ ẋ)s+ ,
k=1
and set
Q0 (ξ̄ ) = max Q(ψ, ξ̄ ). (6.150)
ψ∈M0
Then Theorem 6.26 implies the following theorem.
Theorem 6.43. Let a triple (tf , x, u) afford a Pontryagin minimum in the problem (6.147).
Then the set M0 as in (6.148) is nonempty and
Q0 (ξ̄ ) ≥ 0 ∀ (ξ̄ , x̄) ∈ R. (6.151)
It is clear that the set R, in this necessary condition, can be replaced by its projection
under the mapping (ξ̄ , x̄) → ξ̄ . Denote this projection by % and find out what conditions
specify it. Conditions x̄˙ = Ax̄ and ẍ = Aẋ imply
x̄(t) = eAt c̄(t), ẋ(t) = eAt c(t),
296 Chapter 6. Second-Order Optimality Conditions for Bang-Bang Control
where c̄(t) and c(t) are step functions whose discontinuity points are contained in . Con-
ditions
x̄ 1− = 0, x̄ s+ = ẋ s+ , [x̄]k = [ẋ]k ξ̄k , k = 1, . . . , s,
imply
c̄1− = 0, c̄s+ = cs+ , [c̄]k = [c]k ξ̄k , k = 1, . . . , s.
Therefore
s
s
cs+ = c̄s+ = [c̄]k = [c]k ξ̄k .
k=1 k=1
s
[c]k ξ̄k = cs+
k=1
determine the projection of R under the mapping (ξ̄ , x̄) → ξ̄ , which we denote by %. Since
s
e−Atk [ẋ]k ξ̄k = e−Ats ẋ s+ ,
k=1
which after multiplication by eAts from the left takes the final form
s
eA(ts −tk ) [ẋ]k ξ̄k = ẋ s+ . (6.152)
k=1
Hence % is the set of vectors ξ̄ ∈ Rs satisfying the system of algebraic equations (6.152).
Thus Theorem 6.43 implies the following theorem.
Theorem 6.44. Let a triple (tf , x, u) afford a Pontryagin minimum in the problem (6.147).
Then the set M0 as in (6.148) is nonempty and
Q0 (ξ̄ ) ≥ 0 ∀ ξ̄ ∈ %. (6.153)
It makes sense to use necessary condition (6.153) for investigation of only those
extremals (tf , x, u) for which
α0 := −ψ ẋ = 0 ∀ ψ ∈ M0 , (6.154)
because otherwise, by Theorem 6.39, (tf , x, u) affords an almost global minimum in the
problem. Condition (6.154) guarantees that the set % is nonempty (see Theorem 13.7 in
[79, p. 322]). Theorem 6.44 implies a simple consequence of a geometric nature.
6.8. Time-Optimal Control Problems for Linear Systems with Constant Entries 297
Corollary 6.45. Suppose that (tf , x, u) affords a Pontryagin minimum in the problem
(6.147). Let the vectors ẋ k− and ẋ k+ be different from zero and collinear for some tk ∈ ,
and let the jump [ψ̇ ẋ]k of the product ψ̇ ẋ at the point tk be different from zero for any
ψ ∈ M0 . Then the vectors ẋ k− and ẋ k+ are equally directed.
Theorem 6.46. For an admissible triple (tf , x, u) satisfying the assumptions of Section
6.8.1, let the set M be nonempty and
Then (tf , x, u) affords a strict almost global minimum in the problem (6.147).
for all z̄ ∈ K such that t¯f = 0. Consider an element z̄ ∈ K \ {0} such that t¯f = 0. For this
element, ξ̄ = 0, since otherwise z̄ = 0. Take an arbitrary element ψ ∈ M and put q = −ψ̇ ẋ.
Then [q]k > 0, k = 1, . . . , s. Hence
s
(λ, z̄) = [q]k ξ̄k2 > 0.
k=1
Thus the inequality (6.158) holds for all z̄ ∈ K \ {0}. Therefore by Theorem 6.27, (tf , x, u)
affords a strict strong minimum in the problem (6.147). Moreover, condition M = ∅ in the
problem (6.147) implies that the strict strong minimum is equivalent to the strict almost
global minimum. The last assertion is a consequence of Proposition 13.4 and Lemma 13.1
in [79].
6.8.3 Example
Consider the problem
where x = (x1 , x2 ) ∈ R2 , u ∈ R1 . The minimum principle conditions for this problem are
as follows:
ψ̇1 = 0, ψ̇2 = −ψ1 , u = − sgn ψ2 ,
(6.160)
ψ1 x2 + ψ2 u = −α0 ≤ 0, |ψ(0)| = 1,
where ψ = (ψ1 , ψ2 ). This implies that ψ2 (t) is a linear function, and u(t) is a step function
having at most one switching and taking values ±1.
The system corresponding to u = 1 is ẋ1 = x2 , ẋ2 = 1, and omitting t we have
1
x1 = x22 + C. (6.161)
2
The system corresponding to u = −1 is ẋ1 = x2 , ẋ2 = −1, whence
1
x1 = − x22 + C. (6.162)
2
Through each point x = (x1 , x2 ) of the state plane there pass one curve of the family (6.161)
and one curve of the family (6.162). The condition ψ1 x2 + ψ2 u = −α0 ≤ 0 implies that the
passage from a curve of the family (6.161) to a curve of the family (6.162) is only possible
in the upper half-plane x2 ≥ 0, and the passage from a curve of the family (6.162) to a curve
of the family (6.161) in the lower half-plane x2 ≤ 0. The direction of movement is such that
along the curves of the family (6.161) the point can go to (+∞, +∞), and along the curves
of the family (6.162) to (−∞, −∞). This means that any two points can be joined by the
extremals in no more than two ways. By Theorem 6.39 each extremal with
According to Corollary 6.45 this extremal does not yield a Pontryagin minimum because
it is necessary for Pontryagin minimum that the collinear vectors ẋ 1− and ẋ 1+ are equally
directed. In this case the necessary second-order condition fails.
Thus in this problem an extremal with a switching affords an almost global minimum
iff α0 > 0 for this extremal; i.e., the point x = (x1 , x2 ) at the switching time does not lie on
the x1 axis in the state plane.
For an extremal without switchings there always exists ψ ∈ M0 such that α0 > 0.
Indeed, for this extremal one can set ψ1 = 0, ψ2 = −u. Then α0 = −ψ2 u = 1, and all
conditions of the minimum principle are fulfilled.
More interesting examples of investigation of extremals in time-optimal control prob-
lems for linear systems with constant entries are given in [79, Part 2, Section 14]. The time-
optimal control in a simplified model of a container crane (ore unloader) was discussed
in [63, Section 5]. The optimality of the time-optimal control with three switching times
follows from Theorem 6.39.
Chapter 7
We continue our investigation of the pure bang-bang case. As it was mentioned in introduc-
tion, second-order sufficient optimality conditions for bang-bang controls had been derived
in the literature in two different forms. The first form was discussed in the last chapter.
The second one is due to Agrachev, Stefani, and Zezza [1], who first reduced the bang-bang
control problem to a finite-dimensional optimization problem and then showed that well-
known sufficient optimality conditions for this optimization problem supplemented by the
strict bang-bang property furnish sufficient conditions for the bang-bang control problem.
The bang-bang control problem, considered in this chapter, is more general than that in [1].
Following [99, 100], we establish the equivalence of both forms of second-order conditions
for this problem.
where tˆs+1 = tˆf . Thus, ˆ = {tˆ1 , . . . , tˆs } is the set of switching points of the control û(·) with
tˆk < tˆk+1 for k = 0, 1, . . . , s. Assume now that the set M0 of multipliers is nonempty for the
trajectory Tˆ . Put
ζ = (t0 , tf , x0 , θ ) ∈ V,
299
300 Chapter 7. Bang-Bang Control Problem and Its Induced Optimization Problem
where θ = (t1 , . . . , ts ) satisfies t0 < t1 < t2 < · · · < ts < tf . Define the function u(t; θ ) by
the condition
u(t; θ ) = uk for t ∈ (tk−1 , tk ), k = 1, . . . , s + 1, (7.2)
where ts+1 = tf . The values u(tk ; θ), k = 1, . . . , s, may be chosen in U arbitrarily. For
definiteness, define them by the condition of continuity of the control from the left: u(tk ; θ) =
u(tk −; θ ), k = 1, . . . , s.
Let x(t; t0 , x0 , θ) be the solution of the initial value problem (IVP),
For each ζ ∈ V this solution exists if the neighborhood V of the point ζ̂ is sufficiently
small. We obviously have
We call (7.4) the Induced Optimization Problem (IOP) or simply Induced Problem, which
represents an extension of the IOP introduced in Agrachev, Stefani, and Zezza [1]. The
following assertion is almost obvious.
Theorem 7.1. Let the trajectory Tˆ be a Pontryagin local minimum for the basic control
problem (6.1)–(6.3). Then the point ζ̂ is a local minimum for the IOP (7.4), and hence it
satisfies first- and second-order necessary conditions for this problem.
Proof. Assume that ζ̂ is not a local minimum in problem (7.4). Then there exists a sequence
of admissible points ζ ν = (t0ν , t1ν , x0ν , θ ν ) in problem (7.4) such that ζ ν → ζ̂ for ν → ∞ and
F0 (ζ ν ) < F0 (ζ̂ ). Take the corresponding sequence of admissible trajectories
in problem (6.1)–(6.3). Then the conditions t0ν → tˆ0 , tfν → tˆf , x0ν → x̂0 , θ ν → θ̂ imply that
|u(t; θ ν ) − û(t)| dt → 0, max |x(t; t0ν , x0ν , θ ν ) − x̂(t)| → 0,
ˆ
ν ∩ ˆ
ν ∩
where ν = [t0ν , tfν ]. Moreover, J(T ν ) = F0 (ζ ν ) < F0 (ζ̂ ) = J(Tˆ ). It means that the trajec-
tory Tˆ is not a Pontryagin local minimum for the basic problem (6.1)–(6.3).
We shall clarify a relationship between the second-order conditions for the IOP (7.4) at
the point ζ̂ and those in the basic bang-bang control problem (6.1)–(6.3) for the trajectory Tˆ .
7.1. Main Results 301
Theorem 7.2. Let x̂ be a local minimum in problem (7.5). Then, at this point, the set 0 is
nonempty and the following inequality holds:
max Lxx (μ, x̂)x̄, x̄ ≥ 0 ∀ x̄ ∈ K0 .
μ∈0
Theorem 7.3. Let the set 0 be nonempty at the point x̂ and let
max Lxx (μ, x̂)x̄, x̄ > 0 ∀ x̄ ∈ K0 \ {0}.
μ∈0
These conditions were obtained by Levitin, Milyutin, and Osmolovskii [54, 55]; cf.
also Ben-Tal and Zowe [4].
Lagrange multipliers. Let us define the set 0 ⊂ R1+d(F )+d(K) of the triples μ =
(α0 , α, β) of normalized Lagrange multipliers at the point ζ̂ for the Induced Problem. The
Lagrange function for the Induced Problem is
L(μ, ζ ) = L(μ, t0 , tf , x0 , θ ) = α0 J (t0 , x0 , tf , x(tf ; t0 , x0 , θ ))
+ αF (t0 , x0 , tf , x(tf ; t0 , x0 , θ )) + βK(t0 , x0 , tf , x(tf ; t0 , x0 , θ ))
= l(μ, t0 , x0 , tf , x(tf ; t0 , x0 , θ )), (7.6)
where l = α0 J +αF +βK. By definition, 0 is the set of multipliers μ = (α0 , α, β) such that
where p̂ = (tˆ0 , x̂0 , tˆf , x̂f ), x̂0 = x̂(tˆ0 ), x̂f = x̂(tˆf ) = x(tˆf ; tˆ0 , x̂0 , θ̂ ). Now, let us define the
corresponding set of normalized Lagrange multipliers for the trajectory Tˆ in the basic
problem. Denote by the set of multipliers λ = (α0 , α, β, ψ, ψ0 ) such that
α0 ≥ 0, α ≥ 0, α0 + |α| + |β| = 1, αF (p̂) = 0,
−ψ̇(t) = ψ(t)fx (t, x̂(t), û(t)), −ψ̇0 (t) = ψ(t)ft (t, x̂(t), û(t)),
ψ(tˆ0 ) = −lx0 (μ, p̂), ψ(tˆf ) = lxf (μ, p̂), (7.8)
ψ0 (tˆ0 ) = −lt0 (μ, p̂), ψ0 (tˆf ) = ltf (μ, p̂),
ψ(t)f (t, x̂(t), û(t)) + ψ0 (t) = 0 ∀ t ∈ ˆ \ ,
ˆ
−ψ̇ = ψfx (t, x̂(t), û(t)), ψ(tˆf ) = lxf (μ, p̂) (7.10)
defines ψ(t). Define ψ0 (t) by the equality
ψ(t)f (t, x̂(t), û(t)) + ψ0 (t) = 0. (7.11)
We note that M0 ⊂ holds, because the system of conditions (6.7)–(6.9) and (6.11)
is equivalent to system (7.8). But it may happen that M0 = , since in the definition of
there is no requirement that its elements satisfy minimum condition (6.10). Let us denote
MP
0 := π0 (M0 ), where MP stands for Minimum Principle.
7.1. Main Results 303
Critical cones. We denote by K0 the critical cone at the point ζ̂ in the Induced
Problem. Thus, K0 is the set of collections ζ̄ = (t¯0 , t¯f , x̄0 , θ̄ ) such that
where I is the set of indices of the inequality constraints active at the point ζ̂ . Let K
be the critical cone for the trajectory Tˆ in the basic problem, i.e., the set of all tuples
ˆ satisfying conditions (6.21)–(6.23).
z̄ = (t¯0 , t¯f , ξ̄ , x̄) ∈ Z(),
Proposition 7.6. The operator π1 : (t¯0 , t¯f , ξ̄ , x̄) → (t¯0 , t¯f , x̄0 , θ̄ ) defined by
is a one-to-one mapping of the critical cone K (for the trajectory Tˆ in the basic problem)
onto the critical cone K0 (at the point ζ̂ in the Induced Problem).
We say that the elements ζ̄ = (t¯0 , t¯f , x̄0 , θ̄ ) ∈ K0 and z̄ = (t¯0 , t¯f , ξ̄ , x̄) ∈ K correspond
to each other if θ̄ = −ξ̄ and x̄0 = x̄(tˆ0 ), i.e., π1 (t¯0 , t¯f , ξ̄ , x̄) = (t¯0 , t¯f , x̄0 , θ̄ ).
Now we give explicit formulas for the inverse mapping for π1 . Let V (t) be an n × n
matrix-valued function (n = d(x)) which is absolutely continuous in ˆ = [tˆ0 , tˆf ] and satisfies
the system
V̇ (t) = fx (t, x̂(t), û(t))V (t), V (tˆ0 ) = E, (7.15)
where E is the identity matrix. For each k = 1, . . . , s denote by y k (t) the n-dimensional
vector function which is equal to zero in [tˆ0 , tˆk ), and in [tˆk , tˆf ] it is the solution to the IVP
Proposition 7.7. The inverse mapping π1−1 : (t¯0 , t¯f , x̄0 , θ̄ ) ∈ K0 → (t¯0 , t¯f , ξ̄ , x̄) ∈ K is given
by the formulas
s
ξ̄ = −θ̄, ˙ tˆ0 )t¯0 +
x̄(t) = V (t) x̄0 − x̂( y k (t)t¯k , (7.17)
k=1
μ = (α0 , α, β) ∈ MP
0 and λ = (α0 , α, β, ψ, ψ0 ) ∈ M0
correspond to each other, i.e., π0 λ = μ, and let the elements of the critical cones ζ̄ =
(t¯0 , t¯f , x̄0 , θ̄) ∈ K0 and z̄ = (t¯0 , t¯f , ξ̄ , x̄) ∈ K correspond to each other, i.e., π1 z̄ = ζ̄ . Then the
quadratic forms in the basic and induced problems take equal values: Lζ ζ (μ, ζ̂ )ζ̄ , ζ̄ =
(λ, z̄). Consequently,
for each pair of elements of the critical cones ζ̄ ∈ K0 and z̄ ∈ K such that π1 z̄ = ζ̄ .
Theorems 6.9 and 7.8 and Proposition 7.6 imply the following second-order necessary
optimality condition for the basic problem.
Theorem 7.9. If the trajectory Tˆ affords a Pontryagin minimum in the basic problem, then
the following Condition A0 holds: The set M0 is nonempty and
Theorems 6.10 and 7.8 and Proposition 7.6 imply the following second-order sufficient
optimality condition for the basic control problem.
Theorem 7.10. Let the following Condition B0 be fulfilled for an admissible trajectory Tˆ
in the basic problem:
(a) û(t) is a bang-bang control taking values in the set V = ex U ,
(b) the set M0 is nonempty, and there exists λ ∈ M0 such that D k (H ) > 0, k = 1, . . . , s,
and condition (6.19) holds (hence, u(t) is a strict bang-bang control),
(c) maxμ∈MP Lζ ζ (μ, ζ̂ )ζ̄ , ζ̄ > 0 for all ζ̄ ∈ K0 \ {0}.
0
Then Tˆ is a strict strong minimum.
Under our assumptions, the operator g : V → Rn is well defined and C 2 -smooth if the
neighborhood V of the point ζ̂ is sufficiently small. In this section, we shall derive the first-
order partial derivatives of g(t0 , tf , x0 , θ ) with respect to t0 , tf , x0 , and θ at the point ζ̂ . We
shall use well-known results from the theory of ODEs about differentiation of solutions to
ODEs with respect to parameters and initial values. In what follows, it will be convenient
to drop those arguments in x(t; t0 , x0 , θ ), u(t, θ ), V (t; t0 , x0 , θ ), etc., that are kept fixed.
Consequently, we have
where V (t) satisfies IVP (7.21) along the trajectory (x̂(t), û(t)), t ∈ [tˆ0 , tˆf ].
Proposition 7.12. The vector function w(t; t0 ) is the solution to the IVP
Hence, we obtain
∂x(tˆf ; tˆ0 , x̂0 , θ̂ ) ˙ tˆ0 ).
gt0 (ζ̂ ) := = −V (tˆf )x̂( (7.24)
∂t0
Obviously, we have
∂x(tˆf ; tˆ0 , x̂0 , θ̂ ) ˙
gtf (ζ̂ ) := = x̂(tˆf ). (7.25)
∂tf
Proposition 7.13. For t ≥ tk the function y k (t; tk ) is the solution to the IVP
ẏ k = fx (t, x(t; tk ), u(t; tk )) y k , y k |t=tk = −[f ]k , (7.26)
where [f ]k = f (tk , x(tk ; tk ), uk+1 )−f (tk , x(tk ; tk ), uk ) is the jump of the function f (t, x(t; tk ),
u(t; tk )) at the point tk . For t < tk , we have y k (t; tk ) = 0. Thus, [y]k = −[f ]k , where
[y]k = y(tk +; tk ) − y(tk −; tk ) is the jump of the function y k (t; tk ) at the point tk .
Proof. Let us sketch how to obtain the representation (7.26). For t ≥ tk the trajectory x(t; tk )
satisfies the integral equation
t
x(t; tk ) = x(tk −; tk ) + f (h, x(h; tk ), u(h, tk )) dh.
tk +
from which we get y k |t=tk = −[f ]k and the variational equation in (7.26).
In particular, we obtain
ψ0 (t) is absolutely continuous and the collection λ = (α0 , α, β, ψ, ψ0 ) satisfies all conditions
in (7.8) and hence belongs to the set . The conditions
in the definitions of 0 and are identical. Hence, we must analyze the equation
Lζ (μ, ζ̂ ) = 0 in the definition of 0 which is equivalent to the system
Using the equality lxf (p̂) = ψ(tˆf ) and formulas (7.24), (7.25), (7.22), (7.27) for the deriva-
tives of g with respect to t0 , tf , x0 , tk , at the point ζ̂ , we get
˙ tˆ0 ) = 0,
Lt0 (μ, ζ̂ ) = lt0 (p̂) − ψ(tˆf )V (tˆf )x̂( (7.28)
˙ tˆf ) = 0,
Lt (μ, ζ̂ ) = lt (p̂) + ψ(tˆf )x̂( (7.29)
f f
V̇ = fx V , V (tˆ0 ) = E
with fx = fx (t, x̂(t), û(t)). Then, "(t) := V −1 (t) is the solution to the adjoint equation
˙ = "fx ,
−" "(tˆ0 ) = E.
Consequently, ψ(tˆf ) = ψ(tˆ0 )"(tˆf ) = ψ(tˆ0 )V −1 (tˆf ). Using these relations in (7.28), we get
˙ tˆ0 ) = 0.
lt0 (p̂) − ψ(tˆ0 )x̂(
˙ tˆ0 ) = −ψ0 (tˆ0 ). Hence, (7.28) is equivalent to the
By virtue of (7.11), we have ψ(tˆ0 )x̂(
transversality condition for ψ0 at the point tˆ0 :
Analysis of (7.30). Since ψ(tˆf ) = ψ(tˆ0 )V −1 (tˆf ), equality (7.30) is equivalent to the
transversality condition for ψ at the point tˆ0 :
Proposition 7.14. Let the absolutely continuous function y be a solution to the system
ẏ = fx y on an interval , and let the absolutely continuous function ψ be a solution to
the adjoint system −ψ̇ = ψfx on the same interval, where fx = fx (t, x̂(t), û(t)). Then
ψ(t)y(t) ≡ const on .
˙ k = −[ψ x̂]
ψ(tˆf )y k (tˆf ) = ψ(tˆk )y k (tˆk + 0) = ψ(tˆk )[y k ]k = −ψ(tˆk )[x̂] ˙ k = [ψ0 ]k .
[ψ0 ]k = 0, k = 1, . . . , s,
which means that ψ0 is continuous at each point tˆk , k = 1, . . . , s, and hence absolutely
˙ k that
ˆ = [tˆ0 , tˆf ]. Moreover, it follows from 0 = [ψ0 ]k = −ψ(tˆk )[x̂]
continuous on
where φ(t) = ψ(t)B(t, x̂(t)) denotes the switching function. Finally, differentiating (7.11)
with respect to t, we get
s
ξ̄ = −θ̄ , ˙ tˆ0 )t¯0 +
x̄(t) = V (t) x̄0 − x̂( y k (t)t¯k ,
k=1
where t¯k is the kth component of the vector θ̄ , and put z̄ = (t¯0 , t¯f , ξ̄ , x̄). We shall show
that z̄ is an element of the critical cone K (Equations (6.22) and (6.23)) for the trajectory
7.2. First-Order Derivatives of x(tf ; t0 , x0 , θ ) 309
Tˆ = {(x̂(t), û(t) | t ∈ [tˆ0 , tˆf ] } in the basic problem. Consider the first inequality F0
(ζ̂ )ζ̄ ≤ 0,
where F0 (ζ ) := J (t0 , x0 , tf , x(tf ; t0 , x0 , θ )). We obviously have
F0
(ζ̂ )ζ̄ = (Jt0 (p̂) + Jxf (p̂)gt0 (ζ̂ ))t¯0 + (Jtf (p̂) + Jxf (p̂)gtf (ζ̂ ))t¯f
s
+ (Jx0 (p̂) + Jxf (p̂)gx0 (ζ̂ ))x̄0 + Jxf (p̂)gtk (ζ̂ )θ̄k ,
k=1
where θ̄k = t¯k is the kth component of the vector θ̄. Using formulas (7.24), (7.25), (7.22),
(7.27) for the derivatives of g with respect to t0 , tf , x0 , tk , at the point ζ̂ , we get
˙ tˆ0 ))t¯0 + (Jt (p̂) + Jx (p̂)x̂(
F0
(ζ̂ )ζ̄ = (Jt0 (p̂) − Jxf (p̂)V (tf )x̂( ˙ tˆf ))t¯f
f f
s
+ (Jx0 (p̂) + Jxf (p̂)V (tˆf ))x̄0 + Jxf (p̂)y k (tˆf )θ̄k .
k=1
since V (tˆ0 ) = E, and y k (tˆ0 ) = 0, k = 1, . . . , s. Moreover, using the same definition, we get
s
˙ tˆf )t¯f = V (tˆf )(x̄0 − x̂(
x̄¯f := x̄(tˆf ) + x̂( ˙ tˆ0 )t¯0 ) + ˙ tˆf )t¯f .
y k (tˆf )t¯k + x̂( (7.34)
k=1
or briefly,
J
(p̂)p̄¯ ≤ 0,
where p̄¯ = (t¯0 , x̄¯0 , t¯f , x̄¯f ); see equation (6.21).
Similarly, the inequalities Fi
(ζ̂ )ζ̄ ≤ 0 for all i ∈ I and the equality G
(ζ̂ )ζ̄ = 0 in the
definition of K0 are equivalent to the inequalities (respectively, equalities)
Fi (p̂)p̄¯ ≤ 0, i ∈ I, K (p̂)p̄¯ = 0,
Finally, recall from (7.26) that for each k = 1, . . . , s the function y k (t) is piecewise
continuous with only one jump [y k ]k = −[x̂] ˙ k at the point tˆk and is absolutely continuous on
each of the half-open intervals [tˆ0 , tˆk ) and (tˆk , tˆf ]. Moreover, the function V (t) is absolutely
continuous in [tˆ0 , tˆf ]. Hence, x̄(t) is a piecewise continuous function which is absolutely
continuous on each interval of the set [tˆ0 , tˆf ] \ ˆ and satisfies the jump conditions
˙ k ξ̄k ,
[x̄]k = [x̂] ξ̄k = −t¯k , k = 1, . . . , s.
Thus, we have proved that z̄ = (t¯0 , t¯f , ξ̄ , x̄) is an element of the critical cone K. Similarly,
one can show that if z̄ = (t¯0 , t¯f , ξ̄ , x̄) ∈ K, then putting x̄0 = x̄(tˆ0 ) and θ̄ = −ξ̄ , we obtain
the element ζ̄ = (t¯0 , t¯f , x̄0 , θ̄) of the critical cone K0 .
at the point ζ̂ . We are not sure whether all of them are known; therefore we shall also sketch
the proofs. Here x(t; t0 , x0 , θ) is the solution to IVP (7.3). Denote by gk (ζ ) := xk (tf ; t0 , x0 , θ )
the kth component of the function g.
∂W k
where Ẇ k = ∂t , O is the zero matrix, fk is the kth component of the vector function f , and
Proof. For notational convenience, we use the function ϕ(t, x) := f (t, x, u(t)). By Propo-
0) ∂xi (t;x0 )
sition 7.11, the matrix-valued function V (t; x0 ) = ∂x(t;x
∂x0 with entries vij (t; x0 ) = ∂x0j is
7.3. Second-Order Derivatives of x(tf ; t0 , x0 , θ ) with Respect to t0 , tf , x0 , and θ 311
the solution to the IVP (7.18). Consequently, its entries satisfy the equations
∂ ẋk (t; x0 ) ∂xr (t; x0 )
= ϕkxr (t, x(t; x0 ))
∂x0i r
∂x0i
∂xk (t0 ; x0 )
= eki , k, i = 1, . . . , n,
∂x0i
where eki are the elements of the identity matrix E. By differentiating these equations with
respect to x0j , we get
where (A)ij denotes the element aij of a matrix A, and A∗ denotes the transposed matrix.
Thus, (7.36) and (7.37) imply (7.35).
Proposition 7.16. For t ≥ θ, the function R(t; x0 , θ ) is the solution to the IVP
Ṙ = (y ∗fxx )V + fx R, R(θ; x0 , θ ) = −[fx ]V (θ ; x0 , θ ), (7.39)
where fx and fxx are taken along the trajectory (t, x(t; x0 , θ ), u(t, θ )), t ∈ [t0 , tf ]. Here, by
definition, (y ∗fxx ) is an n × n matrix with entries
n
∂ 2 fk
∗
(y fxx )k,j = yi (7.40)
∂xi ∂xj
i=1
Proof. According to Proposition 7.11 the matrix-valued function V is the solution to the
system
V̇ (t; x0 , θ) = fx (t, x(t; x0 , θ ), u(t; θ))V (t; x0 , θ ). (7.41)
By differentiating this equality with respect to θ , we get the equation
∂ V̇ ∂xi ∂V
= (fx V )
xi + fx ,
∂θ ∂θ ∂θ
i
which is equivalent to
Ṙ = (fx V )xi yi + fx R. (7.42)
i
Upon defining
A= (fx V )xi yi ,
i
the element in the rth row and sth column of the matrix A is equal to
⎛ ⎞
ars = ((fx V )rs )xi yi = ⎝ frxj vj s ⎠ yi
i i j x
i
= yi frxi xj vj s = yi frxi xj vj s
i j j i
∗ ∗
= y fxx v
rj j s
= (y fxx )V rs
,
j
where vj s is the element in the j th row and sth column of the matrix V . Hence we have
A = (y ∗fxx )V and see that (7.42) is equivalent to (7.39). The initial condition in (7.39),
7.3. Second-Order Derivatives of x(tf ; t0 , x0 , θ ) with Respect to t0 , tf , x0 , and θ 313
which is similar to the initial condition (7.26) in Proposition 7.13, follows from (7.41) (see
the proof of Proposition 7.13). The condition R(t; x0 , θ ) = 0 for t < θ is obvious.
Here, V (t) is the solution to the IVP (7.18), y k (t) is the solution to the IVP (7.26) (for
t0 = tˆ0 , x0 = x̂0 , θ = θ̂), and [fx ]k = f (tˆk , x̂(tˆk ), û(tˆk +)) − f (tˆk , x̂(tˆk ), û(tˆk −)), k = 1, . . . , s.
Proposition 7.17. For t ≥ θ the function z(t; θ ) is the solution to the system
ż = fx z + y ∗fxx y (7.45)
In (7.45), fx and fxx are taken along the trajectory (t, x(t; θ ), u(t; θ )), t ∈ [t0 , tf ], and y ∗fxx y
is a vector with elements
n
∂ 2 fk
(y ∗fxx y)k = y ∗fkxx y = yi yj , k = 1, . . . , n.
∂xi ∂xj
i,j =1
are the jumps of the derivatives ft (t, x(t; θ), u(t; θ )) and fx (t, x(t; θ), u(t; θ )) at the point θ
(u2 = u(θ +; θ), u1 = u(θ−; θ )). For t < θ , we have z(t; θ) = 0.
314 Chapter 7. Bang-Bang Control Problem and Its Induced Optimization Problem
Proof. By Proposition 7.13, for t ≥ θ the function y(t; θ ) is the solution to the IVP
By differentiating these equalities with respect to θ at the points θ and θ +, we obtain (7.45)
and (7.46). For t < θ we have y = 0 and hence z = 0.
For the solution x(t; t0 , x0 , θ ) to the IVP (7.3) with an arbitrary s, it follows from
Proposition 7.17 that
where for t ≥ tˆk the vector function zkk (t) satisfies the equation
żkk (t) = fx (t, x̂(t), û(t))zkk (t) + y k (t)∗ fxx (t, x̂(t), û(t))y k (t) (7.48)
Here, for t ≥ tˆk , the function y k (t) is the solution to the IVP (7.26), and y k (t) = 0
for t < tˆk , k = 1, . . . , s. Furthermore, by definition, [ft ]k = ft (tˆk , x̂(tˆk ), û(tˆk +)) −
ft (tˆk , x̂(tˆk ), û(tˆk −)) and [fx ]k = fx (tˆk , x̂(tˆk ), û(tˆk +)) − fx (tˆk , x̂(tˆk ), û(tˆk −)) are the jumps
of the derivatives ft (t, x̂(t), û(t)) and fx (t, x̂(t), û(t)) at the point tˆk . For t < tˆk we put
zkk (t) = 0, k = 1, . . . , s.
Proposition 7.18. For t ≥ t2 the function z12 (t; θ ) is the solution to the system
In (7.50), fx and fxx are taken along the trajectory (t, x(t; θ ), u(t; θ )), t ∈ [t0 , tf ], and
(y 1 )∗ fxx y 2 is a vector with elements
n
∂ 2 fk 1 2
((y 1 )∗ fxx y 2 )k = (y 1 )∗ fkxx y 2 = y y , k = 1, . . . , n.
∂xi ∂xj i j
i,j =1
Proof. By Proposition 7.13, for t ≥ t1 the function y 1 (t; θ ) is a solution to the equation
ẏ 1 (t; θ ) = fx (t, x(t; θ ), u(t; θ))y 1 (t; θ ),
where y 1 (t; θ ) = 0 for t < t1 . Differentiating this equation with respect to t2 , we see that for
1
t ≥ t2 , the function z12 (t; θ ) = ∂y ∂t(t;θ)
2
is a solution to system (7.50). The initial condition
(7.51) is similar to the initial condition (7.26) in Proposition 7.13. For t < t2 , we obviously
have z12 (t; θ ) = 0.
For the solution x(t; t0 , x0 , θ ) of IVP (7.3) and for tk < tj (k, j = 1, . . . , s), it follows
from Proposition 7.18 that
where for t ≥ tˆj the vector function zkj (t) is the solution to the equation
żkj (t) = fx (t, x̂(t), û(t))zkj (t) + y k (t)∗ fxx (t, x̂(t), û(t))y j (t) (7.53)
satisfying the initial condition
zkj (tˆj ) = −[ẏ k ]j = −[fx ]j y k (tˆj ). (7.54)
Here, for t ≥ tˆk , the function y k (t) is the solution to the IVP (7.26), while
y k (t)= 0 holds for t < tˆk , k = 1, . . . , s. By definition, [ẏ k ]j = ẏ k (tˆj +) − ẏ k (tˆj −) and
[fx ]j = fx (tˆj , x̂(tˆj ), û(tˆj +)) − fx (tˆj , x̂(tˆj ), û(tˆj −)) are the jumps of the derivatives ẏ k (t)
and fx (t, x̂(t), û(t)), respectively, at the point tˆj . For t < tˆj we put zkj (t) = 0.
where V (t) is the solution to the IVP (7.18). From (7.27) it follows that
Proposition 7.20. The elements sij (t; t0 , x0 ) of the matrix S(t; t0 , x0 ) satisfy the system
and the matrix S itself satisfies the initial condition at the point t = t0 ,
In (7.65), the derivatives fx and fxx are taken along the trajectory (t, x(t; t0 , x0 ), u(t)),
t ∈ [t0 , tf ], ej is the j th column of the identity matrix E, and, by definition, ẋ(t0 ) = ẋ(t0 ; t0 , x0 ).
By differentiating these equalities with respect to t0 and using Proposition 7.12, we obtain
(7.65). Differentiating the second equality in (7.67) with respect to t0 yields (7.66).
where the elements sij (t) of the matrix S(t) satisfy the system
Here, V (t) is the solution to the IVP (7.18), and the matrix S(t) itself satisfies the initial
condition at the point t = tˆ0 ,
S(tˆ0 ) + V̇ (tˆ0 ) = 0. (7.70)
.
7.3.8 Derivative gtk t0
Consider again the case s = 1 with θ = t1 and define
Proposition 7.21. For t ≥ θ , the function r(t; t0 , θ ) is the solution to the IVP
where y ∗fxx V ẋ(t0 ) is the vector with elements (y ∗fxx V ẋ(t0 ))i = y ∗fixx V ẋ(t0 ), i = 1, . . . , n,
V (θ ) = V (θ; t0 , θ), and
is the jump of the derivative fx (t, x(t; t0 , θ ), u(t; θ )) at the point θ . The derivatives fx
and fxx are taken along the trajectory (t, x(t; t0 , θ ), u(t; θ )), t ∈ [θ , tf ]. For t < θ we have
r(t; t0 , θ) = 0. Then the jump of the function r(t; t0 , θ ) at the point t = θ is given by [r] =
[fx ]V (θ )ẋ(t0 ).
Proof. By Proposition 7.13 we have y(t; t0 , θ ) = 0 for t < θ and hence r(t; t0 , θ ) = 0 for
t < θ. According to the same proposition, for t ≥ θ the function y(t; t0 , θ ) satisfies the
equation
ẏ(t; t0 , θ) = fx (t, x(t; t0 , θ ), u(t; θ))y(t; t0 , θ ).
Differentiating this equation with respect to t0 , we get
∂x
ṙ = fx r + y ∗fxx .
∂t0
According to Proposition 7.12,
∂x(t; t0 , θ )
= −V (t; t0 , θ )ẋ(t0 ),
∂t0
7.4. Quadratic Form for the Induced Optimization Problem 319
ṙ = fx r − y ∗fxx V ẋ(t0 ).
Here V (t) is the solution to the IVP (7.18) and y k (t) is the solution to the IVP (7.26). The
˙ tˆ0 ) has components
vector (y k )∗ fxx V x̂(
˙ tˆ0 ))j = (y k )∗ fj xx V x̂(
((y k )∗ fxx V x̂( ˙ tˆ0 ), j = 1, . . . , n.
s
s
Lζ ζ (μ, ζ̂ )ζ̄ , ζ̄ = Lx0 x0 x̄0 , x̄0 + 2 Lx0 tk x̄0 t¯k + Ltk tj t¯k t¯j
k=1 k,j =1
s
+ 2Lx0 tf x̄0 t¯f + 2 Ltk tf t¯k t¯f + Ltf tf t¯f2
k=1
s
+ 2Lx0 t0 x̄0 t¯0 + 2 Lt0 tk t¯0 t¯k + 2Lt0 tf t¯0 t¯f + Lt0 t0 t¯02 . (7.75)
k=1
320 Chapter 7. Bang-Bang Control Problem and Its Induced Optimization Problem
All derivatives in formula (7.75) are taken at the point (μ, ζ̂ ). Now we shall calculate these
derivatives. Recall the definition (7.6) of the Lagrangian,
Note that all functions V , W k , y k , zkj , S, R k , q, w, r k , introduced in Sections 7.1.2 and 7.3,
depend now on t, t0 , x0 , and θ . For simplicity, we put V (t) = V (t; tˆ0 , x̂0 , θ̂ ), etc.
∂
lx0 (t0 , x0 , tf , x(tf ; t0 , x0 , θ ))x̄0 = x̄0∗ lx0 x0 (t0 , x0 , tf , x(tf ; t0 , x0 , θ ))
∂x0
+ x̄0∗ lx0 xf (t0 , x0 , tf , x(tf ; t0 , x0 , θ ))V (tf ; t0 , x0 , θ ), (7.78)
and
∂
lxf (t0 , x0 , tf , x(tf ; t0 , x0 , θ ))V (tf ; t0 , x0 , θ ))x̄0
∂x0
= x̄0∗ V ∗ (tf ; t0 , x0 , θ ) lxf x0 (t0 , x0 , tf , x(tf ; t0 , x0 , θ ))
+ lxf xf (t0 , x0 , tf , x(tf ; t0 , x0 , θ ))V (tf ; t0 , x0 , θ )
∂
+lxf (t0 , x0 , tf , x(tf ; t0 , x0 , θ )) V (tf ; t0 , x0 , θ )x̄0 . (7.79)
∂x0
From (7.77)–(7.79) and the transversality condition lxf (p̂) = ψ(tˆf ) it follows that at the
point ζ̂ , we have
Lx0 x0 x̄0 , x̄0 = x̄0∗ lx0 x0 (p̂)x̄0 + 2x̄0∗ lx0 xf (p̂)V (tˆf )x̄0 + x̄0∗ V ∗ (tˆf )lxf xf (p̂)V (tˆf )x̄0
∂
+ ψ(tf ) (V (tf ; t0 , x0 , θ )x̄0 )x̄0 . (7.80)
∂x 0 ζ =ζ̂
Using these equations together with the adjoint equation −ψ̇ = ψfx , we obtain
d
ψk W k
= ψ̇k W k + ψk Ẇ k
dt
k k k
∗
= − ψfxk W +
k
ψk V fkxx V + fkxr W r
k k
r
= − ψfxk W k + V ∗ (ψk fkxx ) V + ψk fkxr W r
k k k r
∗
= − ψfxr W + V
r
ψk fkxx V + ψk fkxr W r
r k r k
∗
= − ψfxr W + V (ψfxx )V +
r
ψfxr W r = V ∗ Hxx V .
r r
Proof. Using Propositions 7.22 and 7.23 and the initial conditions W k (tˆ0 ) = 0 for k =
1. . . . , n, we get
∂
ψ(tf ) V (tf ; t0 , x0 , θ )x̄0 x̄0
∂x0 ζ̂
ˆ
tf
= x̄0∗
ψk (tˆf )W (tˆf ) x̄0 = x̄0
k ∗
ψk (t)W (t) x̄0
k
k k tˆ0
tˆf d tˆf
= x̄0∗ ψk W k x̄0 dt = x̄0∗ V ∗ Hxx V x̄0 dt
tˆ0 dt tˆ0
k
tˆf
= (V x̄0 )∗ Hxx (V x̄0 ) dt.
tˆ0
Proof. Using equation (7.44) and the adjoint equation −ψ̇ = ψfx , we get for t ∈ [tˆk , tˆf ],
d
(ψR k ) = ψ̇R k + ψ Ṙ k = −ψfx R k + ψ ((y k )∗ fxx )V + fx R k
dt
= ψj (y k )∗ fj xx V = (y k )∗ ψj fj xx V = (y k )∗ Hxx V ,
j j
where Hxx is taken along the trajectory (t, x̂(t), ψ(t), û(t)). Consequently,
tˆf
ˆ k ˆ ˆ k ˆ
ψ(tf )R (tf ) = ψ(tk )R (tk ) + (y k )∗ Hxx V dt.
tˆk
Formulas (7.87) and (7.88) and the condition y k (t) = 0 for t < tˆk imply the equality
Lx0 tk x̄0 t¯k = x̄0∗ lx0 xf (p̂)y k (tˆf )t¯k + (V (tˆf )x̄0 )∗ lxf xf (p̂)y k (tˆf )t¯k
tˆ
− [Hx ]k V (tˆk )x̄0 t¯k + tˆf Hxx y k t¯k , V x̄0 dt.
0
∂2
l(t0 , x0 , tf , x(tf ; t0 , x0 , θ ))
∂tk2
∂
= lx (t0 , x0 , tf , x(tf ; t0 , x0 , θ )) y k (tf ; t0 , x0 , θ )
∂tk f
∂
+ lxf (t0 , x0 , tf , x(tf ; t0 , x0 , θ )) y k (tf ; t0 , x0 , θ )
∂tk
= lxf xf (t0 , x0 , tf , x(tf ; t0 , x0 , θ ))y k (tf ; t0 , x0 , θ ), y k (tf ; t0 , x0 , θ )
+ lxf (t0 , x0 , tf , x(tf ; t0 , x0 , θ ))zkk (tf ; t0 , x0 , θ ), (7.90)
324 Chapter 7. Bang-Bang Control Problem and Its Induced Optimization Problem
and thus,
∂2
Ltk tk = l(t , x , t , x(t ; t , x , θ ))
∂tk2
0 0 f f 0 0
ζ =ζ̂
= lxf xf (p̂)y k (tˆf ), y k (tˆf ) + lxf (p̂)zkk (tˆf ).
Let us rewrite the last term in this formula. The transversality condition lxf = ψ(tˆf ) implies
tˆf d
lxf (p̂)zkk (tˆf ) = ψ(tˆf )zkk (tˆf ) = (ψzkk ) dt + ψ(tˆk )zkk (tˆk ). (7.91)
tˆk dt
Using this equation together with the adjoint equation −ψ̇ = ψfx , we get
d
(ψzkk ) = ψ̇zkk + ψ żkk = −ψfx zkk + ψfx zkk + ψj ((y k )∗ fj xx y k ) = (y k )∗ Hxx y k ,
dt
j
(7.92)
and thus tˆf
lxf (p̂)zkk (tˆf ) = (y k )∗ Hxx y k dt + ψ(tˆk )zkk (tˆk ). (7.93)
tˆk
(k H )(t) = H (t, x̂(t), ûk+ , ψ(t)) − H (t, x̂(t), ûk− , ψ(t)),
d
D k (H ) = − (k H )t=t + = −[Ht ]k − [Hx ]k x̂( ˙ tˆk +) − ψ̇(tˆk +)[Hψ ]k (7.94)
dt k
Proof. Multiplying the initial condition (7.49) for zkk at the point t = tˆk by ψ(tˆk ), we get
˙ tˆk +) + y k (tˆk ) .
ψ(tˆk )zkk (tˆk ) + ψ(tˆk )ẏ k (tˆk +) = −ψ(tˆk )[ft ]k − ψ(tˆk )[fx ]k x̂( (7.96)
Here, we obviously have the relations ψ(tˆk )[ft ]k = [Ht ]k , ψ(tˆk )[fx ]k = [Hx ]k , and y k (tˆk ) =
[y k ]k . Moreover, equation (7.26) for y k together with the adjoint equation −ψ̇ = ψfx
implies that ψ ẏ k = ψfx y k = −ψ̇y k . Hence, in view of the initial condition (7.26) for y k ,
we find
ψ(tˆk )ẏ k (tˆk +) = −ψ̇(tˆk +)y k (tˆk ) = ψ̇(tˆk +)[f ]k = ψ̇(tˆk +)[Hψ ]k .
From the relations (7.91), (7.93), and (7.95) and the equality y k (t) = 0 for t < tˆk , it
follows that
tˆf
Ltk tk t¯k2 = lxf xf (p̂)y k (tˆf )t¯k , y k (tˆf )t¯k + (y k t¯k )∗ Hxx y k t¯k dt
tˆ0
+D k
(H )t¯k2 − [Hx ]k [y k ]k t¯k2 , k = 1, . . . , s. (7.97)
Let us calculate Ltk tj for k < j . Differentiating (7.89) with respect to tj , we get
∂2
l(t0 , x0 , tf , x(tf ; t0 , x0 , θ ))
∂tk ∂tj
∂
= lxf (t0 , x0 , tf , x(tf ; t0 , x0 , θ )) y k (tf ; t0 , x0 , θ )
∂tj
∂ k
+ lxf (t0 , x0 , tf , x(tf ; t0 , x0 , θ )) y (tf ; t0 , x0 , θ )
∂tj
= lxf xf (t0 , x0 , tf , x(tf ; t0 , x0 , θ ))y k (tf ; t0 , x0 , θ ), y j (tf ; t0 , x0 , θ )
+ lxf (t0 , x0 , tf , x(tf ; t0 , x0 , θ ))zkj (tf ; t0 , x0 , θ ). (7.99)
Thus,
∂2
Ltk tj = l(t0 , x0 , tf , x(tf ; t0 , x0 , θ ))|ζ =ζ̂
∂tk ∂tj
= lxf xf (p̂)y k (tˆf ), y j (tˆf ) + lxf (p̂)zkj (tˆf ). (7.100)
We can rewrite the last term in this formula as
tˆf d
lxf (p̂)zkj (tˆf ) = ψ(tˆf )zkj (tˆf ) = (ψzkj ) dt + ψ(tˆj )zkj (tˆj ).
tˆj dt
By formula (7.53), żkk = fx zkk + (y k )∗ fxx y j for t ≥ tˆj . Similarly to (7.92), we get
k ∗
dt (ψz ) = (y ) Hxx y and thus obtain
d kj j
tˆf
lxf (p̂)z (tˆf ) =
kj
(y k )∗ Hxx y j dt + ψ(tˆj )zkj (tˆj ). (7.101)
tˆj
Using the initial condition (7.54) for zkj at the point θ̂ j , we get
ψ(tˆj )zkj (tˆj ) = −ψ(tˆj )[fx ]j y k (tˆj ) = −[Hx ]j y k (tˆj ). (7.103)
Formulas (7.100)–(7.103) imply the following representation for all k < j :
Ltk tj t¯k t¯j = lxf xf (p̂)y k (tˆf )t¯k , y j (tˆf )t¯j
tˆf
+ (y k t¯k )∗ Hxx y j t¯j dt − [Hx ]j y k (tˆj )t¯k t¯j . (7.104)
tˆ0
Again, we transform the last term in this formula at the point ζ = ζ̂ . Using the adjoint equa-
tion −ψ̇ = ψfx and the transversality condition ψ(tˆf ) = lxf , we get
Consequently,
˙ tˆf )t¯f , x̄0 + lx t V (tˆf )x̄0 t¯f
Lx0 tf x̄0 t¯f = lx0 tf x̄0 t¯f + lx0 xf x̂( f f
˙ (7.105)
+ lxf xf x̂(tˆf )t¯f , V (tˆf )x̄0 − ψ̇(tˆf )V (tˆf )x̄0 t¯f .
∂2 ∂
l(t0 , x0 , tf , x(tf ; t0 , x0 , θ )) = {lx y k }|t=tf
∂tk ∂tf ∂tf f
= {(lxf xf ẋ + lxf tf )y k + lxf ẏ k }|t=tf
= {lxf xf ẋy k + lxf tf y k + lxf fx y k }|t=tf .
We evaluate the last term in this formula at the point ζ = ζ̂ using the adjoint equation
−ψ̇ = ψfx and the transversality condition ψ(tˆf ) = lxf :
Therefore,
˙ tˆf )t¯f , y k (tˆf )t¯k + lx t y k (tˆf )t¯k t¯f − ψ̇(tˆf )y k (tˆf )t¯k t¯f .
Ltk tf t¯k t¯f = lxf xf x̂( (7.106)
f f
7.4. Quadratic Form for the Induced Optimization Problem 327
which gives
˙ tˆf ) + lx x x̂(
Ltf tf = ltf tf + 2ltf xf x̂( ˙ tˆf ), x̂(
˙ tˆf ) + ψ(tˆf )x̂(
¨ tˆf ). (7.107)
f f
Let us transform the last term. Equation (6.11) in the definition of M0 is equivalent to the
relation ψ x̂˙ + ψ0 = 0. Differentiating this equation with respect to t, we get
∂2 ∂
l(t0 , x0 , tf , x(tf ; t0 , x0 , θ )) = {lx + lxf V }|t=tf
∂x0 ∂t0 ∂t0 0
∂x ∂x ∂V
= lx0 t0 + lx0 xf + lxf t0 + lxf xf V + lxf .
∂t0 ∂t0 ∂t0 t=tf
Now, using the transversality condition lxf = ψ(tˆf ), formula (7.24), and the notation ∂V
∂t0 = S,
we get
˙ tˆ0 ) + lx t V (tˆf ) − x̂(
Lx0 t0 = lx0 t0 − lx0 xf V (tˆf )x̂( ˙ tˆ0 )∗ V (tˆf )∗ lx x V (tˆf ) + ψ(tˆf )S(tˆf ).
f 0 f f
The transformation of the last term in this formula proceeds as follows. Using the adjoint
equation for ψ and the system (7.69) for S, we obtain the equation
d
(ψS) = ˙ tˆ0 )∗ V ∗
ψ̇S + ψ Ṡ = −ψfx S + ψfx S − x̂( ψi fixx V
dt
i
= ˙ tˆ0 )∗ V ∗ Hxx V ,
−x̂(
which yields
tˆf
ψ(tˆf )S(tˆf ) = − ˙ tˆ0 )∗ V ∗ Hxx V dt + ψ(tˆ0 )S(tˆ0 ).
x̂( (7.109)
tˆ0
328 Chapter 7. Bang-Bang Control Problem and Its Induced Optimization Problem
Using now the initial condition (7.70) for S at the point t = tˆ0 and the equation V̇ = fx V ,
we get
(ψS)|tˆ0 = −(ψ V̇ )|tˆ0 = −(ψfx V )|tˆ0 = (ψ̇V )|tˆ0 = ψ̇(tˆ0 ), (7.110)
since V (tˆ0 ) = E. Formulas (7.109) and (7.110) then imply the equality
Therefore,
˙ tˆ0 )t¯0 , x̄0 + lx t V (tˆf )x̄0 t¯0
Lx0 t0 x̄0 t¯0 = lx0 t0 x̄0 t¯0 − lx0 xf V (tˆf )x̂( f 0
˙
−lxf xf V (tˆf )x̄0 , V (tˆf )x̂(tˆ0 )t¯0 + ψ̇(tˆ0 )x̄0 t¯0
tˆf
− ˙ tˆ0 )t¯0 dt.
Hxx V x̄0 , V x̂( (7.112)
tˆ0
∂2 ∂
l(t0 , x0 , tf , x(tf ; t0 , x0 , θ )) = {lx y k }|t=tf
∂tk ∂t0 ∂t0 f
k ∗ ∂x ∂y k
= lxf t0 y + (y ) lxf xf
k
+ lxf
∂t0 ∂t0 t=tf
= {lxf t0 y k + (y k )∗ lxf xf w + lxf r k }|t=tf .
Let us transform the last term in this formula. Using the adjoint equation for ψ and the
system (7.73) for r k , we get for t ≥ tˆk :
d
(ψr k ) = ψ̇r k + ψ ṙ k = −ψfx r k + ψfx r k − (yi )∗ ˙ tˆ0 )
ψj fj xx V x̂(
dt
j
∗ ˙ tˆ0 ).
= −(yk ) Hxx V x̂(
It follows that
tˆf
ψ(tˆf )r k (tˆf ) = − ˙ tˆ0 ) dt + ψ(tˆk )r k (tˆk ).
(yk )∗ Hxx V x̂( (7.114)
tˆk
7.4. Quadratic Form for the Induced Optimization Problem 329
The initial condition (7.74) for r k at the point tˆk then yields
˙ tˆ0 ) = [Hx ]k V (tˆk )x̂(
ψ(tˆk )r k (tˆk ) = ψ(tˆk )[fx ]k V (tˆk )x̂( ˙ tˆ0 ). (7.115)
Formulas (7.113)–(7.115) and the condition y k (t) = 0 for t < tˆk then imply the equality
˙ tˆ0 )
Ltk t0 = lxf t0 y k (tˆf ) − (y k (tˆf ))∗ lxf xf V (tˆf )x̂(
tˆf
+ [Hx ]k V (tˆk )x̂( ˙ tˆ0 ) − (yk )∗ Hxx V x̂( ˙ tˆ0 ) dt. (7.116)
tˆ0
Hence,
˙ tˆ0 )t¯0
Ltk t0 t¯k t¯0 = lxf t0 y k (tˆf )t¯k t¯0 − (y k (tˆf )t¯k )∗ lxf xf V (tˆf )x̂(
tˆf
+ [Hx ]k V (tˆk )x̂( ˙ tˆ0 )t¯0 t¯k − ˙ tˆ0 )t¯0 dt.
(y k t¯k )∗ Hxx V x̂( (7.117)
tˆ0
˙ tˆ0 ) + lx t x̂(
Ltf t0 = ltf t0 − ltf xf V (tˆf )x̂( ˙ tˆf )
f 0
which yields
Ltf t0 t¯f t¯0 = ltf t0 t¯f t¯0 − ltf xf (V (tˆf )x̂(˙ tˆ0 )t¯0 )t¯f + lx t (x̂(
˙ tˆf )t¯f )t¯0
f 0
where
∂x ∂w ∂ 2 x
w= , q= = 2.
∂t0 ∂t0 ∂t0
The transversality condition lxf = ψ(tˆf ) yields
Lt0 t0 = lt0 t0 + 2lt0 xf w(tˆf ) + lxf xf w(tˆf ), w(tˆf ) + ψ(tˆf )q(tˆf ). (7.122)
Let us transform the last term using the adjoint equation for ψ and the system (7.58) for q:
d
(ψq) = ψ̇q + ψ q̇ = −ψfx q + ψfx q + ψj (w ∗fj xx w) = w ∗ Hxx w.
dt
j
˙ tˆ0 ), we obtain
Also, using the equality w = −V x̂(
tˆf
ψ(tˆf )q(tˆf ) = ψ(tˆ0 )q(tˆ0 ) + w∗ Hxx w dt
tˆ0
tˆf
= ψ(tˆ0 )q(tˆ0 ) + ˙ tˆ0 ), V x̂(
Hxx V x̂( ˙ tˆ0 ) dt. (7.123)
tˆ0
7.4. Quadratic Form for the Induced Optimization Problem 331
From the equation ẇ = fx w (see Proposition 7.12), the adjoint equation −ψ̇ = ψfx , and
˙ tˆ0 ), it follows that
the formula w = −V x̂(
˙ tˆ0 ).
−ψ ẇ = −ψfx w = ψ̇w = −ψ̇V x̂(
˙ tˆ0 ).
ψ(tˆ0 )q(tˆ0 ) = ψ̇0 (tˆ0 ) − ψ̇(tˆ0 )x̂( (7.127)
Thus we have obtained the following explicit and massive representation of the
quadratic form in the IOP:
= x̄0∗ lx0 x0 x̄0 + 2x̄0∗ lx0 xf V (tˆf )x̄0 + (V (tˆf )x̄0 )∗ lxf xf V (tˆf )x̄0
tˆf
+ (V x̄0 )∗ Hxx V x̄0 dt
tˆ0
s s
+ 2x̄0∗ lx0 xf y k (tˆf )t¯k + 2(V (tˆf )x̄0 )∗ lxf xf y k (tˆf )t¯k
k=1 k=1
s s tˆf
− 2[Hx ]k V (tˆk )x̄0 t¯k + 2Hxx y k t¯k , V x̄0 dt
tˆ0
k=1 k=1
s s
tˆf
+ lxf xf y k (tˆf )t¯k , y k (tˆf )t¯k + (y k t¯k )∗ Hxx y k t¯k dt
ˆ
k=1 k=1 t0
s
s
+ D k (H )t¯k2 − [Hx ]k [y k ]k t¯k2 + 2lxf xf y k (tˆf )t¯k , y j (tˆf )t¯j
k=1 k=1 k<j
tˆf
+ 2(y k t¯k )∗ Hxx y j t¯j dt − 2[Hx ]j y k (tˆj )t¯k t¯j
ˆ
k<j t0 k<j
s
s
+ 2lxf t0 y k (tˆf )t¯k t¯0 − ˙ tˆ0 )t¯0
2(y k (tˆf )t¯k )∗ lxf xf V (tˆf )x̂(
k=1 k=1
s s
tˆf
+ ˙ tˆ0 )t¯0 t¯k −
2[Hx ] V (tˆk )x̂(
k ˙ tˆ0 )t¯0 dt
2(y k t¯k )∗ Hxx V x̂(
ˆ
k=1 k=1 t0
˙ tˆ0 )t¯0 )t¯f + 2lx t (x̂(
+ 2ltf t0 t¯f t¯0 − 2ltf xf (V (tˆf )x̂( ˙ tˆf )t¯f )t¯0
f 0
Again, we wish to emphasize that this explicit representation involves only first-order vari-
ations y k and V of the trajectories x(t; t0 , x0 , θ ).
Type (b): Mixed terms with [Hx ]k connected with the variation t¯k ,
s
s
b := − 2[Hx ]k V (tˆk )x̄0 t¯k − [Hx ]k [y k ]k t¯k2
k=1 k=1
s
− 2[Hx ]j y k (tˆj )t¯k t¯j + ˙ tˆ0 )t¯0 t¯k . (7.132)
2[Hx ]k V (tˆk )x̂(
k<j k=1
334 Chapter 7. Bang-Bang Control Problem and Its Induced Optimization Problem
Since
s
k−1
[Hx ]j y k (tˆj )t¯k t¯j = [Hx ]k y j (tˆk )t¯k t¯j = [Hx ]k y j (tˆk )t¯k t¯j ,
k<j j <k k=1 j =1
s
k−1
1 ˙ tˆ0 )t¯0 t¯k .
b=− 2[Hx ]k V (tˆk )x̄0 + [y k ]k t¯k + y j (tˆk )t¯j − V (tˆk )x̂( (7.133)
2
k=1 j =1
s
x̄(t) = V (t)x̄0 + ˙ tˆ0 )t¯0 .
y k (t)t¯k − V (t)x̂( (7.134)
k=1
Then we have
k−1
x̄(tˆk −) = V (tˆk )x̄0 + ˙ tˆ0 )t¯0 ,
y j (tˆk )t¯j − V (tˆk )x̂(
j =1
since y j (tˆk −) = y j (tˆk ) = 0 for j > k and y k (tˆk −) = 0. Moreover, the jump of x̄(t) at the
point tˆk is equal to the jump of y k (t)t¯k at the same point, i.e., [x̄]k = [y k ]k t¯k . Therefore,
1 k−1
V (tˆk )x̄0 + [y k ]k t¯k + ˙ tˆ0 )t¯0
y j (tˆk )t¯j − V (tˆk )x̂(
2
j =1
1 1
= x̄(tˆk −) + [x̄]k = (x̄(tˆk −) + x̄(tˆk +)) = x̄av
k
.
2 2
Thus, we get
s
k ¯
b=− 2[Hx ]k x̄av tk . (7.135)
k=1
k=1
= 2lxf tf x̄¯f t¯f . (7.142)
336 Chapter 7. Bang-Bang Control Problem and Its Induced Optimization Problem
s
d5 := (V (tˆf )x̄0 )∗ lxf xf V (tˆf )x̄0 + 2(V (tˆf )x̄0 )∗ lxf xf y k (tˆf )t¯k
k=1
s
+ lxf xf y k (tˆf )t¯k , y k (tˆf )t¯k + 2lxf xf y k (tˆf )t¯k , y j (tˆf )t¯j
k=1 k<j
s
˙ tˆf )t¯f , V (tˆf )x̄0 +
+ 2lxf xf x̂( ˙ tˆf )t¯f , y k (tˆf )t¯k
2lxf xf x̂(
k=1
+ lxf xf x̂( ˙ tˆf )t¯f − 2lx x V (tˆf )x̄0 , V (tˆf )x̂(
˙ tˆf )t¯f , x̂( ˙ tˆ0 )t¯0
f f
s
− ˙ tˆ0 )t¯0 − 2lx x V (tˆf )x̂(
2(y k (tˆf )t¯k )∗ lxf xf V (tˆf )x̂( ˙ tˆ0 )t¯0 , x̂(
˙ tˆf )t¯f
f f
k=1
˙ tˆ0 )t¯0 , V (tˆf )x̂(
+ lxf xf V (tˆf )x̂( ˙ tˆ0 )t¯0 . (7.143)
One can easily check that this sum can be transformed into the perfect square
'
s
d5 := lxf xf V (tˆf )x̄0 + ˙ tˆf )t¯f − V (tˆf )x̂(
y k (tˆf )t¯k + x̂( ˙ tˆ0 )t¯0 ,
k=1
s (
V (tˆf )x̄0 + ˙ tˆf )t¯f − V (tˆf )x̂(
y (tˆf )t¯k + x̂(
k ˙ tˆ0 )t¯0
k=1
= lxf xf x̄¯f , x̄¯f . (7.144)
s
d7 := −2ψ̇(tˆf )V (tˆf )x̄0 t¯f − 2ψ̇(tˆf )y k (tˆf )t¯k t¯f
k=1
˙ tˆf ) + ψ̇0 (tˆf ))t¯2 + 2ψ̇(tˆf )(V (tˆf )x̂(
− (ψ̇(tˆf )x̂( ˙ tˆ0 )t¯0 )t¯f
f
s
= −2ψ̇(tˆf ) V (tˆf )x̄0 + ˙ tˆ0 )t¯0 ) t¯f
y k (tˆf )t¯k − V (tˆf )x̂(
k=1
˙ tˆf ) + ψ̇0 (tˆf ))t¯2
− (ψ̇(tˆf )x̂( f
˙ tˆf ) + ψ̇0 (tˆf ))t¯2 .
= −2ψ̇(tˆf )x̄(tˆf )t¯f − (ψ̇(tˆf )x̂( (7.146)
f
˙ tˆf ))t¯2 .
d7 = −2ψ̇(tˆf )x̄¯f t¯f − (ψ̇0 (tˆf ) − ψ̇(tˆf )x̂( (7.147)
f
7.5. Equivalence of the Quadratic Forms 337
This completes the whole list of all terms in the quadratic form associated with the
IOP. Hence, we have
7
Lζ ζ ζ̄ , ζ̄ = a + b + c + d, d= di .
i=1
We thus have found the following representation of this quadratic form; see formulas
(7.131) for a, (7.135) for b, and (7.137) for c:
s
s tˆf
Lζ ζ ζ̄ , ζ̄ = D (H )t¯k2 −
k k ¯
2[Hx ]k x̄av tk + Hxx x̄, x̄ + d, (7.148)
k=1 k=1 tˆ0
where according to formulas (7.138), (7.139), (7.141), (7.142), (7.144), (7.145), (7.147) for
d1 , . . . , d7 , respectively,
d = lx0 x0 x̄0 , x̄0 + 2lx0 tf x̄0 t¯f + ltf tf t¯f2 + 2lx0 t0 x̄0 t¯0 + 2ltf t0 t¯f t¯0 + lt0 t0 t¯02 + 2lxf t0 x̄¯f t¯0
˙ tˆ0 ))t¯2
+ 2lx0 xf x̄¯f , x̄0 + 2lxf tf x̄¯f t¯f + lxf xf x̄¯f , x̄¯f + 2ψ̇(tˆ0 )x̄0 t¯0 + (ψ̇0 (tˆ0 ) − ψ̇(tˆ0 )x̂( 0
˙ tˆf ))t¯2 .
− 2ψ̇(tˆf )x̄¯f t¯f − (ψ̇0 (tˆf ) − ψ̇(tˆf )x̂( (7.149)
f
In (7.148) and (7.149) the function x̄(t) and the vector x̄¯f are defined by (7.134) and (7.140),
respectively. Note that in (7.149),
lx0 x0 x̄0 , x̄0 + 2lx0 tf x̄0 t¯f + ltf tf t¯f2 + 2lx0 t0 x̄0 t¯0 + 2ltf t0 t¯f t¯0 + lt0 t0 t¯02 + 2lxf t0 x̄¯f t¯0
+ 2lx0 xf x̄¯f , x̄0 + 2lxf tf x̄¯f t¯f + lxf xf x̄¯f , x̄¯f = lpp p̄,
¯ p̄ ,
¯ (7.150)
where, by definition,
p̄¯ = (t¯0 , x̄0 , t¯f , x̄¯f ). (7.151)
Finally, we get
¯ p̄
d = lpp p̄, ¯ + 2ψ̇(tˆ0 )x̄0 t¯0 + (ψ̇0 (tˆ0 ) − ψ̇(tˆ0 )x̂(˙ tˆ0 ))t¯2
0
¯ ˙
− 2ψ̇(tˆf )x̄1 t¯f − (ψ̇0 (tˆf ) − ψ̇(tˆf )x̂(tˆf ))t¯ .
2
(7.152)
f
correspond to each other, i.e., let π0 λ = μ hold. Then for any ζ̄ = (t¯0 , t¯f , x̄0 , θ̄ ) ∈ R2+n+s the
quadratic form Lζ ζ ζ̄ , ζ̄ has the representation (7.148)–(7.152), where the vector function
x̄(t) and the vector x̄¯f are defined by (7.134) and (7.140). The matrix-valued function V (t)
is the solution to the IVP (7.18), and, for each k = 1, . . . , s, the vector function y k is the
solution to the IVP (7.26).
338 Chapter 7. Bang-Bang Control Problem and Its Induced Optimization Problem
Theorem 7.29. Let λ = (α0 , α, β, ψ, ψ0 ) ∈ and ζ̄ = (t¯0 , t¯f , x̄0 , θ̄ ) ∈ R2+n+s . Put μ =
(α0 , α, β), i.e., let π0 λ = μ ∈ 0 hold; see Proposition 7.4. Define the function x̄(t) by
formula (7.134). Put ξ̄ = −θ̄ and z̄ = (t¯0 , t¯f , ξ̄ , x̄), which means π1 z̄ = ζ̄ ; see Propositions 7.6
and 7.7. Then the following equality holds:
Proof. By Theorem 7.28, the equalities (7.148)–(7.152) hold. In view of equation (6.21), put
s
˙ tˆ0 ) = V (tˆ0 )x̄0 +
x̄¯0 = x̄(tˆ0 ) + t¯0 x̂( ˙ tˆ0 )t¯0 + t¯0 x̂(
y k (tˆ0 )t¯k − V (tˆ0 )x̂( ˙ tˆ0 ).
k=1
Since y k (tˆ0 ) = 0 for k = 1, . . . , s and V (tˆ0 ) = E, it follows that x̄¯0 = x̄0 . Consequently, the
¯ which was defined by (6.21) as (t¯0 , x̄¯0 , t¯f , x̄¯f ), coincides with the vector p̄,
vector p̄, ¯ defined
in this section by (7.151). Hence, the endpoint quadratic form d in (7.152) and the endpoint
quadratic form Ap̄, ¯ p̄
¯ in (6.35) take equal values, d = Ap̄, ¯ p̄ .
¯ Moreover, the integral
tˆf
terms tˆ Hxx x̄, x̄ dt in the representation (7.148) of the form Lζ ζ ζ̄ , ζ̄ and those in the
0
representation (6.34) of the form coincide, and
s
s
s
D k (H )ξ̄k2 + 2[Hx ]k x̄av
k
ξ̄k = D k (H )t¯k2 − k ¯
2[Hx ]k x̄av tk ,
k=1 k=1 k=1
because ξ̄k = −t¯k , k = 1, . . . , s. Thus, the representation (7.148) of the form Lζ ζ ζ̄ , ζ̄
implies the equality (7.153) of both forms.
Theorem 7.8, which is the main result of this chapter, then follows from Theorem 7.29.
Remark. Theorems 7.8 and 7.29 pave the way for the sensitivity analysis of parametric
bang-bang control problems. Since the IOP is finite-dimensional, the SSC imply that we
may take advantage of the well-known sensitivity results by Fiacco [32] to obtain solution
differentiability for the IOP. The strict bang-bang property then ensures solution differ-
entiability for the parametric bang-bang control problems; cf. Kim and Maurer [48] and
Felgenhauer [31].
Chapter 8
339
340 Chapter 8. Numerical Methods for Solving the Induced Optimization Problem
where tˆ0 = t0 , tˆs+1 = tˆf . Thus, ˆ = {tˆ1 , . . . , tˆs } is the set of switching points of the control
û(·) with tˆk < tˆk+1 for k = 0, 1, . . . , s. Put
For convenience, the sequence of components of the vector ζ̂ in definition (7.1) has been
modified. Take a small neighborhood V of the point ζ̂ and let
ζ = (x0 , θ, ts+1 ) ∈ V, θ = (t1 , . . . , ts ), ts+1 = tf ,
where the switching times satisfy t0 < t1 < t2 < · · · < ts < ts+1 = tf . Define the function
u(t; θ) by the condition
u(t; θ) = uk for t ∈ (tk−1 , tk ), k = 1, . . . , s + 1. (8.6)
The values u(tk ; θ ), k = 1, . . . , s, may be chosen in U arbitrarily. For definiteness, define
them by the condition of continuity of the control from the left, u(tk ; θ ) = u(tk −; θ ) for
k = 1, . . . , s, and let u(0; θ) = u1 . Denote by x(t; x0 , θ ) the absolutely continuous solution
of the Initial Value Problem (IVP)
ẋ = f (t, x, u(t; θ )), t ∈ [t0 , tf ], x(t0 ) = x0 . (8.7)
8.1. The Arc-Parametrization Method 341
For each ζ ∈ V this solution exists in a sufficiently small neighborhood V of the point ζ̂ .
Obviously, we have
x(t; x̂0 , θ̂) = x̂(t), ˆ
t ∈ , u(t; θ̂ ) = û(t), ˆ \ .
t ∈ ˆ
Consider now the following Induced Optimization Problem (IOP) in the space Rn × Rs+1
of variables ζ = (x0 , θ, ts+1 ):
F0 (ζ ) := J (x0 , ts+1 , x(ts+1 ; x0 , θ )) → min,
(8.8)
G(ζ ) := K(x0 , ts+1 , x(ts+1 ; x0 , θ )) = 0.
To get uniqueness of the Lagrange multiplier for the equality constraint we assume that the
following regularity (normality) condition holds:
rank ( Gζ (ζ̂ ) ) = d(K). (8.9)
The Lagrange function (7.6) in normalized form is given by
L(β, ζ ) = F0 (ζ ) + β G(ζ ), β ∈ Rd(K) . (8.10)
The critical cone K0 in (7.13) then reduces to the subspace
K0 = { ζ̄ ∈ Rn × Rs+1 | Gζ (ζ̂ )ζ̄ = 0 }. (8.11)
Assuming regularity (8.9), the SSC for the IOP reduce to the condition that there exist
β̂ ∈ Rd(K) with
Lζ (β̂, ζ̂ ) = (F0 )ζ (ζ̂ ) + β̂ Gζ (ζ̂ ) = 0, (8.12)
Lζ ζ (β̂, ζ̂ )ζ̄ , ζ̄ > 0 ∀ ζ̄ ∈ K0 \ {0}. (8.13)
From a numerical point of view it is not convenient to optimize the switching times tk (k =
1, . . . , s) and terminal time ts+1 = tf directly. Instead, as suggested in [44, 45, 66] one
computes the arc durations or arc lengths
τk := tk − tk−1 , k = 1, . . . , s, s + 1, (8.14)
of the bang-bang arcs. Hence, the final time tf can be expressed by the arc lengths as
s+1
tf = t0 + τk . (8.15)
k=1
Next, we replace the optimization variable ζ = (x0 , t1 , . . . , ts , ts+1 ) by the optimization
variable
z := (x0 , τ1 , . . . , τs , τs+1 ) ∈ Rn × Rs+1 , τk := tk − tk−1 . (8.16)
The variables ζ and z are related by the following linear transformation involving the regular
(n + s + 1) × (n + s + 1) matrix R:
In 0 In 0
z = Rζ, R = , ζ = R −1 z, R −1 = ,
0 S 0 S −1
⎛ ⎞ ⎛ ⎞
1 0 ... 0 1 0 ... 0
⎜ .. . ⎟ ⎜ . . ⎟ (8.17)
⎜ −1 1 . .. ⎟ ⎜ 1 1 . . .. ⎟
S=⎜ ⎜ ⎟ −1
S =⎜ . ⎜ ⎟
.. .. ⎟, ⎟.
⎝ . . 0 ⎠ ⎝ .. . . . . . . 0 ⎠
0 −1 1 1 ... 1 1
342 Chapter 8. Numerical Methods for Solving the Induced Optimization Problem
Denoting the solution to the equations (8.7) by x(t; z), the IOP (8.8) obviously is equivalent
to the following IOP with tf defined by (8.15):
F0 (z) := J (x0 , tf , x(tf ; x0 , τ1 , . . . , τs )) → min,
(8.18)
G(z) := K(x0 , tf , x(tf ; x0 , τ1 , . . . , τs )) = 0.
This approach is called the arc-parametrization method. The Lagrangian for this problem
is given in normal form by
L(ρ, z) = F0 (z) + ρ G(z). (8.19)
It is easy to see that the Lagrange multiplier ρ agrees with the multiplier β in the Lagrangian
(8.10). Furthermore, the SSC for the optimization problems (8.8), respectively, (8.18), are
equivalent. This immediately follows from the fact that the Jacobian and the Hessian for
both optimization problems are related through
Kζ = Kz R, Lζ = Lz R, Lζ ζ = R ∗ Lzz R.
Thus we can express the positive definiteness condition (8.13) evaluated for the variable z as
Lzz (β̂, ẑ)z̄, z̄ > 0 ∀ z̄ ∈ (Rn × Rs+1 ) \ {0}, Gẑ (ẑ)z̄ = 0. (8.20)
This condition is equivalent to the property that the so-called reduced Hessian is positive
definite. Let N be the (nz × (nz − d(K))) matrix, nz = n + s + 1, with full column rank nz −
d(K), whose columns span the kernel of Gz (ẑ). Then condition (8.20) is reformulated as
N ∗ Lzz (β̂, ẑ) N > 0 (positive definite). (8.21)
The computational method for determining the optimal vector ẑ ∈ Rn+s+1
is based on a
multiprocess approach proposed in [44, 45, 66]. The time interval [tk−1 , tk ] is mapped to
the fixed interval Ik = [ k−1 k
s+1 , s+1 ] by the linear transformation
3 4
k−1 k
t = ak + bk r, ak = tk − kτk , bk = (s + 1)τk , r ∈ Ik = , , (8.22)
s +1 s +1
where r denotes the running time. Identifying x(r) ∼ = x(ak + bk · r) = x(t) in the relevant
intervals, we obtain the transformed dynamic system
dx dx dt
= · = (s + 1) τk f (ak + bk r, x(r), uk ) for r ∈ Ik . (8.23)
dr dt dr
By way of concatenation of the solutions on the intervals Ik , we obtain an absolutely
continuous solution x(r) = x(r; τ1 , . . . , τs ) for r ∈ [0, 1]. Thus we are confronted with the
task of solving the IOP
s+1
F0 (z) := J (x0 , tf , x(1; x0 , τ1 , . . . , τs )) → min, tf = t0 + τk ,
(8.24)
k=1
G(z) := K(x0 , tf , x(1; x0 , τ1 , . . . , τs )) = 0.
This approach can be conveniently implemented using the routine NUDOCCCS
developed by Büskens [13]. In this way, we can also take advantage of the fact that
NUDOCCCS provides the Jacobian of the equality constraints and the Hessian of the
Lagrangian, which are needed in the check of the second-order condition (8.13), respec-
tively, the positive definiteness of the reduced Hessian (8.21). Moreover, this code allows
for the computation of parametric sensitivity derivatives; cf. [48, 74, 23].
8.1. The Arc-Parametrization Method 343
with given initial value x(t0 ; x0 , θ) = x0 . The IOP with the optimization variable
τk := tk − tk−1 , k = 1, . . . , s + 1.
Invoking the linear time transformation (8.22) for mapping the time interval [tk−1 , tk ] to the
fixed interval Ik = [ k−1 k
s+1 , s+1 ],
t = ak + bk r, ak = tk − kτk , bk = (s + 1)τk , r ∈ Ik ,
This implies the two switching conditions ψ2 (t1 ) = 0 and ψ2 (t2 ) = 0. Hence, the optimization
vector for the IOP (8.18) is given by
z = (τ1 , τ2 , τ3 ), τ1 = t1 , τ2 = t2 − t1 , τ3 = tf − t2 .
The code NUDOCCCS gives the following numerical results for the arc lengths, switching
times, and adjoint variables:
τ1 = t1 = 1.12051, τ2 = 2.18954, t2 = 3.31005,
τ3 = 0.35813, tf = 3.66817,
ψ1 (0) = −0.122341, ψ2 (0) = −0.082652, (8.39)
ψ1 (t1 ) = −0.215212, ψ1 (t2 ) = 0.891992,
ψ1 (tf ) = 0.842762, ψ2 (tf ) = −0.25, β = ψ(tf ).
8.2. Time-Optimal Control of the Rayleigh Equation Revisited 345
(a) time-optimal control u and switching function ( x 4 ) (b) state variables x1 and x2
6
4
3 4
2 2
1
0
0
-2
-1
-2 -4
-3 -6
-4
-8
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
Figure 8.1. Time-optimal control of the Rayleigh equation with boundary condi-
tions (8.31). (a) Bang-bang control and scaled switching function (×4), (b) State vari-
ables x1 , and x2 .
The corresponding time-optimal bang-bang control with two switches and the state variables
are shown in Figure 8.1.
We have already shown in Section 6.4 that the control u in (8.38) enjoys the strict
bang-bang property and that the estimates D k (H ) > 0, k = 1, 2, in (6.114) are satisfied. For
the terminal conditions (8.31), the Jacobian is the (2 × 3) matrix
−4.53176 −3.44715 0.0
Gz (ẑ) =
−11.2768 −7.62049 4.0
which is of rank 2. The Hessian of the Lagrangian is the (3 × 3) matrix
⎛ ⎞
−10.3713 −8.35359 −6.68969
Lzz (β̂, ẑ) = ⎝ −8.35359 −5.75137 −4.61687 ⎠ .
−6.68969 −4.61687 1.97104
Note that this Hessian is not positive definite. However, the projected Hessian (8.21) is the
positive number
N ∗ L̃zz (ẑ, β̂)N = 0.515518,
which shows that the second-order test (8.20) holds. Hence, the extremal characterized by
the data (8.39) provides a strict strong minimum.
Now we consider a modified control problem, where the two terminal conditions
x1 (tf ) = x2 (tf ) = 0 are substituted by the scalar terminal condition
Figure 8.2. Time-optimal control of the Rayleigh equation with boundary condition
(8.40). (a) Bang-bang control u and scaled switching function φ (dashed line). (b) State
variables x1 , x2 .
x2
P
C
q2
q1
O x1
lower arm OP
Figure 8.3. Two-link robot [67]: upper arm OQ, , and angles q1 and q2 .
necessary conditions, since sufficient conditions were not available. In this section, we show
that SSC hold for both types of robots considered in [21, 37, 81].
First, we study the robot model considered in Chernousko et al. [21]. Göllmann [37]
has shown that the optimal control candidate presented in [21] is not optimal, since the sign
conditions of the switching functions do not comply with the Minimum Principle. Figure
8.3 represents the two-link robot schematically. The state variables are the angles q1 and q2 .
The parameters I1 and I2 are the moments of inertia of the upper arm OQ and the lower
arm QP with respect to the points O and Q. Further, let m2 be the mass of the lower arm,
L1 = |OQ| the length of the upper arm, and L1 = |QC| the distance between the second
link Q and the center of gravity C of the lower arm. With the abbreviations
A = I1 + m2 L21 + I2 + 2m2 L1 L cos q2 , B = I2 + m2 L1 L cos q2 ,
R1 = u1 + m2 L1 L(2q̇1 + q̇2 )q̇2 sin q2 , R2 = u2 − m2 L1 Lq̇12 sin q2 , (8.44)
D = I2 , = AD − B 2 ,
the dynamics of the two-link robot can be described by the ODE system
1
q̇1 = ω1 , ω̇1 = (DR1 − BR2 ),
(8.45)
1
q̇2 = ω2 , ω̇2 = (AR2 − BR1 ),
where ω1 and ω2 are the angular velocities. The torques u1 and u2 in the two links represent
the two control variables. The control problem consists of steering the robot from a given
initial position to a terminal position in minimal final time tf ,
q1 (0) = 0, q2 (0) = 0, ω1 (0) = 0, ω2 (0) = 0,
(8.46)
q1 (tf ) = −0.44, q2 (tf ) = 1.83, ω1 (tf ) = 0, ω2 (tf ) = 0.
Both control components are bounded by
|u1 (t)| ≤ 2, |u2 (t)| ≤ 1, t ∈ [0, tf ]. (8.47)
348 Chapter 8. Numerical Methods for Solving the Induced Optimization Problem
(a) control u1 and switching function 1 (b) control u2 and switching function 2
2 1
1.5
1 0.5
0.5
0 0
-0.5
-1 -0.5
-1.5
-2 -1
Figure 8.4. Control of the two-link robot (8.44)–(8.47). (a) Control u1 and scaled
switching function φ1 (dashed line). (b) Control u2 and scaled switching function φ2
(dashed line). (c) Angle q1 and velocity ω1 . (d) Angle q2 and velocity ω2 .
I3 denotes the moment of inertia of the load with respect to the point P , and ω2 is now the
angular velocity of the angle q1 + q2 . For simplicity, we set I3 = 0. Again, the torques u1
and u2 in the two links are used as control variables by which the robot is steered from a
given initial position to a nonfixed end position in minimal final time tf ,
!
q1 (0) = 0, (x1 (tf ) − x1 (0))2 + (x2 (tf ) − x2 (0))2 = r,
q2 (0) = 0, q2 (tf ) = 0, (8.55)
ω1 (0) = 0, ω1 (tf ) = 0,
ω2 (0) = 0, ω2 (tf ) = 0,
where (x1 (t), x2 (t)) are the Cartesian coordinates of the point P ,
x1 (t) = L1 cos q1 (t) + L2 cos(q1 (t) + q2 (t)),
(8.56)
x2 (t) = L1 sin q1 (t) + L2 sin(q1 (t) + q2 (t)).
The initial point (x1 (0), x2 (0)) = (2, 0) is fixed. Both control components are bounded,
|u1 (t)| ≤ 1, |u2 (t)| ≤ 1, t ∈ [0, tf ]. (8.57)
The Hamilton–Pontryagin function is given by
ψ3
H = ψ1 ω1 + ψ2 (ω2 − ω1 ) + (A(u1 , u2 )I22 − B(u2 )I12 cos q2 )
(8.58)
ψ4
+ (B(u2 )I11 − A(u1 , u2 )I12 cos q2 ) .
The switching functions are computed as
1
φ 1 = H u1 = (ψ3 I22 − ψ4 I12 cos q2 ) ,
(8.59)
1
φ2 = Hu2 = (ψ3 (−I22 − I12 cos q2 ) + ψ4 (I11 + I12 cos q2 )) .
For the parameter values
1
L1 = L2 = 1, L = 0.5, I1 = I2 = , r = 3,
m1 = m2 = M = 1,
3
we will show that the optimal control has the following structure with five bang-bang arcs:
⎧ ⎫
⎪ (−1, 1) for 0 ≤ t < t1 ⎪
⎪
⎪ ⎪
⎪
⎨ (−1, −1) for t1 ≤ t < t2 ⎬
u(t) = (u1 (t), u2 (t)) = (1, −1) for t2 ≤ t < t3 , (8.60)
⎪
⎪ ≤ ⎪
⎪
⎪
⎩ (1, 1) for t3 t < t 4 ⎪
⎭
(−1, 1) for t4 ≤ t ≤ tf
where 0 = t0 < t1 < t2 < t3 < t4 < t5 = tf . Since the initial point (q1 (0), q2 (0), ω1 (0), ω2 (0)) is
specified, the optimization variable in the optimization problem (8.8), respectively, (8.18), is
z = (τ1 , τ2 , τ3 , τ4 , τ5 ), τk = tk − tk−1 , k = 1, . . . , 5.
The code NUDOCCCS yields the arc durations and switching times
t1 = 0.546174, τ2 = 1.21351, t2 = 1.75968,
τ3 = 1.03867, t3 = 2.79835, τ4 = 0.906039, (8.61)
t4 = 3.70439, τ5 = 0.185023, tf = 3.889409,
8.3. Time-Optimal Control of a Two-Link Robot 351
1 1
0.5 0.5
0 0
-0.5 -0.5
-1 -1
-1 0.5
0
-1.5
-0.5
-2 -1
-2.5 -1.5
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
Figure 8.5. Control of the two-link robot (8.53)–(8.57). (a) Control u1 . (b) Con-
trol u2 . (c) Angle q1 and velocity ω1 . (d) Angle q2 and velocity ω2 [17].
The two bang-bang control components as well as the four state variables are shown in
Figure 8.5. The strict bang-bang property (6.19) and the inequalities (6.14) hold in view of
For the terminal conditions in (8.55), the Jacobian in the optimization problem is computed
as the (4 × 5) matrix
⎛ ⎞
−10.8575 −12.7462 −5.88332 −1.14995 0
⎜ 0.199280 −2.71051 −1.45055 −1.91476 −4.83871 ⎟
Gz (ẑ) = ⎝
6.19355 ⎠
,
−0.622556 3.31422 2.31545 2.94349
9.36085 3.03934 0.484459 0.0405811 0
1 1
0.5 0.5
0 0
-0.5 -0.5
-1 -1
Figure 8.6. Control of the two-link robot (8.53)–(8.57): Second solution. (a) Con-
trol u1 . (b) Control u2 . (c) Angle q1 and velocity ω1 . (d) Angle q2 and velocity ω2 .
It is remarkable that this value is identical to the value of the projected Hessian for the first
local minimum. Therefore, also for the second solution we have verified that all conditions
in Theorem 7.10 hold, and thus the extremal (8.63), (8.64) is a strict strong minimum. The
phenomenon of multiple local solutions all with the same minimal time tf has also been
observed by Betts [5, Example 6.8 (Reorientation of a rigid body)].
The parameters have the following meaning: tp , cavity lifetime of the photon; , cavity con-
finement factor; β, coefficient that weights the (average) amount of spontaneous emission
coupled into the lasing mode; B, incoherent band-to-band recombination coefficient; P0 ,
carrier number without injection; q, carrier charge; Gp , gain term; #, gain compression
factor; Ntr , number of carriers at transparency. All parameter values are given in Table 8.1.
The following bounds are imposed for the injected current:
Table 8.1. List of parameters from [29]. The time unit is a picosecond [ps] = [10−12 s].
tp 2.072 × 10−12 s Gp 2.628 × 104 s−1 0.3
# 9.6 × 10−8 Ntr 7.8 × 107 β 1.735 × 10−4
P0 1.5 × 107 q 1.60219 × 10−19 C A 1 × 108 s−1
B 2.788 s−1 C 7.3 × 10−9 s−1 I0 20.5 mA
Imin 2.0 mA Imax 67.5 mA I∞ 42.5 mA
where 0 ≤ Imin < Imax . To define appropriate initial and terminal values for S(t) and N(t),
we choose two values I0 and I∞ with Imin < I0 < I∞ < Imax . Then inserting the constant
control functions I (t) ≡ I0 and I (t) ≡ I∞ into the dynamics (8.65), one can show that
there exist two asymptotically stable stationary points (S0 , N0 ) and (Sf , Nf ) with Ṡ = Ṅ = 0
such that
(S(t), N(t)) → (S0 , N0 ) for t → ∞ and I (t) ≡ I0 ,
(S(t), N (t)) → (Sf , Nf ) for t → ∞ and I (t) ≡ I∞ .
Hence, we shall impose the following initial and terminal conditions for the control process
(8.65):
S(0) = S0 , N(0) = N0 and S(tf ) = Sf , N (tf ) = Nf . (8.68)
When controlling the process by the function I (t), one goal is to determine a control func-
tion by which the terminal stationary point is reached in a finite time tf > 0. But we can set
a higher goal by considering the following time-optimal control problem:
Minimize the final time tf (8.69)
subject to the dynamic constraints and boundary conditions (8.65)–(8.68). For computation,
we shall use the nominal parameters from [29] (see Table 8.1).
For these parameters, the stationary points, respectively, initial and terminal values,
in (8.68) are computed in normalized units as
S0 = 0.6119512914 × 105 , N0 = 1.3955581328 × 108 ,
(8.70)
Sf = 3.4063069073 × 105 , Nf = 1.4128116637 × 108 .
The Hamilton–Pontryagin function is given by
S
H (S, N, ψS , ψN , I ) = ψS − + G(N , S)S + βBN (N + P0 )
tp
(8.71)
I
+ ψN − R(N ) − G(N , S)S ,
q
where the adjoint variables (ψS , ψN ) satisfy the adjoint equations
1
ψ̇S = −HS = ψS − G(N , S) + Gp (N − Ntr )#S
tp
+ ψN G(N , S) − Gp (N − Ntr )#S , (8.72)
ψ̇N = −HN = −ψS (Gp (1 − #S)S + βB(2N + P0 )) + ψN ( A + B(2N + P0 )
+ C(3N 2 + 4N P0 + P02 ) + Gp (1 − #S)S).
8.4. Time-Optimal Control of a Single Mode Semiconductor Laser 355
In view of the control law (8.74), we get the switching condition ψN (t1 ) = 0. Moreover,
since the final time tf is free and the control problem is autonomous, we obtain the addi-
tional boundary condition for a normal trajectory,
H (S(tf ), N (tf ), ψS (tf ), ψN (tf ), I (tf )) + 1 = 0. (8.76)
The optimization variable in the IOP (8.18) is
z = (τ1 , τ2 ), τ1 = t1 , τ2 = tf − t1 . (8.77)
It is noteworthy that the IOP (8.18) reduces to solving an implicitly defined nonlinear
equation: determine two variables τ1 , τ2 such that the two boundary conditions S(tf ) = Sf
and N (tf ) = Nf in (8.70) are satisfied. Thus solving the IOP is equivalent to applying a
Newton-type method to the system of equations. We obtain the following switching time,
terminal time, and initial values of adjoint variables:
t1 = 29.52274, tf = 56.89444 ps,
ψS (0) = −21.6227, ψN (0) = −340.892, (8.78)
ψS (tf ) = −4.6956, ψN (tf ) = 395.60.
The corresponding control and (normalized) state functions as well as adjoint variables are
shown in Figure 8.7. Note that the constant control I (t) ≡ I∞ has to be applied for t ≥ tf in
order to fix the system at the terminal stationary point (Sf , Nf ). Since the bang-bang control
I (t) has only one switch, Proposition 6.25 asserts that the computed extremal furnishes a
strict strong minimum. The computed trajectory is normal, because the adjoint variables
satisfy the necessary condition (6.7)–(6.11) with α0 = 1. Moreover, the graph of ψN (t) in
Figure 8.7 shows that the strict bang-bang property and D 1 (H ) > 0 in (6.14) hold in view of
φ(t) < 0 ∀ 0 ≤ t < t1 , φ(t) > 0 ∀ t1 < t ≤ tf , φ̇(t1 ) > 0.
These conditions provide first-order sufficient conditions. Alternatively, we can use Theo-
rem 7.10 for proving optimality. The critical cone is the zero element, since the computed
2 × 2 Jacobian matrix
0.199855 −0.000155599
Gz (ẑ) =
0.0 −0.00252779
356 Chapter 8. Numerical Methods for Solving the Induced Optimization Problem
40 100
0
30
-100
20
-200
10
-300
0 -400
-20 0 20 40 60 80 0 10 20 30 40 50 60
0
0 100 200 300 400 500 600
Figure 8.8. Normalized photon number S(t) for I (t) ≡ 42.5 mA and optimal I (t) [46].
is regular. Comparing the optimal control approach with the topological phase-space tech-
nique proposed in [29], we recognize that the control structure (8.75) constitutes a translation
of the latter technique into rigorous mathematical terms.
The comparison in Figure 8.8 between the uncontrolled and optimally controlled
laser shows the strength of the optimal control approach: the damped oscillations have been
completely eliminated and a substantial shortening of the transient time has been achieved.
Indeed, it is surprising that such a dramatic improvement is caused by the simple control
strategy (8.75) adopted here.
8.5. Optimal Control of a Batch-Reactor 357
Feed of B
Cooling
It is worth stressing the improvement that the optimal control approach has brought
to the problem. The disappearance of the damped oscillations allows the laser to attain
its final state in a finite time rather than asymptotically. In practical terms, one can set a
threshold value, δ, around the asymptotic level, S∞ , and consider the state attained once
S∞ − δ < S(t) < S∞ + δ holds. The parameter δ can be determined by the amount of noise
present in the system, which we do not consider in our analysis. Visually, this operation
corresponds to saying that the damped oscillation and asymptotic state are indistinguishable
below a certain level of detail, e.g., if t > 500 ps as in Figure 8.8.
With this convention, one can give a quantitative estimate of the amount of improve-
ment introduced by the control function: the asymptotic state is attained at t ≈ 50 ps, even
before the first crossing of the same level which occurs at t ≈ 70 ps (dashed line in Figure
8.8). The improvement is of the order of a factor 10!
x = (MA , MB , MC , MD , H ) ∈ R5 ,
where Mi (t) [mol] and Ci (t) [mol/m3 ] stand for the molar holdups and the molar concen-
trations of the components i = A, B, C, D, respectively. H (t) [MJ] denotes the total energy
358 Chapter 8. Numerical Methods for Solving the Induced Optimization Problem
holdup, TR (t) [K] the reactor temperature and V (t) [m3 ] the volume of liquid in the system.
The two-dimensional control vector is given by
u = (FB , Q) ∈ R2 ,
where FB (t) [mol/s] controls the feed rate of the component B while Q(t) [kW] controls
the cooling load. The objective is to determine a control u that maximizes the molar holdup
of the component C. Hence, the performance index is
Here, rj denote the reaction rates and kj the corresponding Arrhenius rate constants of both
reactions (j = 1, 2):
The reference temperature for the enthalpy calculations is Tref = 298 K and the specific
molar enthalpy of the reactor feed stream is hf = 20 kJ/mol. Initial values are given for all
state variables,
with TR defined as in (8.82). The control vector u = (FB , Q) appears linearly in the control
system (8.80) and is bounded by
The reaction and component data appearing in (8.80)–(8.82) are given in Table 8.2.
8.5. Optimal Control of a Batch-Reactor 359
Calculations show that for increasing tf the switching structure gets more and more
complex. However, the total profit of MC (tf ) is nearly constant if tf is greater than a certain
value, tf ≈ 1600. For these values one obtains singular controls. We choose the final time
tf = 1450 and will show that the optimal control has the following bang-bang structure with
0 < t1 < t2 < tf :
⎧ ⎫
⎨ (10, 0) for 0 ≤ t < t1 ⎬
u(t) = (FB (t), Q(t)) = (10, 1000) for t1 ≤ t < t2 . (8.86)
⎩ (0, 1000) for t ≤ t ≤ t ⎭
2 f
Since the initial point x(0) is specified and the final time tf is fixed, the optimization variable
in the IOP (8.18) is given by
z = (τ1 , τ2 ), τ1 = t1 , τ2 = t2 − t1 .
Then the arc-length of the terminal bang-bang arc is τ3 = 1450 − τ1 − τ2 . The code
NUDOCCCS yields the following arc-lengths and switching times:
We note that for this control the state constraint TR (t) ≤ 520 imposed in [17] does not
become active. The adjoint equations are rather complicated and are not given here explic-
itly. The code NUDOCCCS also provides the adjoint functions, e.g., the initial values
(a) control FB and scaled switching function 1 (b) control Q and scaled switching function 2
10 1000
8 800
6 600
4 400
200
2
0
0
-200
-2
-400
-4
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
molar concentration MA molar concentration MB
9000 400
350
8000
300
7000 250
6000 200
150
5000
100
4000 50
0
3000
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
molar concentrations MC and MD total energy holdup H
4000 700000
3500 600000
3000
500000
2500
400000
2000
300000
1500
1000 200000
500 100000
0 0
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
Figure 8.10. Control of a batch reactor with functional (8.79). Top row: Control
u = (FB , Q) and scaled switching functions. Middle row: Molar concentrations MA and
MB . Bottom row: Molar concentrations (MC , MD ) and energy holdup H .
Figure 8.10 (top row) clearly shows that the strict bang-bang property holds with
Thus, the second-order condition (8.20) is satisfied and Theorem 7.10 tells us that the
solution (8.86)–(8.88) provides a strict strong minimum.
8.6. Optimal Production and Maintenance with L1 -Functional 361
Let us now change the cost functional (8.79) and maximize the average gain of the
component C in time, i.e.,
MC (tf )
minimize J2 (tf , x(tf )) = − , (8.89)
tf
where the final time tf is free. We will show that the bang-bang control
(10, 1000) for 0 ≤ t < t1
u(t) = (FB (t), Q(t)) = (8.90)
(0, 1000) for t1 ≤ t ≤ tf
with only one switching point 0 < t1 < tf of the control u1 (t) = FB (t) is optimal. Since
the initial point x(0) is specified, the optimization variable in the IOP (8.18) is
z = (τ1 , τ2 ), τ1 = t1 , τ2 = tf − t1 .
Using the code NUDOCCCS, we obtain the switching and terminal times
We may conclude from Figure 8.11 that φ̇1 (t1 ) > 0 holds, while the switching function for
the control u2 (t) = Q(t) = 1000 satisfies φ2 (t) < 0 on [0, tf ]. The Jacobian for the scalar
terminal condition (8.84) is
Hence the SSC (8.20) hold and, again, Theorem 7.10 asserts that the solution (8.90)– (8.92)
is a strict strong minimum.
(a) control FB and scaled switching function 1 (b) control Q and scaled switching function 2
10 1000
8 800
6 600
4
400
2
200
0
0
-2
-200
-4
-400
-6
0 100 200 300 400 500 0 100 200 300 400 500
molar concentration MA molar concentration MB
9000 900
800
8500 700
600
8000
500
400
7500
300
200
7000
100
6500 0
0 100 200 300 400 500 0 100 200 300 400 500
molar concentrations MC and MD total energy holdup H
1800 160000
1600 140000
1400
120000
1200
100000
1000
800 80000
600 60000
400 40000
200
20000
0
0
0 100 200 300 400 500 0 100 200 300 400 500
Figure 8.11. Control of a batch reactor with functional (8.89). Top row: Control
u = (FB , Q) and scaled switching functions. Middle row: Molar concentrations MA , MB .
Bottom row: Molar concentrations (MC , MD ) and energy holdup H .
the proportion of good units of end items produced, v is the scheduled production rate
(control), and m denotes the preventive maintenance rate (control). The parameters are
α = 2 obsolescence rate, s = 4 demand rate, and ρ = 0.1 discount rate. The control bounds
and weights in the cost functional will be specified below. The dynamics of the process is
given by
ẋ(t) = y(t)v(t) − 4, x(0) = 3, x(tf ) = 0,
(8.93)
ẏ(t) = −2y(t) + (1 − y(t))m(t), y(0) = 1,
with the following bounds on the control variables:
Recall also that the terminal condition x(tf ) = 0 implies the nonnegativity condition
x(t) ≥ 0 for all t ∈ [0, tf ]. Now we choose the following L1 -functional: Maximize the
8.6. Optimal Production and Maintenance with L1 -Functional 363
under the constraints (8.93) and (8.94). Though the objective has to be maximized, we
discuss the necessary conditions on the basis of the Minimum Principle which has been
used throughout the book. Again, we use the standard Hamilton–Pontryagin function instead
of the current value Hamiltonian:
where ψ1 , ψ2 denote the adjoint variables. The adjoint equations and transversality con-
dition yield, in view of the terminal constraint x1 (tf ) = 0 and the salvage term in the cost
functional,
ψ̇1 = −e−0.1 , ψ1 (tf ) = ν,
(8.97)
ψ̇2 = −ψ1 v + ψ2 (2 + m), ψ2 (tf ) = −10e−0.1 ,
where ν ∈ R is an unknown multiplier. The switching functions are given by
Singular controls for which either φv (t) = 0 or φm (t) = 0 holds on an interval I ⊂ [0, tf ]
cannot be excluded a priori. However, for the data and the final time tf = 1 chosen here, the
application of direct optimization methods [5, 13, 14] reveals the following control structure
with four bang-bang arcs:
⎧ ⎫
⎪ (3, 0) for 0 ≤ t < t1 ⎪
⎨ ⎬
(0, 0) for t1 ≤ t < t2
(v(t), m(t)) = . (8.99)
⎩ (0, 4) for t2 ≤ t < t3
⎪ ⎪
⎭
(3, 4) for t3 ≤ t ≤ tf = 1
z = (τ1 , τ2 , τ3 ), τ1 = t1 , τ2 = t2 − t1 , τ3 = t3 − t2 .
Therefore, the terminal arc-length is given by τ4 = tf −(τ1 +τ2 +τ3 ). The code NUDOCCCS
yields the following numerical results for the arc-lengths and adjoint variables:
0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(c) production control v and switching function v (d) maintenance m and switching function m
3 4
2 3
1
2
0
1
-1
0
-2
-3 -1
-4 -2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Figure 8.12 clearly indicates that the strict bang-bang property (6.19) holds, since in partic-
ular we have
φ̇v (t1 ) > 0, φ̇v (t3 ) < 0, φ̇m (t2 ) < 0.
For the scalar terminal condition x(tf ) = 0, the Jacobian is the nonzero row vector
from which the reduced Hessian (8.20) is obtained as the positive definite (2 × 2) matrix
∗ 20.4789 6.61585
N Lzz (β̂, ẑ)N = .
6.61585 49.6602
Hence the second-order test (8.20) holds, which ensures that the control (8.99) with the
data (8.100) yields a strict strong minimum.
8.7. Van der Pol Oscillator with Bang-Singular Control 365
It follows that the order of a singular arc is q = 1; cf. the definition of the order in [49].
Moreover, the strict Generalized Legendre Condition holds:
3 4
∂ d2 ∂
(−1)q Hu =− φ̈ = 1 > 0.
∂u dt 2 ∂u
366 Chapter 8. Numerical Methods for Solving the Induced Optimization Problem
0.5 0.6
0.4
0 0.2
0
-0.5
-0.2
-0.4
-1
-0.6
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Figure 8.13. Control of the van der Pol oscillator with regulator functional.
(a) Bang-singular control u. (b) State variables x1 and x2 .
The application of direct optimization methods yields the following control structure with
two bang-bang arcs and a terminal singular arc:
⎧ ⎫
⎨ −1 for 0 ≤ t < t1 ⎬
u(t) = 1 for t1 ≤ t < t2 . (8.108)
⎩ ⎭
2x1 (t) − x2 (t)(1 − x1 (t) ) for t2 ≤ t ≤ tf = 4
2
The corresponding control and state variables are shown in Figure 8.13. The code
NUDOCCCS furnishes the Hessian of the Lagrangian as the positive definite matrix
194.93 −9.9707
Lzz (ẑ) = .
−9.9707 0.56653
At this stage we have only found a strict local minimum of the IOP (8.28). Recently, SSC
for a class of bang-singular controls were obtained by Aronna et al. [2]. Combining these
new results with the Riccati approach in Vossen [111] for testing the positive definiteness of
quadratic forms, one can now verify the SSC for the extremal solutions (8.108) and (8.109).
Here, the following property of the switching function is important:
φ(t) > 0 for 0 ≤ t < t1 , φ̇(t1 ) < 0, φ(t) < 0 for t1 < t < t2 ,
φ(t2 ) = φ̇(t2 ) = 0, φ̈(t2− ) < 0, φ(t) = 0 for t2 < t < tf .
Bibliography
367
368 Bibliography
IEEE Conference on Decision and Control and European Control Conference 2005,
pp. 929–933, Sevilla, December 2005. [339]
[11] A. Bressan, A high order test for optimality of bang-bang controls, SIAM Journal
on Control and Optimization, 23, pp. 38–48 (1985).
[12] A. E. Bryson andY. C. Ho, Applied Optimal Control, Revised Printing, Hemisphere
Publishing Corporation, New York, 1975. [219]
[14] C. Büskens and H. Maurer, SQP-methods for solving optimal control problems
with control and state constraints: Adjoint variables, sensitivity analysis and real-
time control, Journal of Computational and Applied Mathematics, 120, pp. 85–108
(2000). [6, 219, 250, 287, 288, 291, 339, 363]
[16] C. Büskens and H. Maurer, Sensitivity analysis and real-time control of para-
metric optimal control problems using nonlinear programming methods, In: Online
Optimization of Large Scale Systems (M. Grötschel, S. O. Krumke, J. Rambau, eds.),
pp. 57–68, Springer-Verlag, Berlin, 2001. [219]
[22] D. I. Cho, P. L. Abad, and M. Parlar, Optimal production and maintenance de-
cisions when a system experiences age-dependent deterioration, Optimal Control
Applications & Methods, 14, pp. 153–167 (1993). [4, 248, 249, 250, 361]
[26] A. V. Dmitruk, Jacobi-type conditions for the Bolza problem with inequalities,
Mat. Zametki, 35, pp. 813–827 (1984). [4, 191, 193]
[30] A. Ya. Dubovitskii and A. A. Milyutin, Theory of the maximum principle, In:
Methods of Extremal Problem Theory in Economics [in Russian], Nauka, Moscow,
pp. 6–47 (1981). [166, 167]
[32] A.V. Fiacco, Introduction to Sensitivity and Stability Analysis in Nonlinear Pro-
gramming, Mathematics in Science and Engineering 165, Academic Press, New
York, 1983. [338]
[34] R. Gabasov and F. Kirillova, The Qualitative Theory of Optimal Processes, Marcel
Dekker Inc., New York, Basel, 1976.
370 Bibliography
[49] A. J. Krener, The high order maximum principle, and its application to singular
extremals, SIAM Journal on Control and Optimization, 15, pp. 256–293 (1977).
[365]
[50] U. Ledzewicz, H. Maurer, and H. Schättler, Optimal and suboptimal proto-
cols for a mathematical model for tumor anti-angiogenesis in combination with
chemotherapy, Mathematical Biosciences and Engineering, 8, pp. 307–328 (2011).
[343]
[51] U. Ledzewicz, H. Maurer, and H. Schättler, On optimal delivery of combination
therapy for tumors, Mathematical Biosciences, 22, pp. 13–26 (2009). [343]
[52] U. Ledzewicz, J. Marriott, H. Maurer, and H. Schättler, The scheduling of
angiogenic inhibitors minimizing tumor volume, Journal of Medical Informatics and
Technologies, 12, pp. 23–28 (2008).
[53] U. Ledzewicz and H. Schättler, Optimal bang-bang controls for a 2-compartment
model in cancer chemotherapy, Journal of Optimization Theory and Applications,
114, pp. 609–637 (2002). [270]
[54] E. S. Levitin, A. A. Milyutin, and N. P. Osmolovskii, On local minimum con-
ditions in problems with constraints, In: Mathematical Economics and Functional
Analysis [in Russian], Nauka, Moscow, pp. 139–202 (1974). [3, 301]
[55] E. S. Levitin, A. A. Milyutin, and N. P. Osmolovskii, Higher-order local min-
imum conditions in problems with constraints, Uspehi Mat. Nauk, 33, pp. 85–148
(1978). [3, 21, 174, 301]
[56] E. S. Levitin, A. A. Milyutin, and N. P. Osmolovskii, Theory of higher-order con-
ditions in smooth constrained extremal problems, In: Theoretical and Applied Opti-
mal Control Problems [in Russian], Nauka, Novosibirsk, pp. 4–40, 246 (1985). [3]
[57] K. Malanowski, C. Büskens, and H. Maurer, Convergence of approximations
to nonlinear optimal control problems, In: Mathematical Programming with Data
Perturbations (A.V. Fiacco, ed.), Lecture Notes in Pure and Applied Mathematics,
Vol. 195, pp. 253–284, Marcel-Dekker, Inc., New York, 1997.
[58] K. Malanowski and H. Maurer, Sensitivity analysis for parametric control prob-
lems with control-state constraints, Computational Optimization and Applications,
5, pp. 253–283 (1996). [219]
[59] K. Malanowski and H. Maurer, Sensitivity analysis for state constrained optimal
control problems, Discrete and Continuous Dynamical Systems, 4, pp. 241–272
(1998). [219]
[60] P. Martinon and J. Gergaud, Using switching detection and variational equa-
tions for the shooting method, Optimal Control Applications and Methods, 28,
pp. 95–116 (2007). [339]
[61] H. Maurer, Numerical solution of singular control problems using multiple
shooting techniques, Journal of Optimization Theory and Applications, 18, pp.
235–257 (1976). [339]
372 Bibliography
[62] H. Maurer, First and second order sufficient optimality conditions in mathemat-
ical programming and optimal control, Mathematical Programming Study, 14,
pp. 163–177 (1981). [25]
[64] H. Maurer and D. Augustin, Second order sufficient conditions and sensitivity
analysis for the controlled Rayleigh problem, In: Parametric Optimization and
Related Topics IV (J. Guddat et al., eds.) Lang, Frankfurt am Main, pp. 245–259
(1997). [219, 339]
[65] H. Maurer and D. Augustin, Sensitivity analysis and real-time control of paramet-
ric optimal control problems using boundary value methods, In: Online Optimization
of Large Scale Systems (M. Grötschel et al., eds.), pp. 17–55, Springer Verlag,
Berlin, 2001. [203, 219, 339]
[66] H. Maurer, C. Büskens, J.-H. R. Kim, and Y. Kaya, Optimization methods for the
verification of second-order sufficient conditions for bang-bang controls, Optimal
Control Methods and Applications, 26, pp. 129–156 (2005). [6, 339, 340, 341, 342,
357]
[68] H. Maurer and H. J. Oberle, Second order sufficient conditions for optimal con-
trol problems with free final time: The Riccati approach, SIAM Journal on Control
and Optimization, 41, pp. 380–403 (2002). [219]
[69] H. Maurer and N. P. Osmolovskii, Second order optimality conditions for bang-
bang control problems, Control and Cybernetics, 32, pp. 555–584 (2003). [267]
[73] H. Maurer and S. Pickenhain, Second order sufficient conditions for optimal
control problems with mixed control-state constraints, Journal of Optimization
Theory and Applications, 86 (1995), pp. 649–667 (1995). [219, 267]
Bibliography 373
[74] H. Maurer and G. Vossen, Sufficient conditions and sensitivity analysis for
bang-bang control problems with state constraints, In: Proceedings of the 23rd IFIP
Conference on System Modeling and Optimization, Cracow, Poland (A. Korytowski,
M. Szymkat, eds.), pp. 82–99, Springer Verlag, Berlin, 2009. [342, 343]
[75] H. Maurer and J. Zowe, First and second-order necessary and sufficient condi-
tions for infinite-dimensional programming problems, Mathematical Programming,
16, pp. 98–110 (1979). [25]
[76] A. A. Milyutin, Maximum Principle in the General Optimal Control Problem [in
Russian], Fizmatlit, Moscow (2001). [179]
[77] A. A. Milyutin, A. E. Ilyutovich, N. P. Osmolovskii, and S. V. Chukanov,
Optimal Control in Linear Systems [in Russian], Nauka, Moscow (1993). [293]
[78] A. A. Milyutin and N. P. Osmolovskii, (1) Higher-order minimum conditions
on a set of sequences in the abstract problem with inequality-type constraints;
(2) Higher-order minimum conditions on a set of sequences in the abstract problem
with inequality- and equality-type constraints; (3) Higher-order minimum condi-
tions with respect to a subsystem in the abstract minimization problems on a set of
sequences, In: Optimality of Control Dynamical Systems [in Russian], No. 14, All-
Union Institute of System Studies, Moscow, pp. 68–95 (1990); English translation:
Computational Mathematics and Modeling, 4, pp. 393–400 (1993). [9, 16]
[79] A. A. Milyutin and N. P. Osmolovskii, Calculus of Variations and Optimal Con-
trol, Translations of Mathematical Monographs, Vol. 180, American Mathematical
Society, Providence, 1998. [1, 2, 3, 4, 5, 30, 32, 71, 102, 103, 104, 105, 106, 107,
108, 109, 115, 127, 128, 130, 131, 207, 225, 229, 232, 255, 293, 294, 296, 297, 298]
[80] J. Noble and H. Schättler, Sufficient conditions for relative minima of bro-
ken extremals in optimal control theory, Journal of Mathematical Analysis and
Applications, 269, pp. 98–128 (2002). [270]
[81] H. J. Oberle, Numerical computation of singular control functions for a two-link
robot arm, In: Optimal Control, Lecture Notes in Control and Information Sciences,
95, pp. 244–253 (1987). [339, 346, 347, 349]
[82] H. J. Oberle and W. Grimm, BNDSCO—A program for the numerical solution of
optimal control problems, Institute for Flight Systems Dynamics, DLR, Oberpfaffen-
hofen, Germany, Internal Report No. 515–89/22, 1989. [204, 273, 287, 288, 291, 339]
[83] H. J. Oberle and H. J. Pesch, Private communication, 2000. [272]
[84] N. P. Osmolovskii, Second-order weak local minimum conditions in an optimal
control problem (necessity, sufficiency), Dokl. Akad. Nauk SSSR, 225, pp. 259–262
(1975); English translation: Soviet Math. Dokl., 15, pp. 1480–1484 (1975). [25]
[85] N. P. Osmolovskii, High-order necessary and sufficient conditions for Pontryagin
and bounded strong minima in the optimal control problems, Dokl. Akad. Nauk
SSSR Ser. Cybernetics and Control Theory, 303, pp. 1052–1056 (1988); English
translation: Soviet Physics Doklady, 33, pp. 883–885 (1988). [1, 2]
374 Bibliography
[104] A. V. Sarychev, First- and second-order sufficient optimality conditions for bang-
bang controls, SIAM Journal on Control Optimization, 35, pp. 315–340 (1997). [279]
[106] J. Stoer and R. Bulirsch, Introduction to Numerical Analysis, 2nd ed., Springer-
Verlag, Berlin, 1992. [206, 339]
[112] G. Vossen, Switching time optimization for bang-bang and singular controls,
Journal of Optimization Theory and Applications, 144, pp. 409–429 (2010). [6, 365]
376 Bibliography
absorbing set of sequences, 9 B, 33, 99, 142, 149, 154, 155, 228,
abstract set of sequences , 9 242, 263, 265, 278, 297
adjacent conditions, 13 B(), 94, 174, 181
admissible function (t, u), 51 Bco (), 94
admissible pair, 28, 164, 184 Bτ , 156
admissible pair (variation) with respect to BV , 149
Qtu , 52 B0 , 266, 304
admissible process, 224, 238, 274 ℵγ , 21
admissible trajectory, 256 cone Ct , 177
almost global minimum, 293 Ct (), 176
arc lengths, 341 C , 68
arc-parametrization method, 342, 343 conjugate point, 188, 206
augmented Hamiltonian, 166 constant CK , 74
augmented Pontryagin function, 150, 166, Cγ (m, g ), 11
178 Cγ (ω0 , K), 25
auxiliary minimization problem, 186 Cγ (0 , K), 22
auxiliary problem V , 138 Cγ (0 , σ
), 22
Cγ (0 , σ
γ ), 22
basic bang-bang control problem, 255 Cγ̄ (C ; S4 ), 89
basic constant, 3, 13, 38 Cγ̄ (C ; S5 ), 92
basic problem, 255, 299 Cγ (C ; o(√γ ) ), 76
batch-reactor, 357
bounded strong γ -sufficiency, 120, 139 Cγ (C ; loc √
o( γ ) ), 79
√
Cγ (M ; o( γ ) ), 101
bounded strong γ1 -sufficiency, 231
bounded strong minimum, 117, 154, 173, Cγ (0 , σ γ ), 13, 22, 38
180, 185, 226, 240 Cγ (˜ 1 ; S1 ), 84
C
bounded strong sequence on Q, 173 Cγ (˜ 1 ; loc√ , 83
C o( γ )
Cγ1 (2C ; S2 ), 85
C-growth condition for H , 173
canonical problem, 27 Cγ2 (2C ; S2 ), 87
canonical representation of a sequence, Cγ2 (3C ; S3 ), 88
41, 42 Cγ (σ , ), 12
closure [· ]2 , 73 convexification of control system, 230
condition A, 33, 115, 137, 153, 173, 180, critical cone, 22, 25, 31, 132, 153, 170,
227, 242, 263, 278, 294 180, 227, 241, 260, 301
Aco , 74 K0 , 261
Aτ , 156 K0 in the induced problem, 303
A0 , 304 K τ , 158
377
378 Index
˜ 1 , 83
local sequence, 39
C
˜ 1λ , 82
Lyusternik condition, 11
σ , 10, 100 Lyusternik normalization, 14
k
u0 , 39 singular control, 365
V ∗ , 39 smooth problem, 11
Vk+∗ , 39 smooth system, 11
Vk−∗ , 39 space L1 (, Rd(u) ), 29
Vk∗ , 39 L2 (, Rd(u) ), 31
V 0 , 39 L∞ ([t0 , tf ], Rr ), 24
%, 296 L∞ (, Rd(u) ), 28
set of active indices, 11, 24, 31, 165, 170, P C 1 (, Rd(x) ), 260
227, 241, 301 P C 1 (, Rn ), 276
set of indices I , 56 P W 1,2 ([t0 , tf ], Rm ), 185
I in the IOP, 303 P W 1,2 ([t0 , tf ], Rd(x) ), 170
I ∗ , 34 P W 1,1 (, Rd(x) ), 60
S
set of sequences , 117 P W 1,2 (, Rd(x) ), 30
, 29 Rn∗ , 24
, 54 W , 24, 28, 164, 184, 224
∗ , 41 W 1,1 ([t0 , tf ], Rn ), 24
∗uk+ , 41 W 1,∞ (, (Rd(x) )∗ ), 30
∗uk− , 40 W 1,1 (, Rd(x) ), 28
∗uk , 41 Z(), 67, 89
∗u , 41 Z2 (), 30, 170, 186, 216, 226
+ , 10 Z(), 260
+ g , 10 Z2 (), 180, 241
0 , 29, 34 standard normalization, 12, 13
S , 116, 173 strict bang-bang control, 228, 259, 276
loc √
o( γ ) , 76
strict bounded strong minimum, 139, 154,
173, 180, 226, 240
loc 1
σ γ , 57 strict minimum on a set of sequences, 10,
loc , 39 29, 127, 165
loc
u , 39 strict minimum principle, 33, 154, 185
loc
σ γ , 54 strict order, 11, 21
0 , 9 strict Pontryagin minimum, 128
g , 10 strict strong minimum, 154, 259, 275
σ γ , 38 strictly -differentiable mapping, 10
σ
, 22 strictly Legendre element, 33, 141, 228,
σ γ , 13 242
o(√γ ) , 76 strong minimum, 116, 154, 173, 259, 275
S 2 , 65 subspace K# , 212
S 3 , 65 K(τ ), 191
S 4 , 66 K0 , 188
S1 , 84 L0 , 68
S2 , 85 L̃2 ([τ , tf ], Rm ), 195
S3 , 88 T , 92
S4 , 89 support of a function, 40
S5 , 91 switching function, 225, 257, 275
simplest problem of the calculus of vari- for the u-component, 239
ations, 1, 183 system S, 10
382 Index
value ρ, 264
Van der Pol oscillator, 365
time-optimal control of a, 286
variable ζ of the IOP, 300
variational problem, 205
¯ 241
vector p̄,
violation function, 10, 38, 100, 139, 174,
230