The Characteristic Method and Its Generalizations

CHAPMAN & HALL/CRC
Monographs and Surveys in

Pure and Applied Mathematics I 01
THE CHARACTERISTIC
METHOD AND ITS
GENERALIZATIONS FOR
FIRST-ORDER NONLINEAR
PARTIAL DIFFERENTIAL
EQUATIONS
TRAN DUC VAN

MIKIO TSUJI
NGUYEN DUYTHAI SON
CHAPMAN & HALL/CRC
Monographs and Surveys in Pure and Applied Mathematics
Main Editors
H. Brezis, Universitéde Paris
R.G. Douglas, Texas A&M University
A. Jeffrey, University of Newcastle upon Tyne (Founding Editor)
Editorial Board
H. Amann, University of ZUrich
R. Aris, University of Minnesota
G.I. Barenblatt, University of
H. Begehr, Freie Universitat Berlin
P. Bullen, University of British Columbia
R.J. Elliott, University of Alberta
R.P. Gilbert, University of Delaware
R. Glowinski, University of Houston
D. Jerison, Massachusetts Institute of Technology
K. Kirchgassner, Universitat Stuttgart
B. Lawson, State University of New York
B. Moodie, University of Alberta
S. Mon. Kyoto University
L.E. Payne, Cornell University
D.B. Pearson, University of Hull
I. Raeburn, University of Newcastle
G.F. Roach, University of Strathclyde
1. Stakgold, University of Delaware
WA. Strauss, Brown University
J. van der Hock, University of Adelaide
CHAPMAN & HALUCRC
IC Monographs and Surveys in

Pure and Applied Mathematics I0I
THE CHARACTERISTIC
METHOD AND ITS
GENERALIZATIONS FOR
FIRST-ORDER NONLINEAR
PARTIAL DIFFERENTIAL
EQUATIONS
TRAN DUCVAN
MIKIOTSUJI
NGUYEN DUY THAI SON
CHAPMAN & HALL/CRC

Boca Raton London New York Washington, D.C.
Library of Congress Cataloging-in-Publication Data
Tran, Due Van.

The characteristic method and its generalizations for first-order nonlinear
partial differential equations / Tran Due Van, Mikio Tsuji. and Nguyen Duy Thai
Son.
p cm. -- (Chapman & Hall/CRC mongraphs and surveys in pure and
applied mathematics 101)
Includes bibliographical references and index.
ISBN 1-58488-016-3 (alk. paper)
I. Differential equations. Nonlinear--Numerical solutions.
I.Tsuji, Mikio. II. Nguyen, Duy Thai Son. 111. Title.
IV. Series.
QA374.T65 1999
519.5'.353--dc2I 99-27321
CIP
This book contains information obtained from authentic and highly regarded
sources. Reprinted material is quoted with permission, and sources are indicated.
A wide variety of references are listed. Reasonable efforts have been made to
publish reliable data and information, but the author and the publisher cannot
assume responsibility for the validity of all materials or for the consequences of
their use.
Neither this book nor any part may be reproduced or transmitted in any form
or by any means, electronic or mechanical, including photocopying, microfilming.
and recording, or by any information storage or retrieval system, without prior
permission in writing from the publisher.
The consent of CRC Press LLC does not extend to copying for general distri-
bution, for promotion, for creating new works, or for resale. Specific permission
must be obtained in writing from CRC Press LLC for such copying.
Direct all inquiries to CRC Press LLC. 2000 N.W. Corporate Blvd., Boca Raton,
Florida 33431.
Trademark Notice: Product or corporate names may be trademarks or regis-

tered trademarks, and are used only for identification and explanation, without
intent to infringe.
© 2000 by Chapman & Hall/CRC
No claim to original U.S. Government works

International Standard Book Number 1-58488-016-3
Library of Congress Card Number 99-27321
Printed in the United States of America I 23 4567890
Printed on acid-free paper
Contents
Contents
Preface
Chapter 1. Local Theory on Partial Differential
Equations of First-Order 1
1.1. Characteristic method and existence of solutions 1
1.2. AtheoremofA.Haar 7
1.3. A theorem of T. 8
Chapter 2. Life Spans of Classical Solutions of Partial
Differential Equations of First-Order 12
2.1. Introduction 12
2.2. Life spans of classical solutions 13
2.3. Global existence of classical solutions 18
Chapter 3. Behavior of Characteristic Curves and
Prolongation of Classical Solutions 22
3.2. Examples 23
3.3. Prolongation of classical solutions 24
3.4. Sufficient conditions for collision of characteristic curves I 26
3.5. Sufficient conditions for collision of characteristic curves II 28
Chapter 4. Equations of Hamilton-Jacobi Type in One
Space Dimension 30
4.1. Nonexistence of classical solutions and historical remarks 30
4.2. Construction of generalized solutions 34
4.3. Semi-concavity of generalized solutions 38
4.4. Collision of singularities 41
Chapter 5. Quasi-linear Partial Differential Equations of
First-Order 44
5.1. Introduction and problems 44
5.2. Difference between equations of the conservation law and
equations of Hamilton-Jacobi type 47
5.3. Construction of singularities of weak solutions 48
5.4. Entropy condition 52
Chapter 6. Construction of Singularities for Hamilton-
Jacobi Equations in Two Space Dimensions 55
CONTENTS
6.2. Construction of solutions 56

6.3. Semi-concavity of the solution u = u(t, x) 61
6.4. Collision of singularities 63
Chapter 7. Equations of the Conservation Law without
Convexity Condition in One Space Dimension 67
7.2. Rarefaction waves and contact discontinuity 68
7.3. An example of an equation of the conservation law 70
7.4. Behavior of the shock S1 74
7.5. Behavior of the shock S2 77
Chapter 8. Differential Inequalities of Haar Type 82
8.2. A differential inequality of Haar type 84
8.3. Uniqueness of global classical solutions to the Cauchy
problem 91
8.4. Generalizations to the case of weakly-coupled systems 95
Chapter 9. Hopf's Formulas for Global Solutions of
Hamilton-Jacobi Equations 103
9.2. The Cauchy problem with convex initial data 105
9.3. The case of nonconvex initial data 113
9.4. Equations with convex Hamiltonians I = f(p) 121
Chapter 10. Hopf-Type Formulas for Global Solutions in
the case of Concave-Convex Hamiltonians 127
10.2. Conjugate concave-convex functions 128
10.3. Hopf-type formulas 136
Chapter 11. Global Semiclassical Solutions of First-Order
Partial Differential Equations 146
11.2. Uniqueness of global semiclassical solutions to the Cauchy
problem 148
11.3. Existence theorems 156
Chapter 12. Minimax Solutions of Partial Differential
Equations with Time-measurable Hamiltonians 161
12.2. Definition of minimax solutions 165
12.3. Relations with semiclassical solutions 173
12.4. Invariance of definitions 176
12.5. Uniqueness and existence of minimax solutions 179
12.6. The case of monotone systems 191
Chapter 13. Mishmash 198
13.1. Hopf's formulas and construction of global solutions via
characteristics I 9R
CONTENTS
13.2. Smoothness of global solutions 205

13.3. Relationship between minirnax and viscosity solutions 208
Appendix I. Global Existence of Characteristic Curves 214
Appendix II. Convex Functions, Multifunctions, and
Differential Inclusions 217
AII.1. Convex functions 217
AII.2. Multifunctions and differential inclusions 222
References 227
Index 236
Preface
One of the main results of the classical theory of first-order partial differential
equations (PDEs) is the characteristic method which asserts that under certain
assumptions the Cauchy problem can be reduced to the corresponding characteristic
system of ordinary differential equations (ODEs). To illustrate this, let us consider
the Cauchy problem for the nonviscid Burger equation:
,3u Ou
—+u----=O,
Ot Ox
t>O,XER,
u(O,x)=h(x), XER.
We try to reduce the problem (1)-(2) to an ODE along some curve x = x(t). More
precisely, let us find x = x(t) such that
x(t)) = x(t)) + u(t, x(t)).
By the chain rule, we may simply require dx/dt = u, and so the characterist2cs
x = x(t) can be defined by
dx
= u(t,x).
Along each characteristic x = x(t) we have du/dt = 0, i.e., ti = u(i,x(t)) takes a
constant value and then the characteristic must be a straight line with slope given
by(3).
Thus, by the initial data (2), the characteristic passing through any given point
(0,s) on the x-axis is
x = a + h(s)t,
on which u has the constant value:
u = h(s).
Hence, if the C1-norm of h = h(s) is bounded, then, by means of the implicit

function theorem and (4), we can get
S = .s(t,x)
PREFACE
for small values of t. Substituting (6) into (5) gives the classical solution (C1-
solution)
u = h(s(t, z)) (7)
to our Cauchy problem (1 )-(2). However, in general, this solution exists only locally
in time. In fact, if h = h(s) is not a nondecreasing function of s, there exist two
points and (0,52) on the x-axis such that
<s2 and h(s1) > h(s2). (8)
Then the characteristic curves beginning from (0, Si) and (0, s2) will intersect at
time
—
Since the solution u = u(t, x) is constant along each of the two curves but has
different values h(si) and h(s2), respectively, at the intersection point, the value
of the classical solution cannot be uniquely determined. Hence, in this case the
Cauchy problem (1)-(2) never admits a global classical solution on {t O}; in fact,
the classical solution will blow up in a finite time no matter how smooth and small
the initial data h = h(s) are.
On the other hand, if h = h(s) is a nondecreasing function of s, then the
characteristics emanating from distinct points (0, and (0,82) on the s-axis wiU
not intersect, and thus the solution u = u(t, x) will exist globally for t 0.
The previous example shows, generally speaking, that for (first-order) nonlinear
partial differential equations or systems, classical solutions to the Cauchy problem
exist only locally in time, while singularities may occur in a finite time, even if the
initial data are sufficiently smooth and small. Therefore, the notions of generalized
solutions or weak solutions have been introduced. In fact, the global existence and
uniqueness of generalized solutions have been well studied from various kinds of
viewpoints.
In the 1950s — 1970s, the theory and methods for constructing generalized so-
lutions of first-order PDEs were discovered by Aizawa, S. [2]-[4], Bakhvalov, Ben-
ton, S.H. 121], Conway, E.D. 1321-1331, Douglis, A. [44]-[45], Evan, C., Fleming, W.H.
[511-1521, Friedman, A. [541, Celfand, I.M., Codunov, S.K., Hopf, E. [63]-[64j,
Kuznetsov, N.N. [95], Lax, P.D. [97]-[98], Oleinik, O.A. [112), Rozdestvenskii, B.L.
[118], and other mathematicians. Among the investigations of this period we should
mention the results of Kruzhkov, S.N. ([87)-[92], [94)), which were obtained for
Hamilton-Jacobi equations with convex Hamiltonian. The global existence and
uniqueness of generalized solutions for convex Hamilton-Jacobi equations were well
studied by several methods: variational method, method of envelopes, vanishing
viscosity method, nonlinear semi-group method, etc.
Since the early 1980s, the concept of viscosity solutions introduced by Crandall
and Lions has been used in a large portion of research in a nonclassical theory of
first-order nonlinear PDEs as well as in other types of PDEs. The primary virtues
of this theory are that it allows merely nonsmooth functions to be solutions of
nonlinear PDEs, it provides very general existence and uniqueness theorems, and it
PREFACE
yields precise formulations of general boundary conditions. Let us mention here the
names: Crandall, M.G., Lions, P.-L., Aizawa, S., Barbu, V., Bardi, M., Barles, G.,
Barron, E.N., Cappuzzo-Dolcetta, I., Dupuis, P., Evans, L.C., Ishii, H., Jensen, R.,
Lenhart, S., Osher, S., Perthame, B., Soravia, P., Souganidis, P.E., Tataru, D.,
Toinita, Y., Yamada, N., and many others (see [51, [10]-[20], [28], [351-139], [471-1501,
[67]-[72], [79], [99]-[1O1], [122]-[123J, [1311, and the references therein), whose con-
tributions make great progress in nonlinear PDEs, and where the global existence
and uniqueness of viscosity solutions have been established almost completely. The
concept of viscosity solutions is motivated by the classical maximum principle which
distinguishes it from other definitions of generalized solutions.
Another direction in the theory of generalized solutions is motivated by dif-
ferential game theory as suggested by A.!. Subbotin. This leads to the notion of
minimax solutions of first-order nonlinear PDEs. As the terminology "minimax
solutions" indicates, the definition of such global solutions is closely connected with
the minimax operations. This definition is based, to some extent, on the so-called
"characteristic inclusions" (a generalization of the classical characteristic system in
this situation). Subbotin and his coworkers (11], [124]-[127], [1291-11301) developed
an effective theory of minimax solutions to first-order single PDEs and gave nice
applications to control problems and differential games. The research of rninimax
solutions employs methods of nonsmooth analysis, Lyapunov functions, dynamical
optimization, and the theory of differential games. At the same time, the research
contributes to the development of these branches of mathematics. A review of
the results on minimax solutions and their applications to control problems and
differential games is given in [124]-[125].
We also want to mention the investigations on PDEs based on the idempotent
analysis and Cole-Hopf transformation, which have been discovered by V.P. Maslov
and his coworkers. Indeed, V.P. Maslov, V.N. Kolokol'tsov, S.N. Sainborskii, and
others developed a nonclassical approach to define the weak global solutions to first-
order nonlinear PDEs, in which by a suitable structure of new function semimodules,
nonlinear operators become "linear" ones. In this direction, based on the methods
and results of the well-developed linear mathematical physics, investigations of
first-order nonlinear PDEs with convex Hasniltonian have been considered (see [82],
[105]).
Concerning the theory of differential inequalities, let us mention that the theory
of ordinary differential inequalities was originated by Chaplygin [30] and Kaznke
181], and then developed by [157]. Its main applications to the Cauchy
problem for (ordinary) differential equations concern questions such as: estimates
of solutions and of their existence intervals, estimates of the difference between two
solutions, criteria of uniqueness and of continuous dependence on initial data and
right sides of equations for solutions, Chaplygin's method and approximation of so-
lutions, etc. Results in this direction were also extended to (absolutely continuous)
solutions of the Cauchy problem for countable systems of differential equations sat-
isfying Carathéodory's conditions. We refer to Szarski [128] for a systematic study
of such subjects.
PREFACE
As for the theory of partial differential inequalities, the first achievements were
obtained by Haar [611, Nagumo [107], and then by [154]. Up to now the
theory has attracted a great deal of attention. (The reader is referred to Deim-
ling [40], Lakshniikanthain and Leela [96], Szarski [1281, and Walter [153), for the
complete bibliography.)
It must be pointed out that the characteristic method gives us the local existence
and uniqueness of classical solutions to first-order nonlinear PDEs. We would like
to use this method as an important basis for setting the global generalized solutions.
This book is devoted to some developments of the characteristic method and mainly
represents our results on first-order nonlinear PDEs.
Our aim in the first seven chapters is to fill a gap between the local theory
obtained by the characteristic method and the global theory which principally de-
pends on vanishing viscosity method. This is to say, we try to extend the smooth
solutions obtained by the characteristic method. Our first problem then is to deter-
mine the life spans of the smooth solutions. Next, we want to obtain the generalized
solutions or weak solutions by explicitly constructing their singularities.
In Chapter 1, we present the classical results which are necessary for our follow-
ing discussions: the characteristic method, existence of local solutions, and Theo-
rems of Haar and on the uniqueness of solutions to the Cauchy problem
in C1 -space.
Chapter 2 is devoted to the life spans of classical solutions of the noncharacter-
istic Cauchy problem. Our method depends on the analysis of the smooth mapping
obtained by the family of characteristic curves. Even if the Jacobian of the mapping
may vanish at some point, we can sometimes extend the classical solution beyond
the point where the Jacobian vanishes. Therefore, we are obliged to consider very
often some properties of the inverse of the mapping in a neighborhood of a singular
point. This is the subject of Chapter 3.
In Chapters 4 and 5, we consider the extension of solutions beyond the singu-
larities of solutions in the case where the dimension of space is equal to one. Then
our principal problem is to construct the singularities of generalized solutions or of
weak solutions.
The theme of constructing the singularities of solutions is picked up again in
Chapter 6 for convex Hamilton-Jacobi equations in two space dimensions. The
difference between Chapters 4-5 and this one is the dimension of space. But the
problem caused by this difference would become much more complicated.
In Chapter 7, we treat the case where equations are not convex and not concave.
We will give only typical phenomena appearing in these cases without developing
a general theory.
In the last six chapters, the first-order PDEs under consideration are always
assumed to satisfy Carathéodory's conditions (or something like them). Chapter
8 introduces the so-called differential ineqtzalities of Haar type. Here we use some
new techniques based on the theory of multifunctions and differential inclusions to
investigate the uniqueness problem. The idea originates from the generalized char-
acteristic method. Roughly speaking, "characteristic differential inclusions" and
PREFACE
"characteristic bundles" are invoked instead of characteristic differential equations

and characteristic curves.
Chapters 9-10 are devoted to the study of Hopf-type formulas for global solutzons
to the Cauchy problem in the case of non-convex, non-concave Harniltonians or
initial data. In Chapter 9, we first consider the case where the initial data can
be represented as the minimum of a family of convex functions, and next the case
where it is a d.c. function (i.e., it can be represented as the difference of two convex
functions). In Chapter 10, the Hazniltonians are concave-convex functions.
The method of Chapters 8-10 allows us to deal with global solutions, the con-
dition on whose smoothness is relaxed significantly. In Chapter 11, we propose the
notion of global semiclassical solutions, which need only be absolutely continuous
in the time variable, and investigate their uniqueness and existence. By the way,
an answer to an open uniqueness problem of S.N. Kruzhkov 193] is given.
In Chapter 12, we extend the notion of Subbotin's minimax solutions to the case
of first-order nonlinear PDEs with time-measurable Hamiltonian. The uniqueness
and existence of such solutions are investigated by the theory of multifunctions and
differential inclusions. Our road here is devious (by some "perturbation technique"
on sets of Lebesgue measure 0), and proceeds via an implicit version of Gronwall's
inequality and via a sharpening of a well-known theorem on the Lebesgue sets for
functions with parameters. The results are new even when restricted to the case
of continuous Hamiltonians. Generalizations for monotone systems will also be
considered.
Finally, in Chapter 13 we examine Hopf's formulas in relations with the con-
struction of global solutions via characteristics and the smoothness of the solutions.
In this chapter, the relationship between minimax and viscosity solutions is also
investigated.
We have to say that this book is not designed as an introduction to, or a guide-
book on, the general theory of first-order nonlinear PDEs. Our goal is not to try to
cover as many subjects as possible, but rather to concentrate on some basic facts
and ideas of the generalized characteristic methods for studying global solutions.
Suitable as a text, the book is self-contained and assumes as prerequisites only
calculus, linear algebra, topology, ODEs, and basic measure theory. In the ap-
pendices at the end of the book we collect necessary facts, mostly of nonsmooth
analysis and the theory of differential inclusions.
The authors are grateful to Professor H. Begehr for his proposal that we prepare
this book for Chapman & Hall / CRC Monographs and Surveys in Pure and Applied
Mathematics.
The work of the first and third named authors was supported in part by the Na-
tional Center for Natural Science and Technology and the National Basic Research
Program in Natural Science, Vietnam. Some parts of this monogragh were written
and typeset at Ohio University (USA) where the third named author was a guest
for two quarters in 1998 by invitation of the Department of Mathematics, College
of Arts and Sciences. We record here our gratitude for all the support and help we
received from these institutions.
Chapter 1
Local Theory on
Partial Differential Equations
of First-Order
§ 1.1. Characteristic method and existence of solutions
Partial differential equations of first-order have been studied from various points of
view: for example, classical mechanics, variational method, geometrical optics, etc.
In this chapter we will always suppose that the equations and solutions are real-
valued. The classical method to solve the equations is the characteristic method.
As this is the fundamental tool in our following discussions, we will give here a brief
explanation of the method. For more detailed results and geometrical meanings,
refer to, for example, R. Courant and D. Hilbert [34] and F. John [80].
First we consider a quasi-linear partial differential equation of first-order as
follows:
+ = ao(t,x,u) in U,
tz(0,x)=4)(x) on
U is an open neighborhood of (i, x) = (0,0). Let V be an open neighborhood
of {(0,x,4)(x)) : x U0} in Assume that a1 = a,(t,x,u) (1 = 0,1,... ,n)
and 4) = 4)(x) are of class C1 in V and U0, respectively. A function is said to be
of class Gil if it is k-times continuously differentiable, and Ck(U) is the family of
functions being of class C" in U. A C"-function means that it is a function of class
Ck.
Characteristic curves of (1.1)-(1.2) are defined by solution curves of the following
system of ordinary differential equations:
!.a1(i,r,v)
J
= ao(t,x,v).
1. LOCAL THEORY
In accordance with (1.2), the initial condition for (1.3) is given by
x,(O) = (i = 1,2,... ,n), v(O) = (1.4)
The ordinary differential equations in (1.3) are called the "characteristic equations"
for (1.1), where we use v = v(t,y) instead of u =u(t,x) to avoid confusion. In the
following discussions, u = u(i, x) is a solution of (1.1) and v v=(i, y) is a solution of

(1.3)-(1.4) which is equal to the value of ii = u(t,x) restricted on the corresponding
solution curve x = x(t,y) of (1.3)-(1.4). As a1 = a1(t,x,v) (i = O,1,...,n) and
= are of class C1 in V and (Jo, respectively, the Cauchy problem (1.3)-(1.4)
has a system of solutions x1 = x1(i, y) (i = 1,2,..., n) and v = v(t, y) which are of
class C1 inaneighborhoodof{(O,y) yEUo}. :
Let us fix our notations on derivatives of functions. A vector x is vertical,

i.e., x = Therefore dx/dt = On
the other hand, given any real-valued function q5 = q5(x), we write grad4(x) =
= qV(x) = ,8c6/Ox,1). For an n-vector valued function
x = x(y) of an n-vector y, we define its Jacobi matrix and Jacobian, respectively,
by
Ox1 Ox1 Ox1
0111 0112 OYn
df
a III
am axn
01,1 0112 01/n
and
Dx del (Ox1
= det
Y \ Y3 ij=1,2 a
We will sometimes write the Jacobi matrix simply by Since x(O, y) = y,
we see that (Dx/Dy)(t,y) = 1 for t = 0, y E Uo. In a neighborhood of {(O,y)

y E Uo}, as (Dx/Dy)(t,y) does not vanish, we can uniquely solve the equation
x = x(t,y) with respect toy and write the solution by y = y(t,x). Putting u(t, x)
v(t,y(t,x)), we will prove that u = u(t,x) satisfies (1.1)-(1.2) in a neighborhood of
the origin.
Theorem 1.1. The Cauchy problem (1.1)-(1.2) has uniquely a solution of class C1
in a neighborhood of the origin.
§1.1. CHARACTERISTIC METHOD AND EXISTENCE OF SOLUTIONS 3
Proof. We use the notations introduced in the above. The following discussions are
true only in the definition domain of = y(t, x). This domain is a neighborhood of
the origin, where the Jacobian (Dx/Dy)(t,y) does not vanish. As x = x(t,y(t,x)),
we have
(p!-) = I (identity matrix), (1.5)

Yj n .
x3 i,j=1,2
Ox Oy —
i,j=1,2 n
As u(t,x) = v(t,y(t,x)), we have
Ou Ov OvOy
= + (1.7)
By (1.7), using (1.5) and (1.6), we get
Ou Ov Ox
= ao(t,x,u(t,x)) — —
\
Ou Ox
= ao(i,x,u(t,x)) —
= ao(t,x,u) —
As u(O,x) = v(O,y(O,x)) = cS(s), we see that u = u(t,x) satisfies the Cauchy

problem (1.1 )-( 1.2) in the above neighborhood of the origin.
We will show the uniqueness of solutions. Let u = u(t, x) be any solution of class
C' of(1.1)-(1.2), and put where x = x(t,y) and v = v(t,y) are
the solutions of (1.3)-(1.4). Then the difference w(t,y) — v(t,y) satisfies
the following Cauchy problem:
= (a,(t,x,v) — + — ao(i,x,v))
J
L. w(O,y) = 0.
As the right-hand side of this differential equation can be estimated by — vi =
Miwl, we get w(t,y) 0, i.e. v(i,y) for any (t,y) in a neighborhood of

the origin. This means that the solution of C'-class is unique along the curves
x = x(t,y). That is to say, as long as the Jacobian (Ds/Dy)(t,y) does not vanish,
the solution of (1.1)-(1.2) is unique in the C'-space. 0
1. LOCAL TREORY
Next we consider the Cauchy problem for general partial differential equations
of first-order as follows:
in U (1.8)
on
U is an open neighborhood of the origin. Let V be an open neighborhood of

x E U0} in R x RTh x R x
: Assume that I = f(t,x,u,p)
and = are of class C2 in V and U0, respectively.
Characteristic strips for (1.8)-( 1.9) are defined as solution curves of the following
system of ordinazy differential equations:
= (i = 1,2,... ,n),
dv " Of
= (1.10)
at
dp1 Of Of
= (z = 1,2,... ,n),
with
x,(0)=y,, v(0)=cb(y), (i=1,2,... ,n). (1.11)
We remark that system (1.10) is called the "characteristic system of differential

equations," or simply "characteristic equations," for equation (1.8). As I and
are of class C2, the Cauchy problem (1.10)-(1.11) has uniquely the solutions x =
x(t,y),v = v(t,y) and p = p(t,y) in a neighborhood oft = 0. Moreover, they are
of class C1 with respect to (t, y). As x(0, y) = y, we have (Dx/Dy)(0, y) = 1 for
any y E Uo. Therefore, there exists an open neighborhood W of {(0, y) y E Uo}
:
where the Jacobian (Dx/Dy)(t, y) does not vanish and the equation x = x(t, y) can
be uniquely solved with respect to y. Denote the solution by y = y(t, x), and put
u(t,x) def .
= v(t,y(t,x)). We will prove that u = u(t,x) is of class C 2
,and that it
satisfies the Cauchy problem (1.8)-(1.9). For this aim, we prepare some lemmas.
Lemma 1.2. For all (t,y) in the existence domain of solutions to (1.10)-(1.11),
we have
§1.1. CHARACTERISTIC METHOD ANI) EXISTENCE OF SOLUTIONS 5
OX1
81/i 01/2
::: ::: ::: ::: , (1.12)
oxn
0Y18Y28Yn
where p(t, y) = (pi(t, y), p2(t, y),... , y)).
Proof. We put
On Ox1
OYi 0112 01/n
deC t9V
z(t,y) = —p(t,y)
01/2 01/n
Using (1.10), we have
y) =
(
z(0,y)=0.
As this is a linear ordinary differential equation concerning z = z(t,y), we get
z(t,y)wO. 0
Remark. Fix any y E Uo, and let J C R be an interval around 0 on which the
solutions x = x(t,y),v = v(t,y) and p = p(t,y) of the characteristic equations
(1.10)-(1.11) exist. Then (1.12) is true for each t E J even if the Jacobian may
vanish at (i,y).
Recall that in W we have (Dx/Dy)(t, y)0, and we can uniquely solve the
equation x = x(t, y) with respect to y. The solution has been denoted by y = y(t, x)
deC
and used to define u(t,x) = v(t,y(t,x)).
Corollary 1.3. In the definition domain of y = y(t,x), we have
= (i = 1,2,... ,n).
Using Lemma 1.2 and its corollary, we get the following:

1. LOCAL THEORY
Theorem 1.4. Suppose f 6 C2 and 6 C2. Then the Cauchy problem (1.8)-(1.9)
has uniquely a solution of class C2 in a neighborhood of the origin.
Proof. We first prove the existence of solutions. Let u(t,x) v(t,y(t,x)) as in
the above notations. By Lemma 1.2 and Corollary 1.3, using (1.10), we have
Ou i9v Oy
x) = y(t, x)) + y(t, x)) . x)
= — f(t,x,v,p)—p(t,y(t,x))
$ 1<i,j<n
1$
= — f(i,x,v,p) —p(t,y).
= —f(t,x,v(t,y(t,x)),p(t,y(t,x)))
Ou
= —f(t, x, u(t, x), x)).
Moreover, by (1.11), u(0,x) = v(0,y(0,x)) = q5(x). It follows that u = u(t,x) is a

solution of (1.8)-(1.9). As = (i = 1,2,... ,n) are of class C1 in W and
y = y(t,x) is of class C' in its definition domain, we see by (1.13) that u = u(t,x)
is of class C2 in the definition domain of y = y(t,x). This domain is actually a
neighborhood of the origin.
Finally, we give the sketch of a proof of the uniqueness of solutions. Let u =
u(i, be a solution of class C2 of(1.8)-(1.9), and z = z(t, y) = (zi(t, y),... , y))
be a solution of
Of Ou
= (, = 1,2 n),
I. z1(0) =
def def
We put w(t, y) = u(t, z(t, y)) and q(t, y) = (Ou/Ox)(t, z(t, y)), then we get
= —f(t,z,w,q) +
= — (z = 1,2,...
with w(0,y) = çb(y) and q(0,y) = qV(y). Hence
(z, w, q) = (z(t, y), w(t, y), q(t, !J))

§1.2. A THEOREM OF A. HAAR 7
satisfies the Cauchy problem (1.1O)-(1.11). By the uniqueness of solutions of ordi-

nary differential equations, we have
r(t, y) = z(t, y), v(i, y) = w(t, y), p(i, y) = q(t, y),
where (x,v,p) = (x(t,y),v(t,y),p(t,y)) is the solution of (1.1O)-(1.11) which has

already appeared in Lemma 1.2. This says that the solution of class C2 is unique
along the curves x = x(i, y). Therefore, as long as the Jacobian (Dx/Dy)(t, y) does
not vanish, the solution of class C2 of (1.8)-(1.9) is unique. 0
§ 1.2. A theorem of A. Haar
We have seen by Theorem 1.4 that the Cauchy problem (1.8)-(1.9) has a solution of
class C2 in a neighborhood of the origin. But, as the equation (1.8) is of first-order,
we would like to discuss the existence and uniqueness of solutions in C1-space. In
the following, we write a solution of class as a C's-solution. C'-solutions will be
sometimes called "classical solutions."
One of our aims is to consider what kinds of phenomena may appear when we
extend the classical solutions of (1.8)-(1.9). In this procedure, we need the unique-
ness of solutions in C1-space. This subject had been well studied by A. Haar [611
and T. Walewski [154]-[155]. Moreover, their results are indispensable to develop
our discussions. But, as it seems to us that they are not well-known, we would
like to introduce them. In this section, we report a theorem of A. Haar [61]. We
consider the Cauchy problem (1.8)-(1.9) in one space dimension, that is x R'.
Let be a triangle defined by
wherea>O, LO, c<dand2La<d—c.

{(u,p) E R2}.
Theorem 1.5. (A. Haar [61]) Suppose that the function I = f(t,x,u,p) satisfies
a Lipschitz condition as follows:
If(t,x,u,p) — f(t,x,v,q)I Lip— qI +Miu — vi

1. LOCAL THEORY
for all (t,x,u,p) and (t,x,v,q) in x Let u3 = u,(t,z) (j = 1,2) be sn

and (u1(t,x),(Ou,/Ox)(t,x)) E for all (t,z) E If u,(t,x) (j = 1,2) satisfy
the equation (1.8) in the domain and ui(O,x) = u2(O,z) for x E [c,d], then
uj(i,x) in
Proof. According to the Lipschitz continuity off = f(t,x,u,p), the difference

z(t,x) — u2(t,z) satisfies the inequality
L + Mlz(t, x)I. (1.14)
Assume that z = z(t,x) takes some positive values in We put w(t,x)

fora > M. As w = w(t,x) is continuous on there exists a point
P E at which w = w(t, x) attains the maximum. The point P is not on the initial
line {t = 0) because w(O,x) 0. Let us consider the derivatives of w = w(t,x) at
P with respect to the directions (—1,--L) and (—1,L). As w = w(t,s) takes the
maximum at the point P, we have
— <0, — <0.
Hence it holds that
9w 8w
Rewriting this for z = z(t,x), we get

Or Oz
az(P) + L
As z(P) > 0 and a > M, this contradicts (1.14). Therefore z(t,x) < 0 for
all (t,x) Since we can similarly prove that z(t,x) 0, we obtain ui(t,x)
0
§ 1.3. A theorem of T. Wazewski
In this section we report a theorem of T. Waiewski [154] which is a generalization

of Haar's theorem to arbitrary space dimensions.
Let us consider the Cauchy problem (1.8)-(1.9) in a pyramid which is defined
by
= {(t,x) O t <a, <d1 —Let (i = 1,2,... ,n)},

§1.3. A THEOREM OF T. WAZEWSKI
wherea>O, L1O, ,n.

is a compact set in {(u,p) : u E R, p R't}.
Theorem 1.6. (T. Waiewski [154J) Suppose that I = f(t,x,u,p) satisfies the
following Lipschitz condition:
if(t,x,u,p) — — + MItz — vi
for (t,x,u,p) and (t,x,v,q) in x u = u(t,x) (j = 1,2) be in

Let
and (u,(t,x),(Ou,/ôx)(t,x)) all(t,x) If u = u,(t,x) (j = 1,2) are
solutions of (1.8) in the domain and ui(t,z) = u2(t,x) on fl {t = 0), then
uj(t,x) u2(t,x) in
As preparation for the proof of this theorem, we give a fundamental lemma

which plays an important role in his many works.
Lemma 1.7. Let be the pyramid defined in the above, and {x E R"
(t,x) E Let u = u(t,x) be in and put xE :
for each t E [0,a]. Then w = w(t) is differentiable from the right on [0,a), and
n
= —
for some x E with w(t) = u(t,x).
Proof. As the first step, consider the case where L1 = 0, cj = —1 and d1 = 1

for all i = 1,2,... ,n. By the definition of w = w(t,x), we easily see that it is
continuous. Let U(t) {x w(t) = u(t,x)} for t E [0,a). Then the sets
U(t) (0 < t < a) are closed, bounded, and non-empty. As u = u(t,x) is in
for any fixed t0 E [0,a), we can pick up a point x° E U(t°) C such that
0) — max —(t° a,)

Ot ZEU(t0)Ot
We will prove that w = w(t) is differentiable from the right at t° with =

For any x E U(t°), we have
w(t°) = tz(t°,x) and

10 1. LOCAL THEORY
Define z(t) [w(t) — w(t°)]/(t — t°), and put

def. def.
/3 = hmsupz(t) and Cl = hminfz(t).
g-4t0
As w(t) — w(t°) u(t,x°) — u(t°,x°), we see that & (Ou/Ot)(t°,x°). By the

definition of/3, we can pick up a sequence {tm}m C (t°,a) with urn tm = t° and
limz(t"t) = /3. For each tm, take arbitrarily xm E U(tm), so that u(tm,xm) =
w(tm). Since the set {(tm,xm) m = 1,2,... } is bounded, we can assume that
:
the sequence {(tm, Xm)}m is convergent to a point (t°, in As the functions

u = u(t,x) and w = w(t) are continuous, this implies that u(t0,°) = w(t°), i.e.,
E U(t°), hence that (Ou/&)(t°,x°). By the definition of
w = w(t), we have w(t°) u(t°, xtm) for all m = 1,2 Using this inequality, we
get
z(tm) = w(tm) — w(t°)

ttm — to
— u(tm,xm) —w(t°)
— tm_to
_u(tO,xm)
< —
— tm_to — Ot'
0 0 0 Ott
with m-Ioo
lim (t ,xm) = ). It follows that /3 < ) < —(t°,x°). There-
Ot at
fore,
<a
So w = w(t) is differentiable from the right at t0, and = (Ou/Ot)(t°,z°).
Next, we consider the same problem in a general pyramid To do so, we take
the transform of coordinates as follows:
{ — — (i = 1,2,... ,n).
Then the set is mapped to the set {(s,y) 0 s < a, —1 < : 1 (i =
1,2,... , n)}. Combining the result obtained in the first step and the above trans-
form of coordinates, we complete the proof of the lemma. 0
§1.3. A THEOREM OF T. WAZEWSKI 11
Proof of Theorem 1.6. We put u(t,x) — u2(t,x), and define w(t)

max{u(t,x) xE For t E [O,a), let r(t) E be such that w(t) = u(t,x(t))
= —
(cf. Lemma 1.7). Then
= x(t)) — — clu(t, . (1.15)

x(t)) }
{
On the other hand, it follows from the hypotheses of the theorem that
+ MIu(t, (1.16)
If we choose o > M, then from (1.15)-(1.16) we get

f (Mlw(t)I —
1. w(O) = 0,
from which it may be concluded that w(t) 0, i.e., u1(t,x) < u2(t,x), for all
(t, x) Since we can similarly prove uz(t, x) u1(t, x), we finally get ui(t, x)
u2(t,x) on 0
We will rewrite Theorem 1.6 in a general form. A function f = f(t,x,u,p) is
said to be locally Lipschitz continuous with respect to (u,p) if, for any compact set
in R x Rx there exist constants L1,L2,... and M such that
—vi
for all (t,x,u,p) and (t,x,v,q) in fi..
Theorem 1.8. Suppose that f = f(t, x, u,p) is locally Lipschitz continuous with
respect to (u,p). If u = u(t,x) and v = v(i,x) are C1 -solutions of (1.8) in a
neighborhood of the origin satisfying u(0,r) v(O,x), then there exists an open
neighborhood of the origin, such that u(t,x) v(t,x) in
We will leave the proof to the readers.
Remark. The uniqueness of C'-solutions is assured by the Lipschitz continuity of

f = f(t,x,u,p) with respect to (u,p). But this condition is not sufficient to get the
solvability of the Cauchy problem (1.8)-(l.9). 11561 has given necessary
and sufficient conditions which guarantee the local existence and uniqueness of
classical solutions of (1.8)-(1.9).
Chapter 2
Life Spans of Classical
Soluti9ns of Differential
Equations of I-irst-Order
§2.1. Introduction
As we have shown in Chapter 1, the Cauchy problems (1.1)-(1.2) and (1.8)-(1.9)

have locally classical solutions. It is well-known that the solutions may generally
have singularities in finite time even for smooth initial data. For some equations of
the conservation law, the life spans of classical solutions have been exactly calcu-
lated, for example by E. Hopf [63], P.D. Lax [97]-[98], E.D. Conway [32], etc. The
principal aim of this chapter is to determine the life spans of classical solutions of
general partial differential equations of first-order. Moreover, using the results on
the life spans of classical solutions, we will give necessary and sufficient conditions
which guarantee the global existence of classical solutions for (1.1)-(1.2) and (1.8)-
(1.9). As an example, we consider a simple equation of the conservation law as
follows:
'I
+ = 0 in {t > 0, x (2.1)
u(O,x) = on {t = 0, x (2.2)
where a1 = a(u) (i = 1,2,... ,n) and = are of class C1 in Rand RTh,

respectively.
The equation treated in E. Hopf [63] is the case where n = 1 and ai(u) = u.
E.D. Conway considered the above equation (2.1)-(2.2) in [32]. In this case, the
characteristic curves which are the solutions of (1.3)-(1.4) are written by
x= y + v(t, y) = (2.3)
wherea(u) def
= (ai(u),az(u),... Then the Jacobian of the mappmgx =
x(t,y) is given by (Dx/Dy)(t,y) = 1 + tA(y) with A(y)
Obviously, the Jacobian (Dx/Dy)(t, y) does not vanish in a neighborhood of (t, y) =
§2.2. LIFE SPANS OF CLASSICAL SOLUTIONS 13
(0,0). Therefore, as we have shown in §1.1, the Cauchy problem (2.l)-(2.2) has a
unique C1-solution u = u(t,r) in a neighborhood of the origin. Since u(t,x) =
x)) where y = y(t, z) is the solution of the equation x = x(t, y), we have
f'Ou 1
24
' 1 + '
Assume A(y°) <0. Then the Jacobian (Dx/Dy)(t°,y°) = 0 for t° l/(—A(y°)).

We see by (2.4) that, when (t, x) goes to (t°, x°) along the curve x = x(t, y°) where
0det
z = x(t o ,yo ), at least one of the first derivatives
.
of u = u(i,x) tends to infixuty.
Therefore, if the Jacobian vanishes somewhere, then the Cauchy problem (2.1)-
(2.2) can not admit a global solution of class C1. In §2.2 and §2.3, we will give
similar results on the life spans of classical solutions for general partial differential
equations of first-order. What we would like to remark here is to point out that there
exist some differences between quasi-linear equations and general partial differential
equations. See Theorem 2.1 and Theorem 2.2. In §2.4, we will consider the global
existence of classical solutions. These existence results are corollaries of theorems
given in §2.2 and §2.3.
§2.2. Life spans of classical solutions
Let us consider a quasi-linear partial differential equation of first-order as follows:
+ = ao(i,x,u) in {t >0, x E R't}, (2.5)
u(0,x) = on {t = 0, x E W'}, (2.6)
where a = a1(t,x,u) (i = 0,1 n) and 4 = are of class C1 in R x R

and respectively. The characteristic equations for (2.5)-(2.6) are written by
I = a(t,x,v) (i = 1,2,... ,n),

(2.7)
= ao(t, x, v),
with the initial conditions
= y1 (i = 1,2,... ,n), v(0) = (2.8)

14 2. LIFE SPANS OF CLASSICAL SOLUTIONS
We write the solutions of (2.7)-(2.8) by r = x(t, y) and v = v(t, y). Then v = v(t, y)
means the value of a solution of (2.5) restricted on the curve x = x(t, y). Here we
assume the following condition:
(A.I) The Cauchy problem (2.7)-(2.8) has uniquely a global solution x = x(t,y),
v = v(t,y) on {t O} for any y E
It is not easy to write down sufficient conditions which guarantee the assump-
tion (A.I). Concerning this subject, see for example B. Doubnov [43J and M.S.
Krasnosel'skii [83J. We will consider this in Appendix I. When we assume (A.!), we
get a smooth mapping x = x(t,y) from R" to RTh for each t 0. The life span of
classical solutions of (2.5)-(2.6) is determined by the following:
Theorem 2.1. Under Cond;t*on (A.!), suppose that (Dx/Dy)(t°,y°) = 0 and

(Dx/Dy)(t,y°) 0 fort < t°. Then at least one of the first derivatives
of the solution u = u(t,x) tends to infinity when t goes to t0 — 0 along the curve
x = x(t, y°).
Proof. : : x=x(t,y°),0<
t <t°}. By the assumption and by the theorem of inverse functions, we can get an
open neighborhood V of L so that the Jacobian (Dz/Dy)(t, y) does not vanish on
V = V Ii {O < t <t o} and that the equation x = x(t,y) can be uniquely solved
dot
with respect to y for any (t,x) in U = {(t,x) x = x(t,y), (t,y) E V}. Let
:
us write the solution by y = y(t,x) and define u(t,x) We already

proved in §1.1 that u = u(t,x) is a C1-solution of(2.5)-(2.6) in the domain U. The
uniqueness of C' -solution of (2.5)-(2.6) is assured by Theorem 1.1. Here we have
. U
on —
Ov
,n V. (2.9)
On the other hand, from (2.7), and (i,j = 1,2,... ,n) satisfy
the following system of linear equations:
d —
Osk Ov
+
dt — Lank 12 10
Oa0Ov
dt Oy,) — Oy. + Ov Oyj
§2.2. LIFE SPANS OF CLASSICAL SOLUTIONS 15
with the initial data

11 if z=j, Ov Oçb
= and =
0 if
By the linearity of (2.10), we have
0
oy.
rank =rank •.. =n forall t0, (2.11)

0
Oy" Oy
Ox, / Ox1 0x1\ Ov /Ov Ov

where we write and —
= 0111
01, , —) 0Y = 0111 (31Jn
Assume that all the components of Ou/Ox remain bounded along the curve C0,
then we can pick up a sequence {tm}m C and one c = (ci,c2,... E
such that
1) lim t'Th = j0, and
2) lim =c.
Ox
From (2.9'), we get
= (2.12)
As (Dx/Dy)(t°,y°) = 0, (2.12) contradicts (2.11). This means that at least one

component of (Ou/Ox)(t, x(i, y°)) tends to infinity when t —0 along C0. 0
Next we consider general partial differential equations of first-order as follows:
in {t>O, xER"}, (2.13)
u(0,x) = on {t = 0, x E (2.14)
where f = f(t,x,u,p) and 4 = are of class C2 in R x Rx and

respectively. The characteristic equations for (2.13)-(2.14) are written by
2. LIFE SPANS OF CLASSICAL SOLUTIONS
= (i = 1,2,... ,n),
— = —f(t,x,v,p), (2.15)
di j1 VP,
dp1 Of Of
= — (z = 1,2,... ,n),
with
x1(O) = y1, v(O) = = (i = 1,2,... ,n). (2.16)
We assume here the following condition:
(A.!)' The Cauchy problem (2.15)-(2.16) has uniquely a global solution on {i O}

for any y E
We denote the solution of (2.15)-(2.16) by x = x(t,y), v = v(t,y), and p =

p(t,y). As (Dz/Dy)(O,y) = 1 for all y E RTh, the Jacobian does not vanish in a
neighborhood of {i = O}. Therefore we can uniquely solve the equation x = x(t, y)
with respect toy, and denote it by y = y(t,x). Define u(t,x) Then
u = u(t, x) is a C2-solution of (2.13)-(2.14) in a neighborhood oft = 0. Moreover,
we can see by Theorem 1.4 that there does not exist another C'-solution of (2.13)-
(2.14) in a neighborhood of {t = 0). When we extend this solution for large t, we
get the following.
Theorem 2.2. Under Condition (A.!)', suppose that (Dx/Dy)(t°,y°) = 0 and

" O2ti
(Dz/Dy)(t,y°) 0 for t < t°. Then E (t,x) tends to infinity when i
i,j=1 UX;UXJ
goes to —0 along the curve x = z(i, y°).
Proof. As the proof is similar to that of Theorem 2.1, we use the same notations
introduced there. Put L {(t, y°) 0: t< Next we choose an open
def
neighborhood V of L so that the Jacobian (Dx/Dy)(t, y) does not vanish on V =
V fl {0 I <I°), and that the equation x = x(t,y) can be uniquely solved with
respect to y for any (t,x) in U {(t,z) x = x(t,y), (t,y) E V} (theorem of
:
inverse functions). We write the solution by y = y(t,x). Here we define u(t,x)

v(t,y(t,x)). Then u = u(i,x) is a C2-solution of (2.13)-(2.14) in the domain U.
S2.2. LIFE SPANS OF CLASSICAL SOLUTIONS 17
Our aim is to show that, when t goes to t0 —0 along the curve x = x(t, y°), at least
one of {Ou/0x1, 02u/Ox1Ox, : i,j = 1,2,... ,n} tends to infinity.
Differentiating (2.15) with respect to y, (j = 1,2,... ,n), we get a system of
linear ordinary differential equations of first-order concerning {0x1/0y1, .3v/Oy,,
i,j = 1,2,... ,n} like (2.10). As this system of equations is linear, we
have
(0z1/Oy)(t) (Oxi/Oy)(O)
(Ox,1/Ôy)(t) (ôXm/Oy)(O)
rank = rank (Ov/äy)(0) = n for any t 0. (2.17)
(0p1/Oy)(t) (Opl/Oy)(O)
(Opfl/Oy)(0)
Since (2.17) is true at (t,y) = (t°,y°), we can choose n vectors in {8x1/äy,...

such that they are linearly independent at
(t, y) = (t°, y°). We denote them by a1(t, y) (8/Oy)b1(t, y) (i = 1,2,. . , n), where .
b1(t, y) is one of {xk(t, y), v(t, y), pk(t, y) : k = 1,2,... , n}. As (Dx/Dy)(t°, y°) =
0, we can find v(t,y) or some pk(t,y) in Here we recall
Corollary 1.2:
=p,(t,y(t,x)) in U (j = 1,2,...,n). (2.18)
On the other hand, we get in U
ai(t, y)
r o 1
a2(t,y)
= ... -(t,x)
1
=
y)
Since {a1 (t, y), a2(i, y),... , a neighborhood of

(t°,y°) and (Dx/Dy)(t°,y°) = 0, it follows that at least one component of the
2. LIFE SPANS OF CLASSICAL SOLUTIONS
matrix tends to infinity when t goes tot0 —o along the

curve x = x(t,y°). But we see by (A.!)' and (2.18) that
y(t, x))} = x(t, y°)) = p,(t,y°) = 1,2

{ z-z(*y°)
remain bounded (though the Jacobian vanishes); therefore, some pa(t, y) must be
contained in {bi(t,y),b2(t,y),... Hence we get Theorem 2.2. 0
Theorem 2.2 says that, if the Jacobian vanishes at a point (t°, y°), then the
second derivatives of classical solutions blow up at time t = t°. But this does
not prevent the existence of C1-solutions even if the Jacobian may vanish. To
understand this situation, we need to know the behavior of characteristic curves in
a neighborhood of the point where the Jacobian vanishes. After having studied this
subject, we will again consider the extension of classical solutions in Chapter 3.
§2.3. Global existence of classical solutions
As we have shown in §1.1, the Cauchy problem for a partial differential equation
of first-order has locally a classical solution. We have also seen in §2.2 that the
solution may generally have singularities in finite time even for smooth initial data.
But in some cases the Cauchy problem admits a global classical solution. In this
section, following Tsuji-Li [1391, we will give necessary and sufficient conditions
which guarantee the global existence of classical solutions. First we give some
comments on this subject.
For the Cauchy problem (2.1)-(2.2), E.D. Conway (32] gave a necessary and
sufficient condition for the global existence of C1-solutions on the half space {t O},
using Hadamard's lemma on the diffeomorphism of Euclidean spaces. Another proof
was given by Li Ta-tsien and Chen Shu-xing [102J. Qin Tie-hu [1141 treated the
case where the functions a (z = 1,2,... ,n) depend on t and x. Li Ta-tsien and
Shi Jia-hong [103] gave certain relations between the global existence of nontrivial
smooth solutions in the whole (t, x)-space and the linear dependence of the functions
aj = aj(u) =
Now we review the results on the diffeomorphism of the Eucidean space
Concerning this subject, we must go back to J. Hadamard, but here we follow W.B.
Gordon's note [58]. Let H be a C1-mapping from R" to R" defined on the whole
§2.3. GLOBAL EXISTENCE OF CLASSICAL SOLUTIONS
space. Then the mapping H is said to be proper if is compact whenever

K is compact. We have:
Lemma 2.3. (J. Hadainard) The mapping H is a diffeomorphssm from R" to

if and only if H is proper and the Jacobian of H does not vanish anywhere on the
whole space.
First we consider the Cauchy problem for a quasi-linear equation, say Prob-
lem (2.5)-(2.6). We assume the following conditions. (Here and subsequently, ri
denotes the Euclidean norm of x E
(A.I) For any y E the Cauchy problem (2.7)-(2.8) has always a unique global
C'-solution x = x(t,y), v = v(t,y) on {t O}.
(A.II) On the solution curves of (2.7)-(2.8), it holds that
a1(t, x(t, y), v(t, v))i +
for any t [O,T], y E and i = 1,2,... ,n where the constant M depends only
on T.
When Conditions (A.1) and (A.Il) are satisfied, we can define, for any t 0, a
C -mapping
1
from R to R by x = x(t,y) def = Ht(y). Then we get:
Lemma 2.4. Under the assumptions (A.I) and (A.!!), the mapping is proper
for any t 0.
Proof. As we have
Ix(t, y)a2(t, x(t, y), v(t,

)I
—nM(1 + z(t,y)l),
we get Ix(t,y)i (1 + — 1. Therefore —* 00 as 00. This

means that is proper. 0
Using Lemma 2.3, we have:
Lemma 2.5. For any fixed t 0, the mapping is a diffeomorphism from to

if and only if its Jacobian does not vanish at any point y €
20 2. LIFE SPANS OF CLASSICAL SOLUTIONS
By means of Lemma 2.5, we obtain the following.
Theorem 2.6. Under the assumptions (A.I) and (A.!!), the Cauchy problem (2.5)-
(2.6) has a global C' -solution uniquely on the domain D {t 0, x E if and
only if the Jacobian (Dx/Dy)(t,y) of the mopping does not vanish for any t 0
and y E RTh.
Proof. When (Dx/Dy)(t,y) 0 for any y E and t 0, we can see from

Lemma 2.5 that H, is a diffeomorphism from R" to R't for any t > 0. Hence the
inverse function y = y(t, x) of x = H,(y) is a function of class C1 defined on D and
u = u(t,x) v(t,y(t,x)) is the unique global C1-solution of (2.5)-(2.6). (For the
uniqueness, follow the proof of Theorem 1.1 or 1.6.)
Next, for the necessity, suppose that (Dx/Dy)(t,y°) = 0 for some t E (0,oo)
and y° E As (Dx/Dy)(O, y) = 1 for all y E R't, we can get a unique C1-solution
of (2.5)-(2.6) in a neighborhood of {t = 0}. Then we see by Theorem 2.1 that this
classical solution blows up at time t0 > 0 (Dx/Dy)(t,y°) = 0}. (Notice
:
thatO<t°<t.) 0
Next we consider the Cauchy problem (2. 13)-(2. 14) for general partial differen-
tial equations of first order. We assume the following hypotheses:
(A.I)' For any y E the Cauchy problem (2.15)-(2.16) has always a unique global
C1-solution x = x(t,y), v = v(t,y), p = p(t,y) on the halfspace {t 0}.
(A.II)' On the solution curves of(2.15)-(2.16), it holds that
x(t, y), v(t, y), p(t, y)) .tvf( 1 + Ix(t,
for any t e [0,TI, y E andi = 1,2,... ,n where the constant M depends only
on T.
For any t 0, we define a C -mapping H, from K

1
to K by x = x(t,y) def
=
H,(y). Then Conditions (A.!)' and (A.!!)' guarantee that the mapping H, is proper.
Therefore, Lemma 2.5 is also true for this H,.
Theorem 2.7. Under the assumptions (A.!)' and (A.!!)', the Cauchy problem
(2. 13)-(2. 14) has uniquely a global C2-solutzon on the domain D {t 0, x E
§2.3. GLOBAL EXISTENCE OF CLASSICAL SOLUTIONS 21
if and only if the Jacobian of the mapping x = Hg(y) does not vanish for any t 0
and y E
Proof. When (Dx/Dy)(t,y) 0 for any t 0 and y E we can uniquely solve

the equation x = with respect toy for any (t,x) E D. Let y = y(t,x) be the
. def
inverse function. A global C 2-solution of (2.13)-(2.14) is given by u = u(t,x) =
.
v(t,y(t,x)). The uniqueness of this C2-solution follows from Theorem 1.6. (It can
also be deduced from the method of proof used in Theorem 1.4.) The necessity of
the above condition comes from Theorem 2.2. 0
Chapter 3
of Characterjstic
Lurves and Prolongation
of Classical Solutions
§3.1. Introduction
This chapter is continued from Theorem 2.2 in §2.2. Let us rewrite the equation
which we will again consider here:
=0 in {t >0, rE (3.1)
u(0,x) = 4(x) on {t = 0, x E R"}, (3.2)
where f = f(t,x,u,p) and = are of class C2 in R x Rx and

respectively. The characteristic equations corresponding to (3.1)-(3.2) are given by
(2.15)-(2.16). Let x = x(t,y), v = v(t,y), and p = p(i,y) be solutions of (2.15)-
(2.16). We consider the Cauchy problem (3.1)-(3.2) in the following situation:
(I) (Dx/Dy)(t°,y°) = 0, and (II) (Dx/Dy)(t,y°) 0 for t < t°.

0def
We put x = x(t o , y o ). Theorem 2.2 says that, when (t, x) goes to (to , x o ) along
the curve r = x(t,y°), one of the second derivatives of the solution u = u(t,x) of
(3.1)-(3.2) tends to infinity. But this does not prevent the existence of C1-solution
in a neighborhood of the point (t°, x°). Our problem is to see whether or not we
can extend the classical solution u = u(t, x) beyond the time t°. On the other
hand, we will show later that, if the characteristic curves meet in a neighborhood
of (t°,x°), then the Cauchy problem (3.1)-(3.2) cannot admit a classical solution
there. Therefore, it is necessary for us to consider whether or not the characteristic
curves meet in a neighborhood of (t°, x°), i.e., whether or not there exist two points
yi and 112 (yi y2) satisfying x(t, yi) = x(t, 112) for some t. In §3.2, we will give two
examples in which characteristic curves do not meet though the Jacobian vanishes.
In §3.3, we will consider the case where we can extend classical solutions of (3.1)-
(3.2) beyond a point where the Jacobian vanishes. In §3.4 and §3.5, we will give
EXAMPLES 23
sufficient conditions so that the characteristic curves meet in a neighborhood of the

point (t°,x°).
§3.2. Examples
Example 1. We consider the Cauchy problem for a quasi-linear partial differential

equation as follows:
(Ou Ou
=u in {t >0, xE R }, 1
(33)
I. u(0,x)=x on {t=0, xER'},
where o(t,u) + /3l(t)e_SmuS and the two functions = (t), $ = fl(t)

satisfy the following conditions:
1) o = (t) arid /3 = /3(t) are in C1(R1).

2) ci(t) 0 for each t, ci(0) = 1, and a(i) = 0 for all t K = constant > 0.
3) i3(t) 0 for each t, /3(0) = 0.
Then the characteristic curves for (3.3) are written by
x = x(t,y) = cl(t)y + /3(t)y3 and v = v(t,y) = ety, (3.4)
from which we easily see that
= (t) +3/3(t)y2, and = 0 for all t K.
But we can also see from (3.4) that the characteristic curves x = x(t, y) do not
meet for all t 0. In this case the solution u = u(t, s) is represented as
u(t, x) = for t K.
This representation says that the solution contains algebraic singularity at x = 0,

and that the singularity of shock type does not appear though the Jacobian vanishes.
24 3. PROLONGATION OF CLASSICAL SOLUTIONS
Example 2. We consider the following Cauchy problem:

(Ots Ou
I
(IX
in
.
{t>O, xER }, 1
(3.5)
on {t=O,xER'},
where
f(i,u,p) + f3'(t)e3tp4 — u
and the functions a = a(t), = are the same functions that we introduced
in Example 1. This example is not of quasi-linear type, and it satisfies Conditions
(A.!)' and (A.!!)' given in §2.3. The characteristic curves for (3.5) are written as
x = x(t, y) = ci(t)y + 4/3(t)y3 and v = v(t, y) = + 313(t)ety4.
Therefore the Jacobian (Dz/Dy)(t,y) = a(t) + 12i3(t)y2 vanishes on L {(i,y)

t K and y = O}. But x = r(t, y) is a bijective mapping defined in a neighborhood
of y = 0 for each t K. In a neighborhood of L, the solution u = u(t, z) is written
by
u(t, x) = for i K.
This says that the solution u = u(t,x) is of class C', but not of class C2, in a
neighborhood of L.
§3.3. Prolongation of classical solutions
As we have shown in §3.2, there exists the case where the characteristic curves do
not meet in a neighborhood of points where the Jacobian vanishes. In this case we
can uniquely extend classical solutions even if the Jacobian may vanish. This is the
problem which we would like to prove in this section. Now let us make clear the
situation under which we consider the Cauchy problem (3.1)-(3.2).
We always assume Condition (A.!)' (Chapter 2) which assures the global ex-
istence of characteristic curves. Let x = x(t,y), v = v(t,y) and p = p(t,y) be
the solutions of (2.15)-(2.16), and define a mapping H from to by
H(t,y) del
= (i,x(t,y)). Suppose:
(I) (Dx/Dy)(i°,y°) = 0,
§3.3. PROLONGATION OF CLASSICAL SOLUTIONS 25
and
(II) (Dx/Dy)(t,y°) 0 fort <1°.

In this section we consider the case where the characteristic curves do not meet.
Therefore, furthermore, we assume the following condition:
(B) The mapping H is bijective from a neighborhood of (t°, y°) to another one of
(t°,z°) where
By Condition (B), we can uniquely solve the equation x = z(t,y) with respect
toy, and denote it by y = y(t,x) (for (t,x) in a neighborhood of (t°,x°), (t,y) in a
neighborhood of (t°, y°)). The function y = y(t, x) is obviously continuous, though
it may not be differentiable. Then we get the following.
Theorem 3.1. Under the hypothesis (A.I)', suppose (I)-(II) and (B). Then the
solution u = u(t,x) remains a C'-solution of(3.1) in a neighborhood
of(t°,x°), though it is not of class C2.
Proof. Let V and U be open neighborhoods of (t°, y°) and (t°, x°), respectively,
such that the mapping H is bijective from V to U. Consider S dcf
= {(t, y) V
(Dz/Dy)(t,y) = 0) and H(S) = {H(t,y) : (t,y) E S}. By Sard's theorem,
the Lebesgue measure of H(S) is zero. Therefore U\H(S) is dense in U. We see
that u = u(t,x) is of class C2 in the domain U\H(S). The reason is as follows.
For any E U\H(S), there exists uniquely a point E V satisfying
0 and = The inverse function y = y(t,x) is of class C1
in a neighborhood of which is contained in U\H(S). Here we recall Lemma
1.2 and Corollary 1.3 in §1.1, and we see that u = u(t,x) is a function of class
C2 satisfying the equation (3.1) in the neighborhood of (t , Next we show that
u = u(t,x) is continuously differentiable in the domain U. We pick up arbitrarily
_o
a point (t , ) in H(S). Then we can choose a sequence {(tm,xm)}m of points in
U\H(S) such that, when m tends to infinity, (tm,xm) is convergent to (t ). As
the mapping H is bijective from V to U, there exists a unique point (t, ym) E V\S
satisfying H(tm,ym) = (tm,xm) for each in. Since y = y(t,x) is continuous in U,
26 3. PROLONGATION OF CLASSICAL SOLUTIONS
0 o
it follows that ym = y(tm,xm) is convergent to y(t , ) when in tends to
infinity. As u = u(t,x) is continuously differentiable at (tm,xm), we have by (1.13)
p(t,y) =
Because p = p(t,y) is continuous on the whole half space {t 0, y E R"}, we get
ôu
lim —(t,x) =
m-.oo Ox
lim p(t,y) =p(i ).
The derivative On/Or of u = u(t, x) can thereby be continuously extended (in such
a way that (Ou/Ox)(t, x) = p(t, y(t, x))) over U. That is to say, u = u(t, r) is of class
C' in the domain U. Moreover, notice that the uniqueness of the above (extended)
C'-solution is assured by Theorem 1.6. 0
Remark. In Theorem 3.1, the assumption (A.I)' is a crucial point. It will be
proved in Chapter 5 that, for quasi-linear equations of first-order, the assumption
(A.!)' is not compatible with the property that the Jacobian vanishes somewhere.
§3.4. Sufficient conditions for collision

of characteristic curves I
For quasi-linear equations of first-order, Theorem 2.1 says that, if the Jacobian
vanishes somewhere, the classical solutions blow up there. Therefore we cannot
extend the classical solutions beyond the time t° when the Jacobian vanishes. This
obliges us to treat weak solutions for t > t°. The typical singularity of weak
solutions is "shock." As it will be shown in Chapter 5, the shock appears by the
collision of characteristic curves. Therefore we try here to give sufficient conditions
so that the characteristic curves meet after the Jacobian vanishes.
In this section we consider quasi-linear equations of first-order in one space
dimension as follows:
On On .
= ao(t,x,u) in {t >0, XE R }, (3.6)
u(0,x) = on {t = 0, r E R'}, (3.7)
where = a1(t,r,u) (i = 0,1) and = are of class C1 in R3 and R1,

respectively. The characteristic curves for (3.6)-(3.7) are defined as solution curves
§3.4. COLLISION OF CURVES 1 27
of (2.7)-(2.8) for n = 1. Here, as in Chapter 2, we again assume Condition (Al)

which assures the global existence of characteristic curves. Let x = x(t, y) and
v = u(t,y) be the solutions of (2.7)-(2.8) for n = 1.
Theorem 3.2. Under Condition (A.!), ioe assume:

(I) = 0, and (II) 0 fort < t°.
If (ôa1/Ou)(t°,x°,v°) $ 0 where x0 x(t°,y°) and v0 v(t°,y°), then the
characteristic curves meet in a neighborhood of (t°, x°).
Proof. This theorem follows immediately if the monotonicity of x = x(t, y) with

respect to y is violated. First we prove that (Ox/8y)(t, y°) is negative for t > 1°
with t — 1° small. Actually, differentiating (2.7) with respect to y, we get the system
of ordinary differential equations (2.10) for n = 1. Then (2.10) is linear with respect
to Ox/Oy and Ov/Oy. As the initial data are not zero, we get
Ox Ot,
(0,0) for any t 0, yE R 1
Hence we have (Ov/Oy)(t°,y°) 0. By the assumption, we get

d (Ox\ — Oaj
(tO 0 0
0
di Oy) — Ou ' Oy'
Moreover, as (Ox/Oy)(t°,y°) = 0 and (Ox/Oy)(t,y°) > 0 for t <t0, we get
<0, i.e., <0.

di Oy (t,y)=(t°,y°) dt Oy
Therefore we get (Ox/Oy)(i, y°) < 0 for t > with t — t° small. That is to say,
(Ox/Oy)(t,y°) changes its sign at I = t°. We now define a function h = h(y) by
setting h(y) inf{t : (OxfOy)(t, y) = 0) for each y. Then h = h(y) is of class
C1 in a neighborhood of y = y0 In fact, as we have (Ox/Oy)(t°,y°) = 0 and
(82x/OtOy)(t°,y°) <0, we can uniquely solve the equation (Ox/Oy)(i,y) = 0 with
respect to y in a neighborhood of (t°, y°). Let us restrict our following discussions
into a small neighborhood of (t°, x0, v°) only. We see by the same reasoning as in
the above that, as a function oft, (Ox/Oy)(t, y) changes its sign at t = h(y) for each
y.
Next, we will show that x = r(t,y) is not monotone with respect to y. Pick up
a point y°. If <t°, then we get
Ox-.. <0 and Ox
>0
)
28 3. PROLONGATION OP CLASSICAL SOLUTIONS
for t As this means that x = x(t,y) is not monotone with respect

to y, the characteristic curves meet in a neighborhood of (t°, x°). When h = h(y)
is constant in a neighborhood of y = y°, it follows that (Ox/ôy)(t°,y) 0 in
a neighborhood of y = y°, i.e., x = x(t°,y) is constant there. This says that
the characteristic curves meet at the point (t°,x°). When there exists a point
y = such that > t°, we can similarly see that x = z(t,y) is not monotone
with respect to y in a neighborhood of y = y° for each t E (t°, Hence the
characteristic curves x = x(t, y) meet in a neighborhood of the point (t°, r°). 0
§3.5. Sufficient conditions for collision

of characteristic curves H
In this section we consider the same problem as in §3.4 for general partial differential
equations of first-order in one space dimension as follows:
in {t>0, rER ), (3.8)
u(O,z) = on {t = 0, x R'}, (3.9)
where f = f(t,x,u,p) and = are of class C2 in R4 and R', respectively. The

characteristic curves for (3.8)-(3.9) are obtained as solution curves of (2.15)-(2.16)
for n = 1. Here we assume Condition (A.I)' (Chapter 2) which assures the global
existence of characteristic curves. Let x = r(t,y), v = v(t,y), and p = p(i,y) be
the solutions of (2.15)-(2.16) for n = 1.
Theorem 3.3. Under Condition (A.I)', we assume:

(I) and (II)
If (02f/8p2)(t°,x°,v°,p°) 0 where x0 x(t°,y°), v° v(t°,y°) and p°
p(t°, y°), then the characteristic curves meet in a neighborhood of (t°, x°).
Proof. As in the proof of Theorem 3.2, we will show that x = r(t,y) is not
monotone with respect to y. We repeat the same discussions in §1.1. When the
Jacobian (ôx/Oy)(t,y) 0, we can solve the equation z = x(t,y) with respect to
§3.5. COLLISION OF CHARACTERISTIC CURVES II 29
y, and denote the solution by y = y(i,x). Then the solution of (3.8)-(3.9) is given
by u(i,x) def
= v(t,y(t,x)) and
p(t, y(t, x)) = x) = x)
= —(t,y(t,x)) . (_(t,y))
When a point (t,x) goes to (t°,x°) along the curve {(i,x) : x=
(Ox/Oy)(i,y(i,x)) tends to zero and p(i,y(t,x)) remains bounded. Therefore it
must be true that (Ov/Oy)(t°,y°) = 0. Differentiating the system (2.15) with re-
spect toy, we get a system of linear ordinary differential equations for Ox/Oy, Ov/Oy
and Op/Oy just like (2.10). As the initial data
(Ox/Oy, Ov/Oy, Op/Oy)(O, y) = (1,
are not zero, it follows that (Ox/Oy,Ov/Oy,Op/Oy)(t,y) (0,0,0) for any t 0.

Hence (Op/Oy)(t°,y°) #0.
We can see that (Ox/Oy)(t, y) satisfies the following differential equation:
d (OX\ — 821 Ox + 82.f Ov + 82f Op
di Oy) — OxOp Oy OuOp Oy Op2 a1i
Using the assumptions and the above results, we get
— " ,x 0 ,v0 ,p ,y°

di — Op2
Moreover, as (Ox/Oy)(t°,y°) = 0 and (Ox/Oy)(t,y°) >0 fort < to, it must be true
that
d Ox
—(—) <0, i.e., —(—) <0.
dt \Oyi (t,p)=(t°,y°) — di Op (i,ii)=(i°4i°)
Hence (Ox/Oy)(t,y°) changes its sign at I = t°. Here we introduce a function h =
def. .
h(y), just as in the proof of Theorem 3.2, by setting h(y) = inf{t
•
(Ox/Oy)(t, y) = :
0) for each y. We consider the problem in a small neighborhood of (t°, x0, v°,p°)
yO)•
where 82f/Op2 0. We pick up a point By the same reasoning as in
the above, we see that (Ox/Oy)(t, changes its sign at I = If > we
have
Ox
and
for I E (t,h(j)). As this means that x = x(t,y) is not monotone with respect to
y, the characteristic curves meet in a neighborhood of (10, x°). In the case where
h = h(y) is constant, we can similarly prove that the characteristic
curves meet in a neighborhood of (t0, x°). 0
Chapter 4
Equations
of Hamilton-Jacobi Type
in One Space Dimension
§4.1.Nonexistence of classical solutions
and historical remarks
In this chapter we will treat general partial differential equations of first-order (3.1)-
(3.2) which satisfy Condition (A.!)' (Chapter 2); that is to say, the characteristic
equations (2.15)-(2.16) have uniquely global solutions x = x(t,y), v = v(t,y) and
p = p(t,y) for all y E The crucial point of Condition (A.!)' is the global
existence of p = p(t, y). Let us sum up the principal results proved in Chapters 1,
2, and 3:
1) The Cauchy problem (3.1)-(3.2) has locally a unique C2-solution.

2) If the Jacobian (Dx/Dy)(t, y) of the mapping x = x(t, y) vanishes somewhere, it
is impossible to extend the C2-solution beyond a point where the Jacobian vanishes.
3) Suppose that the characteristic curves do not meet in a neighborhood of the
point where the Jacobian vanishes. Then the solution keeps being of class C1, but
not of class C2, in the neighborhood of the point.
Therefore, from now on, we consider the Cauchy problem (3.l)-(3.2) under the
following conditions:
(I) (Dx/Dy)(t°,y°) = 0.
(I!) (Dx/Dy)(t,y) 0 for t <t° and y E I where I is an open neighbourhood of

1$ = Y0.
(II!) The characteristic curves x = x(t, y) meet in a neighborhood of (t°, x°) where
x0
§4.1. NONEXISTENCE OP CLASSICAL SOLUTIONS 31
In this section, the space dimension is always equal to one, though some subjects
are similarly discussed even in the case of several space variables. Now let us
consider the case (III).
Proposition 4.1. Suppose the above conditions (I), (II), and (III). Then the
Cauchy problem (3.1)-(3.2) cannot admit a solution of class C1 in the whole space.
Proof. Let x = x(t,y), v = v(t,y), and p = p(t,y) be the solutions of (2. 15)-(2.16)
for n = 1. By the assumption (III), there exist two points (a,a) and (a,/3) such
that x(a, a) = x(a, /3) (a j9). As f = f(t, x, u, p) is of class C2, a Cauchy problem
for (2.15) can not have more than one solution. Therefore we have
(x(a, a), v(a,a),p(a,a)) (x(a,j3),v(a,/3),p(a,/3)).
As x(a,a) = x(a,/9), it holds that
(v(a, a),p(a, a)) (v(a,/9),p(a,/3)). (4.1)
Let us assume that there exists a classical solution u = u(t,x) of (3.l)-(3.2) in

the whole space. A. Ph proved in El 131 that a solution of class C' is generated by
characteristic strips, that is to say
u(i,x(t,y)) = v(t,y), = p(t,y).
As x(a,a) = r(a,/3) def

= x , it follows that
u(a, x°) = v(a, a) = v(a, /3), x°) = p(a, a) = p(a, /3).
This contradicts (4.1).
When there do not exist classical solutions, we must introduce generalized solu-
tions which contain singularities. As the solutions are not of class Ci, the solutions
as regular as possible are required to be continuous. We recall by Condition (Al)'
that p = p(t,y) remains bounded. Since p = p(t,y) is corresponding to the first
derivatives of solutions, the generalized solutions should be Lipschitz continuous.
This suggests to us that, if the equations satisfy (A.I)', we would be able to construct
generalized solutions which are Lipschitz continuous. This is the aim of the present
4. EQUATIONS OF HAMILTON-JACOBI TYPE
chapter. Before advancing further, we had better give briefly a historical survey on
the global existence of generalized solutions for Hamilton-Jacobi equations.
Short historical survey. As an example of equations satisfying (A.!)', we are very

familiar with "classical Hamilton-Jacobi equations." Here, when the function I =
f(t, x,p) is independent of u, we call (3.1) a "classical Hamilton-Jacobi equation."
On the other hand, notice that quasi-linear equations of first-order, for example
equations of the conservation law, do not satisfy (A.!)' (see Proposition 5.1 in
Chapter 5).
A global theory for classical Hamilton-Jacobi equations has been well studied
and there are a lot of papers on this subject. At an early stage it has been considered
under the convexity condition saying that the function I = f(t, x, p) is convex with
respect to p. In this case generalized solutions are defined as "Lipschitz continuous
and semi-concave functions" which satisfy the equations almost everywhere. A
function u = u(t, x) is said to be semi-concave if it satisfies the inequality
u(t,x + y) + u(t,x — y) — 2u(t,x) (4.2)
for any x, y E RTh and t E R where K is a constant locally independent of x, y, and
The global existence and uniqueness of such generalized solutions for convex
Hamilton-Jacobi equations have been well studied by several methods: for example,
by variational method (E.D. Conway and E. Hopf [331), by methods of envelopes
(E. Hopf 1641, S. Aizawa and N. Kikuch.i 141), by difference-differential method
(A. Douglis [44]), especially by vanishing viscosity method (W.H. Fleming [51],
A. Friedman [54], S.N. Kruzhkov [881, [92]). Non-linear semi-group approach has
also been well developed by S. Aizawa [2], Y. Tomita [135]. The global theory for
non-convex Hamilton-Jacobi equations has recently been considered by M.G. Cran-
dall, L.C. Evans, P.-L. Lions ([35], [381, [100]). They have introduced the notion of
"viscosity solutions" to define generalized solutions and characterized their proper-
ties. There are also many interesting works related to the viscosity solutions, for
example, C. Barles [14], H. Ishii [67], P.E. Souganidis [123). By these contribu-
tions, the global existence and uniqueness of generalized solutions were established
almost completely. Concerning more detailed references, refer to P.-L. Lions [100]
and S.H. Benton [211. Historical remarks for equations of the conservation law will
be given in Chapter 5.
§4.1. NONEXISTENCE OF CLASSICAL SOLUTIONS 33
As we have already written, generalized solutions contain singularities. Our

principal aim in this chapter is to see the structure of singularities of generalized
solutions or weak solutions. However, this problem could not be solved by the
above many works. The reason comes from the methods. As the nice existence
theorems for generalized solutions depend generally on vanishing viscosity method,
the generalized solutions are obtained as the limit of solutions of non-linear heat
equations. As the process of taking the limit is complicated, it seems to us that it
would be difficult to get the informations on the singularities of generalized solu-
tions by vanishing viscosity method. But this method is very effective to prove the
existence of solutions. Let us review here many methods which are used to con-
sider the propagation of singularities in linear wave propagation. To consider the
singularities of solutions, one gives representation formulas for solutions, especially
using Fourier-Laplace transform, pseudo-differential operators, or Fourier integral
operators. Next, one tries to analyze their expressions. Though this procedure is
not always true, it is a typical approach to consider the propagation of singularities
in linear wave propagation. Concerning this subject, refer to L. Hörmander 165],
L. Gârding [55], and G.F.D. Duff [461.
How can we construct concretely the solutions of non-linear partial differen-
tial equations of first-order? The method we know is the characteristic one. The
weak point is that it is a local theory. The reason comes from the fact that the
characteristic curves x = x(t, y) cannot be uniquely solved with respect to y in a
neighborhood of a point where the Jacobian vanishes; that is to say, the inverse
function y = y(t, z) takes many values in a neighborhood of a critical point of the
mapping x = x(i, y). Therefore the "solutions" constructed by the characteristic
method also takes many values there. As discussed in the above, we are looking
for one-valued and Lipschitz continuous solutions. However, a generalized solution
is not unique in the space of Lipschitz continuous functions (see Chapters 9 and
11). Therefore we impose a supplementary condition on generalized solutions which
guarantees the uniqueness of solutions. It is the semi-concavity condition (4.2).
What we would like to prove is that, if we assume (A.!)', we can choose only
one value from many values of the "multi-valued solution" in order to get a solution
that becomes one-valued and continuous. Then, if the equation is convex, we can
show that the semi-concavity condition (4.2) is automatically satisfied. This is the
program which we would like to develop in the following discussions.
34 4. EQUATIONS OP HAMILTON-JACOBI TYPE
From the results in this chapter, we will see that the equations with proper-
ty (A.I)' have properties similar to that of classical Hamilton-Jacobi equations.
Therefore we would like to call the equations with property (Al)' "equations of
Hamilton-Jacobi type." Notice that S. Lzumiya [74J considers the differences be-
tween Hamilton-Jacobi equations and first-order quasi-linear partial differential e-
quations from the geometrical point of view.
M. Bony [23] has considered the propagation of weak singularities for non-linear
partial differential equations. His method depends on some techniques elaborated to
study linear problems. We think that the singularities which will be discussed here
could not be treated by his method, because, for the definition of paradifferential
operators which are the approximations of nonlinear equations, solutions must be
at least classical solutions.
§4.2. Construction of generalized solutions
To make clear our problem, we write it again. Consider the Cauchy problem in one
space dimension as follows:
Ou .
in {t>O, rER }, (4.3)
u(O,x) = 4(x) on {t = 0, x R1}, (4.4)
where f = f(t,x,u,p) and = are of class C in R1 x R1 x R' x and

R1, respectively. For the construction of singularities of solutions, we need a little
stronger assumption on the regularity of f = f(t,x,u,p) and 4 = Then the
characteristic curves for (4.3)-(4.4) are defined as the solution curves of (2. 15)-(2. 16)
for n = 1. We write them by x = x(t,y), v = v(t,y), and p =p(t,y).
Suppose (A.!)'. This assures the global existence of x = x(t, y), v = v(t, y), and
p = p(t, y). We consider the Cauchy problem (4.3)-(4.4) under three conditions (I),
(II), and (III) introduced in §4.1. As (III) is not concrete, we assume the following
sufficient condition for it to hold (see Theorem 3.3):
0 (4.5)
where v= v(t o ,yo) and p0def

= p(t o ,y ).
o
§4.2. CONSTRUCTION OF GENERALIZED SOLUTIONS 35
From now on, we will construct the singularities of generalized solutions for 2>
where t—t0 is small. First we solve the equation (ôx/ôy)(t,y) = 0 with respect to
2. As (Ox/Oy)(t, y) > 0 for t < t° and y E I (cf. (H)), it holds that (Ox/Oy)(t°, y)
0 for y E I. Moreover, as = 0, we get (02x/8y2)(t°,y°) = 0. We
assume here (83x/0y3)(t°, y°) 0. This assumption is natural from the generic
point of view. In this case, we see that
>0. (4.6)
In fact, if (4.6) were false, we would have (82x/8y2)(i°,y) < 0 for y > y°, i.e.,
(Oz/Oy)(t°,y) < 0 for y > y°. As (ôs/Oy)(0,y) = 1, there exists 2 < 20 such that
(Ox/Oy)(t,y) = 0. This contradicts (I).
Let us draw a figure for the curve x = s(t,y) (2> t°) (see Figure 4.1).
xl
ya yl y
Figure 4.1
We explain the reason why the curve x = x(t,y) is drawn as in Figure 4.1.
Taking the derivative of (2.15) with respect to y, we get a system of ordinary
4. EQUATIONS OP HAMILTON-JACOBI TYPE
differential equations concerning Ox/Oy, Ov/Oy, and Op/Oy just like (2.10). For
example, the equation for Ox/Oy is written by
dOx =
02f Ox 621 02f
+ + (4.7)
This system of equations is linear with respect to Ox/Oy, Ov/Oy and Op/Oy. Since
((Ox/Oy)(O, y), (Ov/Oy)(O, y), (Op/Oy)(O, y)) = (1, 4'(y), 4/'(y)) (0,0,0), we see
that ((Ox/Oy)(t,y),(Ov/Oy)(t,y),(Op/Oy)(t,y)) (0,0,0) for any (t,y) R2. We
recall here Lemma 1.2:
= (t,y) R2.
As (Ox/Oy)(t°,y°) = 0, it follows that (Ov/Oy)(t°,y°) = 0, hence that
0.
On the other hand, as (Ox/Oy)(t°,y°) = 0 and (Ox/Oy)(t,y°) > 0 for t < t°, the
left-hand side of (4.7) must be non-positive at the point (t°, y°). Using (4.5) and
the above results, we get
= <0. (4.8)
Therefore we can uniquely solve the equation (Ox/ay)(t, y) = 0 with respect to tin
a neighborhood of (t°,y°), and write the solution as a C°°-function t = p(y) of y.
Then we have
d Ox 02x , 02x
= (y) + = 0.
But (62x/0y2)(t°,y°) = 0 and (62x/OtOy)(t°,y°) < 0, it follows that p'(y°) = 0.

Similarly, since
d2 Ox 02x 03x
= o
,y o )pu(yo ) + o
,yo) = 0,
using (4.6) we get p"(y°) > 0. As p'(y°) = 0, it follows that p'(y) > 0 for y > y°
and that p'(y) <0 for y <y°. Thus the function t = p(y) is strictly increasing for
y > y° and strictly decreasing for y < y°, so the equation t = p(y) has two solutions
§4.2. CONSTRUCTION OP SOLUTIONS 37
y = yi(t) andy = y2(t) (y2(t) < y° < y'(t)) fort > = p(y°) where i—ta is small.
Summing up these results, we get the following.
Lemma 4.2. In a nezghborhood of (t°,y°), the equation (Ox/Oy)(t,y) = 0 has two

soluüons, y = y1(t) and y = y2(t); y1(t) > y3(t) for t > 2° with 2 — t° small. The
two solutions y = y1(t) (i = 1,2) are of class C°° fort > and are continuous for
t° (with yi(t° + 0) = y2(t° + 0) = y°).
The proof is obvious by the above discussions. Here we put x(t, y1(t)) (i =
1,2). Then ri(t) <52(2) fort> t° (with +0) = = x(t°,y°) = x°).
Next we solve the equations = x(t,y) with respect toy for x E (x1(t),52(t)), 2 > 2°
(t—t° small). Then we get three solutions y = g,(t,x) (i = 1,2,3) with g1(t,x) <
g2(t,x) < gs(t,x). Define u1(t,x) 1! v(t,g,(t,x)) (i = 1,2,3). This means that
the "solution" constructed by the characteristic method takes three values for x
(xj(t),x2(t)) (2 > t° with 2— t° small). By the assumption (4.5), we treat here the
case where
> 0.
Then we get the following.
Lemma 4.3. i) For any 2 > 20 with 2—2° small and for any x E (xl(t),x2(t)), we
have
ui(t,x) —U2(t,s) <0 and u2(t,x) — >0.

ii) There exists uniquely x = ir(t) (x1(t),52(t)) satisfying
ui(t,7(t)) =
Proof. i) By (4.8) and (4.9), we have (Op/Oy)(t°,y°) <0, i.e., (Op/Oy)(t,y) <0 in
a small neighborhood of (t°,y°). As gi(t,z) <g3(t,x) <gs(t,x), it follows that
p(t, (2, x)) > p(t, g2(t, x)) > p(t, gs(t, x)). (4.10)
Using these inequalities, we get
-(uj(t,z) — u2(t,x)) = p(2,gj(t,x)) — p(t,g2(t,x)) > 0,

38 4. EQUATIONS OF HAMILTON-JACOBI TYPE
with
(ui(t,z) — = 0.
Hence it holds that ui(t,x) — u2(i,x) < 0 for x E Exi(t),x2(t)). In the same way,
we obtain the inequality u2(t,x) — us(t,x) >0 for x E (xi(i),x2(t)j.
ii) Using the results of i, we have
(t, x) — us(t, x)) = p(i, gj (t, x)) — p(t, g3(t, x)) > 0,
with
(ui(t,x) — = (u3(i,x) — >0
z=za(t)
and
(u1(i, x) — = (uj(t, x) — u2(t, < 0.
(t) (t)
Therefore there exists uniquely x = 7(t) E (xi(t),x2(t)) satisfying ui(t,'y(t)) =

us(t,y(t)). 0
We see by Lemma 1.2 that the three functionsu = = 1,2,3) satisfy the
equation (4.3) in their domain where i > t° (t—i° small) and x2(t) x xi(t). As
we are looking for a single-valued and continuous solution, we extend the solution
u = u(t,x) into the above domain by defining
J u1(t,z) if x
u( ,x)
— u3(t,x) if x>
This extended solution is obviously Lipschitz continuous. It satisfies the equation
(4.3) except on the curve {(t,x) t t°, r =
: The next problem is to prove
the uniqueness of generalized solutions. This is the subject of the following section.
§4.3. Semi-concavity of generalized solutions
First we give an example of non-uniqueness of Lipschitz continuous solutions which

is due toY. Tomita (see also Remark 2 of Theorem 11.5 in Chapter 11).
Example. Consider the Cauchy problem:
=0 in
I. u(0,x)=0 on {t=0, xER1}.

§4.3. SEMI-CONCAVITY OF GENERALIZED SOLUTIONS 39
deC
Then it has a trivial solution u = uo(t,x) = 0, and also
deC10
u=ui(t,x)=ç if 0<t<IrI,
I. if
The above example shows the non-uniqueness of generalized solutions in the

space of Lipschitz continuous functions. Therefore we must introduce a supplemen-
tary condition on generalized solutions which guarantees the uniqueness of solutions.
If f = f(t, x, u, p) is convex with respect to p, the supplementary condition would
be the semi-concavity condition (4.2). When I = f(t,x,u,p) is concave, then it
becomes the semi-convexity condition as follows:
u = u(t, x) is said to be semi-convex if it satisfies the inequality
u(t, x + y) + u(t, x — y) — 2u(i, x) (4.12)
for any x, y and t where K is a constant locally independent of x, y and t.

The supplementary condition (4.2) [or (4.12)] is very important, because it is
related to the stability of solutions of the Cauchy problem (4.3)-(4.4). What we
would like to claim here is that the solution constructed by (4.11) in §4.2 satis-
fies automatically the above supplementary condition; that is to say, we get the
following:
Theorem 4.4. Assume > 0 (resp. < 0).

Then the solution u = u(t,x) extended by (4.11) is semi-concave (resp. semi-
convex) in a neighborhood of (t°,x°).
Proof. We consider the convexity case (4.9). It suffices to prove (4.2) in a small
neighborhood of {(t,z) x= : Assume that x = -y(t) and y > 0 (y is small).
Then
u(t, x + y) + u(t, x — y) — 2u(t, x)
= {u(t, x + y) — u(t, x + 0)} — {u(t, x —0) — u(t, x— y)}

On On On On
+Oy) — —0)— —
On On
+0) — — 0)}y,
where 0 < 9,9 < 1.

40 4. EQUATIONS OP HAMILTON-JACOBI TYPE
To estimate the first and second terms, we must calculate the second derivative
of u = u(t,x):
= -(Ou/Oz)(t,x) = _p(t,y(t,x))
iOp lOx
=
p=y(t,r)
By (4.8)-(4.9), (Op/Oy)(t,y) is negative in a neighborhood of (t°,y°). From one of

the assumptions, it follows that (Ox/Oy)(t°,y°) = 0 and (Ox/Oy)(t,y°) >0 fort <
t°. Hence, when (t,x) goes to (t°,x°) along the curve x = x(t,y°), (82u/8x2)(t,s)
tends to —oo, i.e., (02u/0x2)(t, x) is bounded from above in a neighborhood of
(t°,x°).
By (4.10) and the definition of u(t,x), the third term is estimated as follows:
+ 0) — 7(t) —0) = p(t,g3(t, x)) — p(t,g1(t, x)) < 0. (4.13)
Summing up the above results, we see that there exists a constant K satisfying
u(t,x + y) + u(i,x — y) — 2u(t,x)
In the case where (02f/8p2)(t°,x°,v°,p°) is negative, we can similarly prove

that the solution is semi-convex in a neighborhood of (t°, x°). 0
As it is well known that a generalized solution with property (4.2) or (4.12) is
unique in the space of Lipschitz continuous functions, we see that our generalized
solution (4.11) is reasonable.
By (4.11), the singularity of the solution u = u(t,x) lies on the curve {(t,x)
x = 1(t)} where u1 (t, 'y(t)) = us(t, -y(t)). Taking the derivative of this equality with
respect to t, we get
'y(t)) — y(t)) —
{
By the definition of u = u(t, x), it holds that
= — 0), y(t)) = + 0).

§4.4. COLLISION OF SINGULARITIES 41
As u = (i = 1,2) satisfy the equation (4.3) in their definition domain, we

get
— f(t,x,u, + 0)) — f(t,x,u, — 0))
(• 14
-y(t) + 0) — 7(i) —0)
where x def
= -y(t), u
def
= u(t,-y(t)). The equality (4.14) is correspondmg to the
Rankine-Hugoniot jump condition for equations of conservation law.
Remark. We can extend the singularity of the solution for large t by the same
method. But, if (02f/0p2)(t, x, u,p) changes its sign, the solution may generafly
lose the above supplementary condition. Then we must introduce other types of
singularities. This is the reason why the above construction of singularities is local.
This is the subject which will be discussed in Chapter 7.
§4.4. Collision of singularities
In this section, we assume that f = f(t,x,u,p) is convex with respect to p in the

whole space, and consider the case where two singularities of the solution collide
with each other. In this case we see by Theorem 3.3 that, if the Jacobian vanishes
at a point, then the characteristic curves meet in a neighborhood of the point.
Therefore the singularities of the generalized solution can be expressed in a form of
(4.11).
Let W1 = {x = (i = 1,2) be singularities of the solution u = u(t, x).
t >
Assume that W1 and W2 meet at a time t = T; that is to say, let 1'1(t) <y2(t) (t <
T) and = 72(T). By (4.10) and (4.13), we see that
(t) — 0) > (t) + 0) and 72(t) — 0) > + 0)
for t <T. Taking the limit as i -+ T — 0 in the above inequalities, we get
= (4.15)
As! = f(t,x,u,p) is convex with respect top, we have by (4.14)
> i.e., 71(t) > 72(t) for t > T.

42 4. EQUATIONS OF HAMILTON-JACOBI TYPE
Our problem is how to extend the solution u = u(t,x) for t> T.
Figure 4.2
First, see Figure 4.2 which expresses the behavior of the curve x = x(t, y) for i>
T. We use the notations indicated in Figure 4.2. Each part (i = 1,2,3) of the
curve x = x(t,y) can be uniquely solved with respect toy, and we write y = g;(t,x)
(1 = 1,2,3). Put x) v(t,91(t, x)) (i = 1,2,3). As > for t > T,
the solution u = u(t, x) defined by (4.11) takes two values for x E We
def
define I(t,x) = ui(i,x) — us(t,x). By Lemma 4.3, we get
I(t, = uj(t, 'y2(t)) — u,(t, 12(i))

= u1(t, 72(t)) — u2(t, < u1 (t, 71(t)) — u2(t, 11(t)) = 0,
I(t,y1(t)) = u1(i,'71(t)) —
= u2(t,y1(t)) — u,(i,yi(t)) > u2(t,y2(t)) — u3(t,72(t)) = 0.
Moreover, it follows from (4.10) that
a'
—(t,x) =p(t,gi(t,x)) —p(t,gs(t,x)) > 0.
ax
Hence there exists uniquely a, = E satisfying = 0. As

we are looking for a continuous and single-valued solution, we define the solution
u = u(t,x) for x E [72(t),11(t)] as follows:
I u3(t,x) if x
u(t,x)
= 1 uj(t,x) if a,
Then we can similarly show that u = u(t, a,) is Lipschitz continuous and semi-
concave. Taking the derivative of I(t, = 0 with respect to i, we also get the
same jump condition (4.14).
The collision of a finite number of singularities can be treated by the above
method. For example, let {x = -y,(t), t > t} (i = 1,2,... ,k) be singularities of
the solution u =u(t,x). Suppose (I) <72(t) < ... for i <T, and
(II) 11(T) = 12(T) = ... = 7k(T). Then we have
dt'' >
Therefore it follows that
71(i) > > .. . > for t > T.
Next, we can prove that there exists uniquely a, = E satisfying

ui(t,-y(t)) = A new generalized solution may be defined in the same
way as before.
Except the case where an infinite number of singularities collide simultaneously,
we can repeat the above discussions forever.
If (02f/0p2 )(t, x, u, p) changes its sign, the situation may generally become com-
plicated. The phenomena are locally almost the same as the above, but the global
behavior of singularities is generally quite different. We will consider this subject
in Chapter 7.
Chapter 5
Quasi-linear Partial Differential
Equations of First-Order
§5.1. Introduction and problems
In this chapter we consider the Cauchy problem for quasi-linear equations of first-
order in one space dimension as follows:
in {t>0, xER }, (5.1)
u(0,x) = on {t = 0, x E R1}, (5.2)
where = a1(t,x,u) (i = 0,1) and 4 = are of class in R.' x R1 x

and R', respectively. Our principal interest is on the construction of singularities
of weak solutions of (5.1). Therefore we put a little strong assumptions on the
regularity of = (i = 0,1) and = 4(y). On the other hand, there
are many works on the global existence and uniqueness of weak solutions for quasi-
linear equations of first-order, especially for equations of the conservation law. For
example, refer to P.D. Lax (981, J. Smoller (121), A. Majda (104), A. Jeffrey (77].
Though the difference method is very important from the point of view of numerical
analysis, the best result on the global existence of weak solutions was obtained by
the vanishing viscosity method (see O.A. Oleinik [1121, S.N. Kruzhkov [88), (92)).
By the same reason stated in Chapter 4, it seems to us that it would be difficult
to get the informations on the singularities of solutions by the vanishing viscosity
method. Therefore, in this chapter as well, we construct the singularities, especially
of "shock" type, by the analysis of characteristic curves.
The characteristic curves for (5.1)-(5.2) are the solution curves x = x(t,y) and
v = v(t,y) of (2.7)-(2.8) for n = 1. Here, we assume (A.!) (stated in §2.2), which
assures the global existence of characteristic curves. We have seen by Theorem 2.1
that, if the Jacobian of the mapping x = z(t, y) vanishes at a point (i°, y°), then the
classical solutions blow up at a point (i°, x°), where x0 x(t°, po). Therefore we
INTRODUCTION AND PROBLEMS 45
must introduce weak solutions for equation (5.1). To define it, we rewrite equation
(5.1) as follows:
+ = g(t,x,u) (5.1')
where (Of/Ou)(t,x,u) = aj(t,x,u) and g(t,x,u) — (Of/Ox)(t,z,u) = ao(t,x,u). If

g(t,z,u) 0, Equation (5.1') is of conservation law. Let w = w(t,x) be locally
integrable in R2. The function w = w(t,x) is called a weak solution of (5.1)-(5.2)
if, for any k = k(t,x) in it holds that
x) + f(t, x, + g(t, x, w)k] dtdx + x)dx = 0, (5.3)
where {(t,x) t > 0, x R'}. Let w = w(t,x) be a piecewise smooth

:
weak solution of (5.1) which has jump discontinuity along a curve z = 1(t). Then,
by the definition of weak solutions, we get the following Rankine-Hugoniot jump
condition:
dy — f(t, x, u+(t)) — f(t,z,u_(t))
54
dt —
where = u(t, y(t) ± 0) lirnu(t, 1(t) ± e).

From now on, we will consider the Cauchy problem (5.1)-(5.2) in a neighborhood
of(t°,x°) (x° = x(t°,y°)) under the following two assumptions:
(I) (Dz/Dy)(t°,y°) = 0.
(II) (Dx/Dy)(t,y) 0 for t < t° and y I where I is an open neighbourhood of
y = y°.
Our problem is to see what kinds of phenomena may appear for t > t°. As
Example 1 in §3.2 shows, we have to consider two cases as follows:
(III) Though the Jacobian of x = x(t,y) vanishes at a point (t°,y°), the char-
acteristic curves x = x(t,y) do not meet in a neighborhood of (t°,x°), where
x(t°, y°).
(IV) Fort > t°, where (Dx/Dy)(t°,y°) = 0, the characteristic curves x = x(t,y)
meet in a neighborhood of (t°, x°).
46 5. QUASI-LINEAR EQUATIONS OF FIRST-ORDER
In the case (III), we can uniquely solve the equation x = x(t, y) with respect to
y and denote it by y = y(t,x) which becomes a continuous function. Put u(i,x)
v(i,y(t,x)). Then Theorem 2.1 means that u = u(t,x) is not a classical solution of
(5.1), however, we can see that it is a continuous weak solution. The proof is as
follows. The continuity of u = u(t, x) comes from the continuity of v = v(t, y) and
y = y(i,x). Put S {(i,y) (Dx/Dy)(i,y) = 0) and H(S) {(t,x) r=
z(t,y) for (t,y) E S}. Then, by Sard's theorem, the Lebesgue measure of H(S)
is zero. As a corollary of Theorem 2.1, we can see that u = u(t,x) satisfies the
equation (5.1) outside H(S). Hence u = u(t,x) is a continuous weak solution of
(5.1 )-(5.2).
The important problem remaining is to consider the case (IV). Suppose that the
characteristic curves meet, that is to say, x(t,yj) = x(t,y2) where Yl By the
uniqueness of solutions of (2.7)-(2.8), we have (x(t, yi), v(i, yi)) (r(t, Y2), v(t, y2)).
As x(t,yi) = x(t,y2), it follows v(t,yj) v(t,y2). This means that we will not be
able to get continuous weak solutions. Therefore we will look for piecewise smooth
weak solutions. As we will see a little later, the solution u = u(t, x) v(t, y(i, z)) in
the case (IV) takes several values after the Jacobian vanishes. As we are looking for
a single-valued solution, our problem is how to choose one appropriate value from
these so that the solution becomes a single-valued weak solution of (5.1). if one
may jump from a branch of solution to another one, the jump discontinuity must
satisfy Rankine-Hugoniot's equation (5.4). Solving the Cauchy problem for (5.4),
we can get a curve of jump discontinuity. J. Guckenheimer [59] and G. Jennings
[78] took this approach. But they forgot to pay attention on the uniqueness of
solutions for the Cauchy problem to (5.4). In fact, the right-hand side of (5.4) is
not Lipschitz continuous at an initial point (see Lemma 5.3). The aim of §5.3 is to
prove the uniqueness of solutions for the Cauchy problem to (5.4).
The generic property of singularities was studied in D.G. Schaeffer 1120] and
T. Debeneix [41j by applying Thom's "Catastrophe theory." The construction of
singularities in two space dimensions for Hamilton-Jacobi equations was studied
by M. Tsuji [137]. By a similar method, S. Nakane [109]-[11O] constructed the
singularities of shock type in several space dimensions. But, as he treated the
singularities of fold and cusp types only, his results are essentially those in the
case of two space dimensions. Recently S. Izumiya and G.T. Kossioris (see for
§5.2. EQUATIONS OF HAMILTON-JACOBI TYPE AND CONSERVATION LAW 47
example, 1741-1761) give the generic classifications for the bifurcations of singularities
of geometric solutions.
§5.2. Difference between equations of the conservation law

and equations of Hamilton-Jacobi type
In this section we will give some property of equation (5.1). This will characterize
a difference between Equation (5.1) and equations satisfying (A.!)'.
A difference between characteristic Equations (2.7) and (2.15) is that (2.7) does
not contain an equation concerning p = p(t,y). Let r = x(t,y) and v = v(t,y)
be the solutions of (2.7)-(2.8) for n = 1. As p = p(t,y) is corresponding to
(ôu/Ox)(t,x(t,y)), an ordinary differential equation for p = p(t,y) is written as
follows:
= — x, v)p2 + x, v) — x, v)}p + x, v), (5.5)
p(O) = (5.6)
The equation (5.5) corresponds to the last one of (2.15). More concretely, we
see by Corollary 1.3 that, if u = u(t,z) is of class C2, p = p(t,y) is equal to
(ôu/&c)(t, y(t, x)).
Proposition 5.1. Assume (Oz/Oy)(t°,y°) = 0 and (Ox/Oy)(t,y°) 0 fort <t°.

Then the solution p = p(t, y°) of (5.5)-(5.6) tends to infinity when t goes to t0 — 0.
Proof. Remember that Ox/Oy and Ov/Oy satisfy the system of ordinary differential
equations (2.10) for n = 1. Since this system (2.10) is linear with respect to
{Ox/Oy,av/ay}, we get (Ox/Oy,Ov/Oy) (0,0) for any (t,y) E R2. Therefore, if
(Ox/Oy)(t°,y°) = 0, then (Ov/Oy)(t°,y°) 0. Moreover, we have by Lemma 1.2
= (t,y) E R2.
Hence we can easily get this proposition.
This proposition means that, if the Jacobian vanishes somewhere, (A.I)' cannot
be satisfied for quasi-linear equations of first-order.
§5.3. Construction of singularities of weak solutions
In this section we will consider the case (IV) and construct the singularities of weak
solutions which are called "shock" or "shock wave." A sufficient condition which
guarantees the situation (N) is
(5.7)
where v0 v(t°, y°). This is the result of Theorem 3.2. If (5.7) is violated, we
can construct an example in which characteristic curves do not meet though the
Jacobian vanishes (see Example 1 in §3.2). Therefore, the solution of this case
remains to be continuous, though it is not differentiable.
First we solve the equation x = x(t, y) with respect toy for t > t0 in a neighbor-
hood of (t°, y°). This procedure is almost the same as in §4.2. By the assumptions
(I) and (II), we have (02x/0y2)(t°, y°) = 0. We assume (03x/8y3)(t°, y°) 0. This
assumption is natural from the generic point of view. In this case also, we get, by
the same way as the proof of (4.6),
> 0. (5.8)
Then the graph of the curve x = x(t,y) for t > t0 is drawn just as in Figure 4.1
(see §4.2). Therefore we use the same notations given in Figure 4.1. Here the
functions y = y1(t) and y = y3(t) > are the solutions of (Ox/Oy)(t,y) = 0
def
and z1(t) = x(t,y1(t)) (z = 1,2). Then Lemma 4.2 is also true for these y =
y1(t) (i = 1,2). When we solve the equation x = x(t,y) with respect to y for
z E (x1(t),xz(t)), we get three solutions y = g3(t,z) (g1(t,z) < g2(i,x) < gs(t,x))
and define x) x)) (i = 1,2,3). As we are looking for a single valued
solution, we must choose only one value from {u1(t,x) :i = 1,2,3} so that we can
get a weak solution of (5.1). As we have written in §5.1, we can not get continuous
weak solutions under the condition (N). Therefore we look for a weak solution
which is piecewise smooth. If u = u(t, x) is a weak solution of (5.1) which has jump
discontinuity along a curve x = -y(t), it must satisfy the R.ankine-Hugoniot jump
condition (5.4). This suggests us that a nice weak solution jumps from a branch
{ (t,x,u) : u = uj(t,x)} to the other {(t,x,u) u = u3(t,x)} along a curve
:
§5.3. CONSTRUCTION OF SINGULARITIES OF WEAK SOLUTIONS 49
x = -y(t) on which condition (5.4) is satisfied. Therefore our problem is to solve the
Cauchy problem for Rankine-Hugoniot's equation as follows:
( dx f(t, x, uj(t, x)) — f(t, x, us(t, x))

()
—
— > t°
di — uj (t, x) — us(i, x)
I. x(t°) = x0.
where xi(t) < x(t) < x3(t) for I > t°. The function j = j(t,x) is obviously
differentiable in a domain U {(t,x) I > to and xj(t) < x < x2(t)} and
:
is continuous on U (the closure of U). As we will show in Lemma 5.3, it is not

Lipschitz continuous at the end point (I°, y°). As j = j(t, x) is continuous in U, we
can see the existence of solutions by the classical theory. But we can not get the
uniqueness of solutions. In 3. Guckenheimer [59] and (3. Jennings [78], they did not
pay attention to this point. This is one of our problems which we will consider in
this chapter. As we put the hypothesis (5.7), we assume here more concretely
>0. (5.10)
Though j = j(t,x) is not Lipschitz continuous at the point (t°,x°), j = j(t,x)

is in C'(U). Therefore, if the solution could enter into the interior of the domain
U, we can easily extend it further. Hence we restrict our discussions in a small
neighborhood of (10, x°).
Lemma 5.2. i) (Ov/Oy)(t°,y°) <0.

ii) For (i,x) E U, we get
ui(t,x) > u2(t,x) > u3(t,x) and <0, > 0, <0.
iii) When (t,x) goes to (t°,x°) in U, then (0u1/Ox)(i,x) (i = 1,2,3)

tend to infinity.
Proof. i) As (Ox/Oy)(t,y°) > 0 fort < t° and (Ox/Oy)(t°,y°) = 0, we have
= °G1(jO
0>— di '..OyI Ox ' ' Oy
Since ((Ox/Oy)(t,y),(Ov/Oy)(t,y)) (0,0) for all (t,y), it holds that
(Ov/Oy)(t°,y°) 0.
Hence we get i) by (5.10).

ii) By the definition of y = (i = 1,2,3), we have gi(t,x) <
g2(t,x) < gs(t,x) and = v(t,g1(t,z)) (i = 1,2,3). Using the property
i, we get the first half of ii. As g1(t,x) < y2(t) < gz(t,x) < yi(t) < gs(t,x) for
x E (xi(t),x2(t)), we have
(i, a,)) > 0, g2(t, z)) < 0, g3(t, x)) > 0.
As (0u1/Ox)(t,z) = we get the second part

of ii, and also iii. 0
Lemma 5.3. i) The function j = j(t, a,) is continuously differentiable in the domain
U and is continuous on U where U {(t,x) t > t°, r2(t) > x > x1(t)}. But it
is not Lipschitz continuous at the point (t°, a,°).
ii) Fort > t°, j = j(t,x) is decreasing with respect to a,.
Proof. The first part of i is obvious. Taking the derivative of j(t, with
a,) respect
to z, we have
0. 1
—
Of
(5.11)
= —
+ a,, ui) {f(t,x, u1) — f(t,x, us)}} a,)

— U1 —
+ {f(t, x, u1) — f(t, x, u3)} — a,).

[
When (t,x) goes to (t°,x°) in U, the first term of (5.11) is convergent to
(02f/OxOu) (i°,x°,u°) where u0 u(t°,x°). Therefore it is bounded in a neigh-
borhood of (t°,r°). When (i,x) -4 (i°,x°) in U, the coefficient of (Oui/Ox)(t,x)
tends to (02f/02u)(t°,x°,u°)/2 = (Oai/Otz)(t°,x°,u°)/2 > 0. The coefficient of
(8u3/Ox)(t,x) has the same property as (Oui/Ox)(t,x). Here we apply ii and iii of
Lemma 5.2 to (5.11), then we see that, when (t, x) goes to (t°, a,°) in U, (Oj/Ox)(t, a,)
tends to —oc. This means that j = j(t,x) is not Lipschitz continuous at (t°,x°),
and that it is monotonically decreasing with respect to x in U. 0
Lemma 5.4. The functions x1 = x1(t) x(t, (i = 1,2) satisfy the following
properties:
i) = for t > t° (i = 1,2).
CONSTRUCTION OF SINGULARITIES OF WEAK SOLUTIONS 51
ii) <j(t,.r1(t)) and > j(t,x2(t)) for >
Proof. i) Since the functions y = y1(t) and y2(t) (y1(t) > y2(t)) are the solutions
of (Ox/Oy)(t,y) = 0 for t > we see that
= +
= = a1(t,x(t,y1(t)),v(t,y1(t))).
ii) By the definition of j(t, x), we have
j(t, x1(t)) = ai(t, xj(t), u3(t, xj(t)) + O(ui(t, x1) — u3(t, x1 ))), 0 < 0 < 1.
As (Oai/Ou)(t°,x°,v°) > 0, ai(t,x,u) is strictly increasing with respect to u in a

neighborhood of(t°,x°,u°). Moreover we have ui(t,x) > u3(t,x) for (t, x) U and
v(t,y1(t)) us(t,zi(t)). Hence we get
j(t, x1 (t)) > a1 (t, x1(t), us(i, x1 (1))) = (t).
We can similarly obtain the second inequality.
Using the above lemma, we can prove the following.
Proposition 5.5. The Cauchy problem (5.9) has a unique solution in the domain
U {(t,x) : t > x2(t) > r > xj(t)}.
Proof. We extend the definition domain of j = j(t, x) to the whole space keeping
the following two properties: (I) j = j(t,x) is continuous on R2; and (II) j(t,x) is
decreasing with respect to x for t > t°. Then it is obvious that the Cauchy problem
(5.9) has a solution x = -y(t). From ii of Lemma 5.4 and Lemma 5.3, we get easily
x1(t) <y(t) <x2(t) for t > t°. Next we will prove the uniqueness of solutions. Let
x = -11(t) and x = 'V2(t) be two solutions of (5.9). As j = j(t,x) is decreasing with
respect to x, we have
— y2(t)]2 = 2(y1(t) — —j(t,12(t))] 0,
11(t°) — y2(t°) = 0.
Hence we get y1(t) 12(t) for t t°.

Now we define the weak solution of(5.1) in the interval (xj(t),x2(t)) fort > t°
by
ui(t,x) if x <
u(t x) = 1 (5.12)
I u3(t,x) if x > y(t).
Remark. It is impossible to jump from the first branch {u = u1(t,x)} to the

second one {u = us(t,x)} so that Rankine-Hugoniot condition (5.9) is satisfied.
This is corresponding to Lemma 6.2 for Hamilton-Jacobi equations.
§5.4. Entropy condition
We first give an example which shows the non-uniqueness of weak solutions of (5.1)-
(5.2) in the space is the space of measurable functions which are
integrable on any compact set in R2.
Example. Consider the Cauchy problem

(Ow 2
=0 in {t>0,XER},
w(0,x) = 0 on {t = 0, x E R}.
Then this problem has a trivial solution w(t, x) = 0 and a weak solution as follows:
0 on {(t,x) : ri t 0},
w(t,x) = 1 on {(t,x) : t > x > 0},
—1 on {(t,x) : t> —x > 0}.
We get the above example by putting w(t,x) where u = u(t,x) is

the function given in "Example" of §4.3. As weak solutions of the above example are
not unique, we must impose the entropy condition which guarantees the uniqueness
of weak solutions in the space Concerning the entropy condition, we follow
here O.A. Oleinik 1112] and P.D. Lax [97].
Consider the equation (5.1') and let u = u(t,x) be a weak solution of (5.1')
which has jump discontinuity along the curve x = 'y(t). Put
Then the entropy condition is expressed as follows: For any value v between u.k. (t)
and u_(t), it holds that
f(v) 1(u)
S[v,u_] S[u÷,u_] where slv,uI (5.13)
§5.4. ENTROPY CONDITION 53
A jump discontinuity of weak solutions of (5.1) satisfying (5.13) is called a

"shock" or "shock wave." The condition (5.13) is also important from the viewpoint
of stability. For example, B.K. Quinn [115] has proved the following.
Theorem 5.6. Assume that g(t,x,u) defined in (5.1') is identically zero. If ti =

u(t, x) and v = v(t, x) are piecewise continuously differentiable weak solutions of
(5.1') for all x and I > 0 with initial data 110(5) and 110(5) which are piecewise
continuously differentiable and L'-integrable ins, and if u = u(t,x) and v = v(t,x)
satisfy the condition (5.13) along discontinuity curves, then it holds that
IIu(t) — v(t)IILI < — VOIIL1. (5.14)
Conversely, if (5.14) is true, then the condition (5.13) is satisfied.
We would like to show that our solution (5.12) is reasonable in the above sense.
That is to say:
Theorem 5.7. The solution defined by (5.12) satisfies the entropy condition (5.13)
in a neighborhood of the point (t°,x°).
Proof. We consider the case where (5.10) is satisfied. By Lemma 5.2, we have
u_(t) > As the inequality (5.10) means the convexity off = f(t,x,u) with
respect to u in a neighborhood of (t0, v°), we get
f'(t,x,u_) > f'(t,x,u+)
along a jump discontinuity, from which we can easily obtain (5.13). 0

Next we extend the weak solutions for large I. If u) does not
change the sign, we can extend the solutions with singularities by the above method
and also treat the collision of shocks just as in §4.4. But, if (52f/0u2)(t, s, u)
changes its sign, the solutions may sometimes lose the entropy condition. Then we
must introduce other types of singularities , for example "contact singularity." This
is the subject which will be discussed in Chapter 7.
Remark 1. the above discussions, we can say that the essential difference
between equations of conservation law and Hamilton-Jacobi equations is the global
solvability of ordinary differential equation (5.5)-(5.6) with respect to p = p(t,y).
This property determines whether the singularities of generalized solutions, or of

weak solutions, are continuous or not.
Remark 2. We solved the ordinary differential equation (5.9) to construct shocks

for quasi-linear differential equations of first order (5.1). But, for equations of
conservation law in one space dimension, we can reduce the construction of shocks
to the singularities of solutions for Hamilton-Jacobi equations.
Suppose that u = u(t, x) satisfies the following equation of conservation law:
+ = 0. (5.15)
Put tt(t,z) = (Ow/Ox)(t,x). Then w = w(t,x) satisfies
+ f(t,x, = 0. (5.16)
For Hamilton-Jacobi equation (5.16), we can construct the singularities of gener-

alized solutions, as done in §4.2. In this procedure, we do not need to solve ordinary
differential equations. Then we see by (4.9) and (4.13) that u = (Ow/Ox)(t,x) is a
weak solution of (5.15) which has jump discontinuity satisfying locally the entropy
condition (5.13). B.L. Rozdestvenskii had written this idea a little in [118]. But we
cannot apply this idea to quasi-linear equations of first-order which are not of con-
servation law because the above transform u(t,x) = (Ow/Ox)(t,x) does not work
well to arrive at Hamilton-Jacobi equation. Moreover the equations treated in [59]
and [78] do not depend on (t,x), i.e., I = f(u). By these reasons, the discussions
in Chapter 5 are necessary to construct the singularities of shock type for general
quasi-linear partial differential equations of first-order.
Chapter 6
Construction of Singularities
for Hamilton-Jacobi Equations
in Two Space Dimensions
Introduction
Consider the Cauchy problem for a Hamilton-Jacobi equation in two space dimen-
sions as follows:
in {t>O,xER2}, (6.1)
u(O,x) = on {t = 0, x E R2}, (6.2)
where f = f(p) is of class C°° and 4i = is in S(R2). Here, S(R2) is the space of
rapidly decreasing functions defined in R2. We assume that f = f(p) is uniformly
convex, that is to say, there exists a constant C > 0 such that
ffl() de=f 8
C. I > 0. (6.3)
[ api p, J
This chapter is continued from Chapter 4. Our aim is to construct the singular-
ities of generalized solutions of (6.1)-(6.2) in two space dimensions. The difference
between Chapter 4 and this one is the dimension of spaces. The crucial part in our
analysis is to solve the equation x = z(i, y) in a neighborhood of a singular point.
To do so, we must see the canonical forms of singularities of smooth mappings in a
neighborhood of a singular point. Though the singularities of smooth mappings are
simple in the case of one space dimension, this subject is difficult and complicated
in the case of higher space dimensions. Here we apply the well-known results of
H. Whitney [159] to get the canonical forms of singular points of the smooth map-
pings. This is the reason why we restrict our discussions to the case of two space
dimensions. First, let us repeat the definition of generalized solutions of (6.1).
Definition. A Lipschitz continuous function u = u(t, x) defined on x R2 is

called a generalized solution of (6.1)-(6.2) if and only if
56 6. CONSTRUCTION OP SINGULARITIES IN TWO SPACE DIMENSIONS
I) u = u(t,x) satisfies the equation (6.1) almost everywhere in R' x R2 and the
initial condition (6.2) on {t = 0, x E R3}
ii) U = u(t, x) is semi-concave, i.e., there exists a constant K such that
u(t,x + y) + u(i,x — y) — 2u(t,x) for all x,y ER2, t >0. (6.4)
Remark. Put v1(t,x) (Ou/0x1)(t,x) (i = 1,2). Then the equation (6.1) is

written down as a system of conservation law:
a a
+ = 0 (z = 1,2). (6.5)
The inequality (6.4) turns into the entropy condition for (6.5). See a remark in
§6.3.
§6.2. Construction of solutions
Characteristic curves for (6.1)-(6.2) are defined as solution curves of ordinary dif-
ferential equations as follows:
= = —f(p)+ (p,f'(p)), =0 (1 = 1,2),
r1(0) = v(0) = = (i = 1,2),
where f'(p) = and (p, q) is the scalar product of two vectors p

and q. Solving these equations, we have
x =p+ (6.6)
v = v(t, y) = 4'(y) + + (6.7)
Then is a smooth mapping from R2 to R2 and its Jacobian is given by
= det[I +
We write A(y) and denote the eigenvalues of A(y) by

When the space dimension is one, .X1(y) = Since >
§6.2. CONSTRUCTION OP SOLUTIONS
0 and 4 = is in 8(R), it follows that = A1(y) must take negative values at

some points. In this case also, we can prove that
minA1(y) = Aj(y°) = —M < 0.

p
Put t° 1/M. First assume t < t°. As
forany yER2,
we can uniquely solve the equation (6.6) with respect to y and write y = y(t,x).
Thenu = u(t,x) def= v(t,y(t,x))asauniqueclassicalsolutionof(6.1)-(6.2)fort <t o
. . .
.
Our problem is to construct the solution for i > t°.

Suppose that t — t0 is positive and sufficiently small, and consider the equation
(6.6) in a neighborhood of (t°,y°). The Jacobian of vanishes on {y E
R2 : 1 + = 0). Assume the condition:
(S.1) Ai(y) Az(y), gradAi(y) 0 on and is a simple closed curve.
In this case, is paranieterized as = {y = : a E I}, where!

is a closed interval and = yj(S) is in C°°(I) (i = 1,2). Put
{y(s°) E : = 0 at a = s°}.
By the definition of H. Whitney [159], a point in is a fold point of the

mapping i.e.,
tx(t,y(s)) 0 for y(s) E
because it holds that
= = 0. (6.8)
Lemma 6.1. Assume that the number of elements of E7 is two. Then
=0 for y(s) E
Proof. Put
I + tA(y) = , a1(t,y) ER2 (i = 1,2).

58 6. CONSTRUCTION OF SINGULARITIES IN TWO SPACE DIMENSIONS
Then a1(t,y) and a2(t,y) are linearly dependent on As they are smooth in
the interior of they can not take every direction of R2. That is to say, when
(t, y) moves in a small neighborhood of (t°, y°), a small
neighborhood of a1(t°,y°) (s = 1,2) where ai(t°,y°) and a2(t°,y°) are linearly
dependent.
Contrarily, when the point y = y(s) makes a round of (dy/ds)(s) takes all
the directions. Therefore (d/ds)x(t, y(s)) vanishes at least at two points. But we
see by (6.8) that the points where it vanishes are contained in Hence we get
this lemma. 0
Assume here the following condition:
(S.2) E = {Y1,Y2}, i.e., the number of element.s of is two; andY1 (z = 1,2) are
cusp points of H1, i.e.,
0 at y(s) = (i = 1,2).
Remark. Assume:
(C.1) The singularities of A1 = A1(y) are non-degenerate; i.e., ifgradA1(y) = Oat

a point y, then the Hessian of A1 = A1 (y) is regular there.
(C.2) (Ov/Oy)(t°,y°) 0.
Then, for t > t° where t — t0 is small, becomes a simple closed curve and the
number of elements of is two.
We denote the restriction of v = o(t,y) on E1 by yE = VE(t,y). We see by

(6.8) that = takes its extremum on especially if we put
c, (i = 1,2) and suppose c1 <c2, then vE(t,y) takes its minimum at y = Y1 and
its maximum at y = Y2. Denote by D1 the interior of the curve and by the
interior of Then the curve {y E 1(2 v(t,y) = c1} is tangent to D1 at
:
y = Y1 (i = 1,2). Here we apply the results of H. Whitney [159]. According to his

theorem, the canonical forms of fold and cusp points are expressed respectively as
follows:
= x2 = Y2 in a neighborhood of a fold point, (6.9k)
= — z2 = in a neighborhood of a cusp point. (6.92)

§6.2. CONSTRUCTION OF SOLUTIONS 59
This means that the mapping can be regarded as the mappings of (6.9k) and
(6.92) in a neighborhood of a fold point and a cusp point, respectively. Moreover he
proved that any smooth mapping from R2 to R2 can be approximated by smooth
mappings whose singularities are only fold and cusp points. We see by his result
that the curve has the cusps at X1 and X3 where (i = 1,2).
When we solve the equation x = with respect toy for x E the expressions
(6.9i) and (6.92) say that the solution y = y(t,x) takes three values. Write them
by y = (i = 1,2,3). Here we choose y = g2(t,z) as g2(t,x) E for x E
When we write u1(t,x) v(t,g1(t,z)) (1 = 1,2,3), the solution of (6.1)-(6.2) has
three values {u1(t,x) i = 1,2,3)) for x E
: (see Figure 6.1 which shows these
situations).
V Ct. y)
H;
Figure 6.1. Curves and
Lemma 6.2. i) -u;(t,x) = for x (i = 1,2,3).
ii) — (g1(t,x) — g,(t,x)) < 0 for x and

iii) ui(t,x) < u2(t,x) and us(t,x) < u2(t,x) for x E

Proof. i) This is equivalent to Corollary 1.3 in §1.1.
ii) Ftom the definition of gj(t, x), we have
x = 91(t,x) + xE
As g1(t,x) g1(t,x), it follows that (ôu;/Ox)(t,x) (Ou,/c3x)(t,x) for i j.

Using the convexity off = f(p), we get the inequality ii.
iii) We prove the first inequality. Divide the simple closed curve into
two parts joining two cusp points Xi and X2 of and write them by C1 and
C3. Here we introduce the family of solution curves of the following differential
equation:
dx
T =gi(t,x)—gz(t,x).
These curves start from C1 (or from C2) and end at C2 (or at C1, respectively),
and the family of these curves covers the domain On each curve it holds that
d 0u2
— u3(t,x)} (91 —92) < 0.
= (j— —
Since ui(t,x) = u2(t,x) on C2 (or on C1), we get u1(t,x) <u3(t,x) in
We are looking for a continuous solution. The iii of Lemma 6.2 means that we
cannot attain our aim by jumping from the first branch {(i,x,u) ii = uj(t,x)} :
to the second one {(t,x,u) u = uz(t,x)} and also from the second branch to the
:
third one {(t,x,u) : u = us(t,x)}. The last choice is to advance from the first
branch to the third one.
Lemma 6.3. Put I(t, x) ui(t, x) — us(t, x). Then {x I(t, x) = O}
is a smooth curve contained in and it joins two cusp points X1 and X2.
Proof. In this case we introduce the family of curves defined by
dx
T = gi(t,x) — gs(t,x).
These curves also start from C1 and end at Cz, and the family of the curves covers
the domain if we change the index "i" of g1(i,x) (i = 1,2,3), the above solution
curves may start from C2 and end at C1. On each solution curve, it holds that
d Ous\
= (91 —ge) <0. (6.10)
—
§6.3. SEMI-CONCAVITY OP THE SOLUTION u = u(t,z) 61
We see by Lemma 6.2 that the sign of is different from that of

I(t,x) = 0 has a unique solution on each solution curve of
(6.10). bbviously, I(t,x) = 0 at x = X1 and x = X2. As it follows from ii of
Lemma 6.2 that
for
we see that = {x I(t,x) = O} is a smooth curve joining the points Xj
a single valued and continuous solution, we define a solution

def
u = u(t,x) in {(t,x) : xE as follows: Writing = {x E : us(t,x) —
u1(t,x) > 0} and dl{x E u,(i,x) — u1(t,x) <0), we define
:
U (t )
I if x E
— us(t,x) if x
As is smooth, it can be parameterized as = {x = x(s)}. Then

d 1ôu1 OU3\ dx
= = 0.
.
This means that, though the first derivative of the solution u = u(t, x) has jump
discontinuity along the curve it is continuous with respect to the tangential
direction of rg.
§6.3. Semi-concavity of the solution u = u(t, x)
Let n(t, x) be a unit normal of advancing from ag,... to and define at x E
—(t,x±0) def.
Ou OU
= hm —(t,x±en). e-4+O8X
Any C2-function satisfies the semi-convexity condition (6.4). Therefore, for the
proof of (6.4), it suffices to treat the case where x and y = en (e > 0). Then
we have
u(i,x +y)+u(t,x — y) — 2u(t,x) = +sy) — +0)). yds
—0)— — .sy)) yds + + 0)— —0)) . y.

The first and second terms are easily estimated by K1y12. To get the inequality
(6.4), it must be
+ 0) — x— 0)) n(t,x) <0. (6.11)
Contrarily, if (6.11) is true, then we can get (6.4). Hence (6.11) is equivalent to the
semi-concavity property.
On the other hand, we have, by the definition of
us(t,x) — ui(t,x) > if xE
and
u3(t,x) — uj(t,x) < if xE
which show immediately that
+ sn) — ui(t,x + 0.
That is to say,
— . n 0, x E 1's. (6.12)
From the definition of u = u(t, z), for x E flj, we have

Ou OiL 0U3
and
Substituting these relations into (6.12), we get (6.11). Thus u = u(t,x) is semi-
concave. Summing up the above results, we get the following.
Theorem 6.4. After the time t0 where the Jacobian of the mapping vanishes
for the first time, the solution takes many values. But we can uniquely pick up only
one value from them so that the solution becomes single valued and continuous.
Then the condition of semi-concavity is automatically satisfied.
Remark. Putting v(t,x) def

.
= (Ou/Ox)(t,x) in (6.11), we have the condition on the
jump discontinuity of v = v(t, x):
(v(t,x+O)—v(t,x—O)).n<0, XE
This is the entropy condition for the system of conservation law (6.5) given in the
remark of §6.1.
§6.4. Collision of singularities
In this section we consider the collision of two singularities and construct-

ed in §6.2, assuming the hypotheses (S.1) and (S.2). Here we use the notations
E,,g, for (i = 1,2) which correspond to for introduced in
§6.2. We see that there exist three kinds of collision as described in Figure 6.2.
Case (i): Consider the case where and collide as in (i) of Figure 6.2. Then
the solution becomes two-valued on a domain encircled by and By almost
the same discussions as in §6.2, we can uniquely pick up one value from two so that
the solution is single valued and continuous. Then we can prove that the solution
is semi-concave. In this case the new singularity appears as a smooth curve joining
two points where and intersect each other. It is described as a dotted curve
in Figure 6.2.
r.
t
(1) (iii)
Figure 6.2. Collision of singularities

Case (ii): Consider the collision of the type (ii) drawn in Figure 6.2. We put
deC
= and
{y E ri,, and p E (i = 1,2).
Then is a smooth simple closed curve which is tangent to at y = Y1,1 and

l',2 (i = 1,2). When the end point of r2,1 is on A2,g is tangent to
A where A = Y2,1 or A = 1'2,2 (see Figure 6.3).
Yl,'
1A2
Figure 6.3. Relation between and
v(t,.) restricted on
As does not take an extremurn at y = A, we get
(Ov/äy)(i,y) Oat p = A; i.e., the curve CA (p E R2 v(t,y) = v(t,.)I.i} is
:
smooth in a neighbourhood of p = A, and it intersects A transversally.

On the other hand, as v(t,.) restricted on takes an extremuni at p = A, the
curve CA is tangent to at p = A. This is in contradiction with the above.
Hence the case (ii) does not happen.
Case (iii): When and meet at a time t = as shown in (iii) of Figure 6.2,
drawn as in (i) of Figure 6.4. But, as the interior domain of the
U E3,to is
curve = {y E R2 1 + tAi(y) = O} is monotonously increasing, U is
described as in (ii) of Figure 6.4 for t > t°. When it satisfies the condition (S.l)
and (S.2), we can construct the singularity of solution by just the same way as in
§6.2.
Remark on Figure 64. Assume that meet at y = y° = (a,b) and

and
that the singularities of Aj = Ai(y) are non-degenerate. As X1 = Xi(y) does not
take minimum and maximum at y = y°, we can suppose by Morse's lemma
.X1(y) = + (yi — a)2 — (y2 — b)2, 1 + t°Ai(y°) = 0.
Therefore (i = 1,2) have singularities at y = y°. But, for t > t°, the curve
{y R2 1 + tX1 (y) = 0} is smooth in a neighborhood of y = y°.
>2. t0
(ii)
Figure 6.4. Change of E1,2 U with respect to time
Summing up these results, we get the following.
Theorem 6.5. Assume that the assumptions (S.1) and (S.2) are always satisfied.
Then, even if two s;ngularities collide with each other, we can uniquely pick up one
value from two values of solution so that the solution becomes single-valued and
continuous. In this case also, the condition of semi-concavity is satisfied.
Remark. What we have done until this point is the local construction of singular-
ities of generalized solutions or weak solutions. The next problem is to consider the
global behavior of singularities. This subject has been considered in several cases.
For single conservation laws, see D.G. Schaeffer [1201 for n = 1 and B. Gaveau 156)
for n = 2 where n is the space dimension.
Chapter 7
Equations of Conservation Law
without Convexity Condition
in One Space Dimension
§7.1. Introduction
We consider the Cauchy problem for an equation of the conservation law in one
space dimension as follows:
in {t>O,xER'}, (7.1)
u(O,z) = 4)(x) on {t = 0, x E R1}, (7.2)
where f = 1(u) is of class C°°. We assume the initial data 4) = 4)(x) to be in

Cg°(R'), except in §7.2 where the Riemaun problem will be discussed. Even in
the case where f = f(u) is not convex, the global existence of weak solutions of
(7.1)-(7.2) has been well-studied, for example by O.A. Oleinik (112J, S.N. Kruzhkov
[881, [92J. Our interests are in the singularities of weak solutions for general par-
tial differential equations of first-order. In Chapters 4, 5, and 6, we have locally
constructed the singularities of generalized solutions or weak solutions. Our next
problem is to extend the weak solution in the large.
In this chapter, we consider the Cauchy problem (7.1)-(7.2) in one space dimen-
sion. if f = 1(u) is convex with respect to u, we can extend the singularities for
large t, because the entropy condition is always satisfied. Moreover, we can treat
the collision of singularities as in §4.4 and §6.4. But, if f"(u) changes its sign, we
may meet a new phenomenon which does not appear in the convex case. The aim
of this chapter is to explain this situation.
We will not cover here all the results obtained until now on this subject. We
present only fundamental notions in non-linear wave propagation and solve explic-
itly a certain example which shows a new phenomenon caused by the non-convexity
of the equation.
68 7. CONSERVATION LAW WITHOUT CONVEXITY CONDITION
§7.2. R.arefaction waves and contact discontinuity
In the Cauchy problem (7.1)-(7.2), we assume in this section that the initial data
= is given by
=
(a x<0 (7.3)
(b, x>O,
where a and b are constants with a b. The Cauchy problem (7.1)-(7.2) with the
initial data (7.3) is called "R.iemann problem." In this case, characteristic curves
are written as
x=x(t (y+tf'(a), y<O, (7.4)
I y+tf'(b), y>O,
a, y<O,
v =v(t,y)={ (7.5)
b, >0.
Assume that f'(b) — f'(a) > 0. Then the domain D {(t,x) tf'(a) <x <
:
tf'(b)} is not covered by the family of characteristic curves (7.4). In the domain D,
we look for a solution u = u(i,x) of the form u(t,x) = r(x/t). Then the function
r = r(p) (p = x/t) satisfies
Ou
+
8 x , 1,x = 0.
= + I (r)]—r
Therefore, if r = r(p) satisfies f'(r(p)) = p, then u = u(t, r) r(x/t) is a solution

of (7.1). As this means that r = r(p) is the inverse function of u i—* f'(u), we must
assume
f"(u) 0 for all u between a and b.
Now we define the solution u = u(i,x) by
a, x < tf'(a),
u(t,x) r(x/t), (t,x) E D, (7.7)
1
b, x <tf'(b).
The solution defined by u = r(x/i) in the domain D is called "rarefaction
wave." The solution (7.7) is continuous in {i > 0}, but it has jump discontinuity
at (t,x) = (0,0).
Next we consider the case where f'(a) — f'(b) > 0. Then the domain R
{(t, x) tf'(b) < x < tf'(a)} is covered two times by the family of characteristic
§7.2. RAREFACTION WAVES AND CONTACT DISCONTINUITY 69
curves (7.4). As we look for a single-valued solution, we repeat the same discussions
as in §5.3. Let x = -y(t) be the solution of the following differential equation:
— f(b)—f(a)
dt — b—a ' (7.8)
{ = 0.
Then we get easily -y(t) = {f(b) — f(a)}t/(b — a). Here we define a weak solution
u = u(t,x) of (7.1)-(7.2) in the domain R by
Ia,
x = -y(t) satisfies the entropy condition (5.13), then this singularity is just
"shock." But, if f = 1(u) is not convex, the inequality (5.13) is sometimes vio-
lated. In this case we use "contact discontinuity" which we will explain from now.
0 b C Li
Figure 7.1
Assume that a > b, and that the entropy condition (5.13) is not satisfied. Sup-
pose that the graph off = f(u) is drawn as in Figure 7.1. We draw tangent lines
L passing the points (b,f(b)) and (a,f(a)), respectively. Let L and
be tangent to the curve f = 1(u) at (c,f(c)) and (d,f(d)), respectively. See Figure
7.1. Then the slopes of L and are equal to f'(c) and f'(d), respectively. We
denote {(t,x) x = f'(d)i, t 0} and C'
: {(t,x) : x = f'(c)t, t 0}.
Put ci {(t,x) f'(c)t < x < f'(d)t, t > O} which is contained in R. The bound-
:
ary äf is equal to U C—. The shock curve x = comes into the domain ci.
We assume here that f'(u) is monotonously decreasing on [d, ci. For any (t, x) E ci,
we can pick up a value v E [d,c] f'(v) = s/t. If we use the function
r = r(p) introduced in the definition of rarefaction wave, we have v = r(x/t). We
define a weak solution u = u(i,z) of (7.1) in the domain R as follows:
b, f'(a)t > x > f'(d)t,

u(t,x) = r(x/t), f'(d)t > x > f'(c)t, (7.9)
a, f'(c)t > x > f'(b)i.
Though the solution u = u(t, x) has jump discontinuity along the curves
and C, it satisfies the entropy condition (5.13). As in the above example, a
shock whose speed equals the characteristic speed of one side is called a "contact
discontinuity."
§7.3. An example of an equation of the conservation law
Before giving an example, we explain the reason why we consider "contact discon-
tinuity." The propagation of singularities for equations of conservation law without
convexity condition has been studied, for example, by D.P. Ballou [9], J. Gucken-
heimer [59], and G. Jennings [78]. Their method is to construct locally the singular-
ities of weak solutions for the Cauchy problem (7.1)-(7.2) where the initial data are
"piecewise smooth." That is to say, though they started from the "smooth" initial
data (especially in [59] and [78]), they solved essentially the Riemann problem for
(7.1)-(7.2). Though the initial data are smooth in their discussion, a
rarefaction wave appears in the solution of the Cauchy problem (7.1)-(7.2). Let us
recall that the rarefaction wave has been introduced to construct a solution in a
region which is not covered by the family of characteristic curves. Therefore our
first question is whether or not we need to use rarefaction waves for solving the
Cauchy problem (7.1)-(7.2) with the smooth initial data. A typical phenomenon
which does not appear for convex equations is "contact discontinuity" introduced in
§7.2. See Figure 7.1 which explains the situation. Our second question is whether
or not the situation in Figure 7.1 would surely happen even for the Cauchy prob-
lem (7.1)-(7.2) with the smooth initial data. In answer to the above two questions,
§7.3. AN EXAMPLE OF AN EQUATION OF THE CONSEftVATION LAW
we will consider the following example: Assume that the graphs of f = 1(u) and
= are drawn as in Figure 7.2 (i) and Figure 7.3 (i), respectively. The graphs
of their derivatives appear in Figure 7.2 (ii)-(iii) and Figure (ii).
a1 0
U
Figure 7.2 (i)
F,
b2
/71 b1
U
Figure 7.2 (ii)

F"
F'' =
Figure 7.2 (iii)
Figure 7.3 (i)

§7.3. AN EXAMPLE OF AN EQUATION OP THE CONSERVATION LAW 73
45
Figure 7.3 (ii)
In the above case, it holds that 1(a1) = = 0 (i = 1,2), f"(b1) = 0 (1 =

1,2) and = = 0. Moreover, we assume
> b2,
because, if not, our problem is equivalent to a case where I = f(u) is concave on

the whole R1.
For the Cauchy problem (7.1)-(7.2), the characteristic curves are written by
x = x(t, y) = y + v = v(i, y) =
Therefore it follows that

Ox
y) = 1 +
We = f
set h(y) def (y), and assume that the graph of h = h(y) is drawn as
in Figure 7.4. Next, take A1 = (y1,h(y1)) (i = 1,2) where h(y1) <0 and h'(y1) = 0.
Then we have at A1
> 0 and <0,
and at A2
<0 and f"(4'(y2)) > 0.
A2
A
Figure 7.4
We now put t1def

= and X, def
= x(t;,yj) (: = 1,2). As we have shown in
§5.3, we see that a shock appears at each point (t1,X1) (i = 1,2), and we denote
it by (z = 1,2). We have proved in §5.4 that each shock S satisfies the entropy
condition for t > where t —is small. Our problem is to see what kinds of
t1
phenomena may happen when we extend the shocks (i = 1,2).
§7.4. Behavior of the shock
In this section we will extend the shock S1 for large t. To explain the situation, we
repeat briefly how we have constructed the shock Si. As the graph of x =
for t > t1 is drawn as in Figure 4.1, we use the same notations appeared there.
Solving the equation x = x(t,y) with respect to y for x E (x1(t),x2(t)), we get
three solutions y = g(t,x) (i = 1,2,3) with gj(t,x) < < g3(t,x). As
v(t,y) = for all (t,y) ER2, we define u1(t,x) (i = 1,2,3). Then
we have by Lemma 5.2
u1(t,x) <u2(i,x) <us(t,x).

§7.4. BEHAVIOR OF THE SHOCK S1 75
Here we must pay attention to the order of this inequality which is contrary to the
one given in Lemma 5.2. This is caused by the property that is negative,
while it is positive in Lemma 5.2.
As in §5.2, we consider the Rankine-Hugoniot jump condition as follows:
( dx — f(ui(t,x)) — f(u3(t,x))
dt — u1(t,z) — u3(t,x) ' (7.12)
I. x(ti)=Xj.
Proposition 5.5 says that the Cauchy problem (7.12) has a unique solution x =
which is the shock curve of S1. To extend the shock Sj, we must see the behavior
of = (i = 1,3). We put uj(t) u(t,-y1(t) ± 0). Then
= us(t, yj(t)) and u_(t) = uj(t,71(t)).
Lemma 7.1. As long as the entropy condition for S1 is satisfied, the function
t gj(t,-y1(t)) is decreasing and t gs(t,-y1(t)) is increasing.
Proof. By the definition of y = x), we have
x= x) + x))) (i = 1,2,3).
Taking the derivatives with respect to t and x, we have
0= + f'(çi(g1)) ÷
and
1 = 0g1{1 +
Hence it holds that
x) = x) (i = 1,2,3) (7.13)
which leads us to
(t)) = { — (t)))} (t)) (i = 1,2,3).
Here we recall
= > u1(i, -11(t)) = u_(t).
Then the entropy condition (5.13) gives us
= f'(u_(t))
and
f'(us(t,1'j(t))) =
Since > 0 for i = 1 and i = 3, we get
and
The proof is complete. 0

The initial function = ç6(x) has the properties as drawn in Figure 7.3. As long
as the entropy condition (5.13) is satisfied, we see by Lemma 7.1 that
advances toO, and that goes to its maximum. We write here =
and P_(t) = (u_(t),f(u_(i))). The shock Sj enjoys the entropy
condition for I > t1 where the graph of I = 1(u) lies entirely in the upper side of
the chord joining two points and P_(t). If we extend the shock S1
further, then the chord PTP may be tangent to the curve {(u,f(u)) u E R1}
in finite time. We will assume so, i.e., assume that is tangent to the curve
I = f(u) at I = T. Then we see obviously that is tangent to the curve
f = f(u) at the point but not at the point P_(T). Therefore it holds that
dll(T) = (7.14)
Lemma 7.2. i) (P11(T) > 0, ii) =

Proof. i) Using the definition (7.12) of (dyi/dt)(i), we have
d [ f'(u+) f(u+)—f(u_)] du+(1)

I))—
d12 — di di
+ [f(u+) — f(u_) — f'(u_) 1

— u_)2 — u_J di
As u÷(t) = we get by (7.14)
d
—u÷(t)
di t=T
i—+ Ous&yi
= FOILs
101 OrOl
=
— = 0. (7.15)
§7.5. BEHAVIOR OF THE SHOCK S2 77
Put Y1 = As u_(t) = u1(t,-11(t)) = we obtain by

(7.13)
= —
This leads us to
— fl(u_(T))J
= — u_(T)
Here we recall that the starting point of the shock S1 is the point (t1, X1). Therefore
it follows that qS'(Yj) >0, u4.(t) —u_(t) >0 and (Ogi/ôx)(t,x) >0. Hence we get
(&yi/dt2)(T) > 0.
Part ii is easily obtained by (7.15). 0
By Lemma 7.2, we have
> t > T.
This implies that the entropy condition is satisfied. Summing up the above results,
we have the following.
Proposition 7.3. The shock S1 starting from the point (t1, X1) always satisfies
the entropy condition.
§7.5. Behavior of the shock S2
In this section we extend the shock S2 which has appeared at the point (t2,X2).
The graph of x = x(t,y) for t > t2 can be drawn as in Figure 4.1. Though
x = x(t,y) in this section is different from x = x(t,y) in §4.2 and §7.4, we use
the same notations introduced there. As in §7.4, we solve the Cauchy problem
(7.12) with the initial condition x(t2) = X2, and denote the solution by r = '72(t).
We write here also = u(t,72(t) ± 0); that is to say, =
and u_(t) = Put = The shock S2 satisfies
the entropy condition (5.13) for t > t2 where t — t2 is small. In this case, as
> 0, (5.13) says that the graph of f = f(u) lies entirely in the lower side
of the chord
Lemma 7.4. As long as the entropy condition for S2 is satssfled, t

is decreasing and t g,(t,y2(t)) increasing.
We omit the proof, because it is similarly obtained as Lemma 7.1. When t gets
larger, tends to 0 and tends to the maximum =
Therefore we assume that, though the entropy condition is satisfied for t <
T, becomes tangent to the curve {(u,f(u)) u E R'} at t = T. Then it
follows that
(T) = =
Lemma 7.5. i) <0, ii) = 0.
The proof is almost the same as that of Lemma 7.2. By this lemma, we get the
following:
Proposition 7.6. For t > T where t —T is small, it holds that
This proposition means that the entropy condition is violated for t > T. To
overcome this point, we use "contact discontinuity" explained in
F.
V Li Li
Figure 7.5
§7.5. BEHAVIOR OF THE SHOCK S2
Consider the equation
u>v. (7.16)
We solve the equation (7.16) with respect to v, and denote the solution by v =
h(u). See Figure 7.5. To construct a contact discontinuity starting at the point
(t,x) = (T,-yz(T)), we solve the Cauchy problem as follows:
( dx
=
(7.17)
x(T) = 72(T).
We denote a solution of (7.17) by xand put S, def

= {(t,x) x = :
T}. Let W be a domain surrounded by the curve and a characteristic line

x = y° which passes through the point (t, x) = (T, So, the point
y° must satisfy the equation -12(T) = y° + Hence y° =
Figure 7.6 explains this situation.
t
x 72(t)
Figure 7.6
Our problem is how to define a weak solution u = u(t, x) in the domain W. As

the domain W is not covered by the family of characteristic lines whose starting
points are on the initial line {t = O}, we introduce a family of characteristic lines
which start from the curve i.e., we construct the "contact discontinuity." For
any point E we first define by where Next
— ..- — deC .
we determine u.k. = u+(t,x) by = h(u_) and draw a characteristic line which
passes through the point (t , as follows:
On the characteristic line, we define the value of the solution u = u(t, s) by u(t, x) =
As the family of the above characteristic lines covers the domain W, the weak
solution u = u(t, x) is completely defined on the domain W. It would be obvious
that the entropy condition is satisfied (see Figure 7.7).
T))
(T.y (T))
Figure 7.7
Repeating the above discussions, we can extend further the weak solution of
non-convex conservation law (7.1). What we would like to insist in the above
construction is that we did not use "rarefaction waves."
Example. For the Cauchy problem (7.1)-(7.2), we assume that the initial function
= 4S(x) is of a form as shown in Figure 7.8, and that it satisfies max > b2
§7.5. BEHAVIOR OF THE SHOCK 52
and mm 4'(x) < Then, after the collision of two shocks, we will arrive at the
situation in Figure 7.1 (see §7.2).
Figure 7.8
Remark. Recently, S. Izumiya [73] and S. Izumiya & G.T. Kossioris [75]-[761 are s-
tudying geometric singularities of generalized solutions of general partial differential
equations of first-order in the framework of "Legendrian unfoldings." Concerning
the propagation of singularities for non-convex first-order partial differential equa-
tions, the results of [76] would be the best at today's point.
Chapter 8
Differential Inequalities
of Haar Type
§8.1. Introduction
The theory of ordinary differential inequalities was originated by Chaplygin [30]
and Kainke 181] and then developed by [1571. Its main applications to
the Cauchy problem for (ordinary) differential equations concern questions such as:
estimates of solutions and of their existence intervals, estimates of the difference
between two solutions, criteria of uniqueness and of continuous dependence on
initial data and right sides of equations for solutions, Chaplygin's method and
approximation of solutions, etc. Results in this direction were also extended to
(absolutely continuous) solutions of the Cauchy problem for countable systems of
differential equations satisfying Carathéodory's conditions. We refer to Szarski [128]
for a systematic study of such subjects.
As for the theory of partial differential inequalities, first achievements were
obtained by Haar [61], Nagumo [107], and then by Wniewski [154]. Up to now the
theory has attracted a great deal of attention. (The reader is referred to Deimling
[40], Lakshmikantham and Leela 196], Szarski (128), and Walter [153]. In particular,
he is referred to [24]-[27] and the references therein for recent results in functional
setting.) We emphasize here that one of its applications to the Cauchy problem
for first-order partial differential equations, videlicet the uniqueness
criterion formulated in Theorem 1.8, is just for classical solutions and may only be
used locally. (For more details, see the introductory comments in the next section.)
The present chapter provides a new method, based on the theory of multifunctions
and differential inclusions, to investigate the uniqueness problem. This method
allows us to deal with global solutions, the condition on whose smoothness is relaxed
significantly. As we shall show more concretely in Chapter 11, the equations to be
considered satisfy certain conditions somewhat like Carathéodory's, and their global
semiclassical solutions need only be absolutely continuous in time variable.
§8.1. INTRODUCTION 83
The structure of the chapter is as follows. In Section §8.2 we introduce a so-

called differential inequality of Hoar type. (See (8.5) later. Note that (1.14) and
(1.16) were usually referred to as Haar's differential inequalities.) An estimate via
initial values for functions satisfying this differential inequality will be established
(cf. (8.6)-(8.7)). As an application, Section §8.3 gives some uniqueness criteria for
global classical solutions to the Cauchy problem for first-order nonlinear partial
differential equations. In this way, moreover, the continuous dependence on the
initial data of solutions can be examined. Finally, Section §8.4 concerns some
generalizations to the case of weakly-coupled systems.
Most of the results presented here were published in [141], [147]-[149J, and [151].
The relevant material on multifunctions and differential inclusions from [8] and [29]
may be found in Appendix II given at the end of the book.
Throughout this chapter, 0 < T < +oo, and
ciT
The notation 0/Ox will denote the gradient (0/On... Let .j and (.,.)
be the Eucidean norm and scalar product in R", respectively.
Denote by Lip(ciT) the set of all locally Lipschitz continuous functions u =
u(t,x) defined on QT. F'ürther, set Lip([0,T) x x R").
For every function u = u(t, x) defined on cii', we put
Dif(u) {(s,y) E cii' : u = u(i,x) is differentiable at (s,y)}.
We shall be concerned with the following class of Lipschitz continuous functions:
E Lip([0,T) x : [0,T] mes(G) = 0,
Dif(u) J ciT\(G x RTh)).
(Here, "mes" signifies the Lebesgue measure on R1.)

In other words, a function u = u(t, x) in Lip([0, T) x belongs to if and
only if for almost all t, it is differentiable at any point (t, x).
84 8. DIFFERENTIAL INEQUALITIES OF UAAR TYPE
§8.2. A differential inequality of Haar type
First, several comments are called for in connection with the classical uniqueness
Theorem 1.8, whose proof is essentially based on Theorem 1.6 or Haar's
Theorem 1.5. Our comments will concern the Cauchy problem in the large for a
general first-order partial differential equation as follows:
Ou/Ot + f(t,x,u,Ou/Ox) = 0 in (8.1)
u(O,x) = on {t = 0, x E RTh}, (8.2)
where the Hainiltonian f = f(t,x,u,p) is afunction of(t,x,u,p) E 11r x R1 x

and = 4(x) is a given function of x E
Assume the function f = 1(1, x, u, p) to be locally Lipschitz continuous with
respect to (u,p) in the sense specified in §1.3; i.e., for any bounded set K C
x R1 x there exist nonnegative numbers L1,. . , and M such that
.
If(t,x,u,p) — — qd + — UI (8.3)
for all (t,x,u,p) and (i,x,v,q) in K. Let u1 = uj(t,x) and u2 = u2(i,x) be

two C1-solutions on the whole 11T of (8.1)-(8.2). Then Theorem 1.8 assures the
equality ui(t,x) = u2(t,z) in a neighborhood of {t = 0, x E in The
question arises as to whether this equality can be extended to the entire domain
In answer to this question, we must go back to Theorem 1.6. It seems to
us that there is no standard procedure for joining a point (t°, x°) E to the
hyperplane {t = 0, x e by consecutively gluing pyramids of the form
{(t,x) : 01 < t a2, cj +L1t < d1 — L1t (i = 1,...,n)}, (8.4)
where
0<ai<a2<T, (i=1,2,...,n),
so that Theorem 1.6 can simultaneously work therein. The reason is that the re-
lat ions between L1,. . , and. are bilateral and somehow awkward. In fact,
L1,. . , . are to some extent overdetermined by Specifically, for each applica-
tion of the theorem in such a procedure, must be ready-given. Thus, by (8.4), a
§8.2. A DIFFERENTIAL INEQUALITY OF HAAR TYPE 85
tuple (L1,. .. , is predetermined. However, the further condition (8.3) is needed

for (i, x, u,p), (t, x, v, q) E x K, with K being a compact set in R1 x R" such that
(u,(t,x),(ôu,/Ox)(t,x)) E K for (t,x) E j = 1, 2.
Therefore, L1,. . . are necessarily Lipschitz constants with respect to ps,... , Pvi
,
of the function f = f(i, x, u,p) restricted to x K. But these constants might un-
fortunately become large, say, substantially greater than the above-predetermined
values L1 So we would reduce and hence possibly fail to touch the by-
perplane {t = 0, x E in such a procedure emanating from a point (t°, x°) E
All the preceding remarks suggest that we should make an attempt to develop
the theory of first-order partial differential equations. The aim of this section is to
prove the following, which is in fact a generalization of Theorems 1.5 and 1.6. (We
put off discussing its applications to the uniqueness problem to §8.3 and Chapter
11.)
Theorem 8.1. Let u = u(t,x) be a function in If there exist a nonnegative

function p = p(x) locally bounded on and a nonnegative function £ = £(t) in
L' (0, T) such that
8u(t, x)/OtI e(t) . [(1 + Ixl)IOu(t, x)/OxI + p(x)Iu(t, x)IJ (8.5)
for almost every t E (0, T) and for all x E then
u(t,x)I exp [c(x)f £(r)dr] . sup u(O,y)I, (8.6)

£(r)dr—i
where
T
C(x) (Ip(y)I : (1 + xI)exPf £(r)dr — i}. (8.7)
Remark 1. Inequalities (1.14) and (1.16) are usually referred to as Haar's dif-
ferential ones (see [128, Corollary 37.1]). Therefore, we would like to call (8.5) "a
differential inequality of Haar type."
Remark 2. We show by the following example that the Lipschitz continuity of

u = u(t,x) is essential in Theorem 8.1.
86 8. DIFFERENTIAL INEQUALITIES OF HAAR TYPE
Let J C [0, 1J be the Cantor set, i.e., the set of all numbers of the form
(8.8)
where each ek is either 0 or 2. The set J is complete, nowhere dense on R1, and is
of Lebesgue measure 0 (see [42J).
We define a function w = w(t), which is called the Cantor ladder, or the Cantor
function [53, p. 381, in the following way. For t E J given by (8.8), we take
+00
def
w(t) def
= k
where =
If (a,f3) is an open maximal interval in (0, 1)\J, then J and = wCO).

We set for t E w(t) def
= w = w(i)
is continuous on [0, 1] and that dw/dt = 0 almost everywhere in (0, 1). In fact,
dw(t)/dt = 0 fort E (0,1)\J.
Setting for (t,x) we easily see that u = u(t,x) belongs to
C1(((o,1)\J) x flC([0,11 x with u(O,x)
0 V(t,x) ((0,1)\J) x
The function u = u(t, x) thereby satisfies all the conditions of Theorem 8.1 except
for the Lipschitz continuity. This explains why (8.6) does not hold: u(t,x) 0.
Proof of Theorem 8.1. For an arbitrary point (t°,x°) E we must prove that
Iu(t°,x°)I exp
[C(x0)ft £(t)dt] sup u(0,y)[. (8.9)
o
Ivt(1+Ie°I)exp
xI<r},r 0. DenotebyE,(t°,x°)thesetof
all absolutely continuous functions x = x(t) from I [0, into which satisfy
almost everywhere in I the differential inclusion subject to
the constraint x(t°) = x0.
From [29, Theorem VI-13], it follows that Ej(t°, x°) is a nonempty compact set
in The sets Z(t,t°,x°) {x(t) x(.) E Ej(t°,x°)} and I'(t°,x°)
:
{ (r, x) r E I, x E Z(r, t°, x°) } are therefore compact in R" and

: respective-
ly, for all t E I. Moreover, by the converse of Ascoli's theorem, the correspondence
I t '-4 Z(t, t°, r°) C is a continuous multifunction.
§8.2. A DIFFERENTIAL INEQUALITY OF HAAR TYPE 87
We now define a real-valued function g = g(t) on I by setting
g(t) {Iu(t,x)I : xE Z(t,t°,x°)}.
Then according to the maximum theorem (see [8, Theorem 1.4.16]), the fact that
u = u(t,x) is continuous on r(t°,x°) implies that g = g(t) is continuous on I. In
addition, we have:
Lemma 8.2. For an arbitrarij number 9 E (0, t°), the function g = g(t) is absolutely
Contznuou3 on [9,t0].
The following assertion will also be needed.
Lemma 8.3. We have for everij t E I the inclusion
Z(t,t°,s°) C . (8.10)
Proof of Lemma 8.3. For each 0, let
m,7(t) + Ix°I £(r)dr —1.
The function rn,7 = rn,7(t) is absolutely continuous, positive on I with the derivative
dm,7(t)/dt = —i(t). (i + rn,,(t)). To prove (8.10) we have only to show that
Ix(t)I <m,,(t) Vt E I
for all x = x(t) in Ei(t°,x°) and for all

Since > = Ix(t°)I, there exists a number E (0,t°) such that
m,7(t) > Ix(t)l whenever t (t° — & t°].
Assume that (8.11) were false, so that there exists t' [0,t°) such that
x(t')I. Setting t1 {t E [0,t°) rn,(t) < x(t)I} < t0, we would have:
:
Ix(t1 ) I = rn,7(t), Ix(t) I < m,,(t) Vt E (t', i°],
and
drn,7(t)/dt = . (1 + m,7(t)) . (i + Ix(t)I)
<dlx(t)I/dt
almost everywhere in (t',t°). On the other hand,
[0
>
88 8. IMFPERENTIAL INEQUALITIES OP HAAR TYPE
if and only if
m,,(t°) — m,,(t') = m,(t°) — Ix(t1)I > Iz(t°)I — Ix(t1)I.
Hence we obtain a contradiction. This proves Lemma 8.3. 0

Proof of Lemma 8.2. Since u = u(t,x) is locally Lipschitz continuous in
there exists a number L 0 such that
Iu(t1,x1) — u(t2,x2)I — t21 -+ Ir' — x21)

V (i',z1), (t2,r2) E (fO,t°] x
By the absolute continuity of the Lebesgue integral, Lemma 8.2 will be proved
if we can show that
g(t1) — g(t2 )I L [its — I+ (1 + 200.

f
(*1*21
£(t)dt . exp
j £(i)dt]
8 12
Vt',t2 E [8,t°].
Now let
g(t') g(t2) and g(t') = Iu(i', x(t1))I
for some x = x(t) in Ei(t°,z°). Since x(t2) E Z(t2,t°,x°), we have
0 g(t') — g(i2) = Iu(t', x(i' ))I — g(t2)

u(t', x(t1 ))I — Iu(t2, x(t2))I u(t1, x(t')) — u(t2, x(t2))I
L[1t1 — gal + lx(t1) — x(t2)l] =
L [hi — £(t) (i + x(t)l)dt]

+
[*1 ,t2J
Therefore, (8.12) follows from Lemma 8.3. The proof is then complete. 0
Going back to the proof of Theorem 8.1, we set now
for t€(0,Tj.
By Lemma 8.3 and the definition of g = g(t), the inequality (8.9) will be obtained
if we show that
g(t) g(0) .exp[C(x°) . h(i)] Vt E [0,t°]. (8.13)

§8.2. A DIPPERENTIAL INEQUALITY OP HAAR TYPE 89
For every > 0, let
g,(t) [g(0) + exp { [C(x°) + . [h(t) + itt]

}.
To get (8.13), it suffices to prove that
g(t) <g,,(t) Vt E [0,t°]. (8.14)
Let w(t) —g(t), where is temporarily fixed. Then (8.14) is equivalent

to w(t) > 0 Vt E [0,,0). Obviously, w(0) = > 0. We shall show that
w(0) Vt E [0,t°). Assume this is false, so there exists t' (0,t°[ such that w(t') <
w(O).
It is well-known that there exists a set G1 C (0, T) of Lebesgue measure 0 with
the property that
dh(t)/dt = £(t) Vt E (O,T)\G1.
By the hypothesis of Theorem 8.1, we find a set G2 C (0, T) also of Lebesgue
measure 0 such that fZT\(G2 x R") C Dif(tt) and that (8.5) holds for all t E
(O,T)\G3, x E RTh.
Since the image of a null set under an absolutely continuous mapping is also a
null set, Lemma 8.2 implies
mes(w(Gn(9,t0j)) =0 VO€(o,t°),
def
where C = C1 U C2. So
mes (C fl [0, tOl)) = limmes (w(G fl [9, tO])) = 0. (8.15)
From (8.15) and the continuity of w = w(t) on I we conclude that there is a

number A with
max{0,w(t')} < A and A w[0, fl [0, t°]).
Let
:
It is obvious that w(t.) = A, t. E (0,t')\G, and that w(t) > A Vt [0,t.).

Suppose that
g(t.) = Iu(t.,z.)I = . u(t.,x,), e

for some E Z(t.,t°,x°). Then one may find a function x = x(t) in Ej(t°,x°)
so that x(t.) = Choose a unit vector e E with
(e,e = — . (8.16)
The system (of n ordinary differential equations)
has a C'-solution y = y(s) on R1 satisfying the condition y(h(t.)) = x•. Let

x(t) def
= y(h(t)) fortE [0,1']. Of course, x = x(t) is absolutely continuous on [0,1'],
x(t.) = and
= = £(t) (1 + x(t)]) e Vt E (0,T)\G1.
Therefore, the function j = .,x(t) defined by
def I x(t) if 0 < t

.x(t)
= 1.
x(t) E Z(t,t°,x°) Vt E
This implies
u(t,z(t)) < Iu(t,x(t))I < g(t) = — w(t) < g,7(t) — A (8.17)
for all I E I0,t.). Besides that,
u(t., x(t.)) = Iu(t.,x.)I = g(t.) = g,,(t.) — w(t.) = — A. (8.18)
Furthermore, since I. E (0, T)\G, we see that:

(i) u = ts(t,x) is differentiable at
(ii) x = x(t) is differentiable at with
(iii) g,, = g,1(t) is differentiable at I. with
[C(x°)+q] .
§8.3. UNIQUENESS OF GLOBAL CLASSICAL SOLUTIONS
So it follows from (8.17)-(8.18) that
[e u(t, x(t))]
Consequently,
Ou /dx i9u
e
[C(x°)+v7] .g,1(i.).
Hence,
Ou
x.) + £(t.)(1 + Ix.I)
/ •
Ou
x.)
•
[C(x°) + . [€(t,) + . EIu(t.,x.)I + A].
Because 0 and A > 0, the last inequality together with (8.16) implies
> £(t.) [(i + Ix.I) . + C(z°) . . (8.19)
On the other hand, since E Z(t.,t°,x°), Lemma 8.3 yields
Jx,I (1 + Jx°I)exp / —1 (1 + Ix°I)exp I £(r)dr — 1.

ft. Jo
Therefore, the formula (8.7) gives C(x°) lp(x.)l, which shows that (8.19) con-
tradicts (8.5). It follows that there exists no t' E [0,t°] with w(t') < w(0). Thus,
w(O) > 0 for all t [0,t°]; the inequality (8.14) is thereby proved. This
completes the proof of Theorem 8.1. 0
§8.3. Uniqueness of global classical solutions

to the Cauchy problem
The advantage of Theorem 8.1, as we have mentioned in the introduction, is that

it allows us to discuss the so-called global semiclassical solutions, which are just
absolutely continuous in time variable, for first-order nonlinear partial differential
equations with time-measurable Hamiltonians. This will be taken up in Chapter 11,
where an answer to a problem of S.N. Kruzhkov [931 is given. In the present section
we restrict ourselves to the case of C'-solutions, dealing with some applications of
Theorem 8.1 to stability questions concerning the Cauchy problem in the large for
partial differential equations of first-order, namely the problem (8.1 )-(8.2). Even
in this "classical case," using the a priori estimate (8.6)-(8.7) of Theorem 8.1, we
find some new uniqueness criteria (posed on the Hamiltonian I = f(t,x,u,p)) for
global C' -solutions of (8.1)-(8.2). Let us first repeat the definition of solutions to
be considered.
Definition 1. A function u = u(t, x) in C' (liT) fl C([O, T) x is called a global

C'-solution to the Cauchy problem (8.1)-(8.2) if it satisfies (8.1) everywhere in liT
and (8.2) for all x E RTh.
As was shown in the introductory comments of §8.2, for the uniqueness of global
C'-solutions, the following result may be invoked instead of Theorem 1.8.
Theorem 8.4. Suppose I = f(t, x, u, p) satisfies the following condition: there

exist nonnegative numbers L, M such that
lf(t,x,u,p) — L(1 + lxl)ip — qi + — (8.20)
for all (t,x,u,p),(t,x,v,q) E QT x R' x If ui = u,(t,x) and u2 = u2(t,r) are

global C'-solutions to the problem (8.1)-(8.2), then u,(t,x) u2(t,x) in 11T•
del
Proof. Consider the function u = ts(t, x) = u,(t, x) — u2(t, x). Then u(0, x) 0.
Furthermore, by (8.20) and the definition of global C' -solutions, we have
= f(t, x, u,(t, x), x)) — f(t, x, u2(t, x), x))
< L(1 + lxi) x) — + z) — u2(t, x)I
= L(1 + lxi) + Mlu(t, x)i
for all (i,x) E Now it follows from Theorem 8.1 that u(t,x) 0 in This
proves the theorem. 0
The next sharpening (and its corollary) of Theorem 8.4 will give some useful
uniqueness criteria for global C' -solutions with bounded derivatives.
Theorem 8.5. Suppose f = f(t,x,u,p) satisfies the following condition: for

any compact sets K, C R', K2 C R" there exist a nonnegative number LK2
and a nonnegative function = locally bounded on such that
(8.20) with L and M, respectively, holds for all
§8.3. UNIQUENESS OF GLOBAL CLASSICAL SOLUTIONS 93
(t,x,u,p),(t,x,v,q) E IZT x K1 x K2. I/u1 = uj(t,x) and u2 = u2(t,x) are global

C1 -solutions to the problem (8.1)-(8.2) with
sup .._L(t,x) <+00 (j = 1,2),

OX
then ui(i,x) u2(t,x) in 11T•

Proof. Let u = u(t, x) be as in the proof of Theorem 8.4, and let
r max sup <+co, K2 C L (8.21)

(t,r)EflT Ox
Xk x ... x (—k,k)c (k = 1,2,...). (8.22)

n timel
For an arbitrarily fixed T' (0, T), we consider the sequence of the
following parallelepipeds:
: 0<t<T', xEXb}.
Obviously, P1 C P2 C ... c ... and = Next, take
s' max max [_3k,3k] C R1. (8.23)

(t,z)EP
We now define a function p = p(x) by setting
def I
if xE X',
p(x) (8.24)
if xE Xk+1\Xk (fork = 1,2...).
=
It follows that p = p(x) is locally bounded on Moreover, (8.21)-(8.24) together
with the hypothesis of the theorem imply
= f(t, x, uj(t, x), — f(t, x, u2(t, x),
x) x) x) — u2(t, x)l
= L(1 + lxi) +p(x)lu(t,x)l in
(We may check this inequality first for (t,x) in P1, and then for (t,x) in each
Theorem 8.1 therefore shows that u(t,x) 0 in IZT.. Since T' (0,T)
is arbitrarily chosen, the conclusion follows. 0
Corollary 8.6. Let I = f(i,x,u,p) belong to C'(riT x x such that the

funchon
def
ii = &i(t,p) = sup + lxi)
(z,u)Ek" xk1 (/11
zsfinite and continuous on (0, TJ x I/ui = ui(t, x) and u2 = u2(t, x) are global
C'-solutzons to the Cauchy problem (8.1)-(8.2) with
°uj
sup —(t,x) < +oo (j = 1,2),
(t,z)EUT
then ui(t,x) ui(t,x) in 12T•

Proof. For any convex compact sets K1 C R', K2 C we see, by assumption,
that
max zi(t,p) < +oo,
and that the function

def Of
=PKt.K2(X)
= (t,u,p)E(O,TlxKixK3
is continuous, and hence locally bounded on RTh. It is easy to check that (8.20)
with LK3 and K2(x) in place of L and M, respectively, holds for any (t,x,u,p),
(t, x, v, q) E 11r x K1 x 1C2. The corollary thereby follows from Theorem 8.5. 0
We conclude this section with the following result of continuous dependence
on initial data for global C1-solutions. (Here the continuity is with respect to the
topology of uniform convergence on compact sets.)
Theorem 8.7. Suppose I = f(t,x,u,p) satisfies the condition (8.20) in Theorem

8.4. Let u1 = u,(t,x) (j = 1,2) be globalC1-solutions to the equation (8.1) with the
initial conditions
u,(0,x)=cb1(x) on {t=0,
where = = 1,2) are given functions of class C° on Then
lui(t,x) — u2(t,x)( exp(Mt) . sup —

IyI(1+IzI) exp(L*)—1
for all(t,r) E uT.

The proof of this theorem will be left to the reader.
§8.4. GENERALIZATIONS TO WEAKLY-COUPLED SYSTEMS
§8.4. Generalizations to the case of weakly-coupled systems
We now examine how the case of systems of first-order partial differential inequal-
ities or equations can be treated by the preceding method. Let m be a positive
integer. Consider the class
x x V(I1T).
,n time,
Each element of Vm(Ilr) is therefore a vector function, namely
u = u(t, x) = (ui(t, x),. . . , x))
from 11r C into Rm such that u5 = u'(t, x) belongs to V(flT) for every
jE{1,...,m}.
First, the following result may be proved in much the same way as Theorem 8.1.
Theorem 8.8. Let u = u(t,x) be a vector function in Vm(flr). If there a

nonnegative function p = p(x) locally bounded on and a nonnegative function
£ = 1(t) in L'(O,T) such that
max Iuk(t, x)I] (j = 1,... , m)

Iôu,(t, z)/OtI < 1(t). [(1+IzI)lôu,(i, z)/OzI+p(z)k=1
(8.25)
for almost every t E (0, T) and for all x then
max £(r)dr] sup max Iu,(O,y)I,

3=1 0
tohere C(x) is given by the formula (8.7).

Proof. For an arbitrary point (t°,x°) we must prove that
max
<exp [c(x0)JtOi(t)dt]. sup max

m
uj(0,y)I.
0 (8.26)
Let us continue using the notations Ei(t°,x°), Z(.,t°,x°), h(.), G1

introduced in the proof of Theorem 8.1 and then define
g(t) max gk(t) (8.27)

k=1
for t E I, where
glI(j) z E Z(t,t°,x°)} (k = 1,... ,m). (8.28)

{Iuk(t,x)I
It follows from Lemma 8.2 that for any number 6 E (0, t°) each function g" =
is absolutely continuous on [6, t°] and so is the function g = g(t). Moreover, they
are all continuous on the whole I. Still as in the proof of Theorem 8.1, we see
that (8.26) will be obtained if we can show that (8.13) holds. To this end, setting
def
w(t) = g,,(t) — g(t), with q > 0 temporarily fixed and
g,,(t) [g(0) + ,i] exp { [C(x°) + ti] . [h(t) + iii] },
we need only claim that c1i(t) w(0) (= rj > 0) for t E I. On the contrary, suppose
that there exists t' E (O,t°] with w(t') <w(O).
By the hypothesis of the theorem, one finds a set G2 C (0, T) of Lebesgue
measure 0 such that
fZT\(G2 x C (8.29)
and that (8.25) is satisfied for any t E (0, T)\G2, x E From the above, it follows
that (8.15) still holds where G G1 U G2; hence, there is a number A with
max{0,w(t')} <A <w(0) and A Ew[0,t']\w(Gfl[O,t°J).
Now take
(0,T)\G 3 {t E (0,t'] w(t) = A}
and 1 <j <in such that

def
g(t.) = g'(t.) = 1u1(t.,z.)I = e u1(t.,x,), e = signu,(t.,x.) (8.30)
for some E Z(t.,t°,x°). Next, choose a unit vector e with
= — (8.31)
Finally, let y = y(s) be an function continuously differentiable on

R1 such that y(h(t.)) = x. and dy/ds = (1 + $vI) e, and let x(t) y(h(t)) for
E [0, T]. Analysis similar to that in the proof of Theorem 8.1 shows that
u,(t, x(t)) Iu,(t, z(t))I < g(t) = g,1(t) — w(t) < g,,(t) — A
S8.4. GENERALIZATIONS TO WEAKLY-COUPLED SYSTEMS 97
for all t E [0,t.), and that
e u,(t.,x(t4) = Iu,(t.,z.)I = g(t.) = g,,(t.)

—w(t.) = g,7(t.) —
Consequently,
[e u,(t,x(t))}
This would give
> £(t.) [(1 + Ix. I) + C(x°)

k—lm uk(t., )I]
[(1 + Ix.I) + k—lm Iuk(t.,

a contradiction with (8.25). The proof is therefore complete.
Remark. Theorem 8.8 can be used to investigate the stability of global solutions
to the Cauchy problem for weakly-coupled systems, i.e., systems of first-order partial
differential equations of the form
thz,/Ot+f5(t,x,u,Ou,/ox) =0 (j = l,...,rn)
in aT. These systems are of special hyperbolic type because each equation contains
first-order derivatives of only one unknown function. Since (classical) solutions of
elliptic equations do not depend continuously (with respect to the topology of uni-
form convergence on compact sets) on initial data, theorems of the non-stationary
type that we have studied in this chapter cannot be expected to apply to partial
differential equations or inequalities of effiptic type. [First results on second-order
partial differential inequalities of parabolic and hyperbolic types were obtained by
Naguino and Simoda [108] and Westphal 1158].]
For our next discussions, we need to extend the notion of comparison equations
given in Szarski 11281 to the Carathéodory case. Consider an ordinary differential
equation
= p(t,w), (8.32)
where the function p = p(t, w) is defined on (0, +oc) x [0, +oo) = {(t, w)
2 > 0, u' 0}. The following Carathhodory conditions are always assumed.
(1) For almost every t E (0, +oo) the function [0, -I-cc) u' p(t, w) is continuous.
(2) For each w E [0, +oo) the function (0, +cc) 2 p(t, w) is measurable.
(3) For any r E (0, +oo) there exists a function m,. = mr(t) in +oo) with
m,.(t) Vw E I0,r]
for almost every t E (0,+co).
In this situation we call (8.32) a Carathéodory differential equatson on A solu-

tson of it on an interval I C (0, +m), with intl 0, means a function w = w(t) 0
absolutely continuous on each compact interval J C I (absolutely continuous on I
for short) such that
w'(t) = p(t,w(t))
almost everywhere in I. We refer to Coddington and Levinson [311 for what concerns
the local existence of a solution of (8.32) through any given point (t°,w°) E
Moreover, every such solution can be extended (as a solution) over a [left, right]
maximal interval of existence.
Definition 2. A Carathéodory differential equation (8.32), with p(t, w) 0 on

and p(t, 0) 0 for almost all t > 0, will be called a comparison equation if
w = w(t) 0 is in every interval the only solution satisfying the condition
limw(t) = 0.
t—+o
Remark. Let £ = £(t) be a nonnegative function Lebesgue integrable on each

bounded interval (0, y) C R, and = a function of class C[0, +m) such that
r(0) = 0, > 0 as w > 0, and f(1/c(w))dw = +co for every 5 > 0. Then (cf.
[128, Example 14.2])
w' = £(t)u(w) (8.33)
is a comparison equation. In fact, assume the contrary that (8.33) admits a nonzero
solution w = w(t) on some interval (0,7) with limw(t) = 0. Letting w(0) 0,
from this we easily find a nonempty subinterval (t1 , t2) of (0, such that w(t') = 0
and w(t) > 0 for all t E (t1,t2]. It follows that
dv
I
a(v)
=I o(w(t))
dt = I £(t)dt < +m,
Jo
a contradiction. Therefore, (8.33) must be a comparison equation. Motivated by

this fact, we propose the following:
§8.4. GENERALIZATIONS TO WEAKLY-COUPLED SYSTEMS 99
Proposition 8.9. Let c = c(w) be of cla8s C[O, +oo), and £ = 1(t) 0 be Lebesgue
+00
integrable on each bounded interval (O,i') C R with f0 £(t)dt = +00.
(i) If (8.33) is a comparison equation, then so is the equation
= a(w). (8.34)
(ii) Conversely, under the condition essinf 1(t) > 0, if moreover (8.34) is a corn-
IE (0.+oo)
parison equation, then so is (8.33).
Proof. (i) Let w1 = w'(t) be a solution of (8.34) on some interval (O,.y1) with
urn w1 (t) = 0. Find a number .72 > 0 such that
1-30
72
= JoI £(r)dr. (8.35)
Setting w2(t) w2(J'e(r)dr), see that w2 = w2(t) is a solution of (8.33) on

we
(O,-y2) with lirnw2(t) = 0. By assumption, w2(t) 0 on (O,.y2). Hence w1(t) 0
on (O,-y1). This shows that (8.34) is a comparison equation.
(ii) Let (0, +m) t be the inverse of (0, +oo) 3 t j £(r)dr, and w2 =
w2(t) be a solution of (8.33) on some interval with limw2(t) = 0. First,
define a number > 0 by (8.35). Then setting w'(t) w2(1(t)), we also see that
w1 = w'(t) is a solution of (8.34) on (O,-y1) with Iimw'(t) = 0 (cf. [40, Proposition
3.4(c)J). The rest of the proof runs as before. 0
In the sequel, for each function g = g(t) defined and continuous in a certain
interval (O,t°), let P9 denote the open set (t (O,t°) g(t) > O}. Here is an
:
elementary property of comparison equations:
Proposition 8.10. Let (8.32) be a comparison equation and g = g(t) a given

function absolutely continuous on some interval (O,t°) such that lLrng(t) < 0 and
that g'(t) p(t,g(t)) almost everywhere in P,. Then g(t) < 0 for all t (O,t°).
Proof. On the contrary, suppose that there exists t1 (0, t°) with & g(t1) > 0.
Setting g(O) limg(t) and t2 sup{t [0, t1) g(t) = 0}, we see that 0 t2
:
t1, g(t2) = 0 and (t2,t1) C P9. Hence, by assumption,
g'(t) <p(t,g(t)) almost everywhere in (t2,t1). (8.36)

Now take
I p(t,max{0,g(t)}) if t2 < t < t°, w max{0,g(t)},
p(t,w) def
= (8.37)
if t2 <t <to, 0 w <max{0,g(t)}.
The above-mentioned Carathéodory conditions (1)-(3) are clearly satisfied for /3 =

on (t2,t°) x [O,+oo). Let w = w(t) be a solution through (t1,w1) of (8.32)
with /3 in place of p, and let (t3, t'] c (t2, t'] be its left maximal interval of existence.
We next claim that
(0 ) w(t) g(t) Vt E (t3,t1J. (8.38)
Assume (8.38) is false. Then one would find a nonempty interval (t4, t5) C (t3, t1)
such that
w(t) > g(t) Vt E (t4,t5), (8.39)
with
w(t5) = g(t5). (8.40)
It follows from (8.36)-(8.37) and (8.39) that g'(t) < p(t,g(t)) = j5(t,w(t)) = w'(t)
almost everywhere in (t4,t5). Thus (8.40) implies that g(t) w(t) for all t E (t4,t5),
which contradicts (8.39). So (8.38) must hold.
We proceed to show that t3 = i2. In fact, if (0 <)t2 < t3, then (8.37) together
with Carathéodory's condition (3), where r max{g(t) t E It3, t']}, proves that
:
the limit lim w(t) exists and is finite. Hence, w = w(t) could be extended (as a
solution of (8.32) with /3 in place of p) over an interval (t6,t'J J [t5,t'], which is
impossible.
Finally, (8.37)-(8.38) shows that w = w(t) is indeed a solution through (t1,w1)
of (8.32) on (t2,t1] with lim w(t) = g(t2) = 0. Setting w(t) fort E [0,i2}, we
obtain a nonzero solution of (8.32) on (0,t') which tends to 0 as t goes to 0. Thus
we arrive at a contradiction. This completes the proof. 0
We can now combine the method of §8.2 with the technique of Carathéodory
comparison equations and prove the following.
Theorem 8.11. Let u = u(t,z) be a vector function :n Vm(1Zr) vnth u'(O,x) 0

(j = ., m), and (8.32) a comparison equation. If there ezists a nonnegative
.
function £ = £(t) in L1(0,T) suth that
Ou,(t, x)/OtI £(t)(1 + IxI) . Ou,(t, x)/OzI +p(t, Iuk(t, x)I) (j = 1,..., m)
(8.41)
§8.4. GENERALIZATIONS TO WEAKLY-COUPLED SYSTEMS 101
for almost every i (0,T) and for all x E then u,(1,x) 0 in =

1 rn).
Proof. For an arbitrary point (t°, x°) E 11T, it suffices to prove that
max u,(t°,z°)l = 0. (8.42)
We shall continue using the notations I [0,t°J, E1(i°,r°), Z(.,t°,x°), h(.),

G1 introduced in the proof of Theorem 8.1 (and, also, of Theorem 8.8) and letting
g = g(t), g" = gk(t) be as in (8.27)-(8.28). Obviously, (8.42) will be obtained if
one can verify that g(t°) = 0. Since g = g(t) is a nonnegative function absolutely
continuous on (0,t°1, with limg(t) = g(0) = 0 (by assumption), Proposition 8.10
shows that we need only claim that
g'(t) p(t,g(t)) almost everywhere in (0,t°). (8.43)
By the hypothesis of the theorem, one finds a set G2 C (0, T) of Lebesgue

measure 0 such that (8.29) and (8.41) are fulfilled for any t E (0,T)\G2, x E R".
Assume without loss of generality that g = g(t) is differentiable at any point of
(0,t°)\G, where C C1 U C2. Now fix an arbitrary point t. E (0,t°)\G and
take 1 <j < m such that (8.30) holds for some E Z(t,,i°,x°). Next, choose
a unit vector e E R" satisfying (8.31). Let y = y(s) be an RTh-valued function
continuously differentiable on R1 such that y(h(t.)) = x. and dy/ds = (1 +
and let x(t) y(h(t)) for t E [0,1']. Of course (see the proof of Theorem 8.1),
x = x(t) is absolutely continuous on [0,1'], x(t.) = x., and
= £(t) (1 + Ix(t)l) .e Vt E (0,T)\G. (8.44)
Moreover,
x(t) Z(t,t°,x°) Vt E [0,t.].
This together with (8.27)-(8.28) implies
e u'(t, x(t)) Iuj(t, x(i))I < g1(t) < g(i) for all t (0, t.). (8.45)
Besides that, by (8.30),
e = Iu,(t,,x.)I = g'(t.) = g(t.). (8.46)

102 8. DIFFERENTIAL INEQUALITIES OF RAAR TYPE
Therefore, since t. E (O,t°)\G, it may be deduced from (8.45)-(8.46) that
g'(t.) < u,(t,x(t))] .

[e
Consequently, by (8.30)-(8.31), (8.41), and (8.44), we conclude that
g'(t.) (Ou,(t.,x,)/Ot) +
— £(t,)(1 + ix.l) lOu,(t.,x.)/OxI
Iuk(ts,x.)I) = p(t., u,(t.,x.)I) = p(t,,g(t.)).
Finally, because C has measure 0 and t. E (O,t°)\C is arbitrarily chosen, (8.43)

must hold. This completes the proof. 0
Theorem 8.12. Let u = u(t,x) be a vector function in with u(O,x) 0
(j = 1,... ,m), and (8.34) be a comparison equation. If there ezut a nonnegative
function = p(x) locally bounded on RTh and a nonnegative function £ = 1(i) in
L' (0, T) suth that
iOu,(t,x)/Oti 1(t) [(1 + lxi) . iOu,(t,x)/ôxl + (8.47)
(j = 1,...,m) for almost every t E (O,T) and for alix E then u,(t,x) 0 in
QT (j =
Proof. For an arbitrary point (t°, x°) E it suffices to prove (8.42). Let us
continue using the method (and notations) introduced in the proof of Theorem
8.11. We may extend the function £ = 1(t) over the whole (0, +oo) and assume
essinf 1(t) > 0. Then by (8.47) [instead of (8.41)1 we get
g'(t) < almost everywhere in (0,i°)
[instead of (8.43)1 for some positive constant C. By Proposition 8.9(ü), the Carathé-
odory differential equation
WI = C1(t)c(w)
is also a comparison equation. Thus (8.42) is straightforward as before.

Chapter 9
Hopf's Formulas
for GloL?al
of Hamilton-Jacobi Equations
§9.1. Introduction
The aim of this chapter is to present some formulas for explicit global solutions of
the Cauchy problem for Hamilton-Jacobi equations of the form
ôu/Ot + f(t,Ou/Ox) = 0 in {t > 0, x (9.1)
u(0,x) = on {i = 0, x (9.2)
According to Theorem 1.4, the Cauchy problem (9.l)-(9.2) has locally a unique
C2-solution if the Hamiltonian f = f(t,p) and initial function = are of class
C2. However, there is generally no hope to find a global classical solution; there-
fore, one needs to introduce a notion of generalized solutions and to develop theory
and methods for constructing these solutions. (For this, cf. Chapters 4-7.) During
the past five decades, many mathematicians have obtained various global results
by relaxing the smoothness conditions on the solutions. In particular, the glob-
al existence and uniqueness of (generalized) solutions for convex Hamilton-Jacobi
equations were well-studied by several approaches (see §4.3 for the uniqueness of
such a solution).
if the Hanultonian I = 1(p) is continuous and if the initial function =
is globally Lipschitz continuous and convex with the Fenchel conjugate =
E. Hopf [64] proved in 1965 that the formula
u(t, x) = max{(p, x) — — tf(p)} (9.3)

pER"
determines a (generalized) solution of the Cauchy problem (9.1)-(9.2) in the sense

that this solution satisfies (9.1) at every point where it is differentiable. Since the
solution is locally Lipschitz continuous, the well-known Rademacher theorem [116,
Theorem 1.18] shows that (9.1) is then satisfied almost everywhere.
104 9. HOPF'S FORMULAS FOR GLOBAL SOLUTIONS
If the Haniiltonian f = f(p) is strictly convex with lim f(p)/)p) = +oo and
if the initial function = çb(x) is globally Lipschitz continuous, E. Hopf [64] also
established the following formula for a (generalized) solution of (9.1)-(9.2):
u(t,x) = f((x — y)/t)}. (9.4)
The above formulas are often associated with the name of Hopf, although (9.4)
was actually first discovered for n = 1 by P.D. Lax [971 in 1957.
Step by step, certain more general cases of Hopf's formula (9.3) will thoroughly
be dealt with in this chapter under a standing hypothesis like (but somewhat more
strict than) Carathéodory's condition on the Hamiltonian I = f(t,p). Section
§9.2 concerns the case of convex (but not necessarily globally Lipschitz continuous)
initial data. In Section §9.3 we consider the Cauchy problem with nonconvex initial
data: first for the case where qS = can be represented as the minimum of a
family of convex functions, and second for the case where = 4(x) is a d.c. function
(i.e., it can be represented as the difference of two convex functions). Finally,
Section §9.4 discusses Hopf's formula (9.4) in case = is just continuous.
Most of the results presented here were originally published in [1421-[145] and
[1521. (For other results in the field, see for example [l1]-[12], [21], [33], [52], [100),
and [124).) Our method is based on some techniques of multifunctions and convex
functions. The relevant material on these subjects from [8] and [117] (without
proofs) can be found in Appendix II given at the end of the book. This makes our
exposition self-contained.
We shall continue using the notation of Chapter 8. Moreover, in accordance with
Chapter 2, when T = +00, we use both V and to denote the set {0 < t <
+00, x Further, for any G C R, put Va ((0,+oo)\G) x = {(t,x)
V t
: G}. Consequently, V = V0. Let us recall that Lip([0, T) x
fl CoO, T) x R"); accordingly, = Lip(D) fl C(D), where Lip(V) is
the set of all locally Lipschitz continuous functions u = u(t, x) defined on V =
Definition. A function u = u(t, x) in Lip(10, T) x is called a global solution of

the Cauchy problem (8.1)-(8.2) if it satisfies (8.1) almost everywhere in QT and if
u(0,x) = çb(x) for all zE R't.
§9.2. THE CAUCHY PROBLEM WITH CONVEX INITIAL DATA 105
§9.2. The Cauchy problem with convex initial data
In this section we consider the Cauchy problem (9.1)-(9.2), with = ti(x) a finite
convex function on Denote by = the Fenchel conjugate function of
for pER",
and by E the effective domain of =
We assume the following two hypotheses:
(E.I) The Hamiltonian I = f(t,p) is continuous in {(t,p) : t E (O,+oo)\G, p E

R"} for some closed set G C R of Lebesgue measure 0. Moreover, to each N E
(0, +oo) there corresponds a function 9N = gN(t) in such that
sup f(t,p)I gN(t) for almost all t E (O,+oo).

pIN
(E.II) For every bounded subset V of there exists a positive number N(V) so
that
(p,x) —
— j I (r,p)dr <
—
— f f(r,q)dr} (9.5)
whenever (t,z) V, > N(V).
Hypothesis (E.I) implies the i-measurability and p-continuity off = f(t,p) on

{t > 0, p E R"}. Moreover, since = is finite on R", this hypothesis allows
us to define an upper senucontinuous function = ço(t,x,p) from x R" into
[—oo,+oo) by
(p,x) — — ff(r,P)dr, (9.6)
which is, for each p E E, actually finite and continuous in (I, x) on

The next theorem will be fundamental in this section.
Theorem 9.1. Let = be a finite convex function on Assume (E.I)-

(E.II). Then a global solution u = u(t,x) of the Cauchy problem (9.1)-(9.2) is given
by
sup {(p,x) f(r,p)dr} for (t,x) ED.

pER's pER" 0
(9.7)
Remark 1. Requirement (E.II) could in a sense be regarded as a compatible con-
dition between Hamiltonian and initial data for the existence of global solutions of
the Cauchy problem (9.1)-(9.2). To see this, we first rewrite (Eli) in an alterna-
tive form that is essentially equivalent (by a standard compactness argument) but
seemingly more amenable to verification:
(E.II)' For every (t°,x°) there exist positive numbers r(t°,x°) and N(t°,x°)
so that
x,p) < max (t, x) V(t°, x°), > N(t°, x°)

IqIN(t°,z°)
(9.5')
where V(t°,x°) {(t,x) E : t — 201 + Ix — x°I < r(t°,x°)}.
We proceed now to consider, for example, the Cauchy problem
in {t>O, xER},
u(O,x)=x2/2 on {t=O, xER}.

Then the method of characteristics in Chapter 1 gives the unique classical solution
= u(t, s) (9.8)
— 2(t— 1)
in {O <t < 1, x R}. This solution can not be extended continuously beyond
the time t = 1. We should note that in this case the set of all (j0, x°) E V such
that (9.5') holds for some r(t°,x°) > 0 and N(t°,x°) > 0 is precisely the domain
(0 2 < 1, x E R}. Moreover, if we try to apply Hopf's formula (9.3), ignoring the
fact that the initial function here, = x2/2, is not globally Lipschitz continuous,
then we also obtain the same solution as (9.8) in (0 2 < 1, x E R}.
§9.2. THE CAUCHY PROBLEM WITH CONVEX INITIAL DATA
Remark 2. Let = qS(x) be a finite convex function on Assume (E.I). Then

(E.II) is satisfied if ço(t,x,p) tends to (—oo) locally uniformly in (t,x) E V as
tends to (+oo), i.e., if the following holds:
(E.II)" For any E R and any bounded subset V of there exists a positive
number N(A, V) so that (t, x) V, > N(A, V).
Indeed, fix an arbitrary q° E 0 [117, Thm. 12.2]. Since the finite
function V (t,.r) '-+ p(t,x,q°) is continuous, it follows that
deC o
= rnf
.
)> —00
for any bounded subset V of V. Under Hypothesis (E.II)", we certainly find a

number Nv (for each such V) so that
p(t,x,p) <Ay
(t, x) V and > Nv, thus getting the validity of (E.fl).
Remark 3. Condition "gj.j = gN(t) in the hypothesis (E.I) could not

be replaced by "9N = gN(i) E To see this, consider the following Cauchy
problem
au/Ox
Ou/Ot+ 1]I'2
=0 in {t>0, XER},
It
u(0,x)=x on {t=0, xER}.

def
Here, n = 1,
f(t,p) def
7) deC
and = x
= ]t — 1p1/2
forxER, p€Randt>0. Then

• 1+00 if
if p=l;
hence (E.II) holds when N(V) 1. All the assumptions of Theorem 9.1 with
in place of for Hypothesis (E.I) are therefore satisfied here. However, in this
case, (9.7) gives the function
u=u(t,x) def
= x—2—21t—1l 1/2 sign(t-. 1),
which is not Lipschitz continuous in any neighbourhood of a point (1, x°).

For the proof of Theorem 9.1, we need some preparations. We first recall that a
directional derivative is defined as follows. Let = be a finite-valued function
of near a point E Rm and let e E Rm. Denote
inf sup + —
e>0 O<ö(t
inf + — }.
e>O O<6<e
Ii—.I<.
The quantities and are called, respectively, the upper and lower
Dliii semiderivatives of (' = at the point in the direction e. if +oo >
= > —oo for all e E Rm, then (' = is said to be directionally
differentiable at and will be called its derivative
at in the direction e. It can be shown that, when = is Lipschitz continuous
near its upper and lower Dim semiderivatives may also be defined equivalently
by
—
5,1.0
and
lii + 5e) — },
respectively.
Lemma 9.2. Let 0 be on open subset of and w = w(e,p) an upper semscon-

tsnuous function from 0 x to [—oo, +oo) with the following two properties:
(i) There exists a nonempty set E C RTh such that = w(e,p) is finste on 0 x E
and that —oo where EC Moreover, to each bounded
subset V of 0 there corresponds a positive number N(V) so that
< max whenever V, IpI > N(V);
(ii) For every fixed p E, the function w = is differentiable in 0.

Besides that, the gradient Ow/Of = is continuous on 0 x E.
Then one has:
a) sI' = pE is a locally Lipschitz continuous function in
the domain 0;
b) = is directionally differentiable in 0 with
= max 0, e E R)
pEL(t)
where
{p E : = C E. (9.9)
We shall also need the following.
Lemma 9.3. Let Condition (i) in Lemma 9.2 hold for a given function w =
which we assume to be continuous in 0 (whenever p E E) and upper
semicont,nuotzs with respect to the whole on 0 x R". Then (9.9) determines
a nonempty-valued, closed and locally bounded multifunction L = of E 0.
Remark. Lemma 9.3 implies that L = is a compact-valued and upper semi-

continuous multifunction. In this lemma, the set 0 C R is not necessarily open.
Proof of Lemma 9.3. For any bounded subset V of 0, denote Nv N(V)
(see (i)). Then C as E V. This means that L = is locally
bounded. Moreover, = being upper semicontinuous, we can deduce (also
from (i)) that the supremum = pE should always be finite
:
and attained. Consequently, L = C E is a nonempty-valued multifunction of

€0.
To complete the proof, one needs only check the closedness of L = For this
purpose, let ph) be a sequence convergent to a point p°) in Ox R" such
that p" ask = 1,2,3 By the definitions of 1' = and L =
we have
w = is upper semicontinuous in and continuous in (9.10) shows

that
lim for all p.
Thus p° L= is therefore closed.
Proof of Lemma 9.2. a) Let V be an arbitrary compact convex subset of 0

and let N N(V). For any two points V, pick up an element p' in the
nonempty set CEn N) (cf. Lemma 9.3). Then
—
9. HOPF'S FORMULAS FOR GLOBAL SOLUTIONS
hence the mean-value theorem gives
— — A. —
where E C V and A is a (finite) upper bound of over

—
Vx Analogously, A — The function
= is thus locally Lipschitz continuous in 0.
b) For any E 0 and e E Rm, we find two sequences of positive
numbers convergent to zero such that
+ ake) —
= lim
and
lim
Let us take an arbitrary p C E and apply the mean-value theorem to

obtain
+ a'e) — + —
= + e)
where A passage to the limit as k +00 shows that
for any p E Hence
sup
Now choose an element E for each k = 1,2 Since the multifunc-
tion L = is closed and locally bounded (Lemma 9.3), by taking a subsequence
if necessary, we can assume that p° Therefore, a passage to the
limit (similar to the above) in the inequality
+ like) — <w(6° + I3he,p") — +

=
13:, 19k
(where (0,/3:,)) gives
< sup (9.12)

Finally, combining (9.11)-(9.12) yields
= = max
pEL(C )
for any E 0, e E Rm. This implies the directional differentiability of =

and completes the proof. 0
We are now in a position to prove Theorem 9.1.
Proof of Theorem 9.1. It can be verified that the function w =
x, p) (see (9.6)) satisfies all the assumptions of Lemma 9.3 where E dom
def def def—
0[117, Theorem 12.2] andm = n+1, = (t,x). Here, weput0 = Dand conclude
that the definition
L(i,x) d.f
= {p E E : = u(t,z)}
determines a nonempty-valued, locally bounded (and closed) multifunction L =

L(t, x) of (t, x) E V.
Our proof starts with the claim that u = u(i, x) is in To this end,
take arbitrarily an r E (O,+oo) and denote V,. {(t,z) i + lxi < r},
N,. N(V,.) (cf. (E.II)). Let gp,r, = be as in Hypothesis (E.I). Then for any
two points (t1, x'), (t2, x2) V,. we may choose an element p' of the nonempty set
L(t1,x1) C and get
u(t1,x1) — u(t2,z2) — = (p',x' — x2) +f f(r,p1)dr

N,... 1x1 — x21 + s,. —
where s,. ess sup (t). Analogously,

tE(O,r)
u(t2, x2) — tz(i1,x') < N,.. 1x1 — x2i + s,. t1 —
Thus, u = u(i,x) belongs to

Next, let e0 (1,0,0,...,0), e1 (0,1,0,. ..,0),..., (0,0,... ,0,1) E
+1 def
R't . It follows from Lemma 9.2 (we now replace 0Do, = the set G being as in
(E.I)) that u = u(t,x) is directionally differentiable in V0 with
= max{—f(t,p) p L(t,x)},
(9.13)
= max{f(t,p) p L(t,x)}
9. HOPF'S FORMULAS FOR GLOBAL SOLUTIONS
and (for 1 <i <n)

= max{p1 : p L(t,x)},
(9.14)
O_e•U(t,X) = max{—p; : pE L(t,x)}.
On the other hand, according to Rademacher's theorem [116, Theorem 1.18], there
existsaset Q C Dof((n+1)-dimensional) LebesguemeasureOsuchthat u = u(t,x)
is (totally) differentiable with
Ou(t, x)/Ot ôeou(t, x) x),

(9.15)
Ou(t,x)/e9x1 = (3eeU(t,X) =
at any point (t,x) V\Q. Hence, (9.14)-(9.15) show that the equalities (for 1
ôu(t,z)/0z1 = max{p1 p E L(t,x)} = min{p, : p E L(t,x)}
hold outside the null set 2def

= (G x R ) Q; i.e., L(t, x) is precisely the singleton
U
.
{Ou(t,x)/Ox} except on P. Consequently, (9.13) together with (9.15) implies that

(9.1) must be satisfied almost everywhere in V.
Further, by a well-known property of Fenchel conjugate functions [117, Theorem
12.2), (9.7) gives
u(0,z) = max{(p,x) = = for all XE RTh.

pER"
From what has already been proved, we conclude that u u(t, x) is a global
solution of (9.1)-(9.2). 0
Remark. By the proof of Theorem 9.1, all partial derivatives of the global solution
u = u(t, x) exist [(9.1) is then satisfied though the solution may fail to be differ-
entiable] at a given point (t°, x°) E D0 if L(t°, x°) is just a singleton. We shall
investigate the smoothness of u = u(i, x) in greater detail in
Corollary 9.4. Let </ = be a finite convex function on Under Hypothesis

(E.I), suppose that inf f(t,p)/(1 + is locally essentially bounded from below
pER"
in t (0, +oo). Then (9.7) determines a global solution of the Cauchy problem
(9.1 )-(9.2).
§9.3. THE CASE OP NONCONVEX INITIAL DATA 113
del
Proof. Only (E.II) needs verifying. Given any r (O,+oo), denote V,. = {(t,x) E
+ lxi <r} and s,. inf f(t,p)/(1 + p1). Then
f(r,p)dr
(9.16)
— 4*(p) — ts,.(1 + lVI) < r(l + 21s,.l) max{1, —
for all (t,x) E V,., p On the other hand, it being known (cf. Remark 2 of
Lemma 9.13 later) that urn = +00, there exists a number N,. 1 such
that r(1 + 2ls,.I) + 1 whenever N,.. Thus, (9.16) implies
provided (t,x)EV,., lpINr.

Finally, since r (0, +co) is arbitrarily chosen, Condition (E.II)" holds; hence so
does (E.II). 0
Corollary 9.5. Let f = f(t,p) be continuous on V and let = be convex and
globally Lipschitz continuous on Then (9.7) determines a global solution of the
Cauchy problem (9.1)-(9.2).
Proof. Since = is convex and globally Lipschitz continuous, E

should be bounded [117, §13.3] (and nonempty). Independently of (t,x) D, it
follows that
= (p,x) — — I f(r,p)dr = —oo
Jo
whenever is large enough. Hypothesis (E.II) therefore holds while (E.I) is triv-
ially satisfied. 0
Remark. As we have mentioned in the introduction, Corollary 9.5 was proved by
E. Hopf [64] in a different way for the case where I = f(p) depends only on p.
§9.3. The case of nonconvex initial data
In this section we consider the Cauchy problem (9.1)-(9.2) under the more general
assumptions that I = f(t, p) is still t-measurable and p-continuous as in §9.2, while
= 4(x) is now a d.c. function, i.e., the difference of two finite convex functions
on R". The class of d.c. functions plays an important role in the theory
of global optimization. This class is rather large: It contains all the semiconvex
9. HOPP'S FORMULAS FOR GLOBAL SOLUTIONS
or semiconcave functions. We emphasize that it also contains all functions of class

C2 (of the whole with second derivatives bounded either from below or from
above. The reader is referred to [6] and [661 for a sufficiently complete study of d.c.
functions.
We first prove the following result.
Theorem 9.6. Let I be an arbitrary nonempty set and let Ua = ua(t, x) be a global
solution of the Cauchy problem for the same Hamilton-Jacobi equation
u/Ot+f(t,3u/Oz)=O in {t>O, x€R'}, (9.1)
with the initial condition
u(O,x) = on {t = 0, x E (9.2a)
for each a E I. Suppose that to each bounded subset V of 1) there correspond a set
W(V) C V of Lebesgue measure 0, a nonnegative number M(V), and a subset J(V)
of I such that all the functions Ua = ua(t, x) for a E J(V) are Lipschitz continuous
in V unth a common Lzpschitz constant M(V) and satisfy (9.1) at every point of
V\W(V) and that infua(t,z) = mm for (t,x) E V. Then the function
aEI aEJ(V)
u = u(t,x) infua(t,x) is a global solution of the Cauchy problem (9.1)-(9.2)
aEJ
where
aEI
Proof. By assumption, u(t, x) = mm ua(t, r) on each bounded subset V of
aEJ( V)
Moreover, Iua(t',Z1) — Ua(t2,X2)I < M(V) .(It' — + Ix' — x2[) for any a E J(V)
and any fixed (t',x'), (t2,x2) V. Assume u(t',x1) 2 u(t2,z2) = u0o(t2,x2) for
some a° E J(V). Then
0 < u(t1,x') — u(t2,x2) < — Uao(t2,Z2) < M(V) . ([ti — + —
This means that u = u(t,x) is in

Now denote V,, {(t,x) E V : t + xl < k}, J(Vk), W,, W(Vk)
for each k = 1,2,... and let Wo C D be a set of Lebesgue measure 0 such that
u = u(t, x) is differentiable at every point of D\W0 (Rademacher's theorem). It
will be shown that u = u(t,x) satisfies (9.1) except on the null set Q
§9.3. THE CASE OF NONCONVEX INITIAL DATA 115
Indeed, given any (t°, x°) V\Q, we choose a positive integer k > t° + x°l and
some index a0 E Jk so that u(t°,r°) = Obviously,
u(t,x) — u(t°,x°) uao(t,x) — u0o(t°,x°)
for all (t,x) close enough to (t°,x°). Since u = u(t,x) and UaO = Uao(t,X) are both
differentiable at (t°, x°), (9.17) implies
Ou(t°,x°)/Ot s9uao(t°,x°)/Ot and Ou(t°,x°)/Ox = OUao(t°,s°)/OX.
But UaO = Uao(t,X) satisfies (9.1) at (t°,x°); so does u = u(t,z).

On the other hand, it is clear from the hypotheses that
u(O,x) = in! = inf4,0(x) = 4,(x) for all x R".

aEJ aEI
The function u = u(t,x) is thus a global solution of(9.1)-(9.2).
Now suppose that 4, = 4,(x) is given in the form
4,(r) in! 4,a(X) for z R", (9.18)

aEI
with = a finite convex function for every a E I. Combining Theorems
9.1 and 9.6, we obtain the following first results for the representation of global
solutions in the case of nonconvex initial data.
Corollary 9.7. Assume (E.I)-(E.II) for each problem =

a finite convex function, a I. Assume, furthermore, that all the hypotheses of
Theorem 9.6 hold for the solutions
def
= uQ(t,x) = max{(p,x) — 4,a(p)
pER — , f(r,p)dr}
of those problems. Then u = u(t, x) in! ua(t, r) is a global solution of the Cauchy
aEI
problem (9.1)-(9.2) where 4, = 4,(x) is defined by (9.18).
Corollary 9.8. Let 4, = 4,(x) with 4'j = 4,j(x),..., 4,k =

some convex and globally Lipschitz continuous functions. 1ff = f(t,p) is continu-
ous on then a global solution u = u(t,x) of the Cauchy problem (9.1)-(9.2) can
be found in the form
u(t,x) 1iE k}
{(p,z) — g5(p) — jf(r,p)dr} for (t,x) €
Proof. Since I { 1,.. . , k} is a finite set, the conclusion is straightforward from

Corollary 9.5 and Theorem 9.6. 0
Example 1. Consider the Cauchy problem
Ou/Ot+l(Ou/Ox)2—1l=O in {O<t<+oc, xER},

u(O,x) = exp(—IxI) = min{exp(x),exp(—x)} on {t = 0, x E R}.
By Corollary 9.8, a global solution of the problem is
def
u = u(t,x) = mm max{px — h,(p)
.
— tip 2 — 1I}
s1,2 pER
where
Ipinp—p if p>O,
def I
h1(p)
= 0 if p=O,
L+oo if p<O,
(+oo if p>O,
h2(p) = 0 if p=O,
—p ln( —p) + p if p < 0.
The solution can also be rewritten as
u(t,x) = min{max{pz — plnp +p —

po
— + pln(—p) — p — tip2 — lI}}
= max{—pIxI—plnp+ p—tIp2 —
Po
in which we adopt the convention that plnp = 0 if p = 0.
Example 2. Let h = h(o) be a finite-valued continuous function of o on a given

compact set K C R't. We put
min{h(ci) + Icil . ixi} for x R".

aEK
If f = f(t,p) belongs to C(V), depends only on t and and is decreasing with

respect to ipi, then it follows from Corollary 9.7 that a global solution u = u(t, x)
of the Cauchy problem (9.1)-(9.2) can be found by the formula
u(t, x) min{h(a) + max {(p, z) -

aEK IpIIaI fo
f(r,p)dr}}
= min{h(a) + Icii ri j f(r,s)dr} for (t,x) €

—
§9.3. THE CASE OP NONCON VEX INITIAL DATA
We now consider the Cauchy problem (9.1)-(9.2) in the main case of this section
where 4, = 4i(x) belongs to the class DC(RTh); i.e., it has a representation of the
form
4,(x)woi(x)—o2(x) on (9.19)
for some finite convex functions = oi(x) and 02 = 02(X). We call (9.19) a d.c.
representation of 4, = 4,(x). (Of course, there are always an infinite number of
such representations for each d.c. function 4, = cb(x).) The notations =
o(p) will signify the Fenchel conjugate functions of = 01(x) and
02 = a3(z). The effective domains of these conjugates are denoted by E1 and B2,
respectively.
Besides Hypothesis (E.I), we shall also assume the following ones.
(E.III) To any bounded sets V C and E C there corresponds a positive number

N(V,E) so that
(Pa(t,X,P) < max a E E, > N(V,E).

IqIN(V,E)
Here,
'pa(t,X,p) (p,x) —
j f(r,p)dr. (9.20)
—
(E.IV) All the multifunctions = La(t,X) of(t,x) E defined by
La(t,Z) {p E : = (9.21)
are actually single-valued in D\Q where Q is a certain closed set of ((n + 1)-
dimensional) Lebesgue measure 0 and is independent of a E
Remark. Let = o1(x) be a finite convex function on and I = f(t,p) be a

p-convex function on {t > 0, p E Assume (E.I) and (E.III). Then it can be
proved that (E.IV) is satisfied with Q x the set G being as in (E.I), if one
of the following two conditions holds:
(E.IV)' f = f(t,p) is strictly p-convex; more precisely, that is to say, it is strictly

convex with respect to p on for almost every fixed t E (0, +oo).
(E.W)" o = (p) is strictly convex on its effective domain E1 dom o.

9. HOFF'S FORMULAS FOR GLOBAL SOLUTIONS
For strictly convex functions on R", see the proof of Lemma 9.13. The detailed
notion of the strict convexity of a function on a convex set may be found in Appendix
Theorem 9.9. Let q4 = be in the class with a d.c. representa-

tion (9.19) such that 02 = a2(x) is globally Lipschitz continuous on Under
Hypotheses (E.l), (E.Ill), and (E.IV), the formula
u(t,x) def
= . *
on —
D, (9.22)
aCE2 pER
in which E2 dom o, determines a global solutzon u = u(t, x) of the Cauchy

problem (9.1)-(9.2).
Proof. Let a Then =

are obviously convex functions. For each a E2, consider the Cauchy problem
By (E.I) and (E.III), it follows from Theorem 9.1 that the formula
def • —
Ua(t,x) =
PER"
on D (9.23)
determines a global solution Ua = x) of this problem. Moreover, we may

assume that Q D G x the sets Q and G being as in (EIV) and (E.I), and then
see that all the solutions Ua = Ua(t,r) satisfy (9.1) at every point of D\Q. (For
the smoothness of such u8 = x), see Remark after Proof of Theorem 9.1.)
Now, since 02 = r2(x) is globally Lipschitz continuous on R", the (nonempty)
set E2 = domo' should be bounded [117, §13.3]. Given any r (O,+oo), denote
v,. {(t,x) t + Izi < r} and N,. N(V,.,E2) (cf. (E.III)). For any
(t',x'), (t2,x2) in Vr, we can then choose p6 E L6(t1,x1) C and deduce
from (9.20)-(9.21) and (9.23) that
u0(t',x') — u6(t2,x2) —
_x2)+f f(r,p6)dr
— + —
where a,. = (cf. (E.I)). The solutions Ua = u6(t,z) therefore satisfy

CE(O,r)
a Lipschitz condition on V,. with constants N,. and a,., which are independent of
a E2.
§9.3. THE CASE OF NONCONVEX INITIAL DATA
Next, rewrite (9.23) as
= o(a) + — a) (9.24)
pER'
and fix temporarily (t,x) E By (9.20) and Hypothesis (E.I),

a E Hence (see 1117]), the right side of (9.24), being the
supremum of a family of continuous functions, actually determines a lower semi-
continuous function of a from the whole R" into (—oo, +oo] whose effective domain
is precisely the nonempty bounded set E2 C It follows that
+00> inf inf

aEE2 aEB2 pER's
= min{o(a)+ pER's cx)}

aEE3
= min{u(a) + — a)} = rninuQ(t,x) (> —oo).
Finally, since 72(x) = max{(x,a)

aEE2
— a(a)} (see [64, p. 964]), one has
aEE2 oEE2
= Ui(X) — max{(x,a) —
a consequence of Theorem 9.6, u = u(t, x) mm ua(t, x) is therefore a

aCE3
global solution of the Cauchy problem (9.1)-(9.2). 0
Corollary 9.10. Let I = f(t,p) be of class C° on and = /(x) have a
d.c. representation (9.19) such that = o1(x), = are globally Lipschitz
continuous on R". Assume (E.IV). Then the function u = u(t, x) given by (9.22)
is a global solution of Problem (9.1)-(9.2).
Proof. Since the nonempty set E1 is bounded [117, §13.3], Hypoth-

esis (E.III) must hold while (E.I) is trivially satisfied. Hence, the conclusion is
immediate from Theorem 9.9. 0
Example. Consider the Cauchy problem
Ou/Ot + f(Ou/Or) = 0 in {0 <t < +oo, x E R},
u(0,x) = 4(x) on {i = 0, x E R},

120 9. HOFF'S FORMULAS FOR GLOBAL SOLUTIONS
with f(p) (1 + (for p E R) and
f x3/3 if x E [—1, 1],

1 x — (2/3)signz if x [—1,1].
We first note that neither the formula (9.3) of Hopf nor the formula (9.7) of Theorem
9.1 works in this case since the initial function here is not convex. Though the
present Hasniltonian I = f(p) is in fact convex we should also mention that Hopf's
formula (9.4) could not be applied directly to the problem, because
lim =1 +00.
In this case, however, it is easy to check the validity of d.c. representation (9.19)
where
10 if x<0,
x3/3 if rE [0,1],
tx—2/3 if x>1
and a2 = a2(x) cr1(—x) are globally Lipschitz continuous on R. Further, we
may invoke either (E.IV)" or (E.IV)' to deduce that (E.IV) holds. Therefore, by
Corollary 9.10, a global solution u = u(t, x) of the problem can be found in the
form
u(i, x) mm max{z(p — — — — t(1 + —
we have seen, Theorem 9.9 and its Corollary concern the Cauchy problem
As
(9.1).(9.2) in case the initial function 4 = has a d.c. representation
— o2(x) such that is bounded in R't. The following will be devoted to
the case where dom o = R".
Theorem 9.11. Let = be in the class with a d.c. representation

(9.19) such that lim = +oo. Under Hypotheses (E.I), (E.III), and
IrI-.+oo
(E.IV), suppose that there exists a function g = g(t) in with the property
that sup{f(t,p) p E R"}
: g(i) for almost all t E (0,+oo). Then (9.22)
determines a global solution of the Cauchy problem (9.1)-(9.2).
Proof. Since a2 = as(x) is a finite convex function on with lim a2(x)/IxI =

12H++OO
+m, so is its Fenchel conjugate function a = o(p); in particular, E2 domo =

R" (d. Remarks 1-2 of Lemma 9.13 later).
§9.4. EQUATIONS WITH CONVEX HAMILTONIANS f = f(p) 121
We shall continue using the notation Ua (t, z) introduced in the proof of Theorem
9.9. Let r E (0,+oo), v,. {(t,x) E V : t + ri < r}, sup
IzI<r
and .1.; g(r)Idr. Since lim = +00, to any M E (0,+oo) there
corresponds a number N,.,M 1 so that
o(a)/lal r+s,.+/4,.+M as al> N,.,M. (9.25)
By (9.20), it follows that, if(t,x) E V,., then (p,

Therefore, (9.24) and (9.25) imply
a(a)+max{(p,x)
pElt
—o(p)} —rIal —8,.
—s,.
—s,. —rlal —8,.> M,
provided (t,x) E V,. and al > N,.,M. This means that
tim ua(t,x) = +oo locally uniformly in (t,x) E V.
Hence (cf. Remark 2 after the formulation of Theorem 9.1), we may find a positive
number N,. for each r E (0, +oo) such that
jflf = mm ua(t,x) whenever (t,x) E Vr.
(It should be noted that Ua(t, x) is lower semicontinuous in a on the whole IR't.)
Moreover, analysis similar to that in the proof of Theorem 9.9 shows that the
solutions = Ua(t, a,) satisfy a Lipschitz condition on V,. with constants depending
on r but independent of a for al N,.; and that they satisfy (9.1) except on a
common set of Lebesgue measure 0. The proof is thus complete in view of Theorem
9.6.
§9.4. Equations with convex Hamiltonians I = 1(p)
We now consider the Cauchy problem
au/at + f(Ou/Ox) = 0 in V = {t > 0, x E (9.26)

u(O,.r) = on {t = 0, xE (9.27)
under the following two hypotheses.
(F.I) The initial function = cb(x) is of class C° and the Hamiltonian f = f(p) is
strictly convex on with lim =

IpI-*+oo
(F.II) For every bounded subset V of V, there exists a positive number N(V) so
that
+ t . f((x — w)/t)} < + t f((x — y)/t)

.
whenever (t,x) E V, > N(V). Here, f = f(z) denotes the Fenchel conjugate
function off = f(p).
In the sequel, we use the notation
ç(t,z,y) + t f((x — y)/t),

. (9.28)
where (t,x) e V, y E and shall prove the following theorem.
Theorem 9.12. Assume (F.I)-(F.II). Then the formula
u(t. x) inf ((t, x, y) = inf +t .

r ((x — y)/t) }
for (t, x) E V (9.29)
yEtR

The next auxiliary lemma is known ([64j, Theorems 23.5, 25.5, and 26.3]),
but what we would like to insist here is on its simple proof by the use of our Lemmas
9.2-9.3.
Lemma 9.13. Let f = f(p) be strictly convex on with lim = +00.

IpI-++oo
Then f' =
f(z) = (z,8f(z)/Oz) — f(5f*(z)/Oz) for all z E W1. (9.30)
Proof of Lemma 9.13. The strict convexity on of the function I = f(p) says
that this function is everywhere finite and that
f(Ap' + (1— A)p2) + (1— A)f(p2)
for any p', p2 E RTh, .\ [0, 1]; the sign of equality holding if and only if p1 =
E p2 or
E {0, 1}. Accordingly, f = f(p) is continuous.
§9.4. EQUATIONS WITH CONVEX HAMILTONIANS f = f(p) 123
It will now take a simple matter to check that w = w(z,p) (z,p) — f(p)
def def def
satisfies all the conditions of Lemmas 9.2-9.3 where we put E = R , in = Ii, =
z, Rm = and shall deal with the function
def
i,b=i,b(z)=sup{w(z,p) : pER }=f(z).
•
Indeed, since lim = +00, Condition (i) of these Lemmas holds while the
IpI-'+oo
others are almost ready.
As f = f(p) is strictly convex, it can be verified that the snultifunction L = L(z)
defined by
: w(z,p)=f(z)}
is actually single-valued on the whole Therefore, by Part b of Lemma 9.2, all
the partial derivatives Of (z)/Oz, exist, and L(z) = {Of'(z)/Oz}. Property (9.30)
I
is thus coming from the definitions of f = (z) and L = L(z). Further, Lemma
9.3 and its Remark imply the continuity of Of */Oz = Of (z)/Oz. 0
Remark 1. Consider a convex and lower semicontinuous function I = 1(p) on
Assume domf 0 and imf C (—oo,+oo] (the function I = f(p) is then called
proper). It will be shown that
lim = +00 if and only if domf* =
In fact, if lim = +00, then for each z E the supremuin

p1-4+00
f(z) = sup {(z,p) — f(p)}

pER"
is essentially taken over all elements p of just a compact set C and hence
finite. Conversely, let there exist an M E R and nonzero points p1, p2,... in such
that f(p") MIp"I fork = 1,2,... and that -+ +00 ask —* +00. Since is
locally compact, we may suppose that z0 E Putting z (M + 1)z°,
we thus get
f'(z) sup{(z,pk) — f(pk)} I(M + 1)(zo,pk/Iphhl) — M1}
> lim
—
Remark 2. Consider a finite convex function = on R" with the Fenchel

conjugate = 4i'(p). Let = be the Fenchel conjugate of =
Then it is known [641 that çiS = (p) is proper, convex, and lower semicontinuous
on and that 4" = Accordingly, domqS" = = hence, Remark 1
implies urn = +00.
Proof of Theorem 9.12. By (F.I)-(F.II) and Lemma 9.13, (9.28) determines a

continuous function (= ((t,x,y) whose derivatives
x, y)/Ot, OC(t, x, y)/ôxi,..., OC(t, x,
exist and are continuous on the whole V x moreover, one may apply Lemma
9.2 to the function =
•
= —((t,x,y) where p del
def del
= y, E = R and m =
def
n + 1, (t, x), 0 V. Consequently, the function u = u(t, x) defined by (9.29)

is locally Lipschitz continuous and is directionally differentiable in 1), with x)
equal to
mm ((f((x — y)/t) — ((x — — y)/i)/Oz),0f((x — y)/t)/Oz),e).

(9.31)
Here, e, L(t,x) {y ((i,x,y) = u(t,x)} 0 (Lemma 9.3). But,
according to Radernacher's theorem, u = u(t, x) is (totally) differentiable at any
point outside a null set Q C V. Therefore, suitable choices of e in (9.31) give
Ou,x)/ôt = — y)/t) — ((x — y)/t,0f((x — y)/t)/Oz)}

(9.32)
= — y)/i) — ((x — y)/t,0f((x — y)/t)/Oz)}
and
Ou(t,z)/8x1 = nun 0f((x — y)/t)/i3z1 = max 0f((x — y)/t)/0z1, (9.33)

yEL(t.r)
provided (t,x) V\Q and i {1,2,...,n}.

Now, given any (t,x) V\Q, we pick up some y L(i,x). Then it follows from
(9.30) and (9.32)-(9.33) that
Ou(t,r)/Ot = f((x — y)/t) — ((x — y)/t,0f((x — y)/t)/Oz)

= —f (i3f((x — y)/t)/c3z) = —f(Ou(t,x)/Ox).
The equation (9.26) is thus satisfied almost everywhere in V.

§9.4. EQUATIONS WITH CONVEX HAMILTONIANS f = f(p)
As the next step, we claim that
urn (9.34)
for each fixed x0 E Indeed, on the one hand, the definition (9.29) clearly forces
u(t, z) + t . f'(O), hence
liinsup (9.35)
On the other hand, let us first take a sequence {(ik, C V converging to

(O,x°) such that liminf u(i,z) = urn u(tk,xk) and second choose arbitrary
points E L(tk,xk) (for k = 1,2,...). Then it will be shown that x0.
(k-++ao)
In the contrary case, suppose without loss of generality that yk y° R",
(k—H-cc)
where y° x0. (We emphasize here that the sequence C is bounded
by Lemma 9.3.) Since lim f*(z)/Izp = +00 (cf. Remark 2 of Lemma 9.13),
hi-H-co
(9.35) and a passage to the limit as k —* +oo in the equality
u(tk,xk) = + f((x" — yk)/jk) (9.36)
would yield
liminf u(t,x) = lirn u(tk,xk) = + (+oo) = +00,

— k—H-co
a contradiction. This shows that lim = x0. But the continuous function
k-H-co
= is bounded from below since again lim f(z)/IzI = +oo. Therefore,
IzI-4+oo
a passage to the limit as k -+ +00 also in (9.36) implies
liminf u(i,x) = lim u(tk,xk) > (9.37)

k—H-co —
Finally, combining (9.35) and (9.37) gives (9.34), which says that u = u(i,x)
has a (unique) continuous extension over the whole satisfying (9.27). The proof
is thus complete. 0
Remark. Assume (F.I). Then (Fl!) is satisfied if lim ((t,z, y) = +oo uniformly
Ivi-H-oo
in (t, x) on each bounded subset of V.
126 9. FORMULAS FOR GLOBAL SOLUTIONS
In fact, let V c V be bounded, say, V C (O,r) x B(O,r) for some r E (O,+oo).

Put M 1f(O)I + < +m. It follows from (9.28) that mm ((t, x, w)
IzIr
((t,x,z) = + 1. f(O) < M whenever (t,x) E V. Hence, if urn ((t,x,y) =
Ii, 1-4+00
+00 uniformly in (t, x) on each such V, then for a suitable number N(V) r we
have
mm ((t,x,w) < rnin((t,x,w) < M < C(t,x,y) as (t,x) E V, > N(V);

IwIN(V)
i.e., (F.II) is satisfied.
Corollary 9.14. Under Hypothesis (F.I), suppose that

> —rn. (9.38)
Ixl-4+oo
Then (9.29) determines a global solution of the Cauchy problem (9.26)-(9.27).
Proof. By Remark above, it suffices to prove that urn ((t, x, y) = +00 uniformly
hi
in (t, x) on each hounded subset V of V. To this end, let V C V be bounded, say,
V C (O,r) x B(O,r) for some r E (O,+oo) and let M E (O,+oo) be arbitrarily given.
Condition (9.38) says that there exist numbers A, N e (0, +m) such that
2 —AIyI whenever 2 N.
But we certainly find a positive number j with the property that
f(z)/jz[ 2(M + A) as IzI 2 ii.
Putting N(V) + v)}, we therefore deduce from (9.28) that, if

(t,x) E V and 1111 2 N(V), then
+ - y)/t) Ix -
((t,x,y) [- A
I(x—y)/tI
.
Ill
.
because (x — y)/tI 2 [r(1 + v) — r]/r = Is (IyI — 2 1/2.
if = is globally Lipschitz continuous on IR", then (9.38) clearly holds. The

following result of Hopf [64] can thus be considered as a consequence of Corollary
9.14.
Corollary 9.15. If the initial funct;on = 4(x) ss globally Lipschztz continuous

and if the Ham:ltonzan f = f(p) is strictly convex on IR" with lim =
Chapter 10
Hopf-Type Formulas
for Global Solutions
in the case ot
Concave-Convex Hamiltonians
§10.1. Introduction
Consider the Cauchy problem for the simplest Hamilton-Jacobi equation, namely,
Ou/Ot+f(Ou/Oz)=0 in {t>0,
u(0,x) = 4>(x) on {t = 0, x E R't}. (10.2)
In the previous chapter, a global solution of this problem was given by explicit
formulas in some cases a little more general than the following two of Hopf: (a) I =
f(p) convex (or concave) and 4> = 4>(x) largely arbitrary; and (b) 4> = 4(x) convex
(or concave) and f = 1(p) largely arbitrary. It is unlikely that such restrictions,
either on f = f(p) or on 4> = 4>(x), are really vital. A relevant solution is expected
to exist under much wider assumptions. According to Hopf [64], that he has been
unable to get further is doubtless due to a limitation in his approach: he uses the
Legendre transformation globally, and this global theory has been carried through
only in the case of convex (or concave) functions [Fenchel's theory of conjugate
convex (or concave) functions].
In the present chapter, we propose to examine a class of concave-convex func-
tions as a more general framework where the discussion of the global Legendre
transformation still makes sense. Hopf-type formulas for non-concave, non-convex
Hamilton-Jacobi equations can thereby be considered. The method here is a devel-
opment of that in Chapter 9, which involves the use of Lemmas 9.2-9.3 (and their
generalizations). It is essentially different from the methods in [64] and [111]. Also,
the class of concave-convex functions under our consideration is larger than that in
[111] since we do not assume the twice continuous differentiability condition on its
functions.
128 10. HOFF-TYPE FORMULAS FOR GLOBAL SOLUTIONS
Let us continue using the notation of Chapters 8-9. We shall often suppose
that n n1 + n3 and that the variables x, p R" are separated into two as
(x',x"), p (pS,ph?) with x', p' Rn', x", p" Accordingly, the zero-
vector in will be 0 = (0', 0"), where 0' and 0" stand for the zero-vectors in
and Rn2, respectively.
Definition 1. [117, p. 349] A function f = f(p',p") is called concave-convex if it is

a concave function of p' for each p" E Rn2 and a convex function of p"
for each p' E Rfl.
In the next section, conjugate concave-convex functions and their smoothness

properties are investigated. Section §10.3 is devoted to the study of Hopf-type
formulas in the case of concave-convex Hainiltonians.
§ 10.2. Conjugate concave-convex functions
Let f = f(p) be a differentiable real-valued function on an open nonempty subset

A of R". The Legendre conjugate of the pair (A,f) is defined to be the pair (B,g),
where B is the image of A under the gradient mapping z = Of (p)/Op, and g = g(z)
is the function on B given by the formula
g(z) —
It is not actually necessary to have z = Of (p)/Op one-to-one on A in order that

g = g(z) be well-defined (i.e., single-valued). It suffices if
(z,p') — 1(p1) = (z,p2) — 1(p2)
whenever Of(p1)/Op = Of(p2)/Op = z. Then the value g(z) can be obtained

unambiguously from the formula by replacing the set (Of (z) by any of the
vectors it contains.
Passing from (A, f) to the Legendre conjugate (B, g), if the latter is well-defined,
is called the Legendre transfonnation. The important role played by the Legendre
transformation in the classical local theory of nonlinear equations of first-order is
well-known. The global Legendre transformation has been studied extensively for
convex functions. In the case where f = f(p) and A are convex, we can extend
f = f(p) to be a lower semicontinuous convex function on all of R" with A as
§10.2. CONJUGATE CONCAVE-CONVEX FUNCTIONS 129
the interior of its effective domain, if this extended I = f(p) is proper, then
the Legendre conjugate (B, g) of (A, is well-defined. Moreover, B is a subset
of domf (namely the range of Of/öp), and g = g(z) is the restriction of the
Fenchel conjugate f = f(z) to B. (See 1117, Theorem 26.41; cf. also Lemma 9.13
in the previous chapter.) Chapter 9 has thereby proved an important role of the
global Legendre transformation in the global theory of first-order partial differential
equations.
For a class of C2-concave-convex functions, H.T. Ngoan [1111 has studied the
global Legendre transformation and used it to give an explicit global solution to
the Cauchy problem (1O.1)-(1O.2) with I = 1(p) = f(p',p") in this class. He
shows that in his class the (Fenchel-type) upper and lower conjugates (117, p. 389],
in symbols f = T(z',z") and f = f(z',z"), are the same as the Legendre
conjugate g = g(z',z") off = f(p',p").
Motivated by the above facts, we introduce in this section a wider class of
concave-convex functions and investigate regularity properties of their conjugates.
(Applications will be taken up in § 10.3.) All concave-convex functions f = f(p',p")
under our consideration are assumed to be finite and to satisfy the following two
"growth conditions."
lim = +00 for each p' E Rn'. (10.3)

IP"I-4+°° Ip"I
urn = —oo for each p" E (10.4)

Ip'I-++oo Ip'I
Let f*2 = 1s2(ps,zu) [reap. f*i = f'(z',p")] be, for each fixed p' [resp.
p" R'2], the Fenchel conjugate of a given p"-convex [resp. p'-concavel function
I = f(p',p"). In other words,
z") sup {(z",p") — f(p',p")} (10.5)

p"€R"2
[resp. f'(z',p") inf {(z',p') — (10.6)
for (p',z") R" x [reap. (z',p") x R"3]. If f = is concave-

convex, then the definition (10.5) [resp. (10.6)] actually implies the convexity [resp.
concavityl of f*2 = f*2(p',z") [resp. f1 = f'l(z',p")] not only in the variable
z" E R"2 (reap. z' E but also in the whole variable (p',z") x
[reap. (z',p") E x Moreover, under the condition (10.3) (reap. (10.4)1,
_______________
________________ _____________
130 10. HOPF-TYPE FORMULAS FOR GLOBAL SOLUTIONS
the finiteness of f = f(p',p") clearly yields that of f2 = f2(pI, z") (reap. 1' =
I '(z',p")] (cf. Remark 1 of Lemma 9.13) with
f•2(p', z") f'(z',p") =
lim = +00 (resp. lim —00] (10.7)
z"I-4+oo Iz"I Iz'I—l+oo z'(
locally uniformly in p' (resp. p" E Rh'2]. To see this, fix any 0 < r2 <+00.
As a finite concave-convex function, I = f(p',p") is continuous on x 1117,
Theorem 35.1]; hence,
def
= sup f(p',p")l < +00. (10.8)
Ip'Iri
IP"I'2
def
So, with p = r2z ,, Liz"] (reap. p, =def—r1z'/]z']], (10.5) [reap. (10.6)] together with
(10.8) implies
f2(p', z")
inf 4 as z"] —* +co
Ip'Iri z"i ]z"]
sup
f'(z',p") as +oo].
(resp. + —r1 z']
Ip"Ir2 z'
Since r1, r2 are arbitrary, (10.7) must hold locally uniformly in p' [reap.
p" as required.
Remark. If (10.4) Ireap. (10.3)] is satisfied, then (10.5) (reap (10.6)] gives
IZ(p',zh') 1,011)
> -+ —p +00 (10.9)
li/i - li/i
fhl(z',p") I,)
(resp. < —+ —oo as (p"] —+ +001
— li"] (10.10)
uniformly in z" [reap. z'
Now let f = f(p',p") be a concave-convex function on x R'%z. Besides

"partial conjugates" f2 = and = fdl(z',p"), we shall consider the
following two "total conjugates" of I = f(p', p"). The first one, which we denote
by 7 = 7(z', z"), is defined as the Fenchel conjugate of the concave function
R" p' '—* z"); more precisely,
f"(z' z") def

= inf {(z',p') + z")} (10.11)
p' I
for each (z', z") E x The second, = f(z', z"), is defined as the Fenchel
conjugate of the convex function _f*1(zI,pSl); i.e.,
f(z', z") sup {(z",p") + f1(z',pU)} (10.12)

— p"ER"2
for (z',z") E x R"2. By (10.5)-(10.6) and (10.11)-(10.12), we have
f*(zI z") = pmi sup {(z',p') + (z",p") — f(p',p")},

ER I p"ER"2
(10.13)
f(z', z") = sup mi {(z',p') + (z",p") — f(p',p")}.
p"ER"Z p'ER'l (10.14)
Therefore, in accordance with [117, p. 3891,f = T(z',z") and f = f(z',z") will

be called the upper and lower conjugates, respectively, of f = f(p', p"). [Of course,
(10.13)-(10.14) imply j'(zl,zhl) L*(zs,zh?).] For any z' E Rnh, the function
x R'12 (p', z") '-+ h(p', z") (z',p') + f3Q/, z")
is convex. Thus (10.11) shows that T = T(z',z") as a function of z" is the image
i—* (Ah)(z") p', z") : A(p', z") = z"}
of h = h(p', z") under the (linear) projection 3 (p', z") A(p', z")
x z".
It follows that f = f(z',z") is convex in z" [117, Theorem 5.7]. On
the other hand, by definition, T = is necessarily concave in z' E
This upper conjugate is hence a concave-convex function on x Rn2. The same
conclusion may dually be drawn for the lower conjugate f = I (z', z").
We have previously seen that if the concave-convex function f = f(p',p") is
finite on the whole x and satisfies (10.3)-(10.4), its partial conjugates
f*2 = f*2(p',zI') and f*i = f'(z',p") must both be finite with (10.9)-(10.10)
holding. Therefore, Remarks 1-2 of Lemma 9.13 show that f = f(z, z") and
f = f(z',z") are then also finite, and hence coincide by [117, Corollary 37.1.2].
In this situation, the conjugate
1* = f(z', z") = (10.15)

off = f(p', p") will simultaneously has the properties:
= +00 for each z' E Rn', (10.3)

lz"I-9+00 Z
urn
f(z',z") = —oo for each z" E (10.4)
Iz'I—*+oo Iz'I
For the next discussions, the following technical preparations will be needed.
Lemma 10.1. Let f = f(pS,pfl) be a finite concave-convex functzon on x R'%a
with the property (10.3) [resp. (10.4)] holdsng. Then
,lim = +co locally uniformly in p' E

IP' I—I+co Ip' I
(10.16)
[resp. lim = —00 locally uniformly p" E

Ip'I (10.17)
Proof. First, assume (10.3). According to the above discussions, (10.5) determines
a finite convex function = f*2(pI,zhl). Further, [117, Corollaries 10.1.1 and
12.2.11 shows that
f(p',p") = sup {(z",p") — f3(pI, z")} (10.18)

z"ER"2
for any (p',p") E x Let 0 < r,M < +co be arbitrarily fixed. As a finite
convex function, f2 = f2(p',z") is continuous [117, Corollary 10.1.1], and hence
locally bounded. It follows that
sup If*2(pl,ZU)I < +oo. (10.19)

Ip'Ir
Iz"IM
So, with z" Mp"/Ip"I, (10.18)-(10.19) imply that
>M— M
Ip'Ir Ip"I — Ip"I
as Ip"I -4 +oc. Since M > 0 is arbitrary, we have
lim =
Ip"l
for any r E (0, +oo). Thus (10.16) holds. Analogously, (10.17) can be deduced from
(10.4). 0
Definition 2. A finite concave-convex function I = f(p', p") on x Rn2 is said
to be strict if its concavity in p' RTh' and convexity in p" E are both strict.
It will then also be called a strictly concave-convex function on x
Lemma 10.2. Let I = f(p',p") be a strictly concave-convex function on x

with (10.3) [resp. (10.4)1 holding. Then its partial conjugate = f*2(ps,zhl) [resp.
f1 = f*1(ZI,p'l)] defined by (10.5) [resp. (10.6)] is strictly convex ]resp. concave]
in p' E [resp. p" E RV121 and everywhere differentiable in z" R'%2 Iresp. z' E
Besidesthat, the gradient mapping R'hl x (p', z") '-+ Of 2(p', z")/Oz"
[resp. x (z',p") '-+ 0f1(z',p")/Ozhl is continuous and satisfies the
identity
z") (z", z")/Oz") — f(p', 0f2(p', z")/Oz")

(10.20)
[resp. f' (z', p") (z', Of" (z', p")/Oz') — f(Of" (z', p")/Oz', p")].
Proof. For any finite concave-convex function f = f(p',p") satisfying the property
(10.3) [resp. (10.4)], the partial conjugate f'2 = f*2(pS, z") [resp. 1" =
is finite and convex [resp. concave] as has previously been proved.
Now, assume that f = f(p', PS') i5 a strictly concave-convex function on x
Rn2 with (10.3) holding. Then Lemma 9.13 shows that f'3 = f'3(p',z") must
be differentiable in z" and satisfy (10.20). To obtain the continuity of
the gradient mapping R'hl x (p', z") i-+ 0f2(p', z")/Oz", let us go back to
Lemmas 9.2-9.3 and introduce the temporary notations: n n2, E m
+n2, and It follows from (10.16) that the
continuous function
= = w(p',z",p") (z",p") — f(p',p") (10.21)
meets Condition (i) of Lemma 9.2. Therefore, by Lemma 9.3 and its Remark, the
nonempty-valued multifunction
L= = L(p',z") {p" € : =
10. HOPP-TYPE FORMULAS FOR GLOBAL SOLUTIONS
should be upper seinicontinuous. However, since I = f(p', p") is strictly convex

in the variable p" E (10.21) implies that L = L(p',z") is single-valued, and
hence continuous in RnI x But L(p', z") = {Of*2(p', z")/Oz"}, which may be
handled by the same method as in the proof of Lemma 9.13 (we use Lemma 9.2,
ignoring the variable p'). The continuity of RnI x (p', z") —* Of 2(pS, z")/Oz"
has accordingly been established.
Next, let us claim that the convexity in p' E of f9 = f3(p', z") is strict.
To this end, fixO <A <1, z" E andp',q' E Rn'. Of course, (10.5) and (10.21)
yield
f2(Ap' + (1 — A)q',z") = maxw(Ap' + (1 —
<
max z",p") + (1 — A) max

p EW'2
= + (1— A)f2(q',zh').
if all the equalities simultaneously occur, then there must exist a point
p" E L(Ap' + (1 — A)q', z") fl L(p', z") fl L(q', z")
with
w(Ap' + (1 — A)q', z",p") = Aw(p', z",p") + (1 — A)w(q', z",p");
hence (10.21) implies
f(Ap' + (1 — A)q',p") = Af(p',p") + (1 — A)f(q',p").

This would give p' = q?,
and the convexity in p' of = f3(p',z") is
thereby strict.
By duality, one easily proves the remainder of the lemma. 0
We are now in a position to extend Lemma 9.13 to the case of conjugate concave-
convex functions.
Proposition 10.3. Let f = f(p',p") be a strictly concave-convex function on

with both (10.3) and (10.4) holding. Then its conjugate f = (z', :") de- I
fined by (10.11)-(10.15) is also a concave-convex function satisfying (10.3)-(10.4).
Moreover, f = f(z', z") is everywhere continuously differentiable with
Zn)) + (au', zhl)) —. z"), zht))
(10.22)
§10.2. CONJUGATE CONCAVE-CONVEX FUNCTIONS
Proof. For reasons explained just prior to Lemma 10.1, we see that 7(z',z")
(z',z"), hence that (10.11)-(10.15) compatibly determine the conjugate f =
I (z', z"), which is a (finite) concave-convex function on x with (10.3)-
(10.4) holding.
We now claim that f = I (z',z") = I (z',z") is continuously differentiable
everywhere. For this, let us again go back to Lemmas 9.2-9.3 and introduce the
def def def def def
temporary notations: n = n2, E = R m = ni+n2, 0 = Rn' +n 2, = (z ,z ),
g
and
pt-f p". Since (10.10) has previously been deduced from (10.3), we can verify
that the function
w= p) = w(z', z", p") (z", p") + 1'' (z', p") (10.23)
meets Condition (i) of Lemma 9.2, while the other conditions are almost ready. In
fact, as a finite concave function, f*a = f*1(z',p") is continuous (cf. [117, Corollary
10.1.1]) and so is w = w(z',z",p") (cf. (10.23)); moreover, Condition (ii) follows
from (10.23) and Lemma 10.2. Therefore, by (10.2) and (10.15), this Lemma shows
that f = f'(z', z") = f(z', z") should be directionally differentiable in x
with
8(e',e")f'(Z', z") e") + K.T(z',p"), es) } (10.24)

= p"EL(z'.z")
for (z',z") x Rn2, (e',e") E x where
L= = L(z', z") {p" E : w(z', z",p") = f(z1, z") = f(z', z")}

(10.25)
is an upper semicontinuous multifunction (see Lemma 9.3 and its Remark). Howev-
er, because f' = f1(z',ph') is strictly concave in p" E R'12 (Lemma 10.2), it may
be concluded from (10.23) and (10.25) that L = L(z', z") is single-valued, and thus
continuous in R'4 x Consequently, according to (10.24) and the continuity
of the gradient mapping x 3 (z',p") '—i (Lemma 10.2),
the maximum theorem [8, Theorem 1.4.16] implies that all the first-order partial
derivatives of f = I z") exist and are continuous in x (cf. also [128,
Corollary 2.2]). The conjugate f = f(z', z") is hence everywhere continuously
136 10. HOPF.TYPE FORMULAS FOR GLOBAL SOLUTIONS
differentiable. In particular, since L = L(z', z") is single-valued, it follows from

(10.24) that
L(z', z") z")}
{
and therefore that
z
— ôf' ( ,
(z , , z ,, )
0' )
= ôz' ' ôz"
Thus, (10.23) and (10.25) combined give
f*(zl z")) + z"))
Finally, we can invoke (10.20) to deduce that
z") (zi' zu)) + (z', zIP))) +

-
' ( 0f '\ Of
Oz' ' Oz"' '
(cii zfl)) + zIP)) — z"), zIP))
The identity (10.22) has thereby been proved. This completes the proof. 0
§10.3. Hopf-type formulas
We now consider the Cauchy problem
On/Of + f(On/Ox) = 0 in V {t > 0, x = (x',x") E IR" x R't2}, (10.26)
u(0,x) = d(x) on {t =0, x = (x',r") ER"' x (10.27)
An explicit global solution u = u(t,x) = u(t,x',x") of the problem will be found

under the following three standing hypotheses.
(G.I) The initial function = j(x) = r") is of class C° and the Hamiltonian
f= f(p) = f(p',p") is strictly concave-convex on IR"' x with (10.3)-(10.4)
holding.
(G.II) The equality
sup inf = ml sup ((t,x,y)

i, i, ER 2
§10.3. HOPF-TYPE FORMULAS 137
is satisfied in D, where
((t,x,y) +t. — y)/t) (10.28)
for(t,x) = (t,x',z") ED, E Here, f = f(z) = f(z',z")
denote, the conjugate defined by (10.11)-(10.15) of I = f(p',p").
(G.III) To each bounded subset V of V there corresponds a positive number N(V)
so that
mm
ER
sup ((t,x,w',w") < sup ç(txw'y")
%D
w'ER°l w'ER''
max inf C(t, X, w', w") > inf ((t, x, y', w")
*D"ER"3
w'I<N(V)
whenever(t,x) V, E R" x with > N(V).

The main result of this section reads as follows.
Theorem 10.4. Assume (G.I)-(G.HI). Then the formula

u(t,x) def
= sup inf ((t,x,y) = inf sup for (t,x) E D
y'ER"a V"ER"2 p'ER"a
(10.29)
For the proof of Theorem 10.4, we need further generalizations of Lemmas 9.2-
9.3 to the mixed sup-inf cases.
Lemma 10.5. Let 0 be an open subset of Rm and x = = a

continuous function of y) = y', y") E 0 x x with the following three
properties:
(i) The identity sup inf = inf sup holds in 0.
(ii) For every bounded subset V of 0, there exists a positive number N(V) such that
*0
mm sup < sup
Iw"IN(V)
and
max mi > mi
w'ER"l w"ER"2 w"ER"2
Iw'IN(V)
whenever V, y = (y',y") E x > N(V).
(iii) The function x = is differentiable in E 0 with continuous gradient
mapping 0 x x 3
Then one has:

(a) = sup mi = inf sup is a locally Lipschitz
cont2nuous function in 0.
(b) = is directionally differentiable in 0 with
y', e)
=
= miii max 0, e Rm)
v'EL'tt)
where
{y' unf =
1, ER 2
(10.30)
{y" E : sup =
Lemma 10.6. Let Conditions (i)-(ii) in Lemma 10.5 hold for a continuous function
= x x Then (10.30) gives a nonempty-valued, closed
x
and locally bounded multifunction L = x of E 0.
Proof of Lemma 10.6. For any bounded subset V of 0, denote Nv N(V)
(see (ii)). The function
0x ? '-÷ inf
ER'2
y
is upper semicontinuous as the unfimum of a family of continuous ones. It follows

from (ii) that for e V the supremum
= sup ml
y'ER'
should essentially be taken over all y' of just the compact set and
hence be attained. This means 0 L'(e) C (0', Nv) when V. Analogously,
0 L= is nonempty-valued and locally
bounded.
To complete the proof, we need only check the closedness of L = For
this, let be a sequence convergent to a point in
0x x Rn2 such that x as k = 1,2,3,...
Then one has inf = = sup

ER 2
and therefore gets
,0 ,k ,,k
= hm ,y, ,y
0 . k . k
,y ,y ) x(e ,y ,y )
.
liminf sup )
k-4+oo y'ER"
0 1,0
sup ,y'y I
).
y'ER"I
By duality, one also gets inf It follows from (i)

y
and the definition of = that
= sup inf inf

y'ER"l II ER"2 V
sup
V'ER"l
0 , ItO
)
inf sup =
V ER 3y'ER"'
hence that (y'°,y"°) E L= is

thus closed. U
0 ,0 1,0 0
Remark. The above proof also shows that ,y ,y ) = ). Since E 0
is arbitrary, we have lI's y") = whenever (y', y") E L(e). By the way, the
function = takes finite values only.
Proof of Lemma 10.5. (a) Let V be an arbitrary compact convex subset of 0

and N N(V). For any two points e',e2E V, pick up elements y'1 E L'(e') c
(0', N), y"2 E L"(e2) C N) [cf. the proof of Lemma 10.6]. Then
—
= inf — sup
v ER 2 y'ER"l
,,2 2 gi
,y ,y
= —
—
where r C V and A is a finite upper bound of over

Vx N) x r2(O" N). Analogously, A. e2
The function = is therefore locally Lipschitz continuous in 0.
(b) For any E 0 and e E Rm, we find two sequences

and of positive numbers convergent to zero such that
+ a'e) —
k-4+øo
and
+ /3be) —
= urn
k-4+Oo
Let us arbitrarily take an element y'° E and choose a sequence

p11k
convergent to some E with E + abe) for k = 1,2,... (Such a
choice is always possible in view of Lemma 10.6.] Then
—
sup inf
+ abe) pER" p ER"2
ak
+
= + y'°, e)
where E (O,ak). A passage to the limit as k —, +00 shows that
for any y E L, o
), hence that
sup
for some E Consequently,
inf sup (10.31)

p"EL"(C°)
It may dually be concluded that
sup inf (10.32)

)
Finally, combining (10.31)-(10.32) yields
inf sup = =
= sup inf
§10.3. HOFF-TYPE FORMULAS 141
for any 0, e Rm. AU "inf" and "sup" are actually attained because the
nonempty set L(e°) = x is compact (Lemma 10.6) and the mapping
x (y', y") y', e) is continuous. The proof is thereby
complete.
Proof of Theorem 10.4. By (G.I)-(G.III) and Lemma 10.3, (10.28) determines

a continuous function ( = ((t, x, y) whose derivatives
x, y), x, y), . . . , .x, y), x, y), . . . , x,
exist and are continuous on the whole V x IRS' x moreover, one may apply Lem-
def
ma 10.5 to the function x = x(e,y) = where m def
= def
+fl2, e = (t,x),
and 0 V. Thus u = u(t, x) defined by (10.29) is locally Lipschitz continuous
and directionally differentiable in V, with x) equal to either of the following
two quantities:
((i — (ii,
—
Here, e L'(t,x) {y' E Rfli : inf ((t,x,y',y") = u(i,x)}

y ER 2
0, L"(t,x) {y" Rn2 : sup C(t,x,y',y") = u(t,x)} 0 (Lemma 10.6).

But, according to Radamacher's theorem, U = u(t, x) is totally differentiable at any
point outside a null set Q C V. It follows that = Ou(t,x)/Ot =
hence, by (10.33), that
Ou . I
—(t,x)= nun max (s—)
t
at p"€L"(t,z) t t
•fx—y\
= max
i,"EL"(t,z)
nun I
cf i—i
t /
". \ t Or \ t
for (t,z) E V\Q. This would imply the existence of an element L"(t,z) so
that
— /S'—Y'
t
t ' t t t
t oak. t ' t
for all y' E L'(i,x). Analogously, it can be deduced from (10.34) that
Ott
—(t x)= . I •ix—y\
<f i—i
/x—y Ofix—y
Ot
max mm
p'EL'(t,z) p'EL"(t,x) \ t —(-———,—i————
/ \ i Oz \ t
I /z—y Of*,x_y
= ruin max <1 I—J-(———,——I—--—
p"EL"(t,z) I \ t / \ t Oz \ t
for (t,x) D\Q, and that there exists L'(t,x) with
Ou • — — y"\ / x' — Of — — y"
'
/z"—y"
\ t ' Oz" " t t
"(10.36)
for all y" V'(t,x). Finally, by (10.35)-( 10.36), we get
(t,x) V\Q, y (y',y") L(t,x) L'(t,x) x L"(t,x). In the same

manner one may see that
(t,x) = (t,x',x") V\Q, y = (y',y") L(t,r).

Now, given any (t,x) 'P\Q, we pick up some yE L(t,x), then use (10.22) and
(10.37)-(10.38) to obtain Ou(t,x)/Ot = —f(Ou(t,x)/Ox). The equation (10.26) is
thus satisfied almost everywhere in V.
As the next step, we claim that
limsup (10.39)
for each fixed x0 (x'°,x"°) Rn1 x Indeed, let us first take a sequence
{ C V converging to (0,x'°,x"°) such that limsup u(t,x) =
lim and second choose arbitrary points
yk L(t", z?k, ru1b) for k = 1,2,...
Then it will be shown that y" —4 x On the contrary, since

the sequence is bounded (Lemma 10.6), we can suppose without loss of
d
generality that y" y0 x Rn2 with y0 z°. It is clear
from (10.28)-(10.29) and the above definition of L = L(i,x) that
u(ik,rik,xutk) = inf
y" ER"3
.
risk) + tk , o") (10.40)
u(t,x1k,xh11) = sup ((tk,xsk,xhIk,y1,yutk) c(jkx?kxiskxskyisk)

y'ER"l
+ jk .f* (o', (10.41)
We need only consider the following two cases.
Case 1: y'° r'0. Then (10.4) and (10.40) show that
u(tk,xuhl,xIilh)
*
((x ,k ,k k
,0 ii )
Jim Jim I.
=
k-4+oo I(r' — y' )/t"I
—
So, (10.41) implies lirn tnhf*(Ot,(xhlk = —oo, a contradiction with
(10.3*).
Case y"° x"0. Analogously, (10.3) and (10.41) show that
,k risk)
lim u(tk, = +00,
hence that Jim i'f((z" — ysk)/tk 0") = +00, which contradicts (10.4).
l,-++oo
The contradictions, which we have got in both Cases 1 and 2, prove that
urn = x0. Therefore, a passage to the limit as k -+ +00 in (10.40)-(10.41)
would give (10.39). It may dually be concluded that
liminf
for each fixed r° = (z'°,x"°) R" x Thus u = u(t,x) has a continuous

extension over the whole satisfying (10.27). The proof is thereby complete. 0
Remark 1. Hypothesis (G.II) says that a "saddle-point" (y',y") E L(t,s) of the

function R" x p = (y', y") ((t, x, y) [with respect to maximizing over
and minimizing over exists. If f = f(p', p") has a special d.c. representation
f(p',p") 92(P) — 91(p') on x
then we can use the "index of non-convexity" and a classical minimax theorem [7]
to give sufficient conditions for (G.II) to hold. (See [12] concerning this question.)
Remark 2. If nj = 0, (10.29) gives the Hopf formula (9.4) and the formula (9.29).
If n2 = 0, the Hopf formula for concave Hamiltonians [64] will also be obtained.
Corollary 1O.T. Under Hypotheses (G.I)-(G.II), suppose that
> —m locally uniformly in z E R ',

.
lan inf ,,
Ix"I—*+oo IX I
(10.42)
çb(s'
lim sup , <+00 locally unsformly in x E Rn2.
IZ'I-4+00 Ix I
(10.43)
Then (10.29) determines a global solution of the Cauchy problem (10.26)-(10.27).
Proof. Let V C V be bounded, say, V C (O,r) x x for

some r (0,+oo). By (10.3*) and (10.42), analysis similar to that in the proof of
Corollary 9.14 shows that
liixi ((t,x,z',y") = urn {ch(x',y") + t. f' (o', S" — 11k)

} = +00
hi I-*+oe Ii' 1-4+00
uniformly in (t,x) = (t,x',x") E V. Hence, we may deduce that
0>— min{0,_1 + (t,a,)EVW"EIU'2

inf inf > —m.
Of course,
max
w'EIR'i w"EIU'3
inf ((t,x,w',w") >
—
inf
w"ER"2
((t,x,x',w") + 1 (10.44)
1w' I
as (t, x) E V. Further, in view of (10.43), there exist numbers A, N E (0, +m) such
that
< whenever Ix°I r, Iy'I N.
Finally, by (10.4), we certainly find a positive number with the property that
c*(z'
/ 2(q — A) as ii.
'
Putting + v))1 we therefore conclude from (10.28) that,
if(t,x) V and Iy'I N(V), then
. —
inf C(t,x,y',w") <((t,x,y',x") =
w€R2 t
+ f*((xlyS)/t,0h1) lx'—y'l]
[A y'I (10.45)
[A + - A). Iy'I
because I(x' — y')/tI [r(1 + ii) — r]/r = ii, — y'I/Iy'I (Iy'I — r)/ly'I 1/2.
Now, (10.44)-(10.45) show that the second condition in (G.III) holds. Analo-
gously for the first one. Thus the corollary follows from Theorem 10.4. 0
Chapter 11
Global Semiclassical Solutions
of First-Order
Partial Differential Equations
The present chapter is in principle a continuation of the previous three. However,

it was actually originated in the following problem posed by S.N. Kruzhkov [931.
Let a smooth [i.e., of class C'J function w = w(t, x) satisfy in the strip
[0, T} x the inequality
< NI&,(t,x)/Oxj, N= const. 0,

and the initial condition
on {i=0, rER1}.
Then it is easy to show (cf. Haar's Theorem 1.5) that w(t, x) 0 in Hr. Therefore,
the Cauchy problem for the first-order nonlinear equation
f = f(p) is of class C1(R1), cannot have more than one solution in 11T,
say, in the class of smooth functions with bounded derivatives. As Kruzhkov al-
ready remarked, the same conclusion may be drawn without appeal to the dif-
ferentiability of w = w(t,x) (reap. of the solution u = u(t,x)) or the validity of
(11.1) (resp. of the equation) at the points in any given finite union of straight
lines {t = const., x E R} C Hi'. The following question arises naturally: to
what extent can the condition on the smoothness of w = w(t, x) and on the va-
lidity of inequality (11.1) in the entire strip 11T be weakened? For example, the
Cauchy problem for the equation Ou/Ot + (Ou/Ox)2 = 0 with the zero initial con-
dition u(0, z) 0 has a continuum of piecewise smooth global solutions, such as
u = ua(i,x) min{O,alsI — a2t}, = const. 0. Note that each function
= satisfies the corresponding inequality almost ev-

erywhere in HT• Therefore, it is interesting to find intermediate classes (as wide
as possible) between C'(flr) and Lip((0,T) x R'), in which only the zero function
can simultaneously satisfy (11.1) and (11.2). These questions can be generalized to
the multi-dimensional case.
The study of this problem suggests that we should single out the widest class
between the class of continuously differentiable functions and the class of Lipschitz
continuous functions in which the Cauchy problem for a first-order nonlinear partial
differential equation has a unique global solution.
We shall assume 0 < T < +m (though most results here still hold in case
T= and continue using the notation of Chapters 8-9. Our discussions in this
chapter make an appeal to Theorems 8.1,8.8,8.11, and 8.12. The condition on the
validity of inequality (8.5) is clearly much weaker than that of (11.1) in the entire
domains of the corresponding functions under consideration. Moreover, it should
be noted (see §8.1) that
C Lip([0,T) x
The smoothness requirement on functions in is really weak enough: roughly

speaking, these functions need only be absolutely continuous in time variable. By
Chapter 8, the class V(QT) would be nominated as best candidate for our discussion
concerning the above questions for the Cauchy problem
ôu/ôt+f(t,x,u,Ou/Ox) =0 in (11.3)
on {t=0, (11.4)
We therefore arrive at the following definition of generalized solutions:
Definition 1. A function u = u(t,x) in is called a global semiclassical

solution to (ll.3)-(11.4) if it satisfies (11.3) for all x E JR" and almost all 2 E (0,T)
and if u(0,x) =4(x) for allx E
The above definition allows us to deal with the case of time-measurable Hamil-
tonians f = f(t, x, u,p). The next section is devoted to the uniqueness of global
semiclassical solutions in this case. Further, an answer to Kruzhkov's problem will
be given. Section §11.3 discusses the existence theorems, whose proofs are much
based on the results of Chapter 9.
148 11. GLOBAL SEMICLASSICAL SOLUTIONS
Most of the material presented here was published in [1411-11431 and 11501-1151].
The reader is referred to §2.3 and the references therein for the global existence of
classical solutions.
§11.2. Uniqueness of global semiclassical solutions

to the Cauchy problem
The present section deals with stability questions concerning global semiclassical
solutions of the Cauchy problem (11.3)-(11.4). Here, the initial data = is
a given continuous function on The Hamiltonian I = 1(1, x, u, p) is always
assumed to be measurable in t E (0,T) and continuous in (x,u,p) E x x
The proof of the following uniqueness criterion is immediate from Theorem 8.1.
Theorem 11.1. Suppose that f = f(t,x,u,p) satisfies the follotinng condition:

there exist a nonnegative function p = p(x) locally bounded on and a nonnegative
function £(t) in L1(0,T) such that
f(t, x,u,p) — f(t, x, < £(t) [(1 + xI)Ip — qi + p(x)]u — vi] (11.5)
for almost every t E (0,T) and for all (x,u,p),(x,v,q) E x x If ui =

u1 (t, x) and u3 = u2(t, x) are global semiclassical solutions to the Cauchy problem
(11.3)-(11.4), then uj(t,x) u2(t,x) in 11T•
Remark. Condition (11.5) is satisfied if and only if for some positive function
= £(t) in L1(0,T), the function
x R1 x 3 (t,x,u,p) f(t,x,u,p)/[t(t)(1 + x])]
is Lipschitz continuous with respect to p uniformly in (t, x, u) E x R', and is

Lipschitz continuous with respect to u uniformly in (t,x,p) E (0,T) xX x for
every compact set X C R't [i.e., uniformly globally in (t,p) and locally in x].
A useful uniqueness criterion for global semiclassical solutions with essentially

bounded derivatives is given by the next sharpening.
Theorem 11.2. Suppose that I = f(t, z, u,p) satisfies the following condition: for
any compact sets K1 C R1, K2 C there exist a nonnegative function 42 =
42(t) in L'(0,T) and a nonnegative function PK1K2 = PK1K2(X) locally bounded
§11.2. UNIQUENESS OF GLOBAL SEMICLASSICAL SOLUTIONS 149
on R° such that (11.5) with and p, respectively, holds for

almost everyt E (O,T) and for all (x,u,p),(x,v,q) E xK2. Ifui = ui(t,z)
and u2 = u2 (t, x) are global semwlassscal solutions to the problem (11 .3)-( 11.4) with
<+00 (j = 1,2),
(t,z)Eflr
then uj(t,x) u2(t,x) in
Remark. 1ff = f(t,p) depends only on t,p and is of class C1 on [0, T] x R't, then
the condition of Theorem 11.2 is satisfied. In this case, Theorem 11.2 solves the
problem of Kruzhkov (see Corollaxy 11.4 later).
To prove Theorem 11.2 we need the following:
Lemma 11.3. Let = be a locally Lipschitz continuous function of x on

If it is differentiable in the whole then
Oil, Oily
= sup
X3
(z=1,...,n).
zER zER
Proof of Lemma 11.3. Fix any i E {1,...,n}. It suffices to treat the case when
def Oil,
= esssup
UXj
<+m.
Let us write x = (x',z1) instead of x = where
, def
x =
Then for almost all [with respect to the (n — 1)-dimensional Lebesgue measurel
x' we have:
—(x,.) Ox1
<s1.
L°'(R') —
Since the function il'(x',.) is absolutely continuous on each bounded segment, it
follows that
—
I
for almost all [with respect to the (n — 1)-dimensional Lebesgue measurej x'
and for all E R1. From the continuity of il' = il'(x) and from (11.6), we
conclude that
— —
for all E R'1. Therefore,
OtI,
<sj
for all x This proves the lemma. 0

Proof of Theorem 11.2. According to the definition of Lemma 11.3
shows that
ouj ôti,
sup = esssup (j = 1,2; = 1,... ,n)
rER"
for almost all t (0, T). Taking the essential supremum over t E (0, T), we find
that
ouj auj
esssup sup = esssup
zER (t,z)Eflr
Consequently, by assumption,
r def j
= rnaxesssup sup —(t,x) <+00. (11.7)
X
t€(0,T)
Let X" be as in (8.22); K3 LC £(.) For an arbitrarily fixed

T' E (0, T), we consider the sequence of the following paraflelepipeds:
pk x = {(t,x) : 0< t <T', XE Xk}.
Continue using (8.23). Then the function = p(x) given by (8.24) is locally
bounded on
We now consider the function u = u(t,x) ui(t,x) — u2(t,x). Of course,
u(O,x) 0. Moreover, in view of (11.7), the hypothesis of the theorem implies
= f(t, x, ui(t, x), z)) — f(i, x, u2(t, x),

< t(t) . [(1 + — + ,z(x)Iui(t,x) — u2(t,x)I]
= £(t) . [(1 +
x for almost all t E (0,T'). Therefore, Theorem 8.1 shows that

u(t, x) 0 in Since T' (0, T) is arbitrarily chosen, the proof is complete. 0
Corollary 11.4. Let I = f(t,x,u,p) be measurable zn i E (O,T). contmuous in

.r E R", and differentiable in (u,p) E R1 x such that, for any compact set
K C R", the function
£K =€K(t)1+ sup
xK '-'P
is Lebesgue integrable on (0, T), and the function
deC 1
1'K = i/K(x,u) = esssupsup
*€(O,T) pEK
is locally bounded on x JR'. If ui = ui(t,x) and u2 = u2(t,x) are global semi-

classical solutions to the Cauchy problem (11.3)-(11.4) with
<+00 (.7 = 1,2),

('X
then uj(t,x) in 11T•
Proof. Let us introduce the notation
,,K3(x)_ sup vK2(x,u)

uEK1
for any convex compact sets K1 c R1, K2 C R'. Then it is easy to check that
(11.5) with 42 and PK,,K2 in place of€ and p, respectively, holds for almost every
t E (O,T) and for all (x,u,p), (x,v,q) E JR" x K1 x K2. The corollary thereby
follows from Theorem 11.2. 0
We leave it to the reader to prove the following criterion of continuous depen-
dence on initial data for global seiniclassica.l solutions.
Theorem 11.5. Suppose f = f(t,x,u,p) satisfies (11.5). Let u3 = u,(t,x) (j =

1,2) be global semiclassical solutions to (11.3) with
on {t=0,
where = (j = 1,2) are given functions continuous on JR". Then
Iui(t,z) - u2(t,x)l <exp[C(x)Je(r)dTJ.

£(r)dr—1
-
11. GLOBAL SEMICLASSICAL SOLUTIONS
C(x) being defined in (8.7).
Remark 1. The example in Remark 2 following Theorem 8.1 shows that the
Lipschitz continuity of functions in the class V(Qr) also plays an essential role in
the definition of global semiclassical solutions. The zero solution aside, this example
gave no other global semiclassical solution to the Cauchy problem
Ou/Ot=O in Q1,
u(O,x)=O on {t=O,
Remark 2. Consider the Cauchy problem
Ou/Oi + (au/Ox)2 = 0 in {0 <t < T, x E R1},
u(0,x)=0 on {t=0, x€R1}.
Figure 11.1
By definition, if u = u(t, x) is a global semiclassical solution to the problem, then

for almost every t (0, T), the function u(t,.) is differentiable on R'. Obviously,
(1 1.8)-( 11.9) has a continuum of piecewise smooth global solutions, such as (see
Figure 11.1)
def 10 if
UA,e = =
— A2(t — T+e)} if T— e t <T,
where A 0, 0 <e <
T. For A > 0 the differentiability of the function UA,g(i,.)
fails somewhere [at x = ±A(t — T + e) and at x = 0] if and only if belongs to the t
interval (T — e, T), whose Lebesgue measure is precisely e (positive but as small

as we please). Thus, the zero function = uo,1(t,x) [i.e., A = 0] is the unique
global semiclassical solution to (11.8)-(11.9) in the class of functions with essentially
bounded derivatives (cf. the remark following Theorem 11.2).
Remark 3. The requirement that £ = £(t) be Lebesgue integrable on (0, T) is also

essential to the above uniqueness theorems. To see this, we consider the Cauchy
problem (even for a linear equation):
Ou lOu u 1_ in 'Tj' rE R'
u(0,x)=0 on {t=0, XER1}.
Here, for 4x) 1, the Hazniltonian f = f(t,u,p) 24P) —

clearly satisfies
(11.5) with £(i) But we can easily check that the present problem admits the
following two global (semi)classica.l solutions:
u=uj(t,x) def
= t and u=u2(t,x) def
= i(l+e ).
The explanation for this event is that the function £ = £(i) in this case is not
Lebesgue integrable on any (0, T). (We could not choose a Lebesgue integrable
£ = £(t) for (11.5) to hold.)
By the use of Theorems 8.8, 8.11, and 8.12, the results in this section can be
generalized to the case of the Cauchy problem for weakly-coupled systems:
0u1/Ot + =0 in fir (i = ,m), (11.10)
u(0,x) = on {t = 0, x (11.11)
Here, the initial data qS = = ,4m(x)) is a given vector function

continuous on Each Hamiltonian = f1(t,x,u,p) is always assumed to be
measurable in t E (0,T) and continuous in (x,u,p) E R7t x x First, we give

the definition of global semiclassical solutions for the problem.
Definition 2. A vector function u = u(t,x) in is called a global serniclas-

sical solution of (11.10)-(1l.11) if it satisfies (11.10) for all r E and almost all
t E (0,T) and if u(0,x) = for ails E
We can now formulate some stability results for global semiclassical solutions of
the problem (l1.l0)-(1l.11) and leave the proofs to the reader.
Theorem 11.6. Suppose f, = (i = 1,... ,m) satisfy the conditions as

follows: there exzst a nonnegative function p = p(x) locally bounded on and a
nonnegative function e = e(t) in L1 (0, T) such that
x,v, < £(t) . [(1 + +p(s) k=i

max Iuk (11.12)
for almost every t E (0,T) and for all (x,u,p), (x,v,q) E x Rm < (i =
m). Let u = n(t, x), ii = x) be global semiclass*cal solutions to (11.10)
with the following corresponding initial conditions:
u(0,x)=4(x), ii(0,x)=cf4x) on {t=0,

where = x), q5 = x) are given vector functwns continuous on R'. Then
— <exp[C(x)f e(r)dr} x
x sup max
.=1 m
C(s) be;ng defined (8.7).
Corollary 11.7. Suppose f, = (i = 1,... ,m) satisfy the conditions

(11.12). If ii = u(t,r) and v = v(t,x) are global semiclassical solutions to the
Cauchy problem (11.10)-(11.11), then u(t,x) v(t,x) in ciT.
Theorem 11.8. Let (8.32) be a comparison equation. Suppose f1 =

(i = 1 m) satisfy the following conditions: there exists a nonnegative function
= £(t) in L1(0,T) such that
s,u,p) — < e(t)(1 + IsI)Ip — qi + p(t, max Iuk — vkl)

for almost every t E (O,T) and for all (x,u,p), (x,v,q) E x x (i =

1,. , m). If u = u(t, x) and v = v(t, x) are global semiclassical solutions to the
. .
Cauchy problem (11.1O)-(11.11), then u(i,x) v(t,x) in fZT.
Theorem 11.9. Let (8.34) be a comparison equation. Suppose = f1(t,x,u,p)

(i = 1,... m) satisfy the following condition,,: there exist a nonnegative functzon
,
p = p(x) locally bounded on and a nonnegative function £ = e(t) in L1(O,T)

such that
x, u,p)—f1(t, x, v, £(t) Iuk—vkl)] (11.13)
for almost every t (O,T) and for all (x,u,p), (x,v,q) x x R" (z =
1, . , m).
. . If u = u(t, x) and v = v(t, x) are global semiclassical solutions to the
Cauchy problem (11.lO)-(11.11), then u(t,x) v(t,x) in
In the case of systems, a useful uniqueness criterion for global semiclassical

solutions with essentially bounded derivatives is given by the next sharpening.
Theorem 11.10. Let (8.34) be a comparison equation. Suppose f6 =

(i = 1, . ,m) satisfy the following conditions: for any compact sets K1 C R, K2 C
. .
there ex;st a nonnegative function £K2 = (t) in L' (0, T) and a nonnegative
function PK1.K2 = PKI,K3(x) locally bounded on such that (11.13) with and
in place of€ and p, respectively, hold for almost every t (0,T) and for all
(x,u,p), (x,v,q) x K1 x K2 (i = 1,... ,m). If u' = u1(t,z) and u2 = u2(t,x)
are global semiclassical solutsons to the problem (11.10)-(11.11) with
max max <+00,

3=1,2 t=1 m
then u1(t,x) tz2(t,x) in
With mw in (8.34) [cf. the remark preceding Proposition 8.91, Theorem

11.10 gives:
Corollary 11.11. Let = f1(t,x,u,p) be measurable in t (O,T), continuous

in x R't, and differentiable in (u,p) Rm x such that, for any compact set
K C R, the function
= £K(t) i + max sup +

OP
11. GLOBAL SEMICLASSICAL SOLUTIONS
is Lebesgue ;ntcgrable on (O,T), and the function
def
esssup sup Ii
= vK(x,u) = k;max 1 m *E(O,T) pEK
is locally bounded on x R". If u1 = u1(t,x) and u2 = u2(t,x) are global

semsclass;cal solutions to the Cauchy problem (11.1O)-(11.11) with
max• max <+00,

j1,2 t1
then u1(t,x) u2(t,x) in cT.
§11.3. Existence theorems
This section is devoted to the existence of global semiclassical solutions to the

Cauchy problem for a single Hamilton-Jacobi equation:
ôu/Ot + f(t,Ou/Ox) = 0 in (11.14)
u(O,x) = on {t = 0, x E R"}, (11.15)
with = a finite convex function on Explicit representation of such

solutions will be obtained by the use of Chapter 9. For definiteness, let us rewrite
the two standing assumptions (E.I)-(E.II) of §9.2 on the Hamiltonian I = f(t,p)
and initial function = in a specified alternative form (when T < +co] as
follows. (The Fenchel conjugate function of = is here denoted by 4' =
as usual.)
(E.I) The Hamiltonian I = f(t,p) is continuous in {(i,p) t E (O,T)\G, p :
for some closed set C C [0, TI of Lebesgue measure 0. Moreover, to each N E

(0, +co) there corresponds a function = gN(t) in T) such that
sup f(t,p)( for almost all t E (0,T).

IpIN
(E.H) For every compact subset V of [0,T) x there exists a positive number
N(V) so that
(p,x) — q4(p) — f f(r,p)dr < max{(q,x) — — jf(r,q)dr}

§11.3. EXISTENCE THEOREMS 157
whenever (t,z) E V, > N(V).
As we have mentioned in Chapter 9, (E.I) implies the t-measurability and p-

continuity of I = f(t,p) on {0 <t <T, p In addition, since 4' = 4'(x) is
finite on R", this hypothesis allows us to define an upper semicontinuous function
= z,p) from [0, T) x x into [—oo, +oo) by taking
x,p) (p, x) — 4'(p) j f(r,p)dr. (11.16)

—
We shall use the notation E dom 4' 0 and deal with the function u = u(t, z)
and multifunction L = L(t,x) of (t,x) [O,T) x determined by the formulas:
pEE
L(t, x) {p E E : z,p) = u(t, x)}.

def def
Lemma 9.3, wzthm = n+1, = (t,x), 0 def
= [0,T)xR andw =
def
shows that L = L(t,x) is a nonempty-valued, closed and locally bounded

multifunction of (t, x) [0, T) x R". This multifunction is therefore upper semi-
continuous. An analogue of Theorem 9.1 says that (11.17) gives a global solution
to the Cauchy problem (11.14)-(11.15).
Next, set
((0,T)\G) x = {(t,x) QT : t G}
for any G C [0,T]. In this section we have:
Theorem 11.12. Let 4' = 4'(x) be a finite convez function on R". Assume (E.I)-
(E.II). Suppose that the multifunction L = L(t,x) given by (11.18) is continuous in
G being as in (E.I). Then (11.17) determines a global semiclassical solution to
the problem (11.14)-(11.15).
Proof. According to the maximum theorem [8, Theorem 1.4.16), the formulas
(9.13)-(9.14) show that the derivatives ôu/Ot, Ou/Oxi,..., exist and are
continuous in (see also [128, Corollary 2.2)). Hence, the global solution u =
u(t, x) is continuously differentiable in Since C is of Lebesgue measure 0, we
conclude that u = u(t,x) belongs to An argument similar to that in the
proof of Theorem 9.1 now gives us the validity of (11.14) in C]
158 II. GLOBAL SEMICLASSICAL SOLUTIONS
Remark. It can he proved that in all partial derivatives of the global solution
u = u(t, x) exist [(11.14) is then satisfied though the solution may fail to he dif-
ferentiable] if and only if the multifunction L = L(t,x) is single-valued. See §13.2
later for the smoothness of v = u(t, x).
Corollary 11.13. Let 4 = be a finite convex function on R?Z. Assume

(E.I)-(E.II). Then (11.17) determines a global semiclassical solution to the problem
(11.14)-(11.15) if one of the following two condztions holds:
(i) = is strictly convex on its effective domain E
(ii) f = f(t.p) is strictly p-convex; more precisely, it is strictly convex with respect
to p on E for almost every fixed t E (0, T).
Proof. Each of (i)-(ii) implies the strict concavity of p = p(t,x,p) with respect
to p on E for every (t, x) E [0, T) x It follows that the upper semicontinuous
multifunction L = L(t, x) is indeed single-valued, and hence continuous. 0
In case n = 1. an existence result is established as follows:
def *
Theorem 11.14. Let = be a finite convex function on IR, E = domç5
and let = be of class C2 in mtE. Assume:
(i) The Hamiltonian f = f(t,p) belongs to C°(((0,T)\G) x with
esssuplf(t,p°)[ <+00
iE(O,T)
for some p° E E and some closed set C C IR of Lebesgue measure 0; it is, moreover,
twice continuously differentiable in p E mt E with
max{esssup sup Of (t,p)/Opj, esssup sup 102f(t,p)/0p21} <+m

tE(O,T) pEint E tE(O,T) E
E(0, +oo).
(ii) To each compact subset V of [0, T) x IR there corresponds a positive number N(V)
so that x, p) < max x, q) whenever (t, x) E V, p E p1 > N(V). Here
pEE
qIN(V)
is defined by (11.16) withn=1, (p,x)=px.
(iii) For every (t,p) E ((0,T)\G) x intE, it holds that
d24
(11.19)
§11.3. EXISTENCE THEOREMS
Then (11.17) determines a global semiclassical soltttson to the problem (11.14)-

(11.15) with n = 1.
Proof. We first note that E is a nonempty convex set in R', and is therefore
an interval with end-points, say, a,b E [—oo, +oo], a < b. Consequently, [117,
Corollary 7.5.1] implies the continuity of the restriction of = (p) to
Condition (i) shows that I = f(t, p) is t-measurable and p-continuous on {O <
t < T, p E Moreover, we conclude from this condition that (11.16) gives a
continuous function = on [0,T) x R x with values in [—oo,+oo); it
being actually finite when p E E and twice continuously differentiable in p E mtE
with the derivatives
= z
d4 ('ôf
— 7—(p)
—
=
— f (r,p)dr. (11.20)
Under Hypothesis (ii), the method of Lemma 9.3 proves that the multifunction
L = L(t,x) defined by (11.18) is nonempty-valued and upper semicontinuous on
[0, T) x R. It may be shown that
ess sup sup If(t,p)I <+00

tE(O,T)
IpIN
for all N E (0,+oo), hence that u = tt(t,x) is in Lip([0,T) x R) (cf. Proof of

Theorem 9.1).
Finally, according to (11.19)-(11.20), for each (t,x) E the maximum set
L(t,z) = cannot contain more than one point. Therefore, L = L(t,x)
is really continuous in Analysis similar to that in the proof of Theorems 9.1
and 11.12 shows that u = u(t,x) belongs to and that (9.13)-(9.15) [with
n = 11 hold in Thus (11.17) determines a global semiclassical solution to the
problem (11.14)-(11.15). 0
Corollary 11.15. Let = be a finite convex function on R, E def
= •
and let = be of class C2 in intE. Under Conditions (i)-(ii) of Theorem

11.14, assume that
>0 (11.21)
for all p intE and almost all t (0,T). Then (11.17) determines a global
semiclasszcal solution to the problem (ll.14)-(11.15) with a = 1.
Proof. Since (11.21) implies (11.19), the conclusion follows.
We conclude this chapter with the following example and remark.
Example. Let J c [0,1) be the Cantor set and w = w(t) be the Cantor ladder
(see Remark 2 after the formulation of Theorem 8.1). Define
def
g(t) = min{It—sI s€J}.
Clearly, g = g(t) is Lipschitz continuous on (0, 1) (with Lipschitz constant 1), and
hence differentiable almost everywhere in (0,1); its derivative dg/dt = dg(t)/di
being Lebesgue measurable. Next, consider the Cauchy problem
On dg . Ou
+ = 0
.
in {0 < t <1, XE R },
1
on {t=0,rER1}.
Then we may invoke Theorem 11.1 to deduce that the function u = u(t,x)
—g(t) + is the only global semiclassical solution to tius problem. Further, this
problem has no classical solution even in the small [i.e., even in the local sense].
Remark. By the contributions on viscosity solutions, the global existence and

uniqueness of generalized solutions to first-order partial differential equations have
been established almost completely. (Look at the short historical survey in §4.1; see
also the historical remarks for equations of conservation law in Chapter 5.) We shall
show later in Chapter 12 that each global semiclassical solution is also a minimax,
hence a viscosity solution since the last two are equivalent. The a priori estimates
in Theorems 11.5-11.6 are certainly of much interest from various view-points (as
we have partly seen in this book) with regard to numerical application inclusive.
However, these estimates could be expected to apply only to sufficiently regular
generalized solutions as they could in the case of global semiclassical solutions.
Chapter 12
Minima.x
ot Par•tial Ditterential
Equations with
Time-measurable Hamiltonians
We have seen in the previous chapters that the main theorems in the classical
theory of first-order nonlinear partial differential equations are valid only local-
ly, in sufficiently restricted domains. The example in the preface shows that the
Cauchy problem might fail to possess a global C'-solution even if the Hamilton-
ian and initial data are analytic. Therefore, the need for introducing generalized
or weak solutions has arisen in the theory of first-order nonlinear equations and
its applications. Such solutions have been investigated by many mathematician-
s in the past 50-70 years ([21-14], [21], [32]-[33j, 1441-145], [51]-(521, [54J, [63]-[64], [87]-
The notion of global solutions in Chapter 9,
[92], [94], [951, [97]-[98], 1112], [1181).
which was probably first used by Hopf [64], affords the existence but not unique-
ness question. Some other supplementary condition like the semi-concavity or
semi-convexity condition (see Chapter 4) must be introduced for the uniqueness
of a global solution (in the case of convex or concave Hamilton-Jacobi equations).
In a similar situation, entropy condition (see §5.4) is needed for the uniqueness
question of weak solutions. In recent years the development of the theory of gen-
eralized solutions has been based to a significant degree on the concept of viscos-
ity solutions, which was introduced by Crandall and Lions. Within this theory,
uniqueness and existence theorems have been developed for various types of equa-
tions and boundary-value problems, and also some applications to control problems
and differential games have been studied ([51,1101-120], [28], [35]-[391, [47]-[50], 167]-
1721, [79], [991-11011, [122]-[i231, [131]). The concept of viscosity solutions is motivat-
ed by the classical maximum principle which distinguishes it from other definitions
of generalized solutions.
In the present chapter, we develop another approach that can be considered
as a nonclassical characteristic method, according to which a generalized solution
162 12. MINIMAX SOLUTIONS
is assumed to be flow invariant with respect to the so-called characteristic inclu-

sjons. This direction has been suggested by Subbotin in [124]-[125J, and leads
to the notion of minimaz solutions. The above term originates from the theory
of differential games. It is justified by permanent presence of mm-max operations
in investigations of these solutions, including the well-known Hopf formulas (see
(9.3)-(9.4)), and in investigations based on idempotent analysis, which have been
implemented in recent years by Maslov, Kolokol'tsov, and Samborskii [821, [105j.
To facilitate access to the topics from motives of the characteristic method, let
us go back to and consider the following "terminal" Cauchy problem (data
being given at the end point t = T of time interval):
Ou/Ot+f(t,x,Ou/Ox)=O in
u(T,x) = on {t = T, x (12.2)
Here, = is a given C1-function on After a change of variables (r, x)

(T — t,x), from (12.1)-( 12.2) we indeed get the usual Cauchy problem (data being
given at the initial point r = 0 of time interval). However, for the construction in
this chapter, it will be more convenient to consider the terminal condition (12.2).
We first assume that the Hanultonian I = f(t, x,p) is of the form
f(t,x,p) (a(t, x),p), (12.3)
with a = a(t,x) (a1(t,x),... an function of class C1 on

Assume, further, that
a(t,x)I <L(1 + IxI) V(t,x) (12.4)
for some constant L E (0, -4-no). The characteristic system (1 .3)-( 1.4) of the Cauchy
problem (12.1)-( 12.2) in this homogeneous linearity case is reduced to the system
( dx
I
dv — (12.5)
0,
Ldt
with the condition
x(T) = y, v(T) = o(y). (12.6)
It is known that the problem (12.5)-( 12.6) has uniquely a solution
(x,v) = (x(t,y),c(y))
on some interval (T — e,T1 for each y E R". Under the hypothesis (12.4), the
characteristic curves x = x(t, y) can be extended over the whole segment 0 t T
(cf. Appendix I). Moreover, an easy argument (adapted from [62, Chapter V,
Peano's Theorem 3.1, Corollary 3.1 and its Remark]) shows that the Jacobian
Dr/Dy of the C1-mapping RTh 3 y i-+ H1(y) does not vanish for any y
RTh and 0 < t < T. Hence Lemmas 2.3-2.5 imply that x = is a diffeomorphism
from R" to itself for any t [0,T). Therefore, the inverse function y = y(t,x) is
of class C on
1 —
and u = u(t,z) def
= v(t,y(t,x)) = o(y(t,x)) is the only global
C1-solution of (12.1)-(12.2) in the homogeneous linearity case (12.3).
The above definition of the solution u = u(t, x) is based on the structure of
characteristics. In such a homogeneous linearity case, the family {(t, x(t, y)) : 0
< of characteristic curves covers all of fZT and no two of them could
intersect in c17; in other words, there exists one unique characteristic curve passing
a given point (t, x) E For a nonhomogeneous linear Hamilton-Jacobi equation
Ou/Ot + (a(t,x),âu/Ox) +g(t,x) = 0, (12.7)
with g = g(t, z) a function in C1 (fir), we can continue using the above family,
{(t,x(t,y)) : 0 t
of characteristic curves and prove that
def I
u = u(t,x) = a(y(t,x)) + j g(r,x(r,y(t,x)))dr
is a unique global C1-solution of (12.7) and (12.2).
Now, turn to the general case. Consider the nonlinear equation (12.1), with
I = f(t, x, p) and = some twice continuously differentiable functions. In
this case, the characteristic system (1.10) may be reduced to
81
I
(z=1,...,n),
I (128
Of
(z=1 n).
We shall assume that for any y there has uniquely a solution
(x,p) = (x(t, y), p(i, y))

of (12.8) on the whole segment 0 t T satisfying the condition
x(T) = y, p1(T) = (i = 1,... ,n).
Moreover, suppose that the family {(t, x(t, y)) : 0 < t < of characteristic
curves covers and that no two of them could intersect in QT. Assume, further,
that the inverse y = y(t, x) is of class C2 on Qr. Then it can be claimed that
u = u(t,x) +
— y(t, (r, x(r, y(t, x)),p(r, y(t, x)))] dr
is the only global C1-solution of the Cauchy problem (12.1)-(12.2) [cf. Theorem
2.71.
Wemust note here that for a general nonlinear equation (12.1) the above as-
sumptions are not always automatically satisfied as they are in the linearity case
(12.3). Sufficient conditions for them to hold could also hardly be given explicitly
because the structure of characteristic curves is then via a projection (z, p) '—i x
from x into R". Without these assumptions the preceding method of finding
a global C1-solution breaks down. A relevant solution, which is expected to exist,
should be understood in some generalized sense; say, one needs to relax the smooth-
ness condition on it as usual. Subbotin's method of defining a minimax solution is
based on a construction in the theory of positional differential games. According to
this theory, the value function of a differential game is characterized by the prop-
erties of being u-stable and v-stable simultaneously (see (84]-[86]). It is also known
that at the points of differentiability the value function satisfies a first-order partial
differential equation (called the Isaacs-Belhnan equation). The u- and v-stability
properties can be expressed in various ways and used to introduce the notion of mm-
imax solutions. We shall see more concretely in the next section that one of these
ways is, to some extent, a relaxation of the classical characteristic method: The
ordinary differential equations (12.8) to determine the characteristics x = x(t, y)
can be slightly separated into two characteristic differential inclusions (see (12.14)
and (12.16) later). The first-integral condition, which means that u = u(t,x) is
constant along each characteristic curve x = x(t, y), is accordingly separated into
two inequalities (see (12.17) and (12.19) in the definition of a minimax solution).
§12.2. DEFINITION OF MINIMAX SOLUTIONS 165
Our aim here is to extend Subbotin's notion (given in [11 and (124]) of minimax
solutions of first-order partial differential equations with continuous Hamiltonians to
the case of time-measurable Hainiltonians and present the uniqueness and existence
theorems for such solutions. The results in this chapter are new even when restricted
to the case of continuous Hamiltonians. Almost of them were published in [132]-
[1331 and 1146].
The outline of the chapter is as follows. In Section §12.2 we give the definition of
minimax solutions to the (terminal) Cauchy problem for a general nonlinear evolu-
tion partial differential equation with time-measurable Hamiltonian. We investigate
some properties of multivalued mappings which play a decisive role in the definition
of minimax solutions. Section §12.3 is devoted to the relations between minimax
solutions and global (semi)classical ones. We prove that a global semiclassical so-
lution is also a minimax solution, and conversely, a minimax solution satisfies the
equation in the classical sense at the points of differentiability. Further, in Sec-
tion § 12.4 we discuss the invariance of the definition with respect to the choice of
concrete multivalued mappings. In Section §12.5 we establish the main theorem of
this chapter on the uniqueness and existence of minimax solutions. Finally, Section
§ 12.6 concerns some generalizations to the case of monotone systems of first-order
partial differential equations. Our method is based on the theory of multifunctions
and differential inclusions, and on a sharpening of a well-known theorem on the
Lebesgue sets for functions with parameters. We also use an implicit version of
Gronwall's inequality. Throughout, 0 < T < +00, S {p E : = 1},
def
B,. = {p E R : <r} (r > 0), B def
= B1.
§12.2. Definition of minimax solutions

10 Formulation of the Cauchy problem
Let us consider the Cauchy problem of the form
=0 in QTdI{0<i<T, xER°}, (12.9)
u(T,x)=o(z) on {t=T, (12.10)
Assume that the terminal data c = o(z) is of class C0 on and that the Hamil-
tonian f = f(t,x,u,p) depends on (t,x,u,p) E xRx with the following
properties.
a) Conditions:
a!) For almost every (in the sense of Lebesgue measure) fixed t E (O,T), the
function Rx 3 (x,u,p) i—+ f(t,x,u,p) is continuous.
a2) For each (x,u,p) E R" x R x S, the function (O,T) 3 t '—* f(t,x,u,p) is
measurable.
b) For any bounded sets D c R" and E C R, there exists a function A1,,8 = AD,E(i)
in L1(O,T) with
A1,,e(t)ix—x'i (12.11)
for almost alit E (O,T) and for allx,x' ED, u E E, p ES.

c) There exists a function £ = 1(t) in L1(o,T) such that
sup{If(t,x,u,p) — — 1(t). (1 + — : p,q E B} SO (12.12)
for almost all t E (0, T) and for all (x, u) x R.
d) For almost all t E (0,T) and for all (x,p) R
f f(t,x,u,p) is positively homogeneous in p E R"; i.e.,
f(t,x,u,s = s . f(t,x,u,p) Vs 0 (12.13)
for almost alit (0,T) and for all (x,u,p) E R" x R x S.
Note that Conditions a2), c), and e) together imply:
a3) For all (x,u,p) Rx the function t '—* f(t,x,u,p) zs Lebesgue

integrable on (0, T).
2° Differential inclusions for supersolutions and subsolutions

In the present section and the next one we shall often use the notations:
F(t,x) . (1 + lxi) . B,
F,j(t, x, u, a) {z E F(t,x) : (z, a) f(t, x, u, a)}, (12.14)
FL(t, x, u, {z E F(t,x) : (z, < f(t, x, u,

for (t,x) QT, u R, o,f3 E S.
§12.2. DEFINITION OF MINIMAX SOLUTIONS 167
Remark. For any multifunction G(t,x) C and any E

denote by the set of all absolutely continuous functions x = x(t) from
[0, T) into which satisfy almost everywhere in (0, T) the differential inclusion
E G(t,x(i)) subject to the constraint = We shall always assume
that G = G(t, x) is nonempty convex compact valued, measurable in t, upper
semicontinuous in x, and that
IG(t, x)l sup {IzI : z E G(t, x)} < c(t) . (1 + rI) (12.15)
on with c = c(t) a function in L'(O,T). From [29, Theorems 11.20, 111.15, and
VI.13j (see also [40, Theorems 5.2 and 7.1]), it follows that XG(i,,x.) is then a
nonempty compact subset of C([0,T],lfr). Now let U XQ(t,x), 0
C 11T• If C [0, TI x is compact, then by Lemma 8.3, C Xo({0} x
Br) with r (r° + 1)ffc(f)di — 1. The compactness of x B,.), which
follows from the upper semicontinuity of the compact valued multifunction R"
r Xo(0,x) c (see [40, Theorem 7.1] or [29, Theorems 11.25 and
VI.13]), therefore implies that of
Now it can be seen that the multifunctions
(1,x) F(t,x) C
x IR x S (t,x,u,a) '-+ C
R xS (i,z,u,8) '-4 FL(t,x,u,(3) C
in (12.14) are nonempty convex compact valued, measurable in i, upper sernicon-

tinuous in x. (It can be shown that they are indeed continuous in x.) As was
mentioned in the above, the sets
def
u,cs) = Xpu(.,.u,a)(i*, xe),
(12.16)
def
XL(:t., u, = x,)
are always nonempty and compact for all (tm, z,) E So we may conclude the
following:
Definition 1. A supersolution of Problem (12.9)-(12.10) is a finite lower semicon-

tinuous function u = u(t, x) on which satisfies the condition
sup mm [u(r,x(r)) — u(i,x)J <0 (12.17)

QES
for all 0 <t < r < T, x E and also the condition
VxERTh. (12.18)
Definition 2. A subsolution of Problem (12.9)-( 12.10) is a finite upper semicon-

tinuous function u = u(t, x) on which satisfies the condition
inf max [u(r,x(r)) — u(t,x)] >0 (12.19)

BES —
for all 0 < t < r <T, x E R", and also the condition
u(T,x)a(z) YxER". (12.20)
The sets of all supersolutions and subsolutions of (12.9)-( 12.10) will be denoted
by Solu and SOIL, respectively.
Definition 3. A function u = u(t,x) in SOIU Soli. is called a minimaz solution

of the Cauchy problem (12.9)-( 12.10).
30 Further properties of F0 = Fu(t, x, u, a) and FL = FL(t, x, u, /3)
It will be shown (see [124, p. 16]) that
Fu(t,x,u,a)flFL(t,x,u,fl)$O for all a,/3E S. (12.21)
In fact, it follows from (12.13) that f(t,x,u,0) = 0, hence from (12.12) that
if(t,r,u,p)i (1 + lxi) Vp E S. (12.22)
Therefore, if (a,/3) = 1, then a = and
z f(t,x,u,a) . a = f(t,x,u,/3) . /3 e Fu(t,x,u,a)fl FL(t,x,u,/3).
Further, let 0 < (a,/3) < 1. Setting e (1 — (a,/3)2)_h/2 . ((a,[3) a — /3), we

get
lel = 1, (e,cx) = 0, (e,/3) = —(i (a,/3)2)h/2 < 0. (12.23)
Next, choose z f(t,x,u,a) a + £(t). (1 + lxi) e and conclude from (12.22)-

(12.23) that lzi2 = [f(t, x, u, a)]2 + [e(t) (1 + lxi)]2 < (1 + lxl)12, i.e., .
that z E F(t,x). On the other hand, (12.23) implies (z,a) = f(t,x,u,a), thus
DEFINITION OF MINIMAX SOLUTIONS 169
Z E Fu(t,x,u,a). In addition, since (a,/3) 0, it follows from (12.13) that (z,/3) =

f(t,x,u, (a,/3) a) + £(t) (1 + si) . (e,f3). But, according to (12.12) and (12.23),
2 1/2
(a,/9) a — /3( = (1— (a,/9) ) = —(e,/3),
f(t, x, u, (a, 13) a) f(i, x, u, /3) — £(t) (1 + si) (e, /3).
Therefore, (z,fl) < f(t,r,u,/3); i.e., z E FL(t,x,u,/3). Hence, z E Fu(t,x,u,a) fl

FL(t, s, u,fi).
Finally, let (a,/3) <0. Setting F(t,x) z (1 + Iii) . (a — /3)/(a —/31,
we obtain (z, a) = £(t) . (1 + (1— (a,/3))h/2 2 £(t) . (1 + si), because Ia — /3( =
— (a,/3))1/2 (as al = $1 = 1). Thus, (12.22) gives (z,a) 2 f(t,s,u,a);
i.e., z E Fu(t,s,u,a). Analogously, z E FL(t,s,u,fl), and in consequence, z E
Fu(i,s,u,a)flFL(i,x,u,/3). The equality (12.21) is thereby proved (for almost all
t E (0, T) and for all (x, u) E x RI. It follows that
f(t,s,u,p) = sup mm (z,p) = mi max (z,p) (12.24)

a€S $ES zEPL
for almost all t E (0,T) and for all (x,u,p) R x R". Indeed, let p €5. By
(12.14), we easily see that
sup mm (z,p) mm
aES
Moreover, by (12.21) and again (12.14), we also have
mm (z,p) < mm (z,p) f(t,s,u,p)

zEF0(*,x,v,a) ZEFU(i,z,u,a)flFL
for any a E S. Thus, in view of (12.13), the first equality in (12.24) is satisfied, no
matter whether p E S or not. Similarly for the second one.
From the monotonicity and continuity in u off = f(i,x,u,p), it may also be
deduced that
1)
U>,,
= Fu(i,s,v,a), fl = FL(t,s,v,/3) (12.25)
for almost all t (0,T) and for all a,j3 E S, V E R, 5 E Another property
of the multifunctions = and FL = FL(t,s,u,/3) is given in the
following.
12. MINIMAX SOLUTIONS
Proposition 12.1. One has
1
supliminf inf — I mm (z,p)dr>f(t,x,u,p),
aES ft :EFu(r,z(r),u,a) —
(12.26)
1
inflimsup sup — I max (z,p)dr<f(t,x,u,p)
ö\O Jt ZEPL(r,z(r),u,.8)
for almost all t E (0, T) and for all x E U E R, p E
The proof of Proposition 12.1 will be based on the next two lemmas, which are
sharpenings of a well-known theorem on the Lebesgue sets (see [119, p. 158]) for
functions with parameters.
Lemma 12.2. For any r E (0, +oo), there exists a set A(r) C (0, T) of Lebesgue
measure 0 such that
1
f(r,x,u,p)dr—f(t,x,u,p)=0 (12.27)
for alit E Ac(r) x E B,. C R", u E,. C R, p

Proof of Lemma 12.2. According to b) and e), we find a function A = A(t) deC
=
AB,,E,(t) in L1(0,T) and a null set Aj(r) C (O,T) so that
f(r,x,u,p) — f(r,x',u,p)] < . — x'I A(r) (12.28)
for 8.11 X,X' E Br, u E E,., p E rE (0,T)\Aj(r). By al), there is a

null set A2 C (0,T) such that the function Rx ? (x,u,p) f(t,x,u,p) is
everywhere continuous for each t A2 = (0,T)\A2.
In virtue of Lebesgue's theorem (see, for instance, [119, p. 158]), we see that
for any g = g(t) in T) the set Leb(g) of all t E (0, T) satisfying
1
g(r)dr—g(t)=0
is of measure 7'. Let Q be the countable set of all rational real numbers and take
A(r) Aj(r) U A2 U (f(,,'p)) fl fl Leb(A))

DEFINITION OF MINIMAX SOLUTIONS 171
[cf. a3) and c)]. Then mes(A(r)) = 0. The only point remaining is to prove
(12.27) for any fixed t Ac(r), x Br, u p RTh. To this end, let e > 0 be
an arbitrary number. Since t Ac(r) C Leb(A) fl we have
t+S
1 i 1 )
sup — I £(r)dr, sup — I A(r)dT + 1 <+oo. (12.29)
ft J
Because t there must be > 0 such that
If(i,x',u',p') — f(t,x,u,p)I < e13 (12.30)
whenever max{Ix — r'I, Izt — u'I, p— <). Now choose

p' E
so that — x'I, — p'I}

— <u <u + 5(2), where
e/[6M(r + 1)1, + i)]}. (12.31)
Since t (k = 1,2), we could find 5(e) > 0 with the following

property:
1
t+6
f f(r, — < e/3 k = 1,2.

(12.32)
Finally, setting > 0, we have only to show that
1
t+o
f(r,x,u,p)dr — f(i,x,u,p) <e
f (12.33)
For k = 1,2 one writes
t+6
1
f(r,x,u,p)dr — f(t,x,u,p)
f t+a
1
[f(r,z,u,p)—f(r,x',u,p')]dr
=
+
1
j [f(r,x',u,p') — (12.34)
+
f —
f(t, u, p)].
12. MINIMAX SOLUTIONS
Let (5E The first summand in the right-hand side of (12.34) is estimated
as
(+6 (+6
1
f [f(r,x,u,p) — f(r,x', <

g+6
Ix —
f
1 e e e
by (12.12)-(12.13), (12.28)-(12.29), (12.31) and by the choice of The

third and fourth summands are estimated as (12.30) and (12.32) (cf. (12.31)).
Therefore, taking k = 1 in (12.34), we get
1 C
f f(r,x,u,p)dr — f(i,x,u,p) <3 = e, (12.35)
because the second suminand in the right-hand side of (12.34) is then nonpositive
by Condition d). Analogously, letting k = 2 implies
1 C
f(r,x,u,p)dr—f(t,x,u,p)> —3. = —e, (12.36)
since the second summand in the right-hand side of (12.34) is then nonnegative.
Combining (12.35) and (12.36) gives (12.33), which completes the proof. 0
Lemma 12.3. There exists a set A C (0, T) of Lebesgue measure 0 such that
1
(+6
f(r,x(r),u,p)dr = f(t,x,u,p) uniformly in x(.) E Xp(t,x) (12.37)
for alliE Ac,x

Proof of Lemma 12.3. It suffices to take A where the sets A(k) are
constructed in Lemma 12.2 and its proof. In fact, it follows easily from (12.27)-
(12.29) that (12.37) must hold for all i E Ac, x E u E R, p E because the
function family Xp(t,z) is uniformly bounded and is equicontinuous at t. 0
Proof of Proposition 12.1. The integrals in (12.26) exist (for reasons explained
just prior to (12.16); cf. also Remark 4 in Section §12.4 later). By the positive
homogeneity in p off = f(t,x,u,p) (see (12.13)) and of(z,p), we need only prove
§12.3. RELATIONS WITH SEMICLASSICAL SOLUTIONS 173
(12.26) for almost all i (O,T) and for all x E U E R, p E S. Indeed, according
to (12.14), one then obtains
fj
1
supliminf ml — I mm (z,p)dr>
aES J\O zEFrj(r,g(r),u,a) —
1
> liminf in! — I mm (z,p)dr
— i\O z(.)EX,(i,z)S ft zEF0(r,zfr),u,p)
1
t4J
= liminf
6\O
in! — I f(r,x(r),u,p)dr,
ft
in! lim sup sup — I max (z,p)dr <
PES ft ZEFL(r,xfr),u.P) —
1
limsup sup — / max (z,p)dr
.\O Jt ZEFL(r,r(r),u,p)
1
=limsup sup — /
f(r,x(r),u,p)dr.
ö\O z(.)EXp(t,z)'5 Jt
Therefore, by Lemma 12.3, the proof is complete. 0

Corollary 12.4. We have
1 d
supliminf in! f(i,x,u,p),
aES ö\Ü
—
ft/ dt
in! limsup
,SES
sup
di
f(t,x,u,p)
for almost alit E (0,T) and for alix E U ER, p E R'.

Proof. The proof is immediate from (12.26). 0
§12.3. Relations with semiclassical solutions
The following result says that the notion of minimax solutions indeed generalizes
"correctly" that of classical ones.
Theorem 12.5. Assume a)-e). Then

i) every global semiclassical solution u = u(t, x) of (12.9)-( 12.10) 30 that the gradient
mapping x i—* Ou(t, r)/Ox is continuous for almost all t (0, T) must also be
a minimax solution of the same problem,
ii) there ezists a null set A C (0, T) such that, at each point (t, x) E ((0, T)\A) x
where a certain minimax solution of (12.9)-(12.1O) is differentiable, the equation
(12.9) must be satisfied.
Proof. i) Let u = u(i,z) be a global semiclassical solution of Problem (12.9)-
(12.10) such that z Ou(t,z)/Ox is a continuous function for almost every
t E (0,T). Based on (12.14), for (t,z,ci) E x S, we define
F°(t,x,a) {z° E Fu(t,x,u(t,x),ci)

(12.38)
(z°,Ou(t,x)/Ox) = mm (z,Ou(t,x)/Ox)}.
xEFrj
Since (t,x) '-+ Ou(t,x)/Ox is measurable in t, continuous in x and the

multifunction Fu = x, u, ci) is indeed measurable in t (0, T), continuous in
(x,u) E RTh x R, it follows from [8, Theorems 8.2.8, 8.2.11) and Berge's maximum
theorem [22, p. 123] that the nonempty convex compact valued multifunction
F° = F°(t,x,s) is measurable in t E (0,T) and upper semicontinuous in x E
deC
Given any ci E S, a certain estimate of the form (12.15) holds for G(.,.) = F (., .,o)
because F°(.,.,a) C F(.,.). Hence, as was mentioned in the Remark
2°, for arbitrarily fixed 0 t < r T, x we could find at least one element
By (12.24) and (12.38), it is then easy to see that the
continuous function [0, TI i—* u is Lipschitz on every segment [e, T—eJ
(0 < e < T/2) with the derivative
= +
+ = 0
almost everywhere in (0,T). It is therefore decreasing on [0,T]. Thus, according

to (12.25), if a function x(.) on [0,T] is chosen so that = and
= for some z(2)() E z,u(t,z), ci), then
z(.) E
Moreover, u(r,x(r)) — u(t,z) = u(r,z(')(r)) — u(t,x(1)(t)) 0. Hence, by def-

inition, u = u(t,x) is a supersolution of (12.9)-(12.10). Analogously, it is also a
subsolution, and is therefore a minimax solution of the same problem.
ii) Take A C (0,T) to be a null set satisfying (12.37) of Lemma 12.3.
By [119, p. 158], we may assume that N C with £ = 1(t) the function
§12.3. RELATIONS WITH SEMICLASSICAL SOLUTIONS 175
mentioned in Condition c). Let u = u(t, x) be a minimax solution, differentiable at

some AC x R't, of (12.9)-(12.10).
Firstly, choose a S and 0 < s R so that ôu(t(°),x(°))/Ox = s a. Since
u = u(t,x) is in Solij, for any 5€ (0,T—t(°)), there exists an
XS(.)
such that
+ + 5)) — 0. (12.39)
Let
def
= dxö
(0
= + 5) —
Because AC is a Lebesgue point oU = £(i) and the function family C

is uniformly bounded, one has
+oo>C def
1
= sup £(r)(1+Ixo(r)I)dr sup —IyoI. (12.40)
.SE(o,T_g(°)) 1(°)
In accordance with (12.40) and with the differentiability at of u = u(t, x),

it follows from (12.39) that
+ (t(°),x(°))) < (1 + C) .5. limes = 0. (12.41)
By the definitions (12.14) of the inultifunction Fri = Fu(t,x,u(t(°),x(°)),a) and

(12.16) of the set Xu(t(°),x(°),u(t(°),x(°)),a), we now see that
/ 1
=
1 1
— I f(r,xo(r),u(t(°),x(°)),a)dr,
hence, by (12.13),
(1
S Ox
I
Jt(O) Ox
This together with (12.37) and (12.41) implies
+ f(t(°),x'°),u(t(°),x(°)), < a.
Ox
Analogously, we have
+ f(t(°),x(°),u(t(°),x(°)), 0,
provided that u = u(t,x) is in SOiL. The equation (12.9) must therefore be satisfied
at (t(°),x(°)). Theorem 12.5 is completely proved. 0
§12.4. Invariance of definitions
There is a certain indefiniteness in the definitions (12.14) of Fu = Fu(t,x,u,a)

and FL = FL(t,x,u,/3). For example, we have a large choice = £(t). However,
this indefiniteness does not influence the basic Definitions in §12.2. As will be
shown in the remainder of the chapter, instead of (12.14) one can, in fact, use any
= Fu(t,x,u,a) Iresp. FL = FL(t,x,u,13)] in the following common families of
multifunctions.
Let P and Q be arbitrary nonempty sets. We consider any multifunctions
xRxP (t,x,u,a) '-* Ftj(t,x,u,a) C R",

fly. x R x Q 3 (t,x,u,f3) FL(t,x,u,13) C
satisfying the general conditions below.
(i) The multifunctions Fu = Fo(t,x,u,a) and FL = FL(t,x,u,/3) are nonempty

convex compact valued, measurable in t and upper semicontinuous in x. Moreover,
x, u, cx) U FL(t, x, u, (3) C F(t, x)
for almost all t E (0, T) and for all x E where c = c(t) in L' (0, T) is independent
of a, (3, u.
(ii) For almost alit (0,T) and for dlx a P, Q, V ER one has
fl Fu(t,x,u,a) = Fu(t,x,v,a), 11 F,(t,x,u,f3) =FL(t,x,v,/3).
(iii) For almost all t (0,T) and for all (x,u,p) E R x R" the following
inequalitzes hold:
sup mm (z,p) <f(t,x,u,p) < mi max (z,p), (12.42)

cxEP — — PEQ ZEFL
§12.4. INVARIANCE OF DEFINITIONS 177
sup liminf inf f(t,x,u,p),

aEP dt
rt+o
inf limsup sup / at
f(t,x,u,p),
5\o
where, still as in (12.16), we let
def def
= XPu(.,.,u,a)(t,X), XL(t,x,u,fJ) =
The family of all multifunctions = (resp. FL = FL(t,x,u,f3))

satisfying Conditions (i)-(iii) will be denoted by .Iu(f) (resp. FL(f)). If the Hainil-
tonian f = f(t,x,u,p) satisfies a)-e), then .Fu(f) 0, 0. In fact, under
such hypotheses, we can choose any P, Q C with the property that
{s.a cEP,sO}={s./3
and then use (12.14) to define a concrete pair of multifunctions =
and FL = FL(t, x, u, This pair would satisfy (i)-(iii). (See Part 30 of §12.2 for
the case where Pdot
= 5, Q def
= S.)
Remark 1. By the Hahn-Banach theorem, it follows from (i) and (12.42) that
Fu(t,x,u,Q)fl Fj(t,x,u,/3) 0 (12.44)
for almost all t E (0, T) and for all E F, E Q, (x, u) E x R.
Remark 2. Condition (12.43) holds if Fu = Fu(t,r,u,a) and FL = FL(t,x,u,/3)

satisfy the inequalities
1
p4+6
supliminf in! — I mm (z,p)dr>f(t,x,u,p),
aEP fg
(12.45)
1
in! limsup sup — / max (z,p)dT<f(t,x,u,p)
flEQ 6\o
for almost all t (0, T) and for all x E U E R, p E
Remark 3. In the case where I = f(t,x,u,p) is a continuous function of its ar-

guments, Subbotin [124] and Adiatullina and Subbotin [1] do not assume (12.43)
for Fu = Fu(t,x,u,a), FL = FL(t,x,u,P); they assume the upper semicontinu-
ity in (t, x) of these niultifunctions (instead of the measurability in t and upper
semicontinuity in x as we do) and assume the equalities in (12.42). However, it

can be proved that (12.43) holds then. In fact, under such assumptions, for all
e > 0, (t, x) E QT, U E R, p E there exists a E P so that
mm (z,p) f(t,x,u,p) — e/2.

zEFu (*,z,u,a)
Given any x(.) Xp(i,x), the multifunction (O,T) 3 r Fu(r,x(r),u,a) is

upper semicontinuous. From this and the maximum theorem, it follows that the
function (0,T) r '—* miii (z,p) is lower semicontinuous. Hence, when
5 > 0 is small enough, one has
— nun (z,p)dr — nun (z,p) — e/2 dr

ZEFv(r,x(r).ii,a) J*
= mm (z,p) — e/2 f(t,x,u,p) — e
zEF0 (t,z,U,Q)
(independently of the choice of x(.) E Xp(t, x) because the function family Xp(t, x)
is eqwcontinuous at i). In other words, the first inequality of (12.45) must be true.
Dual arguments give the second one. Therefore, by Remark 2, (12.43) holds.
Remark 4. Let Fu = x, u, a), FL = FL(i, x, u, 13) be Carathéodory in (t, x) E

QT [i.e., measurable in t E (0, T) and continuous in x E satisfying Condition
(i). Given any x(.) E Xp(t,x), by [8, Theorem 8.2.14] (cf. [29, Theorem 111.151),
the measurability in r of the multifunctions
(0,T) 3 1. Fu(r,z(r),u,a),
(0,T) 3 r FL(r,x(T),u,[3)
(which follows from (8, Theorem 8.2.8]) implies that of the functions
(0,T) 3 r '—+ mm (z,p), (0,T) 3 r max (z,p).

zEFt
Therefore, the integrals in (12.45) exist. Moreover, if (12.42) and (12.45) hold, then
one has in fact the equalities in (12.45), because by Lemma 12.3, (12.42) clearly
forces
1
t+6
supliminf inf — I miii (z,p)dr<
aEp ö\O r()€Xp(t,z)45
1
<liminf inf — I sup mm (z,p)dr
ö\O z( .)EXp (t,) ft aEPZE Ptj (r,z(r),u,Q)
1
<limuni
ö\O
inf .j j f(r,x(r),u,p)dr=f(t,x,u,p),
(JEXF(t,z)o Jt
§12.5. UNIQUENESS AND EXISTENCE OF MINIMAX SOLUTIONS
1
,t+ö
inflimsup sup — I max (z,p)dr
LNo zEFt(r,x(r),u,$)
p*+6
1
lim sup sup — I inf max (z,p)dr
S\O
1
limsup sup — / f(r,x(r),u,p)dr=f(t,x,u,p)
Jt
for almost all t E (0,T) and for all x E u E R, p R".
For any = in FL = FL(t,x,u,13) in we have:
Definition 4. The set Solu(Fu) of supersolutions (relative to Fo) of the Cauchy

problem (12.9)-(12.10) is defined to be the set of all finite lower semicontinuous
functions u = u(t, r) on 11T which satisfy (12.18) and the condition
sup mm [u(r,x(r)) — u(t,x)1 0 (12.46)

aEP
for any0<t<r<T,xER".
Definition 5. The set SO1L(FL) of subsolutions (relative to FL) of the Cauchy
problem (12.9)-(12.10) is defined to be the set of all finite upper semicontinuous
functions u = u(t,x) on which satisfy (12.20) and the condition
inf
flEQ
max [u(r,x(r)) — 0 (12.47)
In Section 5 we shall prove that, for any Fu = Fu(t,x,u,a) in and

FL = FL(t,x,U,/3) in IL(f), the intersection Solu(Fu) fl consists of a u-
nique function, which is independent of Fij = Fu(t, x, u, a) and FL = FL(t, x, ti, f3).
Therefore, in complete concord with the previous definitions, we may also give the
following.
Definition 6. A function u = u(t,x) in Solu(Ftj) fl Soli.(Fz.) is called a minimax

solution (relative to and FL) of the Cauchy problem (12.9)-(12.10).
§12.5. Uniqueness and existence of minimax solutions
The aim of this section is to establish the following main theorem.

Theorem 12.6. Let a = a(.r) be continuous on and f = f(t,x,u,p) satisfy

all Gonditzons a)-e). Then there exists one unique minimax solution of the Gauchy
problem (12.9)-(12.1O).
For the proof of Theorem 12.6 we need to make some preparations. Given any
locally bounded function n = u(t,x) on clT, we define:
def
ii
—
(t,x) def
= — limiuf
.
.
u(r,y) and u + (t,x) = limsup u(r,y)
for (t,x) E Obviously, u = u(t.x) [resp. u+ = u+(t,x)] is then the largest

lower seinicontinuous [resp. the smallest upper semicontinuous] function on ciT
which is at most [resp. at least] u = u(t,x).
From now on, fix any = in .Tu(f), FL = FL(t,x,u,Ø) in
FL(f) and let c = c(t), F = F(t,x) be as in Condition (i) relative to these Fu
Fu(t,x,u,o), FL = For (t.,x.) e r E [O.T], u E R, o E F, /i E
Q we set
D(t., x., r) {x(r) : x(.) E x.)},
def
r,ce) = {x(r) : r(.) E .Atj(t.,x,,u, o)},
DL(t,, r, {.r(r) : x(.) E AL(t., u,
and shall be concerned in the sequel with the following two functions:
(t,x) : yE D(t,x,T)},
124
(t,x) i-+ : yE D(t,x,T)}.
Lemma 12.7. = is a supersolution (relatwe to Pb) of the C'auchy

problem (12.9)-(12.10). Moreover, both the functions = i,1'(1)(t,x), =
are continuous in 11T with o(x).
Proof of Lemma 12.7. By the Remark in Part 2° of one can easily prove
the upper sernicontinuity of the multifunction ciT (t, x) D(t, x, T) C
On the other hand, the lower semicontinuity of this nonempty compact valued
multifunction follows from Filippov's theorem (see [8, Theorem 10.4.1] and [40,
Lemma 8.3]). Thus, by the maximum theorem we see that = =
are in C(ciT). All that is left to show is that = satisfies
(12.46) [with u replaced by To this end, let 0 < t < r < T, x E s E P and
x(.) E C Since D(r,x(r),T) C D(t,x,T), (12.48)
§12.5. UNIQUENESS AND EXISTENCE OF MINIMAX SOLUTIONS 181
implies < i.e., (12.46) with in place of u holds. Lemma

12.7 is proved. 0
Lemma 12.8. If u = u(t,x) is a locally bounded fimction on satisfying (12.18)
and the condition
sup inf [u(r,x(r)) — u(t,x)] < e (12.49)

aEP x(.)EXu (t,z,u(t,x)+e,a) —
for alle > 0, (t,x) E r E [t,T], then u = u(t,x) is a supersolution (relative

to Fu) of the Cauchy problem (12.9)-(12.1O).
Proof of Lemma 12.8. Let x e > > ... > e(h) 0. Then there
exists a sequence C converging to (t,r) so that u(T,x) =
urn u(t(k),x(k)). By (12.49), for any fixed a E P one can find functions
k-+-foo
E + e(k),a)
with
u(t Ic = 1,2,... (12.50)
def
Since = {(T, x), (t , x ),... , (t ,x ), . . . } is compact and since
we may (by the remark in Part 20 of §12.2) assume that x(k)(.) x(.) Xp(T, x)
in Letting Ic -+ +00 in (12.50), we get a(x) < u(T,x). Thus,
= u(t,x) satisfies (12.18) [with u replaced by u].
Fix now 0 < t < r ( T, x E R", a P. Then there exists a sequence
C [0,r) x
converging to (t,x) so that tC(t,x) = lim We choose >

> e(k) and find, by (12.49), functions x(k)(.) E +
6(k),a) such that
u(r, z(k)) + Ic = 1,2,... (12.51)
For every e > 0, since lim [u(t(k), x(k))+e(k)] =

k-4+oo
u(t, x), it follows from Condition
(ii) that x(k)(.) Xu(t(k),z(k),u(t,x) + e,a) when k is large enough. Therefore,
again by (ii) and the Remark in Part 20 of §12.2, we may assume that x(")(.) —*
x(.) E flXu(t,x,u(t,x) +e,a) = Xu(t,x,u(t,x),&). Letting k +00 in
(12.51), we get u(r,x(r)) <iC(t,x). Hence,
u in place of uJ; i.e., by definition, it is a supersolution (relative to F0) of the
Cauchy problem (12.9)-(12.10). The proof is then complete. 0
Now let us define a function A = A(t, r) on 11T by the following formula:
A(t,x) : u(.,.) Solu(Frj)}. (12.52)
Since Du(t, x, u, T, C D(t, x, T) for all (t, x) E 11r, u E R, P, it is easily seen

that u(t,x) for all supersolutions (relative to F0) u = u(t,x)
on
of the Cauchy problem (12.9)-(12.10). Therefore, A(t,x)
on IT [by Lemma 12.7, = is a supersolution (relative to F0) of
the Cauchy problem (12.9)-(12.10)1. This together with the continuity of =
= implies that
x) A(t, x) A(t, x) x) (12.53)
on Moreover, we have:
Lemma 12.9. A = A(t,x) is a lower semicontinuous supersolution (relative to F0)

of the Cauchy problem (12.9)-(12.10).
Proof of Lemma 12.9. By (12.52)-(12.53) we need only show that A = A(t,x)

is a supersolution (relative to Pb) of the Cauchy problem (12.9)-( 12.10). Moreover,
since A = A(t,x) is a lower semicontinuous function equal to = i(x) on
{t = T, x (cf. (12.53) and Lemma 12.7), all that is left to prove is that
(12.46) with u replaced by A holds.
0. Choose a sequence C [0,r) x converging to (t,x) so that

A(t,z) = lim A(t(k),x(k)). Then by (12.52), there exists a sequence of functions
= u(k)(t,x) in Solu(Fu) with
s(")) < + (12.54)
hence, by (12.46) and (12.52), there exist x(k)(.)

such that
A(r, < A(r, x(k)(r)) < < u(k)(t(k), k = 1,2,.
(12.55)
Since urn uUI)(tUo),z(")) = A(t,x), we may assume, in the same way as at the
b-÷+oo
end of the proof of Lemma 12.8, that the sequence converges to some
s(.) E flXü(t,x,)C(i,r) + e,Q) = Xu(t,x,A(t,x),Q). Letting k -9 +00 in
(12.54)-( 12.55), one gets
k-++oo
A(t,x);
i.e., = X(t,x) satisfies (12.46) [with in place of ul. Lemma 12.9 is proved.
0
For any GE [O,T] and E Q, let us now consider a function =
which is given on x R by:
: yE — (12.56)
It is then easy from (12.53) that
: y E D(t,x,O)} — t,b <

: yE D(i,x,9)}
(12.57)
and from Condition (ii) that pe,p = x, is strictly decreasing in Beside
the function = we shall also consider another function
x) defined on by:
E R : p9,fl(i,z,1j) O}. (12.58)
In virtue of (12.57)-( 12.58) and of the monotonicity in of = x, u'), one

must have:
: yE D(t,x,O)} < < : y E D(t,x,6)}.

(12.59)
Finally, let us investigate an auxiliary function = z) by the formula:
aer I if (t,x) E [0,9] x

(12.60)
=1 A(t,x) if (t,x) E [O,T] x
(Note that, by (12.56) and (12.58), ER : A(6,x) — 0)
x).)
The following lemma gives us an important property of = x).
184 12. M1NIMAX SOLUTIONS
Lemma 12.10. = is a supersolution (relative to of the Cauchy

problem (12.9)-(12.10).
Proof of Lemma 12.10. The nonempty compact valued multifunction 3

(t,x) D(t,x,9) C RTh is locally bounded. Therefore by (12.59), taking account
of the continuity of = = (Lemma 12.7), we see that
= and hence, = are also locally bounded. Moreover,
it follows from (12.53), (12.60), and Lemma 12.7 that o(x). Thus,
according to Lemma 12.8, we have only to prove (12.49) with in place of u.
only to consider the following three cases:

— Case 1: 9 <t. By (12.60) and Lemma 12.9, there exists
x(.) E C Xu(i,x,p,p(t,x) + e,a)
so that = A(t,x) = (12.49) [with u replaced by

is thus satisfied.
— Case 2: r 9. By (12.60), we have pe,p(t,x) = hence by (12.58) and
by the monotonicity in of = p,,p(t, x, we get pe,p(i, x, x) + e) <0.
Because of (12.44), there exists at least one
z(.) E +e,cx)flXL(t,x,/Ja,p(t,x)
Of course, DL(T, x(r), x)+E, 0, j3) C DL(i, x, pep(t, x)+e, 0, f3) and therefore,
by (12.56), it may be seen that
+ e) + e) < 0,
This together with (12.58) and (12.60) implies = e9,s(r,z(r))

x) + e; (12.49) [with in place of uJ is thus satisfied, too.
— Case 3: t < 9 < r. By Case 2, there exists x(')(.)
such that
+ e. (12.61)
By Case 1, there exists x(2)(.) such that
pe,p(r, x(1)(O)) (12.62)

S12.5. UNIQUENESS AND EXISTENCE OF MINIMAX SOLUTIONS
Choose now x(.) E C([0,T],R") with
(0,81
—
— (0,8)
and —
Yl (8,T( —
By (12.61)-(12.62) and Condition (ii), we then see that
x(.) Xu(i,x,,.te,,9(t,x)
and that i.e., (12.49) is also satisfied with u def

=
This completes the proof. 0
Lemma 12.11. = is a subsolutson (relative to FL) of the Cauchy
problem (12.9)-(12.10).
Proof of Lemma 12.11. By (12.53) and Lemma 12.7, A = A(t,x) is locally

bounded with A(T,x) a(s). Let 0 < t < B < T, x /3 E Q, e > 0.
We consider the auxiliary function = given in (12.60). Accord-
ing to Lemma 12.10, = is a supersolution (relative to of the
Cauchy problem (12.9)-(12.10); hence by (12.52) and (12.60), one has A(t,x)
= Thus, by (12.58) and the monotonicity in of
= one also has p(t,x,A(t,x) — e) > 0. So, it follows from (12.56)
that there exists x(.) E XL(t, s, A(t, x) — e, /3) such that A(O, x(O)) > A(t, x) — e.
In so doing, we are led to another property of A = A(t, x), that is,
inf sup [A(9,x(O)) — A(t,x)] —e

fiEQ Z(.)EXL
for all e > 0, (t,s) E 8 [t,T}. Therefore, by arguments dual to the ones in
the proof of Lemma 12.8, we conclude that A+ = A+(i, x) is a subsolution (relative
to FL) of the Cauchy problem (12.9)-(12.10). 0
The next lemma will play a crucial role in proving our main theorem. Adiatullina
and Subbotin's method of proof ([1] and 1124]), based rather directly on Gronwall's
inequality, seems to break down in the case of time-measurable Hamiltonians. Our
road to this result here is devious (by some "perturbation technique" on sets of
Lebesgue measure 0), and proceeds via an implicit version of Cronwall's inequality.
Lemma 12.12. Let 17 = ii(t, x) and = x) be super- and sub-solutions (relative

to Fü and FL) of the Cauchy problem (12.9)-(12.10), respectively. Then ii(t, x)
!L(t,x) on T.
Proof of Lemma 12.12. Assume the contrary, that
< (12.63)
for some E Let D B,. C be such that x(1) E D for all

x(.) E t [0,7.'); and A = A(t) where AD,E = AD,g(i) is
the function (in L'(o,T)) existing in Condition b) corresponding to the bounded
sets D and E def
= Im (1) , m'2' ) with
< m(i) = —
def def
nun u(t,x) <rn = max < +oo. (12.64)
(i,x)E(O,TIXD (,z)E(O,TIXD
There exists a set A of measure 0 such that (12.13) and (12.43) hold for all
U E R, p E R't, x R", t E AC. We may assume that AC C Leb(c) fl Leb(A), that
the function R u f(t, x, u, p) is decreasing for all p E xE t E AC, and
that (12.11) is true whenever u E, p S, x, y D, i AC.
For every e > 0, the set A can be contained in an open set U1 C (0, T) with
J8r(1 + r)c(t)dt <e. (12.65)
Let A(1)(t) 2A(t) + e and let = be the function given by
if t U1,
I18r(1+r)c(i)+e if tEll1.
(12.66)
For any t E [t(°), T) we consider the following (compact) set:
E x
.
—
J A(1)(r)dr),
g(O)
< < (12.67)
Of course, N(')(t(°)) 0. Setting sup{t [t(°),T) : 0}, we

see that N(')(t(')) 0 because Xp(t(°),x(°)) is compact and = i&(t,x) (resp.
II = ii(i,x)) is upper (resp. lower) semicontinuous.
Let us prove that = T. To this end, find and assume
again the contrary, that t(1) <T. We need only consider the following two cases:
- Case 1: i(') E Ue. Take S > 0 small enough so that + 5) C 14. By defini-
tion, for any fixed o P, /3 Q there exist E
with
+ + 5)) < + + 5))

(12.68)
where
< (12.69)
Let
2 def I if 0 t t(e), 2 deC I if 0< t <

x
ift(d) < t T, ift(e) < t < T.
(12.70)
Then E Xp(t(°),x(°)) xXp(t(°),x(°)). In addition, since
N(e)(t(e)), it follows from (12.67) and (12.70) that
+ 5) — + 5)12
= — + — —
< . exp(J +J
J
f \
( / / j .exp(/
\Ji(O)
J
.exp(J
J g(O)
(12.71)
(Here, the first inequality follows also from the Bunhiakovskii-Schwartz inequality
and the fact that Idx(t)/dtl < c(t)(1 + Ix(t)I) < (1 + r)c(t) a.e. in (0,T) for all
x(.) E Thus, since again E we see, by (12.67)-
(12.71), that 0 N(e)(t(e) + 5) a contradiction.
— Case 2: j(o) Eu:. Let us use (12.69) and — Since j(c) Leb(c),
def
one has in = sup c(t)dt < +oo, and hence
d
sup
1
—/ dt
dt sup
1
—/ c(t)(1+Ix(t)I)dt
öE(O,T—t())
z(.)EXF(t(°) ,z(°))
<(1+r)m.
(12.72)
Moreover, because E Leb(A), there exists 6(1) > 0 so that, for all 6 E (0,6(')],
we have
1
exp(_ 2A(t)dt) > —
6 J
exp(f A1(t)dt) > exp(26A(t(1))) 1 + 26 (12.73)
On the other hand, the compactness of Xp(t(°),z(°)) implies the existence of 6(2)
with the property that
Ix(t) — x(t(1))I < (12.74)

16(1 + r)(1 + in)
whenever z(.) E Xp(t(°),x(°)), It — < 6(2). Finally, since (12.43) holds at t

one can find a E P, /3 E Q, and > 0 so that
inf ! I mm > —e/8,

6 zEFu(t,z(t),ii(°),a) —
sup — / max + e/8

Jt()
(12.75)
as 0 <6
If we choose now 6 min{6(1),6(l),6(S),T — i(t)} > 0, then by definition,
there exist E E such that
(12.68) is fulfilled. Let E x Xp(i(°),x(°)) be as in

(12.70). It follows from (12.75) that
— E/8,
6
— J dt
(12.76)
f f(t +
By (12.76), setting p(t) t E [O,T], we have
dt
< f(t(e),x(O),u(O),p(O)) — f(t
6
(12.77)
Combining (12.72), (12.74), and (12.77) gives
6 f di
<I — + e/2,
and therefore,
+ 8)12 —
—
+ e/2
A(t(e)) +e/2 = . +e/2
by (12.11), (12.13), (12.64), (12.68), and by the monotonicity in u of the function

f = f(t,x,u,p). Thus, taking account of (12.66), (12.67),(12.73), and of the fact
that E N(t)(i(1)), we see that
+ 8)12 [1 + . + e6
. exp(J A1(i)dt) +
(j .exp(f
t(O)
A(M(t)dt)) •exp(f A'(t)dt)+
+1
<J A(o)(t)dt).
(12.78)
Again, by (12.67)-(12.70) and (12.78), we get 0 N(s)(t(e) + 6)
a contradiction.
The contradiction8 we have got in both Cases 1 and 2 show that = T, i.e.,
that for any e > 0 there exists at least one E It then
follows from (12.67) that
. exp(JT
— A(*)(t)dt). (12.79)
<Jt(o)
Since Xp(t(°), x(°)) is compact, one can find a sequence > > ... > —* 0
such that —+ in C([0, T), Letting k —+ +00 in

(12.79) with e(k) in place of e, we obtain — = 0 (cf. (12.65)-(12.66)).
Hence, a passage to the limit as lv -+ +00 in the inequality
< <
[notice that = ii(t, x) (resp. = x)) is lower (resp. upper) semicontinuous]

gives
< < < <
where x = This is a contradiction, which says that (12.63) must be

wrong. Lemma 12.12 is so completely proved. 0
We are now in a position to prove Theorem 12.6.
Proof of Theorem 12.6. Since the Hamiltonian I = f(i,x,u,p) satisfies a)-e),
Part 30 of Section §12.2 shows that Fu(f) 0. Let F0 = Fu(t,x,u,o)
and FL = FL(t,x,u,[3) be in F0(f) and .FL(f), respectively, and A = A(t,x)
be the function defined on IZT by (12.52). In virtue of Lemmas 12.9 and 12.11,
A = A(t,z) A(t,x) is in = is in SolL(FL). Therefore,
Lemma 12.12 together with (12.53) implies
A(t,x) A(t,x).
Thus A = A(t,x) belongs to SolU(FU)flSolL(FL). Moreover, if =

and = are also in .F0(f) and FL(f), respectively, and =
A(°)(t,x) belongs to fl then Lemma 12.12, applied to F0 =
F,j(t,x,u,cm) and = gives .X(t,x) .\(°)(i,x) on 11T• Another
application of Lemma 12.12 to = and FL = FL(t,x,u,3) gives
A(°)(t,x) A(t,x), hence A(°)(t,z) A(t,z). The proof is complete. 0
Remark. Condition e) in Part 10 of Section §12.2 (positive homogeneity of the
Hainiltonian I = f(t,x,u,p) with respect to p) can be omitted slightly by the
method proposed in [1], (124] as follows. Let
1
for
f(t,x,y,u,p,q) deC
—
= lim(q. f(i,x,u — y,p/q)) for q= 0.
q\0
Here, y R and q R are two supplementary variables, and the limit is assumed
to exist. It should be noted that the auxiliary Hamiltonian I = f(t,x,y,u,p,q) is
positively homogeneous with respect to the variable (p, q) We also assume
that the Haxniltonian f = f(t, x, u, p) satisfies conditions under which the positively
homogeneous auxiliary Hamiltonian 7 = 7(t, x, y, u,p, q) has the properties indicat-
ed in Part 10 of Section §12.2. Then it can be shown that the minimax solution of
the auxiliary Cauchy problem
011- in {O<t<T, x€R ,
on {t=T, yER}
§12.6. TEE CASE OP MONOTONE SYSTEMS 191
satisfiesthe identity u(t, x, y) u(t, x, 0) + y. The function u = u(t, x) def_

= u(t, r, 0)
can be considered as the minimax solution of(12.9)-(12.10) with non-homogeneous
Hainiltonian f = f(t,x,u,p). For this case, an existence and uniqueness theorem
(like Theorem 12.6) can also be established.
§12.6. The case of monotone systems
In this section, we extend the notion of minimax solutions to the case of systems of
first-order nonlinear partial differential equations with time-measurable Hamiltoni-
ans. We will only formulate the main results on the existence and uniqueness of
such solutions for those systems satisfying a certain monotonicity condition (under
which our systems are called monotone) and their relations with the (semi)classicai
solutions. The detailed proofs can be found in Thai Son, N.D., Liem, N.D., and
Van, T.D. [133]. Our method here is to combine that of the previous four sections
with some other techniques of monotonicity.
Inthesequel,misapositiveinteger. Givenu = (Ui,... ,um), v = (Vj,...,Vm) E
1,...,m. Moreover,ifu
u3 = v3 for some index 1 <j <rn, then we write u v. An Rm-valued function
= = iscalledlower(resp. upper) semicontinuousif
its components = x) (1 <i <m) are all lower (resp. upper) semicontinuous.
Similarly, it is differentiable if its components are all differentiable.
10 Formulation of the Cauchy problem

Let us consider the Cauchy problem for a weakly-coupled system of the form
=0 in Qa, (i = 1,...,m), (12.80)
u(T,x) = o(x) on {t = T, x E (12.81)
Assume that the terminal data = cr(x) = .. ,Cm(x)) is a given vector

function continuous on R", and that the Hamiltonians f1 = x, u, p) depend on
(t, x, u, p) E x x with the following properties.
(a) 's Conditions:

(al) For almost all (in the sense of Lebesgue measure) t E (0, T), the functions
x x (x,u,p) i.-* f.(t,x,u,p) are continuous (1 i m).
(a2) For any (x,u,p) E x x S, the functsons (O,T) i f1(t,x,u,p)

are measurable (1 <i <m).
(b) For any lounded sets D C RTh and E C Rm, there exists a function AD,E =
AD,E(t) in L'(O,T) with
x, u, p) — f1(t, x', is, AD,E(t) Ix — x'I
for almost all t E (0, T) and for all x, x' D, u E, p E 5, i E { 1 m}.
(c) There exists a function £ = £(t) in L'(O,T) such that
sup{jf;(t,x,u,p) —f1(t,x,u,q)i —e(t) .(1 + Ip—qI : p,q E B) <0

for almost alit E (0,T) and for all(x,u) x Rm, tE {1,.. . ,m}.
(d) Monotonicity Condition (cf [72, §4, (A.1)&(A.3)]): For almost alit (0,T)
andforallj E {1 rn}, (x,p) E u,v E Rm, if v,—u, = max(v,—u1)
0, then f,(t,x,u,p) f,(t,x,v,p). — —
(e) The functions p f,(t,x,u,p) are positively homogeneous; i.e.,

f1(t,x,u,s .p) = s f1(t,x,u,p) Vs 0
for almost allt (O,T) and for all (x,u,p) E x x S (i = 1,.. .,m).
Note that the monotonicity condition here implies:
(dl) Quasi-monotonicity Condition (ci. [72, Definition 2.2, Lemma 4.81): For
almost all t (O,T) and for allj E (1,... ,m}, (z,p) x S, u,v E if
u v, then f1(t,x,u,p) f1(t,z,v,p).
(d2) The functions R 3 u1 '.4 . . . . ,u,,.,,p) are de-
creasing for almost all t E (0, T) and for all (u1,. . . urn)
20 Differential inclusions for supersolutions and subsolutions

First, for (t,x) uE a,/3 ES, 1 i m, let
F(t, x) . (1 + lxi) . B,
{z E F(t,x) : (z,a) f5(t,x,u,ci)}, (12.82)
{z F(t,x) : f1(t,x,u,$)).
§12.6. THE CASE OF MONOTONE SYSTEMS 193
Further, for any Rm-valued function = = . . defined

on Qp, we set
z, u1, a) (Fü)s(t, x, x), a),
(12 83)
x, /3) (FL)1(t, x, x), /3),
where E R1 and
x) (t, x), . . . , (t, x), (t, x), . , Prn(t, r)) E Rm. (12.84)
Then it can be seen that the multifunctions (1 <i <m)
3 (t,x) i—* F(i,x) C

x x S 9 (t,r,u,c,) C
x Rm x S (t,x,u,fJ) (FL)1(t,x,u,/3) C
are nonempty convex compact valued, measurable in t and continuous in (x, u). On
the other hand, it follows from (dl)-(d2) that
(Fu),(t,x,u,a) C if u > v,
if u v,
x, u, a) C x, u1, . . . , + e, . . . a) if 0,
X, IL, i3) C Z, U1,..., — 6, . . /3) if 0.
Therefore, by [133, Proposition 4.1], if = 92(t,x) is lower (resp. upper) semicon-

tinuous, one can check the measurability in t and the upper semicontinuity in x
of the multifunctions (Fu)T(.,.,u1,a) (resp. (FL)T(.,.,u$,/3)) for 1 < m, u1 E
R1, E S. And hence, according to the Remark 2°, the sets
def
(Xe), (t.,x.,u1,a) =
(12 85)
def
(resp. (XL)1 (t.,z.,u,,/3) =
are all nonempty and compact for (t., x.) So, we can make the following:
Definition 7. A supersolution of Problem (12.80)-( 12.81) is a lower semicontinuous

function u = u(t,x) = (u1(t,x),. . .,um(t,X)) on which satisfies the
condition
max sup mm [u1(r,r(r)) — u1(t,x)J < 0

1<i<rn oES
forallO<t<r<T,
u(T,x) a(s) Vs E (12.86)
Definition 8. A subsolution of Problem (12.80)-(12.81) is an upper semicontinuous

Rm-valued function u = u(t,x) = (uj(t,x),. ..,Um(t,x)) on which satisfies the
condition
mm mi
i<s<m$ES
max Ius(r,x(r)) — 0
forall0<t<r<T,
u(T,x)c(x) (12.87)
The sets of all supersolutions and subsolutions of ( 12.80)-( 12.81) will be denoted
by Solu and SoiL, respectively.
Definition 9. An Rm-vaiued function u = u(t,x) in SOIIJflSO1L is called a mintmax

solution of the Cauchy problem (12.80)-( 12.81).
3° Invariance of definitions
Just as the invariance of definitions that has already been discussed in §12.4 for
the case of single equations, instead of (12.82), we can now use any tuple Fu =
[resp. FL =((FL)1,...,(FL)m)] in the following common fam-
ilies of tuples of multifunctions.
Let P and Q be arbitrary nonempty sets. We consider any muitifunctions
x xP (t,x,u,a) '— C
x Rm x Q (t,x,u,8) (FL),(t,z,u,$) C
satisfying the general conditions below (1 i m).
i) The multifiinctions = and = (FL),(t,x,u,3) are

nonempty convex compact valued, measurable in t and upper semicontinuous in
(x,u). Moreover,
(Fu)8(t, s,u, a) U (FL ),(t, x, u, 0) C F(t, x)

§ 12.6. THE CASE OF MONOTONE SYSTEMS 195
for almost alit E (O,T) and for aU(x,u) E R' x Rm, where c = c(t) in L'(O,T) is
independent of i,a,
ii) For almost alit €(O,T) andforollx ER', a E P, /3 E Q, u =(ui,...,um),v =
(vi,...,vm) E one has
(Fu)1(t,x,tL,a) C if > v,
J if u > v,
x, u, a) C x, Ui,.. . , U1 + . . ,Urn, a) zf e > 0,

(FL)1(t, x, u,/9) C (FL)1(t,x, u1, . . . , — . , u,,,/3) if > 0.
iii) For almost all t E (O,T) and for all (z,u,p) e R.' x Rm x the following
inequalities hold:
sup mm (z,p)<f1(t,z,u,p)< mi max (z,p),

— —
1 d
sup liminf mi — I p'dr > f,(t,x,u p)
aEP J\O \ di I —
1 \
mf limsup sup —
ç—(r),p1dr <
$EQ di
where
x, u, a)def
= X(F,J ), x),
clef
(XL)$(i, x, u, /9) = X(pL ).(.,.,',fl)(t, x).
The family of all tuples (rasp. FL =((FL)1,...,(FL)m))

satisfying Conditions i)-ill) will be denoted by .Tu(f) (resp. FL(f)). If the Harnil-
tonians = (1 <i <m) satisfy (a)-(e), then .Fu(f) 0, 0.
In fact, under such hypotheses, we can choose any F, Q C with the property
that
{s.cs : aEP,s0}={s.13 :
and then use (12.82) to define a concrete pair of tuples = ((Fu)j,. . .
and FL = ((FL)j,... ,(FL)m). This pair would satisfy i)-iii).

From now on, let us fix any (Fu, FL) E Fu(f) x .FL(f) and continue using
the notations (12.83)-(12.85). It is [133, Proposition 4.1] (see aiso the Remark of
§12.2-Part 2°) that makes the following definitions allowable.
Definition 10. The set Solu(Fu) of supersolutions (relative to Fu) of the Cauchy
problem (12.80)-(12.81) is defined to be the set of all lower semicontinuous Rm-
valued functions u = u(t,x) = (ui(t,x),.. . on which satisfy (12.86)
and the condition
max sup mm — u(t,x)] <0

1<i(m aEP —
for
Definition 11. The set SOIL (FL) of subsolut:ons (relatzve to FL) of the Cauchy
problem (12.80)-(12.81) is defined to be the set of all upper semicontinuous Rm-
valued functions u = u(t,x) = (ui(t,x),. . . on 11T which satisfy (12.87)
and the condition
mm inf max [u1(r,x(r)) — u1(t,x)] > 0

I PEQ —
for any0<t<r_<
It can be proved that, for any FL) E .Tu(f) x the intersection
Solu(FrJ)flSolL(FL) consists of a unique (Rm-valued) function, which is independent
of FL). Therefore, in complete concord with the previous definitions, we may
also give the following.
Definition 12. An Rm-valued function u = u(t, x) in SolU(FU)flSolL(FL) is called

a mlnimaz solution (relative toFu and FL) of the Cauchy problem (12.80)-(12.81).
4° Main results
In this part, we formulate the main results on the existence and uniqueness of
minimax solutions for monotone systems of first-order nonlinear partial differen-
tial equations with time-measurable Hamiltonians and their relations with the (se-
mi)classical solutions (see Definition 2 in Chapter 11).
Theorem 12.13. Assume (a)-(e). Then

(i) every global semiclassical solution u = u(t,x) of (12.80)-(12.81) so that the
gradient mappings x i-+ 0u1(t,x)/Ox (1 <i <in) are continuous for almost
all t (0, T) must also be a minimaz solution of the same problem,
§12.6. THE CASE OF MONOTONE SYSTEMS 197
(ii) there exists a null set A C (0, T) such that, at each point (t, x) E ((0, T)\A) x R'
where a certain minimax solution of(12.80)-(12.81) is differentiable, the equations
(12.80) must be satisfied.
Theorem 12.14. Let the initial data a = a(x) be continuous on and the
Hamiltonians = f,(t,x,u,p), 1 < j < m, satisfy all Conditions (a)-(e). Then
there exists one unique minimax solution of the Cauchy problem (12.80)-( 12.81).
Example. Now we present an example of monotone systems (12.80) of first-order

nonlinear partial differential equations with time-measurable Hamiltonians. (The
case of single equations is obtained when rn = 1.) It is left to the reader to check
all the hypotheses (a)-(e) for those systems.
Let = G1(t,x,A) be of class C1 on x R such that R? .\
is increasing for all t E (0,T), x E (1 < i < m). Further, let h1 = he(s) be a
bounded, decreasing, and globally Lipschitz continuous function of s E R for any
i {1,. . . ,m} (we can take, for instance, he(s) arctan(s) with a certain
negative constant for each index i).
Define
(1 + 3x1) . +
for andi€{1,...,rn}. is in L' (0,T) and

= x) are given C'-functions satisfying on the conditions
0, and, c13(t,x) <0 if i j.

Chapter 13
Mishmash
In this final chapter we present various things, more or less "for completeness."
§13.1. Hopf's formulas and construction of global solutions

via characteristics
First, consider the Cauchy problem
Ou/*3t+f(Ou/Ox)=O in (13.1)
u(O,x) = on {t = 0, x E (13.2)
under the following two hypotheses.
(P.1) The initial function = is of Class C° and the Hamiltonian I = 1(p) is

strictly convex on with lim = +00.
IpI—'+oo
(F.II) For every bounded subset V of D, there exists a positive number N(V) so
that
+ t f((x — w)/t)} + t . f((x — y)/t)
whenever (t,x) V, > N(V). Here, f = f(z) denotes the Fenchel conjugate
function of f = 1(p).
By Theorem 9.12, a global solution of (13.1)-(13.2) exists (and is found by

Hopf's formula (9.29)) while the classical theory furnishes only its restriction to a
neighborhood of {t = O} in which it is smooth if the Hainiltonian and initial data
are supposed to be sufficiently smooth. In fact, as we have seen in Chapter 1, if
f = f(p) and = are of class C2 on Cauchy's classical theory furnishes
the solution by fitting characteristics of (13.1) to the initial data = These
characteristics form a hypersurface x = x(t, y), v = v(t, y) in txv-space above lx-
space such that x(0, y) = y, v(0, y) = 4'(y). For sufficiently small t > 0, this
§13.1. CONSTRUCTION OF GLOBAL SOLUTIONS VIA CHARACTERISTICS 199
hypersurface has a simple projection upon tx-space (the tx-projections x = x(t, y)

of the characteristics issuing from different points y at t = 0 do not intersect). Then
one can eliminate y from
x = x(t,y),
(13.3)
I. v=v(t,y),
and the hypersurface appears in the form v = u(t, x) with u = u(t, x) single-valued
and smooth. This is the classical solution. Outside of the classical tx-domain,
however, the hypersurface has, in general, multiple projection upon tx-space or, in
other words, v = u(t, x) becomes multivalued. This multi-valued function exists
for all t > 0. It is natural to ask [64] whether our single-valued global solution
of (13.1)-(13.2) can be determined directly from the several values of v in (13.3)
above the same point (t,x). In the first case of Hopf's formulas, I = 1(p) convex
(or concave), = largely arbitrary, the answer can be foreseen by (4.11) and
Lemma 4.3 as follows: The value at (t, x) of the global solution is either always
the smallest or always the largest of the values of the many-valued v = u(t,x),
depending on the convexity character off = 1(p).
For definiteness, let us suppose I = 1(p) is convex. Recall that the characteris-
tic equations for the Cauchy problem (13.1)-(13.2) can be written by (1.10)-(1.11).
Therefore, in this particular case (the case of Hamilton-Jacobi equations), the char-
acteristic strips are defined as
x = x(t,y) = y+i
v = v(t,y) = +t. — (13.4)
I. p=p(t,y)=4'(y).
df , dq5
Here and subsequently, we write x, y, p, f = — and 4) = — as horizontal vectors
dp dx
in and (.,.) denotes the (Euclidean) inner product in As was mentioned
in the above, in general, we could not completely eliminate y from (13.4) to get
a single-valued function v = u(t,x) (i.e., in this form, u = u(t,x) might become
multi-valued). Anyway, if we try to do so, we may rewrite v = v(t,y) in (13.4) as
v = ((t,x(t,y),y)
where
((t,x,y) +t. f((x — y)/t). (13.5)
200 13. MISHMASH
Then the hypersurface (13.4) projects upon the entire half {t 0) of tx-space, and
the global solution (9.29) of (13.1)-(13.2) equals, at each point (t, x) of {t >0), the
smallest v-coordinate of the hypersurface in the following sense.
Theorem 13.1. Let = be of class C' in R", and let I = f(p) be str*ctly
convex and of class C1 in with lim = +00. Assume (F.lI). Then the
value u(t(°), x(°)) at each (t(°), x(°)), > 0, of the global solution u = u(t, x) given
by Hopf's formula (9.29) of the Cauchy problem (13.1)-(13.2) can be determined as
the smallest values of v(t(°),y) = ((t(°) ,x(t(°),y),y); the minimum being taken over
all y such that the characteristic curves {(t,x(t,y)) : t 0} starting from (O,y)
meet each other at (t(°),x(°)).
Proof. By hypotheses, f = I (z) is the same as the Legendre conjugate of
I= .f(p) (see Lemma 9.13 and [117, Theorem 26.61); moreover, both 1' and f' are
one-to-one mappings from onto itself with I = (f') Therefore, it follows
from (13.5) that ( = ((t,x,y) is differentiable in y with
= — (13.6)
It has been shown in the proof of Theorem 9.12 that (9.29) determines a global
solution u = u(t, x) of (13.1)-(13.2), and that the infimum
= inf (13.7)
pER
is attained at some E Clearly, must be a stationary point of the

function ((t(°>,x(°), .); i.e.,
= 0.
Thus (13.6) implies 4,'(y(O)) = f'((x(°) — y(°))/t(°)). Since f' = (f')1, we have
= or, = + =
(
for every such stationary point So, the characteristic curve {(t,x(t,y(°)))
0}, which is starting from must pass through (t(°),x(°)). Because the
minimum in (13.7) is not affected if it is taken not for all y but only for those
stationary y, the proof is then complete. 0
We now turn to second case of Hopf's formulas, qS = convex (or concave),

I = f(t, p) largely arbitrary, and continue investigating the construction of global
solutions via characteristics. Consider the Cauchy problem
.3u/Ot+f(t,ôu/Ox) = 0 in V {t>0, XE R"}, (13.8)
u(0,x) = on {t = 0, x (13.9)
as in Chapter 9, under the following standing hypotheses:
(E.I) The Hamiltonian I = f(t,p) is continuous in {(t,p) t E (O,+oo)\G, p E

:
R of Lebesgue measure 0. Moreover, to each N E

(0, +oo) there corresponds a function gjq = gri(t) in such that
sup If(t,p)I gN(t) for almost all t E (0, +oo).

IplN
(E.II) For every bounded subset V of D, there ezzsts a positive number N(V) so
that
(p,x) — — ff()d — — ff(r,q)dr}

whenever (t,x) V, I;'I > N(V).
It was shown in the proof of Theorem 9.1 that, if = is a finite convex

function on a global solution of (13.8)-(13.9) exists and is found by Hopf's
formula (9.7). We need to know how this global solution can be constructed by
means of characteristics. For this, suppose the Hamiltonian and initial data are
of class C2. The characteristic strips for (13.8)-( 13.9) in this case are defined (cf.
(1.10)-(1.11)) as
x =x(t,y) =v+f
(13.10)
v = v(t,y) = + — f(r,cb'(y)))dr,
p = p(t,y) =
In general, we can not completely eliminate y from (13.10) to get a single-valued
function v = u(t,x). Anyway, if we try to do so, we may rewrite v = v(t,y) in
(13.10) as
v= + (4'(y),x(t,y) — y) j f(r,qS'(y))dr.
—
202 13. MISHMASH
From now on, let us assume = to be convex on Using [117, Theorems

23.5 and 25.11, we can then further rewrite (13.11) as
v=
where
x,p) (p.x) — ç&(p)

— f f(r,p)dr for (t,x,p) E x (13.12)
It will be proved in Theorem 13.3 that the hypersurface (13.10) projects upon the
entire half {t 0) of tx-space, and the global solution (9.7) of (13.8)-(13.9) equals,
at each point (t, x) of {t 0}, the largest v-coordinate of the hypersurface.
First, let
u(t,x) dcf
= sup (y)), (13.13)
u(t,x) def
= sup (13.14)
pEdom #
for (t,x) We have:
Proposition 13.2. ü(t,x) u(t,x) on V.

Proof. Let E : yE It follows from Corollary 26.4.1 in [117] that
ri(domq5) c E c (13.15)
where ri (dom q5') is the relative interior of dom Therefore, (13. 13)-( 13.14) yield
ü(i,x) < u(t,x). We now prove u(t,x) ü(t,x). Fix any e > 0,
and choose q E dom such that
u(t, x) x, q) + €.
On setting p(m) (1 — 1/rn)q + we see, by (13.15) and [117, Theorem

6.11, that C E for in = 1,2,3,... Since the Fenchel conjugate
= is a lower semicontinuous proper convex function, [117, Corollary 7.5.1)
shows that = lim Hence, it follows from (13.12) that
u(i,x) +e
= lim co(t,x,p(m))+e
<ropp(t,x,p)+e
pEE
= ü(i,x)+ e.
Because e > 0 is arbitrarily chosen, the proof is then complete.
Remark. Let 4) = 4)(x) be convex and of class C' on Assume (E.I)-(E.II). It

has been proved in Chapter 9 that (9.7) then coincides with (13.14), and that the
supremum is always attained (at some point p E dom4)). The situation becomes
different for the supremum in (13.13). We show this by the following example.
Let n def
= 1
(1, p>l,
def I
1(P) = p—plnp, pE[O,1],
ISO, p<O,
and
def I eZ, x<0,
cb(r) =
Lx+1, x>0.
Then
I plnp—p, pE[O,11,
+co,
and
ü(t,x) = sup{x — 4)(e') — tf(e")}.
It is easily seen that the value ü(1, —2) = sup = 0 can not be attained in
(13.13) at any point yE R1.
In the remainder of this section, we make the following assumption:
(E.V) For each (t,z) 1.), the supremum in (13.13) is attained at some point
yE
Remark. Let 4) = 4)(x) be convex and of class C' on R". Assume (E.I)-(E.II).
Then (E.V) is fulfilled, for example, if is an open set (cf. (13.15) and the
Remark above). In particular, this appears when 4) = 4)(x) is a co-finite strictly
convex function, i.e. when it is strictly convex with lim [4)(.Xx)/AJ = +co for all
A-+-4-oo
See [117, Theorems 26.5-26.6].
Theorem 13.3. Let I = f(t,p) and 4) = 4)(x) be of class C2 on {t 0, p

R respectsvely, such that 4) =
, s always posstwe-
definite. Assume (E.II) and (E.V). Then the value u(t(°),x(°)) at each (t(°),x(°)) €
13. MISHMASH
V of the global solution u = u(t,x) given by Hopf's for,nula (9.7) of the Cauchy
problem (13.8)-(13.9) can be determined as the largest values of
i,.) = y), 4,'(y));
the maximum being taken over ally such that the characteristic curves {(t, x(t, y))
0} starting from (O,y) meet each other at (t(°), x(°)).
Proof. Since 4" is positive-definite, 4' = 4'(x) is convex on RTh. Notice that
(El) is satisfied trivially. From Proposition 13.2 and its Remark, we also see that
u(t,x) ü(t,x) on Moreover, it has been known (cf. (13.11)-(13.12)) that
= 4'(y) ÷ (4,'(y),x — y)
— f f(r,4"(y))dr;
hence,
—cp(t,x,4"(y)) = (x — y
— f (13.16)
In view of (E.V), the supremum
= = sup 4,'(y)) (13.17)

pER'
is attained at some y(°) E R't, which must be a stationary point of 4,'(.)).

We then have
P.. (t(O) 4,'( Y )) —0
Because is positive-definite, it follows from (13.16) that
—
— j = 0;
therefore, (13.10) implies x(t(°),yt0)) = Notice that the maximum in (13.17)

is not affected if it is taken not for all y but only for those stationary y. This
completes the proof. 0
Remark. To our knowledge, Hopf [64] was one of the first to ask whether the
global solution (9.29) of (13.1)-(13.2) and the solution (9.7) of (13.8)-(13.9) could
be determined by means of characteristics, i.e., whether they could be determined
directly from the several values of v in (13.4) and (13.10), respectively. The answer
§13.2. SMOOTHNESS OF GLOBAL SOLUTIONS 20$
(essentially Theorem 13.1) was in the former case considered by Hopf [64] himself
while he had no answer in the latter.
§13.2. Smoothness of global solutions
As we have seen (for example, by Theorems 2.1-2.2, Proposition 4.1 ... ), in general,
classical solutions to the Cauchy problem for nonlinear partial differential equations
exist only locally in time, and global (generalized) solutions contain singularities.
In the definition of most kinds of generalized solutions, such as weak, minimax,
or viscosity solutions, the regularity condition is relaxed significantly. (A mimmax
or viscosity solution needs only be continuous.) To get further information about
the solution, it is therefore very important to investigate its smoothness. Theorems
2.6-2.7 and 3.1 can be considered, to some degree, as investigations in this direction.
In particular, the theme of constructing the singularities of generalized solutions,
or of weak solutions, is thoroughly studied in Chapters 4-7.
In this section, we investigate the smoothness of the global solutions that are
given by Hopf's formulas. Consider the Cauchy problem (13.8)-(13.9), =
being a convex function on Let (E.I)-(E.II) hold. It has been shown in Chapter
9 that (9.7), or equivalently (13.14), yields a global solution u = u(t, x) of this
problem. Recall that the definition
L(i,x) def
= {pE R : = u(t,x)} C
determines a nonempty-valued upper semicontinuous multifunction L = L(t, x) of

(t,x) E V. Here, = is defined in (13.12). We now prove:
Theorem 13.4. Let 4 = be a convez function on Assume (E.I)-(E.II).

Then the global solution u = u(t, x) given by Hopf's formula (9.7) of the Cauchy
problem (13.8)-(13.9) is continuously differentiable in an open set U C D0
{(t,x) E V : t G} if and only if L(t,x) is a singleton for all (t,x) E U;
the set C being as in (E.I).
Proof. By the proof of Theorem 9.1, if u = u(t,x) is differentiable at (t(°),x(°)) E
D0, then (9.14)-(9.15) give (for 1 i n)
= :pE = min{p1 : pE
206 13. MISHMASH
So, is precisely the singleton {ôu(t(°),x(°))/Ox}. (Consequently, (9.13)

together with (9.15) implies that (13.8) must be satisfied at (t(°),x(°)).)
Conversely, suppose now that L(t, x) is a singleton for all (i, x) in an open set
U C V0. Being single-valued, the mapping U (t, x) —+ L(t, x) is continuous.
Hence, according to the maximum theorem [8, Theorem 1.4.161, (9.13)-(9.14) show
that the derivatives Ou/Ot, Ou/Oxi,..., Ou/Ox,, exist and are continuous in U (see
also [128, Corollary 2.21); i.e., the global solution u = u(t, x) is continuously differ-
entiable in U. 0
Remark. A special case of Theorem 13.4, when n equals 1, 4, = 4,(x) is globally
Lipschitz continuous, was considered by Kruzhkov, S.N. and Petrosyan, N.S. [94,
Lemmas 1 and 3]. There, instead of the multi-valued function L = L(t, x), they
used the single-valuedp =p(t,x)diinf sup L(t,x),
and established some interesting properties of these single-valued functions. In
particular, we have:
Proposition 13.5. [94, Lemma 51 For n = suppose 4, = 4,(x) is convex and

1,
globally Lipschitz continuous, I = f(p) ss of class C1 on R. Let (t(°), x(°)) E

(O,oo) x R, E L(t(°),x(°)), C {(t,x) x = + (t — t(°))f'(p(°)), 0 i
:
t(°)}. Then L(t,x) for all (t,z) C. Further, if L(t(°),x(°)) is a singleton,

so is L(t,x) for (t,x) E C.
Corollary 13.6. Under the hypotheses of Proposition 13.5, suppose the global
soluhon u = u(t,x) def *
= max{px —4, (p) — tf(p)} of(13.8)-(13.9) ss at
pER
(t(°),x(°)). Then it is also differentiable at (t,x) E C for all t E (0,t(°)).
Proof. Since u = u(t,x) is differentiable at (t(°),x(°)),the proof of Theorem
13.4 shows that L(t(°),x(°)) is a singleton. By Proposition 13.5, L(t,x) is also a
singleton for all (t, x) C. Therefore, its Remark,

by the proof of Theorem 9.1 and
the partial derivatives of the solution exist at (t,x) E (O,t(°)). Notice

C for all t
that u = u(t,x) = max{px — qS(p) — tf(p)} is convex, because it is the maximum

pER
of a family of affine functions. Hence, according to [117, Theorem 25.21, it is
differentiable at the above-mentioned points (t, x). 0
ExampLe 1. Investigate the smoothness of the global solutio

the problem
On On
in {t>0,xER},
S13.2. SMOOTHNESS OF GLOBAL SOLUTIONS 207
u(0,x)=x2/2 on xER},
where
(0, p0,
def I
s
p, O<p<l,
1, p>1.
Note that 4) = x2/2 is convex, but not globally Lipschitz continuous; and we can
not use Hopf's formula (9.3) for n = 1. In this case, however, (E.I) and (E.II)"
hold trivially. (See Remark 2 following the formulation of Theorem 9.1.) Hence, by
(9.7), a global solution of this problem is
u = u(i,x) max{px —p2 /2—tI

pER
More clearly, we have
u(t,x) = max {max{px — p2/2}, max{px — p2/2 — tp}, rnax{px — p2/2 — t}},
or 2
if (t,x)ED1,
0 if (t,x)EV2,
u(t, x) = (x — t)2
(t,x)ED3,
2
if (t,x)ED4.
Here,
zER}\(D1UV2UV3).
This solution is continuously differentiable in {t > 0, x E R}\(C), where the

"singularity" curve (C) is given by
5 x2/2, x 2,
(C) .
—
12(x—1)
This fact can be foreseen by noticing that L(t,x) = {0,x} if t = x2/2, x 2,
and L(t,x) = {x — t,x} if t = 2(x — 1), 1 < x < 2, while L(t,x) is a singleton
208 13. MISHMASH
if (t,x) (C). (Direct computation shows that L(t,x) = {x} for (t,x) V1,
L(t,x) = {0} for (t,x) V2\(C), L(t,x) = {z — i} for (t,x) E D3\(C), and
L(i, x) = {x} for (t, x) E V4\(C).)
Example 2. We consider the Cauchy problem
=0 in {t>0,x€R},
on {t=0, x€R}.
A global solution of this problem is:
u = u(i,x) — +t(1 +p2)112}.

pER 2
By computing and relying on Theorem 13.4 we recognize that u = u(t,.x) is

continuously differentiable in {t> 0, z R)\{(t,0) t 1). Using the method of
:
characteristics, we see that when i > 1, characteristic curves intersect. Concretely,

the two curves {(t,x(t, 1)) t 0} and {(t,x(t,2))
: t O}, where :
ty
x(i,y)=y—
+ y2
starting from (0,1) and (0,2), respectively, meet each other at the point
Nevertheless, the differentiability of the solution is not broken down in some neigh-
borhood of this point.
§13.3. Relationship between minimax and viscosity solutions
As we have mentioned, since the early 1980s, the concept of viscosity solutions in-
troduced by Crandall and Lions has been used in a large portion of research in a
nonclassical theory of first-order nonlinear partial differential equations as well as in
other types of partial differential equations. The primary virtues of this theory are
that it allows merely nonsmooth functions to be solutions of nonlinear partial differ-
ential equations, that it provides very general existence and uniqueness theorems,
and that it yields precise formulations of general boundary conditions. (See 151,
§13.3. RELATIONSHIP BETWEEN MINIMAX ANT) VISCOSITY SOLUTIONS 209
I10]-[20], [281, 135]-[39], [47]-[50], [67]-[72], [79], [991-1101], [122J-[123), [131], and the
references therein.) These contributions make great progress in nonlinear partial
differential equations, where the global existence, uniqueness, and well-posedness
of generalized solutions have been established almost completely.
The concept of viscosity solutions is motivated by the classical maximum prin-
ciple which distinguishes it from other definitions of generalized solutions, and it is
based on replacing the equations by pairs of differential inequalities. In the research
of both minimax and viscosity solutions, much attention is given to the construction
of subgradients, subdifferentials, generalized derivatives, and so on, all of which are
used in work with nonsmooth functions. For equations with continuous Hamiltoni-
ans, Subbotin and his coworkers ([1], [124]) have shown that minimax solutions can
be defined by the use of inequalities for directional derivatives that are equivalent
in essence to inequalities for subgradients and supergradients used in the defini-
tion of viscosity solutions. We now examine the relationship between minimax and
viscosity solutions in the case of equations with time-measurable Hamiltonians. In
our approach, the definition of minimax solutions needs not be via inequalities for
directional derivatives. It is enough to use only those given in Chapter 12.
Let = be a finite-valued function of near a point E Rm and let
e 6 Rm, -y 6 Rm. Recall that the upper and lower Dini semiderivatives of =
at the point e are defined as:
—
inf sup +
e>O O<J(e
(13.18)
—
inf +
e>O O<6(e
We introduce the following notation:
—
inf sup — —
(13.19)
tf inf —
— — —
{i' Rm : 0},
(13 20)
Dt/'(e(°)) {-y E Rm : 0}.
The mapping Rm e '—' 6 [—co, oc] is of course positive-homogeneous,
i.e., = A (A >0). Fore = 0, the relation O(°)) {0,oo}
210 13. MISHMASH
holds. It also follows directly from (13.18) that this mapping is upper semicon-
tinuous. Analogous properties are enjoyed by the mapping Rm e '—p
it is lower semicontinuous and satisfies the relations E {O, —oo} and

= A . (A > 0).
The sets are called the superdifferential and the sub-
and
differential, respectively, of= at the point The elements of these sets
are called supergradients and subgradients, respectively. These sets are also called
the upper and lower Dini semidifferentials. If = is differentiable at then
= = in this case,
—
= + — + o(e
where — — -4 0 as —+ According to (13.19)-(13.20),
— —
+ + o(e
for any supergradient -y E A subgradient y E also satisfies a

one-sided but opposite estimate:
— —
+ + o(e
We can verify that
= {y E Rm : (y,e) Ve
1321 )
={ Rm : (,e) Ve Rm}.
In fact, let -y e E Rm\{0}. Consider the function =

+ for 0 and E Rm. Of course, — =
as ej eI/2. Consequently, (13.18)-(13.20) imply
0 inf sup — —
—
—
mi sup +
—
—
e>0 O<S(e
Ii—eI<e
= [mi sup +
—
—
3
i—el<e
= —
§13.3. RELATIONSHIP BETWEEN MINIMAX ANT) VISCOSITY SOLUTIONS 211
from which we get Ye E Rm\{O}. This still holds even fore = 0,

since otherwise, 0, one would have
inf sup + — = = oc;

S>OO<6<e
hence,
inf sup + — =
e>O
It follows that
0 inf sup +
—
—
e>O O<ö(e
inf
e>O
sup + — — h1
= 00 — il = 00,
a contradiction.
In the converse direction, suppose (y,e) Ye E Rm. Using (13.19),
we can choose a sequence C Rm, 0 — -4 0 as k co, such that
urn —
— —
= (13.22)
—
k-+oo
We may assume without loss of generality that e(k) (e(k) —

e Rm (lel = 1) as k —+ oo. Since = + k)e(k), with
— -4 0, (13.22) together with (13.18) implies
= lim + — —
k-+oo
= lim + e)
k—+oo
— (i',e) 0,
i.e., -y E by (13.20). So the first equality in (13.21) has been completely

proved. The proof of the second is similar.
We are now in a position to give the definition of viscosity solutions to the
Cauchy problem of the form
=0 in (13.23)
212 13. MISHMASH
u(T,z)=c(x) on {t=T, xER'1}. (13.24)
Here, 0 < T < oc. Assume that the terminal data = u(x) is of class C° on
and that the Hamiltonian f = f(t,x,u,p) is measurable in t E (0,T) and
continuous in (x,u,p) E Rx In accordance with [35] and [39], we propose
the following:
Definition. A viscosity solution of (13.23)-(13.24) is defined to be a function

u = u(t,x) continuous on QT satisfying the terminal condition (13.24) and the pair
of inequalities
a + f(t,x,u(t,x),b)0 V(a,b) (13.25)

a +f(t,x,u(t,x),b) <0 V(a,b) Du(t,x) (13.26)
for almost all t E (0,T) and for all x
If the Hamiltonian f = f(t,x,u,p) satisfies the conditions a)-e) indicated in

Part 10 of § 12.2, then it is known (Theorems 12.5-12.6) that a solution
of (13.23)-(13.24) exists and is unique. For almost all t (O,T), the niinimax
solution satisfies (13.23) at each point (t, x) where it is differentiable. Further, if
a classical solution of (13.23)-(13.24) exists, then it coincides with the minimax
solution. In this section, the main result on the relationship between minimax and
viscosity solutions (in the case of equations with time-measurable Hamiltonians) is
as follows.
Theorem 13.7. Under the hypotheses of Theorem 12.6, the m;nimaz solution of
(13.23)-(13.24) is also a viscosity solution.
Proof. Take A C (0, T) to be a null set satisfying (12.37) of Lemma 12.3. By [119,
p. 158], we may assume that AC (0, T)\A C Leb(t), with £ = £(t) the function
mentioned in Condition c) (Part 10 of §12.2), and Leb(i) the set of all t E (0,T)
1
satisfying lim £(r)dr — £(t) = 0.
J
Let u = be the minimax solution of (13.23)-(13.24) that exists and is
unique by Theorem 12.6. Actually, we have shown more, namely that u = u(t, x)
does not depend on the choice of the multivalued mappings = x, u, cs)
and FL = FL(t,x,u,/3) (in Fu(f) and FL(f), respectively). From now on, we
shall particularly use (12.14) for a concrete pair of such multivalued mappings and,
accordingly, use Definitions 1-3 in Chapter 12.
§13.3. RELATIONSHIP BETWEEN MINIMAX AND VISCOSITY SOLUTIONS 213
Let us first prove (13.26) for t Ac, x To this end, suppose

Ac, R", (a,b) E Du(t(°),x(°)). Choose a E S and 0 s R so that
b=s•cx.
x8(.)
with
+ + 5)) — 0. (13.27)
+ 5) — Because E AC is a
=f
Let yo j(°)
1(0)
point of £ = £(t) and the function family {xö(.)}ö C x(°)) is uniformly
bounded, (12.14) yields
1
sup —IyoI sup I £(r)(1 + x6(r)I)dr < oo.
5
Consequently,
1
urn =y (13.28)
5
for some y E R" and > 5(e) > ... > 5(k) 0. But, by (12.16) and (12.14), we
now see that
= a)dr
f f(r,xo(r), u(i(°>, x(°)),cx)dr,
hence, by (12.13), that
J
This together with (12.37) and (13.28) implies (y,b) f(t(°),x(°),u(t(°),x(°)),b).
Therefore, it follows from (13.27)-(13.28) and (13.18)-(13.21) that
0 sup mf { [u(t(°) + + — u(t(°),x(°))]/5}

e>00(J<1
1•a + (y,b)
a +
The inequality (13.26) has been proved for almost all t E (0, T) and for all x E
Similarly for the inequality (13.25). The proof is thus complete. 0
Appendix I
Giobal Existence
of Lharacteristic Curves
In this appendix, we will give sufficient conditions which guarantee the global solv-
ability of the Cauchy problem (2.7)-(2.8) and (215)-(2.16). In [43], B. Doubnov
gave sufficient conditions for the global solvability of Hamilton's equation. First,
we report his problem and results.
Let f = f(t,x,p) be a C2-function defined on = {(i,x,p) t E R, x E
:
pE Consider the Cauchy problem:
= (1 = 1,2,... ,n),
1 (All)
I (i=1,2
x(0) = p(O) = p°. (AI.2)
Assume the following conditions:
I p = 0(11), = 0(1 + IpI') for Ii'I

(AI.3)
'Of Of
I
' C'! I'
Op' Op — 'Op
with C = C(s) a certain continuous function defined on [0,oo), and rk 1. In
(AI.3), all 0-estimations are uniform with respect to t and x. That is to say,
p = 0(11) for -÷ 00 if and only if there exist positive constants R and K
independent of (t, x) such that
[p1 for p[ R.
B. Doubnov [43] proved that, if (AI.3) is satisfied, the Cauchy problem (AI.1)-
(AI.2) has uniquely a global solution on {t 0} for all (x°,p°) E x We
give here a brief explanation for this result. As it is well-known that there exists
a local solution of (AL1)-(AI.2), we show that the solution does not blow up in a
GLOBAL EXISTENCE OF CHARACTERISTIC CURVES 215
finite time. Let x = x(t) and p = p(t) be solutions of (AI.1)-(AI.2). Then it holds
that
=
Since rk 1 in (AI.3), we get the estimate
If(t,x(t),p(t))I <
where L and M are constants. As p = o(fk), p = p(t) does not blow up in a finite
time. Using the third inequality in (AI.3), we see that x = x(t) does not blow up
in a finite time. Therefore, we can say that the seminal idea of B. Doubnov 143)
comes from the following example: The Cauchy problem
(d/dt)x(t) = x(t)k, with x(O) = y,
is globally solvable for any y E R if and only if k 1.

Now we consider the Cauchy problem (2.15)-(2.16). By almost the same rea-
soning as in [43), we assume the following conditions for (2.15).
(C.1) There exists a constant N > 0 such that
Of 01
I NIfI.
(C.2) For any constant L > 0, if we put DL def

•
= {(t,x,u,p) : If(t,x,u,p)I < L},

then there exists a constant M > 0 such that
and
+
on DL.
Proposition AI.1. Suppose (C.1) and (C.2). Then the Cauchy problem (2.15)-
(2.16) has a global solution x = x(t,y), v = v(t,y), p = p(t,y) for any initial
data.
We leave the proof to the readers.

216 APPENDIX I
We conclude this appendix with another remark on [43]. B. Doubnov wrote

that, if I = f(t, x, p) was an algebraic function of p such that lim f(t, x, p) = 00,
1,1—Poe
Condition (AI.3) would be satisfied. However, when S. Ouchi, A. Kaneko, and M
Murata translated [43] into Japanese, they already pointed out that this is not true.
Their counter example is as follows:
for n=2.
Though this example does not satisfy (AI.3), the Cauchy problem (AI.1)-(AI.2) has
a global solution for any initial data. Therefore, we would like to give an example
where the Hamiltonian I = f(x,p) is a polynomial of p with lim f(x,p) = 00,
but the corresponding Cauchy problem (AI.1)-(AI.2) cannot be solved globally with
respect to t.
Example. Let n = 1, and let a = a(x) be of class C°° on R such that a(x) = on
{(xI l}, and that a(x) C = constant > 0 on {IxI 1}. Put f(x,p) a(x)p2.
Then the Cauchy problem (AI.1)-(AI.2) cannot have a global solution for some
initial data. For example, if x° > 1 and p° > 0, then it does not admit a global
solution.
Appendix II
Convex Functions,
Mu Itifu nctions,
and Differential Inclusions
In this appendix, for the convenience of the reader, we summarize without proofs
the relevant material on convex functions, multifunctions, and differential inclusions
that we have used in an essential way since Chapter 8. For the proofs we refer the
reader to [8], [22], [29], [40], [64], and [117].
§AII.1. Convex functions
Throughout, is the usual vector space of real n-tuples x = (x1,. . , . The

Eucidean norm and inner product in it are denoted by .[ and (., .).
A subset D of is called affine (respectively, convex) if (1 — A)x + Ay E D for
any x E D, y E D and A E R (respectively, A (0,1)).
Obviously, the intersection of an arbitrary collection of affine sets is again aifine.
Therefore, given any D C there exists a unique smallest affine set containing
D, namely, the intersection of the collection of alfine sets M such that M J D.
This set is called the affine hull of D and is denoted by affD.
The relative interior of a convex set D in which we denote by riD, is defined
as the interior which results when D is regarded as a subset of its affine hull affD.
In other words,
(x+eB)fl(affD)CDforsomee>0},
where B stands for the unit ball (centered at the origin) in
Theorem AII.1. [117, Theorem 6.11 Let D be a convex set in Let x riD
andyED. Then (1—A)x+AyEriDforoA<1.
218 APPENDIX II
Let 4 = 4(x) be a function whose values are real or ±00 and whose domain is
a subset D of IR". The set
epi4,={(x,p):xED,pER,p4,(x)}
deC
is called the epigraph of 4, = 4,(x). We define 4, = 4,(x) to be a convex function (on

D) if epi 4, is convex as a subset of A concave function is a function whose
negative is convex. An affine function is a function which is finite, convex, and
concave.
The effective domain of a convex function 4, = 4,(x) on D, which we denote by
dom 4,, is the projection on R" of the epigraph of 4, =
def
dom4, = {x : (x,p) E epi4, for some p E IR} = {x : 4,(x) < +m}.
This is a convex set in IR", since it is the image of the convex set epi 4, under a
linear transformation. if im4, {4,(x) x E D} C (—co,
: and dom4, 0,
then 4, = 4,(z) is called proper. Trivially, the convexity of 4, = 4,(x) is equivalent
to that of the restriction of 4, = 4,(x) to dom 4,. All the interest really centers on
this restriction, and D itself has little role of its own. Moreover, one could limit
attention to functions given on all of R", since a convex function 4, = 4,(x) on D
can always be extended to a convex function on all of R" by setting 4,(x) +00
for x D. Therefore, by a "convex function," we shall henceforth always mean
a "convex function with possibly infinite values which is defined throughout the
space IR"," unless otherwise specified. The convexity condition can be expressed in
several different ways. For example, we have:
Theorem AII.2. [117, Theorem 4.1] Let 4, = 4,(x) be a function from D to

where D is a convex set (for example D = JRtt). Then 4, = 4,(x) is
convex on D if and only if
— A)x + Ày) < (1 — A)4,(x) + A4,(y) whenever 0 < A < 1
for any x and y in D.
Theorem AII.3. [117, Theorem 4.2] Let 4i = çb(x) be a function from R" to
[—co, +oo]. Then 4, = 4,(x) is convex if and only if
— A)x + Ày) < (1 — A)a + Afi for any A E (0,1)

§AII.1. CONVEX FUNCTIONS 219
whenever 4,(x) <cs and q5(y) </3.
Theorem AII.4. [117, Theorem 5.7] Let 3 x i—+ Ax E Rm be a linear trans-

formation. Then, for each convex function 4, = 4,(x) on the function a = a(y)
defined by
a(y) def.
= inf{4,(x) : Ax = y}
is convex on
The function a = a(y) in Theorem AII.4 is called the image of 4, = 4,(x) under
A, in symbols, a = A4,. The inequality in Theorem AH.2 is often taken as the
definition of the convexity of a function 4, = 4,(x) from a convex set D to (—oo, +00].
(This approach causes difficulties, however, when 4, = 4,(z) can have both +00 and
—oo among its values, since the expression oo — oo could arise.) In this approach,
furthermore, if 4, = 4,(x) is finite and if the inequality is strict for any two different
points x and y in D, then the function 4, = 4,(x) is called strictly convex on D. Of
course, the condition in Theorem AII.3 could be used as the definition of convexity
in the general case, but the definition given via epigraph seems preferable because
it emphasizes the geometry which is fundamental to the theory of convex functions.
Here are some elementary topological properties of convex functions.
Theorem AII.5. [117, Corollary 7.5.1] For a lower semi continuous proper convex
function 4, = 4,(x), one has
cb(y) = lirn4,((1 — \)x + .\y)
for every x E dom4, and every y.
Theorem AII.6. [117, Corollary 10.1.1] A convex function finite on all of is

necessarily continuous.
The (Fenchel) conjugate of a convex function 4, = 4,(x) on which we denote

by qS = 4,(p), is another convex function on defined by the formula
4,*
4, = 4,(x) from to [—00, +ool
can be defined by the same formula as above. It is actually a lower semicontinuous
220 APPENDIX II
convex function.) The following main facts about conjugate convex functions are
from [117, Theorem 12.2, Corollaries 12.2.1-12.2.2, Theorems 23.4-23.5], or partly
from (64, Theorem 4.1 and the addition].
Theorem AII.T. (i) Let = be a convex function. The conjugate function

= is then a lower sernicontinuous convex function, proper if and only if
= q5(x) is proper.
(ii) The conjugacy operation induces a symmetric one-to-one correspon-
dence in the class of all lower semicontinuous proper convex functions on For
any = in this class, the supremum in (AH.1) is attained (maximum) in
domq' zfpEri(domqS*).
(iii) For any convex function = on one actually has
: x pER's.
Further, we have:
Theorem AII.8. [117, Corollary 13.3.3] Let = be a proper convex function.

In order that dom be bounded, it is necessary and sufficient that = be
finite everywhere and that there exist a real number j.z 0 such that
— — yj Vx,y.
Let= 4(x) be a proper convex function that is (finite and) differentiable at

some y E Then [117, Theorems 23.5 and 25.1] implies
= —
This equality suggests that the conjugacy operation is closely related to the
classical Legendre transformation in the case of differentiable convex functions. (See
§ 10.2 for the definition of the Legendre transformation.) In fact, this relationship
is in detail as follows.
Theorem AII.9. [117, Theorem 26.4] Let = be any lower seznicontinuoiz.s

proper convex function such that D mt (dom is non-empty and = is
differentiable on D. The Legendre conjugate (B, i) of (D, gS) is then well-defined.
Moreover, B is a subset of domq5 (namely, the range of the gradient mapping
x i—÷ and a = a(p) is the restriction = to B.
§AII.1. CONVEX FUNCTIONS 221
Corollary AII.1O. (cf. [117, Corollary 26.4.1]) Let 4, = 4,(x) be any differentiable
convex function on R". Then the Legendre conjugate (B, a) of 4,) is well-
defined. One has
: xER't}Cdom4,.
Furthermore, a = a(p) is the restriction of 4, = 4,(p) to B, and a = a(p) is
strictly convex on every convex subset of B.
In general, the Legendre conjugate of a differentiable convex function need not

be differentiable or convex, and we cannot speak of the Legendre conjugate of
the Legendre conjugate. As will be shown in the theorem below, the Legendre
transformation does, however, yield a symmetric one-to-one correspondence in the
class of all pairs (D, 4,) such that D is a non-empty open convex set and 4, = 4,(x)
is a strictly convex function on D satisfying:
(1) 4, = 4,(x) is differentiable throughout D.
(ii) lim 4,t(x(k))I = +00 whenever x(2),... is a sequence in D converging to
a boundary point of D.
For convenience, a pair (D, 4,) in the class just described will be called a convex
function of Legendre type. By [117, Corollary 26.3.1], a lower semicontinuous proper
convex function 4, = 4,(x) has x '—+ 4/(x) one-to-one if and only if the restriction of
4, = 4,(x) to D mt (dom4,) is a convex function of Legendre type.
Theorem AII.11. [117, Theorem 26.5) Let 4, = 4,(x) be a lower semicontinuous

def. •def.
convex functwn. Let D = int(dom4,) and D = int(dom4, ). Then (D,cb)
a convex function of Legendre type if and only if (DC, 4,) is a convex function of
Legendre type. When these conditions hold, (D,4,') is the Legendre conjugate of
(D,4,), and (D,cti) is in turn the Legendre conjugate of (D,çb). The gradient
mapping x 4,'(x) is then one-to-one from the open convex set D onto the open
convex set D, continuous in both directions, and cb' = (4,S)_1.
We now describe the case where the Legendre transformation and the Fenchel
conjugacy correspondence coincide completely. A finite convex function 4, 4,(x)
on R" is said to be co-finite if epi 4, contains no non-vertical half-lines, and this is
equivalent (by [117, Corollary 8.5.2)) to the condition that
lim [4,(\x)/A] = +oo for all x E RTh\{O},

222 APPENDIX II
or (by [117, Corollary 13.3.1]), to the condition that 4, = q5(p) is finite everywhere.
Theorem AII.12. [117, Theorem 26.61 Let 4, = be a (finite) differentiable

4,(x)
convex function on In order that x '— 4,'(x) be a one-to-one mapping from
onto itself, it is necessary and sufficient that 4, = çb(x) be strictly convex and
co-finite. When these conditions hold, j = çh(p) is likewise a differentiable convex
function on which is strictly convex and co-finite, and = 4,'(p) is the same
as the Legendre conjugate of 4, = 4i(x), i.e.,
= (p, (4,')1(p)) —
4, 4, =
the following fact about differentiability of convex

functions.
Theorem AII.13. [117, Theorem 25.2] Let 4' = 4,(x) be a convex function on
and let y be a point at which 4' = 4'(x) is finite. A necessary and sufficient
condition for 4' = cb(x) to be differentiable at y is that the n two-sided partial
derivatives 84i/0x1 exist at y and are finite.
§AII.2. Multifunctions and differential inclusions
Let X 0 be a set. Then is the family of all subsets of X. Given another

set 0 0, a correspondence 0 i-+ E will be called a multifunction
(a set-valued map, or a multivalued function). Sometimes, we permit ourselves to
write briefly L = The sets C X are the values of the multifunction;
allowing = 0 is (very seldornly) convenient for purely formalistic reasons only.
Talk of "the single-valued case" means = on 0. A function p =
with on 0 will be called a selection of L = We refer to
L(0) and graph(L) : 0, p
as the range and the graph of L =

Throughout the book, X R" and 0 0 C Y Rm; both R" and Rm being
endowed with the corresponding Euclidean metrics. Let P be a property of a subset
of a metric space (for instance, closed, measurable). We shall say as a general rule
SAIL2. MULTIPUNCTIONS AND DIFFERENTIAL INCLUSIONS 223
that a inultifunction satisfies P if and only if its graph satisfies P. For instance, a
multifunction is said to be closed (respectively, measurable) if and only if its graph
is closed (respectively, measurable) in the product metric space Y x X. (Whenever
we deal with measurability, we consider on Y = Rm the cr-algebra of all Lebesgue
measurable subsets and on X = R?i the cr-algebra of Borel subsets.) if the values
of a multifunction are closed, bounded, compact, and so on (in X), we say that it
is closed-valued, bounded-valued, compact-valued, and so on. Of course, L = is
called locally bounded if is bounded for any bounded subset fl of
0.
We shall consider only nonempty-valued multifunctions. Ftirther, as a rule,
multifunctions investigated in this book are always compact-valued. For such miii-
tifunctions, we can use:
Definition 1. A multifunction L = is said to be upper semicontinuous if
is (relatively) closed in 0 whenever A C W is closed.
Definition 2. A multifunction L = is said to be lower semicontinuous if

L1(V) is (relatively) open in 0 whenever V C R't is open.
Definition 3. A multifunction L = is said to be continuous if it is simulta-

neously upper and lower semicontinuous.
Evidently, upper (respectively, lower) semicontinuity is nothing else than conti-

nuity if L = is single-valued. Here are useful tests for upper semicontinuity:
Proposition AII.14 (ci. [40, Propositions 1.1-1.2]) Let L =
L= is upper semicontinuous and 0 is closed, then L =

L is closed L= is upper semicontinu-
otis.
Proposition AII.15. (cf. [29, Theorem 11.20]) Let L = have compact convex
values. Then it is upper semicontinuous if and only if the function
sup(p,x)
,EL(t)
224 APPENDIX II
is upper semicontinuous for every x
Proposition AII.16. (cf. [29, Theorem 11.251) Let L = be an upper semi con-
tinuous multifunction with compact values. Then L(Q) is compact for any compact
subset Q of 0.
Let L = be given. We associate with any real-valued function w = w(e,p)

defined on 0 x R" the following marginal function:
= sup for 0. (AII.2)
The maximum sets will be denoted by
{p for 0. (AII.3)
Berge's maximum theorem concerning the above marginal function can be formu-
lated (see [22, p. 123], or [8, Theorem 1.4.161) as follows.
Theorem AII.1T. (Maximum Theorem) Let L =

L = and u = are lower semicontinuous, so is the marginal
function;
(ii) if L = and w = are upper semicontinuoua, so is the marginal
function;
(iii) if L = and w = w(e,p) are continuous, so is the marginal function; more-
over, = defined by (AII.3) is then an upper semicontinuous (nonempty-
valued) multifunction.
We are now concerned with calculus of measurable multifunctions. As we have

mentioned earlier in this section, a multifunction is measurable if and only if its
graph is measurable. Notice that measurable multifunctions with closed values can
also be defined in the following way:
Definition 4. A closed-valued multifunction L = on a non-empty Lebesgue

measurable set 0 C Rm is said to be measurable if L—1(V) is (Lebesgue) measurable
whenever V C is open.
Especially, the above definition agrees with the classical one for measurable
functions if L = is single-valued.
§AIL2. MTJLTIFUNCTIONS AND DIFFERENTIAL INCLUSIONS 225
Theorem AII.18. (cf. [8, Theorem 8.2.8]) Let 0 E Rk be a measurable

single-valued map, and let G = z) be a multifunction measurable in E 0
and continuous in z E IRk (with compact values in IRA). Then the multifunction
0 is measurable.
Theorem AII.19. [8, Theorem 8.2.11]) Let

(cf. = be a real-valued
functson measurable in E 0 and continuous in p E and let L = be a
measurable multifunction on 0 with closed (non-empty) values in Then the
margmal function = defined by (AI1.2) is measurable. Furthermore, the
multifunction L° = defined by (AII.3) is also measurable.
Theorem AII.20. (cf. [8, Theorem 8.2.14] and [29, Theorem 111.15]) Let 0 be a
Lebesgue measurable subset of If L = is a measurable inultif unction on
O with closed (non-empty) values in then the function
sup(p,x)
is measurable for every x E The converse statement holds true if the values of
L= are convex and bounded.
In the remainder of this appendix, given J [0, T] C R, a multifunction

Jx (t, x) i-+ G(t, x) C R", we are looking for absolutely continuous solutions
of the differential inclusion
E G(t,x(i)) almost everywhere on J. (AII.4)
For any E J x RTh, denote by Xo(t.,x.) the set of all absolutely continuous
functions x = x(t) from J into which satisfy (AII.4) subject to the constraint
x(t.) = Topological properties of the solution sets will be investi-
gated in the Banach space C(J, of continuous functions given on J with values
in R". (The norm in C(J, is the usual "max" one.) We have:
Theorem AII.21. [40, Theorems 5.2 and 7.1]) Let G = G(t, x) have non-
(cf.
empty closed convex values and be measurable in t E J, upper semicontinuous in
x 6 R" such that
G(t,x)I : z E G(t,x)} < c(t)• (1 + ri) for t 6 J, x E

226 APPENDIX II
with c = c(t) a function in L'(J). Then Xcg(t.,x.) is a non-empty compact subset

of C(J,R") for each EJx Further, the multifunctzon 3x
Xo(0,x) C C(J, is upper semtcontinuous.
We conclude this appendix with Filippov's theorem for differential inclusions,

which is as important as Gronwall's lemma for ordinary differential equations.
Theorem A1L22. (Filippov) [8, Theorem 10.4.1] (cf. [40, Lemma 8.3]) Let
G = G(t, x) have closed non-empty values and y = y(t) be an function
absolutely continuous on J, and let 8 > 0. Assume:
(i) G = G(t,x) is measurable in t E J;
(ii) there exist an r > 0 and a nonnegative function £ = £(i) integrable on J such
that
G(t, r1) C G(t, x2) + £(t)Ix' — x2JB Vx', x2 E y(t) + rB
for almost all t E J;
(iii) the function t f—* 1(t) — p E G(t, y(t))} is integrable on J.
p(t) exp(j £(r)dr) and (8 +f

If i7(T) < r, then for every E y(0) + SB, there exists a solution x = x(t) in
of(AII.4) such that
VtEJ
Ix'(t) — y'(t)I < £(i)zj(t) + -y(t) almost everywhere on .J.

References
[1] Adiatullina, R.A. and Subbotin, A.I., Generalized solutions of equations of the
type Oço/Ot + = 0, Differential Equatwns 28 (1992), 643-649.
[2] Aizawa, S., Global solutions of partial differential equations of first order, Sugaku
(Mathematics) 21 (1968), 11-23. (in Japanese)
[3] —, A semi-group treatment of the Hamilton-Jacobi equation in several space
variables, Hiroshima Math. J. 6 (1976), 15-30.
[4] Aizawa, S. and Kikuchi, N., A mixed initial and boundary value problem for
the Hamilton-Jacobi equation in several space variables, Funkcialaj Ekvacioj 9
(1966), 139-150.
[5] Aizawa, S. and Tomita, Y., On unbounded viscosity solutions of a semilinear
second-order elliptic equation, Funkcialaj Ekvacioj 31 (1988), 141-160.
[6] Alexandrov, A.D., On surfaces which may be represented by the difference of
convex functions, Dokl. Akad. Nauk SSSR 72 (1950), 613-616. (in Russian)
[7] Aubin, J.-P. and Ekeland, I., Applied nonlinear analysis, A Wiley — Interscience
publication, John Wiley & Sons, New York, 1984.
[8] Aubin, J.-P. and Frankowska, H., Set-valued analysis, Birklsi.user, Base!, 1990.
[9] Ballou, D.P., Solutions to nonlinear hyperbolic Cauchy problems without con-
vexity conditions, Trans. Amer. Math. Soc. 152 (1970), 441-460.
[10] Barbu, V. and Da Prato, G., Hamilton-Jacobi equations in Hubert spaces, Re-
search Notes in Math., Vol. 86, Pitman, Boston, MA, 1983.
[11] Bardi, M. and Evans, L.C., On Hopf's formulas for solutions of Hamilton-Jacobi
equations, Nonlinear Analysis, Theory, Methods elApplications 8 (1984), 1373-
138 1.
[12] Bard, M. and Faggian, S., Hopf-type estimates and formulas for non-convex
non-concave Hamilton-Jacobi equations, SIAM J. Math. Anal. 29 (1998), 1067-
1086.
[13] Bardi, M. and Osher, S., The non-convex multidimensional Riemaim problem
for Hamilton-Jacobi equations, SIAM J. Math. Anal. 22 (1991), 344-351.
[14] Barles, G., Existence results for first-order Hamilton-Jacobi equations, Ann.
Inst. H. Poincaré - Analyse non-linéaire 1 (1984), 325-340.
[15] —, Uniqueness and regularity results for first-order Hamilton-Jacobi equations,
Indiana Univ. Math. J. 39 (1990), 443-466.
[16] —, Regularity results for first-order Hamilton-Jacobi equations, Differential &
Integral Equations 3 (1990), 103-125.
[17] —, Solutions de viscosité des equations de Hamilton-Jacobi, Springer — Verlag,
Berlin — Heidelberg, 1994.
228 REFERENCES
[18] Barles, G. and Perthame, B., Exit time problems in optimal control and van-
ishing viscosity method, SIAM J. Control Optim. 26 (1988), 1133-1148.
[19] Barron, EN. and Jensen, R., Generalized viscosity solutions for Hamilton-
Jacobi equations with time-measurable Hamiltonians, J. Differential Equations
68 (1987), 10-21.
[20] —, Optimal control and semicontinuous viscosity solutions, Proc. Amer. Math.
Soc. 133 (1991), 397-402.
[21] Benton, S.H., The Hamilton-Jacobi equation: a global approach, Academic Press,
New York, 1977.
[22] Berge, C., Espaces topologiques. Fonctions multivoques, Dunod, Paris, 1966.
[23] Bony, M., Calcul symbolique et propagation des singularités pour les equations
aux dérivées partielles non-linéaires, Ann. Sci. Ecole Normale Sup. série 14
(1981), 209-246.
[24] Brancli, P. and Ceppitelli, R., Existence, uniqueness and continuous dependence
for hereditary differential equations, J. Differential Equations 81 (1989), 317-
339.
[25] Brandi, P. and Marcelli, C., Haar inequality in hereditary setting and applica-
tions, Rend. Sem. Mat. Univ. Padova 96 (1996), 177-194.
[261 —, Extremal solutions and Gronwall inequality in functional setting, Boll. Un.
Mat. It. (7) 10-B (1996), 233-254.
[27] —, Uniqueness and continuous dependence for Carathéodory solutions of func-
tional Cauchy problem, Atti Sem. Mat. Fis. Univ. Modena XLIV (1996),
379-391.
[28] Cappuzzo-Dolcetta, I. and Lions, P.-L., Viscosity solutions of Hamilton-Jacobi
equations and state constraints, Trans. Amer. Math. Soc. 318 (1990), 643-683.
[29] Castaing, Ch. and Valadier, M., Convex analysis and measurable multifunctions,
Lecture Notes in Math., Vol. 580, Springer—Verlag, New York, 1977.
[30] Chaplygin, S.A., Collected papers on mechanics and mathematics, Moscow,
1954. (in Russian)
[31] Coddington, E.A. and Levinson, N., Theory of ordinary dzfferentzal equations,
McGraw-Hill, New York, 1955.
[32] Conway, ED., The formation and decay of shocks for a conservation law in
several dimensions, Arch. Rational Mech. Anal. 64 (1977), 47-57.
[33] Conway, E.D. and Hopf, E., Hamilton's theory and generalized solutions of the
Hamilton-Jacobi equation, J. Math. Mech. 13 (1964), 939-986.
[34] Coura.nt, R. and Hilbert, D., Methods of mathematical phys;cs I-Il, Interscience
Publishers (1953, 1962).
[35] Crandall, M.G., Evans, L.C., and Lions, P.-L., Some properties of viscosity
solutions of Hamilton-Jacobi equations, Trans. Amer. Math. Soc. 282 (1984),
487-502.
[36] Crandall, M.G., Ishii, H., and Lions, P.-L., Uniqueness of viscosity solutions of
Hamilton-Jacobi equations revisited, J. Math. Soc. Japan 39 (1987), 581-596.
[37] —, User's guide to viscosity solutions of second-order partial differential equa-
tions, Bull. Amer. Math. Soc. 27 (1992), 1-67.
REFERENCES 229
[38] Crandall, M.G. and Lions, P.-L., Condition d'unicité pour les solutions gdnéral-
isées des equations de Hamilton-Jacobi du premier ordre, C.R. Acad. Sci. Paris
SCr. I Math. 292 (1981), 183-186.
[39) —, Viscosity solutions of Hamilton-Jacobi equations, Thins. Amer. Math. Soc.
277 (1983), 1-42.
[40] Deimling, K., Multivalued differential equations, De Gruyter, 1992.
[41) Debeneix, T., Certains systèmes hyperboliques quasi-linéaires, Universite' de
Paris VII, Preprint, 1980.
[42] Dieudonné, J., Foundations of modern analysis, Academic Press, New York,
1960.
[43] Doubnov, B., Sur l'existence globale des solutions des equations d'Hamilton,
Supplement dans Théorie des perturbations et inéthodes asymptotiques (par
Maslov, V.P.), Dunod, Paris, 1972 (in French). Japanese transl. by Ouchi,
S., Kaneko, A., and Murata, M., Iwanami Shoten, Tokyo, 1976.
[44] Douglis, A., Solutions in the large for multi-dimensional nonlinear partial d-
ifferential equations of first order, Ann. Inst. Fourier, Grenoble 15 (1965),
1-35.
[45] —, Layering methods for nonlinear partial differential equations of first order,
Ann. Inst. Fourier, Grenoble 22 (1972), 141-227.
[46] Duff, G.F.D., Hyperbolic differential equations and waves, in Boundary value
problems for linear evolution partial differential equations, pp. 27-155, D. Reidel,
1977. (Garnir, H.G., Ed.)
[47] Engler, H. and Lenhart, S., Viscosity solutions for weakly-coupled systems of
Hamilton-Jacobi equations, Proc. London Math. Soc. 83 (1991), 212-240.
[48] Evans, L.C., Some mm-max methods for the Hamilton-Jacobi equations, Indi-
ana Univ. Math. .J. 33 (1984), 31-50.
[49] —, Partial differential equations, Berkeley Mathematics Lecture Notes Series,
Vol. 3, 1993.
[50] Evans, L.C. and Souganidis, P.E., Differential games and representation formu-
las for solutions of Hamilton-Jacobi-Isaacs equations, Indiana Univ. Math. J.
33 (1984), 773-797.
[51] Fleming, W.H., The Cauchy problem for degenerate parabolic equations, J.
Math. Mech. 13 (1964), 987-1008.
[52] —, The Cauchy problem for a nonlinear first-order partial differential equation,
J. Differential Equations 5 (1969), 515-530.
[53] Folland, G.B., Real analysis: modern techniques and their applications, A Wi-
ley — lnterscience publication, John Wiley & Sons, New York, 1984.
[54) Friedman, A., The Cauchy problem for first-order partial differential equations,
Indiana Univ. Math. J. 23 (1973), 27-40.
[55] Gb.rding, L., Hyperbolic differential operators, in Perspectives in Math., An-
niversary of Oberwolfach 1984, pp. 215-247, Birkhãuser Verlag, Basel.
[56] Gaveau, B., Evolution of a shock for a single conservation law in 2+1 dimensions,
Bull. Sc. math. 113 (1989), 407-442.
[57) Golubitsky, M. and Schaeffer, D.G., Stability of shock waves for single
vation laws, Adv. in Math. 15 (1975), 65-71.
230 REFERENCES
[58] Gordon, W.B., On the diffeomorphism of Euclidean space, The American Math-
ematical Monthly 29 (1972), 755-759.
[59] Guckenheimer, J., Solving a single conservation law, Lecture Notes in Math.,
Vol. 468, 1975, PP. 108-134 (Springer - Verlag).
[60] —, Shocks and rarefactions in two space dimensions, Arch. Rational Mech.
Anal. 59 (1975), 281-291.
[61] Haar, A., Sur l'unicité des solutions des equations aux dérivées partielles, C.R.
Acad. Sci. Paris 187 (1928), 23-26.
[62] Hartman, P., Ordinary differential equations, John Wiley & Sons, New York -
London — Sydney, 1964.
[63] Hopf, E., The partial differential equation Ut + = Commun. Pure
AppI. Math. 3 (1950), 201-230.
[64] -__, Generalized solutions of nonlinear equations of first order, J. Math. Mech.
14 (1965), 951-973.
1651 Hórmander, L., The analysis of linear partial differential operators I-IV, Sprin-
ger, Berlin, 1983.
[66] Horst, R. and Tuy, H., Global optimization: deterministic approaches, Springer,
Berlin, 1996. (3., rev, and enl. ed.)
[67] Ishii, H., Uniqueness of unbounded viscosity solutions of Hamilton-Jacobi equa-
tions, Indiana Univ. Math. .J. 33 (1984), 721-748.
[68] —, Existence and uniqueness of solutions of Hamilton-Jacobi equations, Funk-
cialaj Ekvacioj 29 (1986), 167-188.
[69] —, Perron's method for Hamilton-Jacobi equations, Duke Math. J. 55 (1987),
369-384.
[70J —, A simple direct proof for uniqueness for solutions of the Hamilton-Jacobi
equations of eikonal type, Proc. Amer. Math. Soc. 100 (1987), 247-251.
[71J —, Representation of solutions of Hamilton-Jacobi equations, Nonlinear Anal-
ysis, Theory, Methods &Applications 12 (1988), 121-146.
[72] Ishii, H. and Koike, S., Viscosity solutions for monotone systems of second-order
elliptic PDEs, Commun. in PDEs 16 (1991), 1095-1128.
1731 Izumiya, S., Geometric singularities for Hamilton-Jacobi equations, Ado. Stud-
ies in Pure Math. 22 (1993), 89-100.
[74[ —, Characteristic vector fields for first-order partial differential equations, Pre-
print.
[75] Izuiniya, S. and Kossioris, G.T., Semi-local classification of geometric singu-
larities for Hamilton-Jacobi equations, J. Differential Equations 118 (1995),
166-193.
[76J —, Bifurcations of shock waves for viscosity solutions of Hamilton-Jacobi equa-
tions of one space variable, Bull. Sc. math. 121 (1997), 619-667.
[77] Jeffrey, A., Quasilinear hyperbolic systems and waves, Research Notes in Math.,
Vol. 5, Pitman, 1976.
[78] Jennings, G., Piecewise smooth solutions of single conservation laws exist, Adv.
in Math. 33 (1979), 192-205.
REFERENCES 231
[791 Jensen,R. and Souganidis, P.E., Regularity results for viscosity solutions of
Hamilton-Jacobi equations in one space dimension, Trans. Amer. Math. Soc.
301 (1987), 137-147.
[80] John, F., Partial differential equations, Springer, Berlin, 1975.
[81] Kainke, E., Zur Theorie gewohnlicher Differentialgleichungen II, Acta Math. 58
(1932), 57-85.
[82] Kolokol'tsov, V.N. and Maslov, V.P., The Cauchy problem for a homogeneous
Bellman equation, Doki. Acad. Nauk SSSR 296 (1987), 796-800. (in Russian)
[83] Krasnosel'skii, M.S., The operators of translation along the trajectories of dif-
ferential equations, AMS Translations of Math. Monograph, Vol. 19, 1968.
[84] Krasovskii, N.N. and Subbotin, A.I., An alternative for game problem of con-
vergence, Prild. Mat. Mekh. 34 (1970), 1005-1022 (in Russian). English transi.
in J. App!. Math. Mech. 34 (1971), 948-965.
[85] —, On the structure of game-theoretical problems of dynamics, Prikl. Mat.
Mekh. 35 (1971), 110-122. (in Russian)
[86] —, Positional differential games, Nauka, Moscow, 1974. (in Russian)
[871 Krushkov, S.N., The Cauchy problem in the large for certain nonlinear first-
order equations, Soviet Math. Dokl. 1 (1960), 474-477.
[88] —, Generalized solutions of nonlinear first-order equations with several inde-
pendent variables, Math. USSR Sb. 1 (1967), 93-116.
[89] —, Generalized solutions of the Cauchy problem in the large for nonlinear e-
quations of first order, Soviet Math. DokI. 10 (1969), 785-788.
[90] —, On the mm-max representation of solutions of first-order nonlinear equa-
tions, Functional Anal. Applic. 2 (1969), 128-136.
[91] —, First-order nonlinear equations and the differential games connected with
them, Uspekhi Mat. Nauk 24 (1969), 227-228. (in Russian)
[92] —, First-order quasilinear equations in several independent variables, Math.
USSR Sb. 10 (1970), 217-243.
[93] —, F\xnction theory and differential equations, Moscow Univ. Math. Bull. 46
(1991), 30-35.
[94] Kruzhkov, S.N. and Petrosyan, N.S., Asymptotic behaviour of solutions of the
Cauchy problem for first-order nonlinear equations, Uspekhi Mat. Nauk 42
(1987), 3-40. (in Russian)
[95] Kuznetsov, N.N., Precision of certain approximate methods of calculation of
weak solutions for quasi-linear first-order equations, Zhurn. Vychisl. Mat. i
Mat. Fiz. 1 (1975), 1489-1502.
[96] Lakshmikantham, V. and Leela, S., Differential and integral inequalities I-If,
Academic Press, New York, 1969.
[97] Lax, P.D., Hyperbolic systems of conservation laws II, Commun. Pure App!.
Math. 10 (1957), 537-566.
[98] —, Hyperbolic systems of conservation laws and the mathematical theory of
shock waves, SIAM Regional Conference Series in AppI. Math., Vol. 11, 1973.
[99] Lenhart, S., Viscosity solutions for weakly-coupled systems of first-order PDEs,
J. Math. Anal. AppI. 131 (1988), 180-193.
232 REFERENCES
1100] Lions, P.-L., Generalized solutions of Hamilton-Jacobi equatwns, Research Notes

in Math., Vol. 69, Pitman, Boston, MA, 1982.
1101] Lions, P.-L. and Rochet, J.-C., Hopf formula and multitime Hamilton-Jacobi
equations, Proc. Amer. Math. Soc. 96 (1986), 79-84.
[102] Li Ta-tsien and Chen Shu-xing, Regularity and singularity of solutions for non-
linear hyperbolic equations, to appear.
[103] Li Ta-tsien and Shi Jia-hong, Globally smooth solutions in the whole space for
first-order quasi-linear equations, Fadan J. 21 (1982), 36 1-366.
[104] Majda, A., Compressible fluid flows and systems of conservation laws in several
space variables, Springer — Verlag, Berlin, 1984.
[1051 Maslov, V.P. and Samborskii, S.N., Existence and uniqueness of solutions of
time-independent Hamilton-Jacobi-Bellman equations: a new approach, Doki.
Acad. Nauk 324 (1992), 1143-1148. (in Russian)
[106] Mirica, S., Extending Cauchy's method of characteristics for Hamilton-Jacobi
equations, Stud. Cert. Mat. 37 (1958), 555-565.
[107] Nagumo, M., Uber die Ungleichung Ou/Ox > f(x,y,u,ôu/Oy), Jap. J. Math.
15 (1938), 51-56.
[108] Nagumo, M. and Simoda, S., Note sur l'inégalité différentielle concernant les
equations du type parabolique, Proc. Japan Aced. 27 (1951), 536-539.
[109] Nakane, S., Formation of shocks for a single conservation law, SIAM J. Math.
Anal. 19 (1988), 1391-1408.
[110] —, Formation of singularities for Hamilton-Jacobi equations in several space
variables, J. Math. Soc. Japan 43 (1991), 89-100.
[111] Ngoan, H .T., Hopf's formula for Lipschitz solutions of Hamilton-Jacobi equa-
tions with concave-convex Harniltonian, Acta Math. Vietnam. 23 (1998), 269-
294.
[112] Oleinik, O.A., Discontinuous solutions of non-linear differential equations, Us-
pekhi Mat. Nauk 12 (1957), 3-73 (in Russian). English transi. in Amer. Math.
Soc. Transl. 26 (1963), 95-172.
[113] A., Characteristics of non-linear partial differential equations, Bull. Aced.
Polon. Sci. 2 (1954), 419-422.
[1141 Qin Tie-hu, On the existence of global smooth solutions for Cauchy problems
to first order quasi-linear partial differential equations, J. of Fudan Univ. 22
(1983), 41-48.
1115] Quinn, BK., Solutions with shocks, an example of an L1-contractive semi-group,
Commun. Pure Appl. Math. 24 (1971), 125-132.
[116] Robert, R.P., Convex functions, monotone operators and differentiability, Sprin-
ger, Berlin, 1989.
[117] Rockafellar, R.T., Convex analysis, Princeton Univ. Press, 1970.
[118] Rozdestvenskii, B.L., Discontinuous solutions of hyperbolic systems of quasi-
linear equations, Russian Math. Surveys 15 (1960), 53-111.
[119] Rudin, W., Real and complex analysis, McGraw—Hill, New York, 1966.
[120] Schaeffer, D.C., A regularity theorem for conservation laws, Adv. in Math. 11
(1973), 358-386.
REFERENCES 233
[1211 Smoller, J., Shock waves and reactzon-diffusion equations, Springer— Verlag,
Berlin, 1983.
[122] Souganidis, P.E., Existence of viscosity solutions of Hamilton-Jacobi equations,
[123] —, Approximation schemes for viscosity solutions of Hamilton-Jacobi equations,
[124] Subbotixi, A.I., Minimax inequalities and Hamilton-Jacobi equations, Nauka,
Moscow, 1991. (in Russian)
[1251 —, Generalized solutions of first-order PDEs: the dynamical optimization per-
spective, Birkkauser, Boston, 1995.
[126] Subbotin, A.I. and Shagalova, L.G., A piecewise linear solution of the Cauchy
problem for the Hamilton-Jacobi equation, Russian Acad. Sci. DokI. Math. 46
(1993), 144-148.
[127] Subbotina, N.N., Cauchy's method of characteristics and generalized solutions
of Hamilton-Jacobi-Bellman equations, Doki. Acad. Nauk SSSR 320 (1991),
556-561. (in Russian)
[1281 Szarski, J., Differential inequalities, Monografle Matematyczne, Warszawa, 1967.
(2., rev. ed.)
[129] Tarasyev, A.M., Approximation schemes for the construction of the generalized
solution of the Hamilton-Jacobi (Bellman-Isaacs) equation, Prikl. Mat. Mekh.
58 (1994), 22-36. (in Russian)
[1301 Tarasyev, A.M., Uspenskii, A.A., and Ushakov, V.N., Approximation operators
and finite-difference schemes for the construction of generalized solutions of
Hamilton-Jacobi equations, Izv. RAN Tekhn. Kibernet. 3 (1994), 173-185. (in
Russian)
[131] Tataru, D., Viscosity solutions of Hamilton-Jacobi equations with unbounded
nonlinear terms, J. Math. Anal. AppI. 163 (1992), 345-392.
[132] Thai Son, N.D., Liem, N.D., and Van, T.D., Minimax solutions of first-order
nonlinear partial differential equations with time-measurable Haxniltonians, in
Dynamical systems and applications, pp. 647-668, World Sci. Ser. Appl. Anal.,
Vol. 4, World Sci. Publishing, River Edge, NJ, 1995. R.P., Ed.)
[133] —, Minimax solutions for monotone systems of first-order nonlinear partial
differential equations with time-measurable Hamiltonians, Funkcialaj Ekvacioj
40 (1997), 185-214.
[134] Thom, R., The two-fold way of catastrophe theory, Lecture Notes in Math., Vol.
525, 1976, pp. 235-252 (Springer—Verlag).
[135] Tomita, Y., On a mixed problem for the multi-dimensional Hamilton-Jacobi
equation in a cylindrical domain, Hiroshima Math. J. 8 (1978), 439-468.
[136] Tsuji, M., Solution globale et propagation des singularités pour l'équation de
Hamilton-Jacobi, C.R. Acad. Sci. Paris 289 (1979), 397-400.
[137] —, Formation of singularities for Hamilton-Jacobi equation II, J. Math. Kyoto
Univ. 26 (1986), 299-308.
[138] —, Prolongation of classical solutions and singularities of generalized solutions,
Ann. Inst. H. Poincard-Analyse non-linéaire 7 (1990), 505-523.
234 REFERENCES
[139] Tsuji, M. and Li Ta-tsien, Globally classical solutions for nonlinear equations of
first order, Coinmun. zn PDEs 10 (1985), 1451-1463.
[140] —, Remarks on characteristics of partial differential equations of first order,
Funkctalaj Ekvacioj 32 (1989), 157-162.
(141] Van, T.D., Hap, L.V., and Thai Son, N.D., On some differential inequalities
and the uniqueness of global semiclassical solutions to the Cauchy problem for
weakly-coupled systems, Journal of Inequalzties and Applzcatwns 2 (1998), 357-
372.
[142] Van, T.D. and Hoang, N., On the existence of global solutions of the Cauchy
problem for Hamilton-Jacobi equations, SEA Bull. Math. 20 (1996), 81-88.
[143] Van, T.D., Hoang, N., and Gorenflo, R., Existence of global quasi-classical so-
lutions of the Cauchy problem for Hamilton-Jacobi equations, Dzfferentsial 'nyc
Uravneniya 31 (1995), 672-676. (in Russian)
[144] Van, T.D., Hoang, N., and Thai Son, N.D., On the explicit representation of
global solutions of the Cauchy problem for Hamilton-Jacobi equations, Acta
Math. Vietnam. 19 (1994), 111-120.
[145] Van, T.D., Hoang, N., and Tsuji. M., On Hopf's formula for Lipschitz solution-
s of the Cauchy problem for Hamilton-Jacobi equations, Nonlinear Analysis,
Theory, Methods & Applications 29 (1997), 1145-1159.
[146] Van, T.D., Liem. N.D., and Thai Son, N.D., Minimax solutions for some sys-
tems of first-order nonlinear partial differential equations with time-measurable
Hamiltonians, in Structure of solutions of differential equatsons (Katata/Kyoto,
1995), pp. 499-511, World Sci. Publishing, River Edge, NJ, 1996. (Morimoto,
M. and Kawai, T., Eds.)
[147] Van, T.D. and Thai Son, N.D., On the uniqueness of global classical solutions
of the Cauchy problem for Hamilton-Jacobi equations, Acta Math. Vietnam.
17 (1992), 161-167.
[148] _-, On the uniqueness of global classical solutions of the Cauchy problems for
nonlinear partial differential equations of first order, Acta Math. Vietnam. 18
(1993), 127-136.
[149] —, On a class of Lipschitz continuous functions of several variables, Proc. Amer.
Math. Soc. 121 (1994), 865-870.
[150] _, Uniqueness of global semiclassical solutions to the Cauchy problem for first-
order nonlinear partial differential equations, Differential Equations 30 (1994),
659-666.
(151] Van, T.D., Thai Son, N.D., and Hap, LV., Partial differential inequalities of
Haar type and their applications to the uniqueness problem, Vietnam J. Math.
26 (1998), 1-28.
[152] Van, T.D., Thanh, M.D., and Hoang, N., On the representation of Lipschitz
global solutions of the Cauchy problem for Hamilton-Jacobi equations, in Proc.
of Intl. Conference on Appl. Anal. ei Mech. of Cont. Media, Ho Chi Minh City
(12/1995), pp. 428-436.
[153] Walter, W., Differential and integral inequalzties, Springer, Berlin, 1970.
[154] T., Sur l'unicité et Ia limitation des intégrales des equations
dérivées partielles du premier ordre, Rend. Acc. Lsncei 17 (1933), 372-376.
REFERENCES 235
[155] —, Sur l'unicité et la limitation des intégrales de certains systèmes d'équations

aux dérivées partielles du premier ordre, Annali di Matematica pura ed applicata
15 (1936-37), 155-158.
[156] -__, Uber die Bedingungen der Existenz der Integrale partieller Differentialgie-
ichungen erster Ordnung, Math. Z. 43 (1938), 522-532.
[1571 —, Systèmes des equations et des inégalités différentielles ordinaires aux deuxi-
èmes membres monotones et leurs applications, Ann. Soc. Polon. Math. 23
(1950), 112-166.
[158] Westphal, H., Zur Abschätzung der Lösungen nichtlinearer parabolischer Dif-
ferentialgleichungen, Math. Z. 51 (1949), 690-695.
[159] Whitney, H., On singularities of mappings of Euclidean spaces I, Ann. Math.
62 (1955), 374-410.
Index
Cantor function (ladder), 86 equation of the conservation law, 12, 45

Cantor set, 86
Carathéodory differential equation, 98 Filippov's theorem, 180, 226
Carathéodory's conditions, 97, 166, 191 fold point, 57
characteristic
curves, 1 generalized solution, 32, 55
equations, 2, 4 general PDE of first-order, 4
strips, 4 global C'-solution, 92
system of differential equations, 4 global semiclassical solution, 147, 154
1 global solution, 104
C"-solution, 7
collision Haar's differential inequality, 85
of characteristic curves, 26, 28 Haar's theorem, 7
of singularities, 41, 63 Hadamard's lemma, 19
comparision equation, 98 Hopf's formulas, 104
concave-convex function, 128
conjugate Jacobian, 2
(Fenchel) concave-convex, 131 Jacobi matrix, 2
(Fenchel) convex, 105, 219-222
Legendre, 128, 220-222 Lebesgue's theorem, 170
contact discontinuity, 70 Legendre transformation, 128, 220-222
cusp point, 58 life span of classical solutions, 12
d.c. function, 113 marginal function, 224

d.c. representation, 117 maximum theorem, 224
differential inclusion, 225 minimax solution, 168, 179, 194, 196
differential inequality of Haar type, 85 monotonicity condition, 192
Dini serniderivative, 108, 209 rnultifunction, 222
Dini semidlifferential, 210 closed, 223
directional derivative, 108 closed-valued, 223
continuous, 223
effective domain, 105, 218 locally bounded, 223
entropy condition, 52 lower semicontinuous, 223
epigraph, 218 measurable, 223-224
equation of Hamilton-Jacobi type, 34 upper semicontinuous, 223
INDEX 237
non-degenerate singularity, 58 strict convexity, 118, 122, 219

subdifferential, 210
proper, 123, 218 subgradient, 210
subsolution, 168, 179, 194, 196
quasi-linear PDE of first-order, 1 superdifferential, 210
quasi-monotonicity condition, 192 supergradient, 210
supersolution, 167, 179, 193, 196
Rankine-Hugoniot's condition, 41
rarefaction wave, 68 viscosity solution, 212
relative interior, 217
Wniewski's theorem, 9
semi-concavity, 32, 39, 56 weakly-coupled system, 97
shock wave, 53 weak solution, 45
lhc ch.tracteristic meihod s tekls the local theory classical solutions
to lirst-ordcr nonlinear partial ditfereniial equations. The global theory
has principally depended on the vanishing s iscosity method. ftc authors
ol this bridge the between the local and global theories
by using the characteristic method as a basis for settiliC a theoretical
framework for the studs of global generaltied solutions. That is. they
extend the smooth solutions obtained by the characteristic method. Within
such a ork. they oIler material on the life span of classical
solutions, the construction of simzularities of generali,ed solutions, new
existence and uniqueness theorems on minimax solutions.
inequalities of' Haar pe and their application to the uniqueness t global
semi—classical solutions, and iiopi explicit iornnilas tor global
solutions.
i'liis soltirne represents a comprehensixe exposition 01 the authors' works

ox'er the last decade. concentrating on sonle basic liicts and ideas of the
general lied characteristic methods br study tug global si lutions. Suitable
as a text. the book is sell-contained and assumes as prerequisites only
basic iileasurc I het . tt . and ordinary differential equations. lhe
appendices pros ide necessary material, primarily on nonsmooth analysis
and the t he ry of di fb'ere it ial inclusions.
Reiu/ers/iij,: \latheniaticians. sicists. and engineers: in

n in I near partial di I Icrent al equations, ill terent ,il inequalities. multix aiucd
analysis. diltcrential gatnes. and related topics in applied analysis:
upper- les ci undergraduate and graduate students in these disciplines.
irait l)uc is l)iicctor of the Ilaititi Institute of \latliciii:ities and

I 'rol'essor. I )epait ilicilt of Partial 1)1 fferential I ions
\Iikio l'suji is a Professor ol' \lathcitiatics at the laLtilty of Sciences ol

Ky oto ot crslty
\gti%tiI Ihai Son is .in Associate i'rolcssor of \Ijthcnuttcs. College

ol Sciences, line tnixersity.
: :1
CHAPMAN & HALL/CRC

The Characteristic Method and Its Generalizations

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

The Characteristic Method and Its Generalizations

Загружено:

Авторское право:

Доступные форматы

CHAPMAN & HALL/CRC

Monographs and Surveys in

METHOD AND ITS

TRAN DUC VAN

IC Monographs and Surveys in

METHOD AND ITS

CHAPMAN & HALL/CRC

Tran, Due Van.

Trademark Notice: Product or corporate names may be trademarks or regis-

© 2000 by Chapman & Hall/CRC

No claim to original U.S. Government works

6.2. Construction of solutions 56

13.2. Smoothness of global solutions 205

x(t)) = x(t)) + u(t, x(t)).

Hence, if the C1-norm of h = h(s) is bounded, then, by means of the implicit

<s2 and h(s1) > h(s2). (8)

"characteristic bundles" are invoked instead of characteristic differential equations

In accordance with (1.2), the initial condition for (1.3) is given by

x,(O) = (i = 1,2,... ,n), v(O) = (1.4)

following discussions, u = u(i, x) is a solution of (1.1) and v v=(i, y) is a solution of

Let us fix our notations on derivatives of functions. A vector x is vertical,

We will sometimes write the Jacobi matrix simply by Since x(O, y) = y,

we see that (Dx/Dy)(t,y) = 1 for t = 0, y E Uo. In a neighborhood of {(O,y)

(p!-) = I (identity matrix), (1.5)

By (1.7), using (1.5) and (1.6), we get

As u(O,x) = v(O,y(O,x)) = cS(s), we see that u = u(t,x) satisfies the Cauchy

the following Cauchy problem:

As the right-hand side of this differential equation can be estimated by — vi =

Miwl, we get w(t,y) 0, i.e. v(i,y) for any (t,y) in a neighborhood of

U is an open neighborhood of the origin. Let V be an open neighborhood of

x,(0)=y,, v(0)=cb(y), (i=1,2,... ,n). (1.11)

We remark that system (1.10) is called the "characteristic system of differential

::: ::: ::: ::: , (1.12)

Using (1.10), we have

Corollary 1.3. In the definition domain of y = y(t,x), we have

Using Lemma 1.2 and its corollary, we get the following:

Moreover, by (1.11), u(0,x) = v(0,y(0,x)) = q5(x). It follows that u = u(t,x) is a

with w(0,y) = çb(y) and q(0,y) = qV(y). Hence

(z, w, q) = (z(t, y), w(t, y), q(t, !J))

satisfies the Cauchy problem (1.1O)-(1.11). By the uniqueness of solutions of ordi-

r(t, y) = z(t, y), v(i, y) = w(t, y), p(i, y) = q(t, y),

where (x,v,p) = (x(t,y),v(t,y),p(t,y)) is the solution of (1.1O)-(1.11) which has

§ 1.2. A theorem of A. Haar

wherea>O, LO, c<dand2La<d—c.

If(t,x,u,p) — f(t,x,v,q)I Lip— qI +Miu — vi

for all (t,x,u,p) and (t,x,v,q) in x Let u3 = u,(t,z) (j = 1,2) be sn

Proof. According to the Lipschitz continuity off = f(t,x,u,p), the difference

L + Mlz(t, x)I. (1.14)

Assume that z = z(t,x) takes some positive values in We put w(t,x)

Rewriting this for z = z(t,x), we get

§ 1.3. A theorem of T. Wazewski

In this section we report a theorem of T. Waiewski [154] which is a generalization

= {(t,x) O t <a, <d1 —Let (i = 1,2,... ,n)},

wherea>O, L1O, ,n.

for (t,x,u,p) and (t,x,v,q) in x u = u(t,x) (j = 1,2) be in

As preparation for the proof of this theorem, we give a fundamental lemma

for some x E with w(t) = u(t,x).

Proof. As the first step, consider the case where L1 = 0, cj = —1 and d1 = 1

0) — max —(t° a,)

We will prove that w = w(t) is differentiable from the right at t° with =

w(t°) = tz(t°,x) and

Define z(t) [w(t) — w(t°)]/(t — t°), and put

As w(t) — w(t°) u(t,x°) — u(t°,x°), we see that & (Ou/Ot)(t°,x°). By the

the sequence {(tm, Xm)}m is convergent to a point (t°, in As the functions