A NNALI DELLA
S CUOLA N ORMALE S UPERIORE DI P ISA Classe di Scienze
H ITOSHI I SHII
A boundary value problem of the Dirichlet type for Hamilton-Jacobi equations
Annali della Scuola Normale Superiore di Pisa, Classe di Scienze 4
esérie, tome 16, n
o1 (1989), p. 105-135
<http://www.numdam.org/item?id=ASNSP_1989_4_16_1_105_0>
© Scuola Normale Superiore, Pisa, 1989, tous droits réservés.
L’accès aux archives de la revue « Annali della Scuola Normale Superiore di Pisa, Classe di Scienze » (http://www.sns.it/it/edizioni/riviste/annaliscienze/) implique l’accord avec les conditions générales d’utilisation (http://www.numdam.org/conditions). Toute utilisa- tion commerciale ou impression systématique est constitutive d’une infraction pénale.
Toute copie ou impression de ce fichier doit contenir la présente mention de copyright.
Article numérisé dans le cadre du programme Numérisation de documents anciens mathématiques
http://www.numdam.org/
for Hamilton - Jacobi Equations
HITOSHI ISHII
0. - Introduction
In this paper we consider the
boundary
valueproblem
for Hamilton -Jacobi
equations:
in
or on ~ i
Here 11 is an
open subset
and h : an --~ Rare
given functions, u : fi
--~ R is the unknown and Du denotes thegradient
ofu.
The
meaning
of(solution of) problem (HJ) - (BC)
will be madeprecise
inSection 1
by modifying
the notion ofviscosity solution,
introducedoriginally by
M. G. Crandall - P.L. Lions
[7],
so that theboundary
condition isappropriately
taken into account. There are
already
several attempts of similar modifications of the notion ofviscosity
solution. For this we refer to R. Jensen[15],
P.L.Lions
[18],
M.G. Crandall - R. Newcomb[8],
H.M. Soner[20],
P.E.Souganidis [21],
1.Capuzzo
Dolcetta - P.L. Lions[4]
and G. Barles - B. Perthame[1].
Ourintroducing
theboundary
condition(BC)
isstrongly
motivatedby
theanalogy
to the Neumann condition for Hamilton - Jacobi
equations
in P.L. Lions[18].
Problem
(HJ) - (BC)
isimportant
in deterministicoptimal
controltheory.
The value functions of exit time
problems satisfy (HJ) - (BC),
where the firstorder PDE in
(HJ)
is the so-called Bellmanequation,
in theviscosity
sense, andthey
are characterized to beviscosity
solutions of(HJ) - (BC),
underappropriate hypotheses.
We prove these in Sections 3 and 5where, denoting H(x,
u,p) - u again by H (x, u, p),
we treat theproblem:
in
Pervenuto alla Redazione il 25 Novembre 1986 e in forma definitiva il 7 Febbraio 1989.
and
or on
The
boundary problem
for(HJ),
with the Dirichlet condition u = h, is not solvable ingeneral.
Acompatibility
condition on h isrequired
to be satisfied in order that the Dirichletproblem
for(HJ)
is solvable(see
P.L. Lions[16]),
andthe condition on h is
usually
hard to check. See also the recent work[9] by
H.Engler.
On the otherhand, problem (HJ) - (BC)
issolvable,
in a weak sense, underquite general hypotheses (see Proposition
1.3 in Section1).
Moreover theuniqueness
and existence of a continuousviscosity
solution of(HJ) - (BC)
isestablished under suitable
assumptions
in Sections 2 and 4. Thereforeproblem (HJ) - (BC)
seems, at least for theauthor,
a naturalreplacement
of the Dirichletproblem
for(HJ).
Problem(HJ) - (BC)
also arisesnaturally
in connection withelliptic singular perturbation problems
with the Dirichletboundary
condition orwith the
vanishing viscosity
method. In this direction seeProposition
1.2 andrefer to H. Ishii - S. Koike
[14]
for anapplication
of our results toelliptic singular perturbation problems. Finally
we remark that G. Barles - B. Perthame[2]
and P.L. Lions[19]
haverecently
treated(HJ) - (BC) independently.
1. -
Viscosity
solutionsIn this section we define the
viscosity
solution of theboundary problem
for a second order PDE:
in and
or on
and then present some of its
properties.
Here n is an open subset of is a subset of theboundary
F : n u E x R x x - R, where~N
denotes the space of real N x N
matrices,
and aregiven functions,
u : n u E -~ R is theunknown,
andD2 u
denotes the Hessian matrix of u. When F isindependent
of the lastvariables,
i.e.F (x,
r, p,~)
=F (x,
r,p),
and B =
F, problem (1.1) - (1.2)
will besimply
written ason
Our main interest here is to
study problem (HJ) - (BC). However,
it seems natural to see the connection betweenviscosity
solutions of(HJ) - (BC)
andthe
vanishing viscosity
method inlight
of thestability
property ofviscosity
solutions of
(1.1) - (1.2).
Let
S,
T and U be subsets ofsatisfying
S c U and S c T cS’.
Fora function
f :
U --; R u{ - oo, oo ~,
we defineby
sup
This function
(liS, T)*
is upper semicontinuous(u.s.c.
inshort)
on T, and we call it the u.s.c.envelope of f
restricted to S on T. We note /1 that(flS,T)*(x) ~! f (x)
for x E S andthat,
iff
is u.s.c. on S, thenI
( f ~S, T)# (x)
= for x E S. We remarkthat,
if( f I S, T)* (x)
oo forx E T, then
We define the lower semicontinuous
(l.s.c.
inshort) envelope
of f
restricted to S on Tby (fI8, T ) * = - (- flS, T ) * .
When S = T =U,
wewrite
f *
andf * , respectively,
for(fI8, T ) *
and(f I S, T ) * .
We call a function u : 11 u E -~ R Li a
viscosity
subsolution ofproblem ( 1.1 ) - (1.2)
ifu* ( x )
oo, for x E fl u E, and whenever p Eand u* - p attains its local maximum at a
point y En
uE,
thenand
or if
Similarly,
we call a function aviscosity supersolution
of
( 1.1 ) - (1.2)
ifu* ( ~ ~
> - oo, for x andwhenever p
Eand u* - p attains its local minimun at a
point
y E 11 u E, then’
/
if and
A
viscosity
solution of(1.1 ) - (1.2)
is defined to be a function on 11 u E which is both aviscosity
sub- andsupersolution
of( 1.1 ) - (1.2).
Note thatthis definition makes sense even for functions
F, B
which take values in R u(-cxJ, cxJ),
and that ifF(x, r, p, ç)
isindependent
of~,
then the spaceof "test
functions",
can bereplaced by
the spaceWe say also that u satisfies 0 in n and
B ( x, u, Du)
0 orF ( x,
u,Du, D2 u)
0 on E(resp., F ( x,
u,Du,
0 in H andB ( x, u, Du) >
0or 0 on
E),
in theviscosity
sense, if u is aviscosity
subsolution
(resp., supersolution)
of(1.1) - (1.2).
Of course, this definition is a modification of the
original
one introducedby
M.G. Crandall - P.L. Lions[7]
andby
P.L. Lions[17]
to theboundary
valueproblem.
Theboundary
condition(1.2)
is motivatedby
the work[18] by
P.L.Lions
concerning
the Neumannproblem
for Hamilton - Jacobiequations.
In ournotation,
the Neumann condition for(HJ),
in[18],
can be written asor on
where v denotes the outward unit normal of an
(assuming
itexists).
Weshould also remark that G. Barles - B. Perthame
[2]
and P.L. Lions[19]
haveindependently
introduced the same notion ofviscosity
solutions for(HJ) - (BC)
as ours. We refer to H.M. Soner
[20],
I.Capuzzo
Dolcetta - P.L. Lions[4]
foranother type of the
boundary
condition for Hamilton - Jacobiequations
and toM.G. Crandall - H. Ishii - P.L. Lions
[6],
H. Ishii[10,12,13]
and G. Barles -B. Perthame
[1]
for the work related to discontinuousviscosity
solutions andHamiltonians.
PROPOSITION 1.1. Let E be an open subset
of
Let S be afamily of viscosity
subsolutionsof (1.1) - (1.2).
Setu(x) =
sup{ v ( x) :
v ES }, for
x e n u E. Assume u is
locally
boundedfrom
above on fi U E. Then u is aviscosity
subsolutionof ( 1.1 ) - (1.2).
This is a
generalization
of M.G. Crandall - L.C. Evans - P.L. Lions[5, Prop. 1.4]
and H. Ishii[13, Prop. 2.4].
Theproof
of thisproposition
is similarto that of
[13, Prop. 2.4],
and we omitgiving
it here. Ananalogous
assertionholds for
supersolutions.
To seethis,
observe that u is aviscosity
subsolution of(1.1) - (1.2)
if andonly
if v = -u is aviscosity supersolution
of(1.1) - (1.2),
with F and Breplaced, respectively, by
the functionsand
PROPOSITION 1.2.
(Stability of viscosity solutions).
Let E be an open subsetof
an. For 0 E 1, letbe
given functions.
Assumethat, for
each 0 e 1, Us is aviscosity
subsolutionof ( 1.1 ) - (1.2),
withF,
andB,
inplace of
F and B,respectively.
Setfor
and.for
. Assume u islocally
boundedfrom
above on 11 u E.Then u is a
viscosity
subsolutionof ( 1.1 ) - (1.2)
with these F and B.This
generalizes
a result of G. Barles - B. Perthame[1,
TheoremA.2].
Of course, we have a
proposition
similar to the above forsupersolutions.
Wecan prove this
proposition
as in theproof
of[1,
TheoremA.2],
and we do notgive
the details here.The
following
observation is useful inapplications
ofProposition
1.2 toelliptic perturbation problems
or thevanishing viscosity
method.are continuous and that u E
u £)
n satisfiesF ( x,
u,Du, D2 u)
0 inn and
B(x, u, Du)
0 on E, in the classical sense.Assume,
inaddition,
that- F is
elliptic,
i.e.for
( x,
r, p,ç) E nxR xMN and p
EC2(n)
suchthat p
attains its maximumat x, and that B satisfies the condition:
B ( x, r, p) +D~(~c)) > B ( x, r, p) ,
for(x, r, p)
E E x R x Eu £)
suchthat p
attains its maximum at x. Then u is aviscosity
subsolution of( 1.1 ) - (1.2).
FromProposition
1.2 we see, for
instance,
thefollowing.
andh : all --~ R be continuous. For each 0 e 1, let Us E
C (fi)
nsatisfy -
+ 0 inn,
where A denotes the N dimensionalLaplacian,
onafi, in the classical sense. Set = lim supr10
0 e
r},
and assumeu(x)
oo, for xe fi.
Then u is aviscosity
subsolution of
(HJ) - (BC).
PROPOSITION 1.3.
(Existence of viscosity solutions).
Let E be an open subsetof
an. Letf
and gbe, respectively, viscosity
sub- andsupersolutions of
( 1.1 ) - (1.2).
Assumef
s g on 11 U E and thatf
and g arelocally
bounded onn U E.
by
= sup
viscosity
subsolutionof ( 1.1 ) - (1.2),
v g on n UE).
Then u is a
viscosity
solutionof ( 1.1 ) - (1.2).
This extends Theorem 3.1 of H. Ishii
[13].
We leave it to the reader to prove thisproposition
as theproof
is asimple
modification of that of[13,
Theorem
3 .1 ] . "
2. -
Comparison
ofviscosity
solutions. Hereafter we
study problem (HJ) - (BC).
We need thefollowing assumptions
on H and fl.(H 1 )
For each(x, p)
xthe
function u ~H (x, u, p)
isnondecreasing
on R.
(H2)
There is a continuousnondecreasing
function rral :[0,oo)
-->~0, oo), satisfying
rral(0)
= 0, such thatfor and
p E JR. N.
(H3)
There is a continuousnondecreasing
function m2 :[0, 00)
-[0, 00), satisfying m2 (0)
= 0, such thatand
(H4)
There is a continuousfunction n : fi -_+
and a constant b > 0 such thatfor and
(where B (x, 6)
is the usualsphere
centred at x withradius 6).
THEOREM 2.1. Assume 11 is bounded and that
(H 1 ) - (H4)
hold. Let uand v
satisfy, respectively,
in
on
where a > 0 is a constant, and
in
on
in the
viscosity
sense. Then:(i) If
u and vare continuous
atpoints of
a11 and h is continuous on an, then u von fi.
(ii) If
u(resp., v)
is continuous atpoints of a11,
h is l.s.c.(resp., u.s.c.)
on5n and u h
(resp.,
v >h)
holds on an, then u von-a.
REMARK 2.1. As the
proof
belowshows, the
abovetheorem
is still valid if theinequality
in(H3)
is satisfiedonly
ina
subsetof 0 x
R xR N
of theform U x R x where U is a
neighbourhood
of a11 in n.REMARK 2.2. Of course, it is
possible
to formulate acomparison
theoremlike Theorem 2.1 for unbounded domains.
The
proof
below is a modification of the arguments in theproof
of H.M.Soner
[20,
Theorem2.2]
to the currentproblem.
PROOF.
Replacing
u and vby
u* and v*, we may assume that u isu.s.c. and v l.s.c. on
fi. First,
we prove assertion(i).
To thisend,
we supposemax(u - v )
> 0 and will obtain a contradiction in each of thefollowing
threen
possible
cases. We will use the notation:and
CASE 1. Set
r E R and p, q E
Then H
is l.s.c.by (H 1 )
and w satisfies in xn,
in theviscosity
sense.Since
R=,(~r,p) > limH(x,s,p)
slr and~*(a:,r,p) limJ?(a:,5,p),
air forby (ill) - (H3),
we have’
for r, s e R
satisfying
r > s andLet
2 ,
2 and setA=max(u-v)
a ’ andfor Then we have
for and
by (H2). Also,
we havefor I
for and
for witt
Now we fix 6 so small that
max(u - v )
> A + 26 and then e so small thatand if
and Ix - yl e (this
ispossible
since w is u.s.c. andw(x,
A, for x Ean).
Observe that w - p attains a
positive
maximum at somepoint (~, y) E ~ (~)
since w is u.s.c. on
A(e)
andmax
0 A (s)(w - ~p)
0. Thus(2.3)
and(2.5), together
with
(2.4), yield
a contradiction.
CASE 2.
max(u - v) = (u - v) (z)
andv (z) h (z),
for some z E af2. Letand be as in
(H4).
Define 1by
where e is a small
positive
number. Let(~, y)
be apoint of fi x n
such that~~ (x, y) -
max4D,.
Let us assume e b. Then z +by (H4),
andhence
~~ (~, y) > ~~ (z -~ ~r~ (z), z) . 5x5
Thisyields
where m3 is the modulus of
continuity
of u and Thereforeand so
for some constant
C2
> 0.Using again (2.6),
we getFrom
this,
we seeand
for some real-valued continuous function m4 on
~0, oo) satisfying m4(0) =
0.Since
by (2.8) and, choosing
0 b smallenough, ~r~ ( y) - ~ (z) ~ 1+ M4 (s) -
b, for0 e
b 1,
in view of(2.9),
we see from(H4)
that Y E fi, for 0 eb 1.
Since u satisfies
(2.1),
in theviscosity
sense, we thus findfor
Using (2.6), (2.7)
and(2.9),
we findHence, selecting
0b 2
b smallenough,
we see that if 0 eb 2 and y E
Since v solves(2.2),
in theviscosity
sense, we now seeChoose 0
b3
min so small that for0
b3.
We henceforth assume 0 eb3.
Now(2.6)
guaranteesu(()
>V(Y),
for 0 eb 3 . Setting 6
= andcombining (2.10),
(2.11)
and(2.4),
we getThus, using (H3), (H2), (2.7), (2.8)
and(2.9),
we findthis
yields
a contradictionby selecting
e smallenough.
CASE 3.
max(u - v) = (u - v) (z)
andu(z)
>h(z),
for some z E Set_
n
_
h = -h on afi, onil,
andfor
As
remarkedbefore,
u and v,solve, respectively, (2.1 )
and(2.2)
with Hand h
inplace
of H andh,
in theviscosity
sense. Weeasily
see that(HI) -
(H3)
are satisfied for H = H and thatmax(v - 5) = (u - v) (z)
andv(z) h(z).
n
Thus the present case is reduced to Case
2,
and we have a contradiction.Now we turn to the
proof
of assertion(ii).
If we assumemax(u - v)
> 0 nand argue as in the
proof
of assertion(i),
then theassumption u
h on(resp.,
v > h onan) eliminates,
from the arguments, thepossibility
thatCase 3
(resp.,
Case2)
occurs. Meanwhile weonly
need thefollowing continuity
of u, v and h: the
continuity
of u(resp., v)
atpoints
of an and the lower(resp., upper) semicontinuity
of h in the arguments of Cases 1 and 2(resp.,
Cases 1and
3).
DA more
precise
result is needed in Section 4.THEOREM 2.2. Let a be a
positive
number and h : R be l.s.c..Assume fl is
bounded
and that(HI) - (H3)
hold. Let u and v,respectively,
beu.s.c. and l.s.c. on
fi,
andsatisfy (2.1)
and(2.2),
in theviscosity
sense. Letr be an open subset
of
an, and assume:(i)
u v, onanBr,. (ii) u
h, onr; and
(iii) for
each z E r, there is apositive
number b, a bounded sequence and a sequence C(0, 00) converging
to 0, such thatfor
and lim n-00
u(z
+ =u(z).
Then u v onfi.
The
proof
of this theorem is similar to that of Theorem 3.1. Wejust
present its outline here and leave the details to the reader.
OUTLINE OF PROOF.
Suppose max(u - v )
> 0. Let ze n
be apoint
j n
where the maximum of u - v is attained.
By assumption, z
e H U F. Ifmax(u -
anv ) max ( u -
nv),
then the sameargument,
as in Case 1 of theproof
of Theorem
2.1, yields
a contradiction. If z E r, then we go as in Case 2 of theproof
of Theorem2.1,
except that we now usewith
instead of and will get a contradiction. Thus i
COROLLARY 2.1. Let S~ be bounded and conditions
(H 1 ) - (H4)
besatisfied.
Let u and v :
0 -+
R be,respectively, viscosity
subsolution andsupersolution of (HJ)’ - (BC)’.
Then the same assertions, as(i)
and(ii) of
Theorem2.1,
hold.REMARK 2.3. We have an
assertion,
similar to thiscollorary,
whichcorresponds
to Theorem 2.2. This remark alsoapplies
toCorollary
2.2. below.-
PROOF. Let e > 0 and define u E
by
Then Ussolves
in
on
in the
viscosity
sense. From Theorem 2.1 we find that v on0
in all cases.Sending e 10,
we conclude theproof.
DCOROLLARY
2.2. be bounded and(HI) - (H4)
besatisfied.
Let u andv : 0 -+
Rbe, respectively, viscosity
sub- andsupersolutions of (HJ) - (BC).
Assume moreover that p --> H
( x,
r,p) is
convex onR N for
each( x, r)
En
x R,and that there is a
function 0
E cl(0)
and a constant a > 0for
whichfor
Then the same assertions, as
(i)
and(ii)
in Theorem 2.1, hold.PROOF. Note that h is bounded below on afi, in all cases. Therefore we may
h(x),
for x Eby adding
anegative
constant to h, if necessary.Then 0
is aviscosity
subsolution of(HJ) - (BC)
and hence themaximum of u
and 0
takenpointwise,
is aviscosity
subsolution of(HJ) - (BC) by Proposition
1.1. Now let 0 6 1 and set =8 u V ~ ( x ~ -+- ( 1- 8 ) ~ ( ~ ~ ,
for x E
n. By
theconvexity
of H, we see that wo satisfiesin
on
in the
viscosity sense.
To checkthis, let p
E and suppose attains its maximumat yEn. Suppose
further0
(this implies if y E
otherwise we are done. Thenattains its maximum at y, and hence
Thus
by (2.12);
this proves(2.13).
NowTheorem 2.1
guarantees v, onfi.
Sending
0T
1, we see that v > uV 0
> u onfi.
D REMARK 2.4.If
u(resp., v)
isLipschitz
continuous onfi,
in assertion(ii) of
Theorem2.1,
then we have the same conclusion with the weakerassumption (2.14)
below inplace of (H2)
and(H3).
(2.14)
For each R > 0, there is a continuousfunction
mR :[0, 00)
-[0, 00), satisfying
mR(0)
= 0, such thatfor and ;
Analogous
remarks are valid for Corollaries 2.1 and 2.2.PROOF OF REMARK 2.4. Assume all the
hypotheses
of Theorem2.1,
except(H2) - (H3). Assume
further(2.14) and,
fordefiniteness,
that u isLipschitz
continuous on
n,
withLipschitz
constant C > 0. Set R = C + a, and defineSet
for E n x Then G satisfies
(H2)
and(H3),
withappropriate
functions mi and m2.
Also,
u and vsatisfy, respectively, (2.1)
and(2.2),
withG in
place
of H, in theviscosity
sense. Theorem 2.1 now guarantees ourassertions. D
3. - Value functions of exit time
problems
In this section we will show that value functions of exit time
problems,
in deterministic
optimal
controltheory, satisfy
the associated Hamilton - Jacobi(Bellman) equations.
Let 0
be,
asusual,
an open subset A a compacttopological
space,f
a real-valued function xA, g
a function x Ainto RN and h
a real-valued function on an. We assume:
(A 1 )
Functionsf, g
and h are continuous. Moreover, there is a constant L > 0 such thatfor and
Let ,~ denote the set of
controls,
i.e.For X E and a
E A,
we consider the initial valueproblem
The
unique
solution of(3.1 )_will
be denotedby
=X ~t;
x,a).
The exittimes T from n and 7 from 11 are
defined, respectively, by
and
Associated with these are cost functionals:
and
With these cost functionals at
hand,
the main purpose ofoptimal
control isstated
as follows: Find controls or sequences of controls which minimizea)
and
J ( x, a )
over A. Our interesthere, however,
is restricted tocharacterizing
value functions:
and on
~
as
viscosity
solutions of theboundary
valueproblem
in
on
where
The
first order PDE in(3.3)
is called the Bellmanequation.
Note thatH : 0
x I1~ satisfies(H 1 ) - (H3),
underassumption (Al ).
LEMMA
3.1. (Dynamic programming principle).
Assume(A 1 ).
For any t > 0 one hasand
where,
for
any subset Bof A,
1B denotes the characteristicfunction of
B.We refer to P.L. Lions
[16]
for aproof
of this lemma.THEOREM 3.1. Assume
(A 1 ).
Thefunctions
V and V,defined by (3.2),
areboth
viscosity
solutionsof
theboundary
valueproblem (3.3),
with Hdefined by (3.4).
PROOF.
Weonly
prove that V is aviscosity
solution of(3.3).
Theproof
for V is
similar,
andwe leave
it to the reader. To see that V is asupersolution
of
(3.3),
let p ECl (0),
yE 0 and V* - ~p
attains its minimum at y. We mayassume
min(V* - p)
= 0by adding
a constant to p. We may also assume eithern
y E fl, or y E an and
V* ( y) h(y),
because otherwise we havenothing
toprove.
We suppose
and will
get
a contradiction.By continuity,
there is apositive
constant e > 0 such thatMoreover we assume
and
Select M > 0 so that
I g (x, a)
for(x, a)
x A. Note that, ifthen
X(s) = X(s; x, a)
EB(y, e),
for 0 ~ s2M.
Sett =
E ,
and choose2 a S
> 0 so thate(l - e - t )
> 26.By
Lemma3.1,
for anyx E
B (y, ±) nO,
there is an a = a ~ such thatif if
Therefore,
if t r, then we haveAlso,
if t > r, then we haveIn both cases, we thus obtain
a contradiction.
Similarly
we can check that V is a subsolution of(3.3),
and we omitgiving
theproof
here. D4. -
Continuity
ofviscosity
solutionsWe continue to
study problem (HJ)’ - (BC)’
for Bellmanequations.
Let0, A, f, g
and h be as in§3,
and .H be the function definedby (3.4).
Wepresent
a sufficient condition under which
(HJ)’ - (BC)’
has a continuousviscosity
solution. In order to state our
assumptions,
we need the notation: For a subsetS we write for the closed convex hull of S.
Throughout
this paper, the term "cone" will stand for a cone with vertex at theorigin.
For a subset Sof R N and s > 0, we denote
by cn, S
the e - convex conicneighbourhood
ofS, i.e. the set
We note that
cneS
is a closed convex cone We setfor
The
following
arepieces
of the sufficient condition mentioned above. Let z Ean and A c aQ .(A2)
There is an open convex cone K and a constant e > 0 for whichG (z)
c Kand
(z
+K)
nB(z,,6)
n H = 0.(A3)
There is an open convex cone K and a constant c >0
such thatG(z) nK=l 0
and(x+K)
c fl, for xc-B(z,,E) n?i.
(A4)
There is an open convex cone K and a constant e > 0 for which for and(
for
(A5)
There is an open convex cone K and a constant c > 0 such thatG(z)
and(z+K)
nil =0.
It is easy to check that if z is an interior
point (relative
to of A, then condition(A4)
isequivalent
to(A4)’
There is an open convex cone K and aconstant
e > 0 such thatG(z)
c Kand
(~ + ~) nB(z,e)
c0,
for x EB(z,e) n n.
THEOREM 4.1. Assume
(A 1 ).
Let r and E be,respectively,
an open anda closed subsets
of
ao such that ao =Let fo
be anopen subset
of
r. Assume that(A2), (A3), (A4),
with A =ro,
and(A5) hold, respectively, for
z EE,
z Er,
z Ero
and z ErBr o. If
u is aviscosity
solutionof (HJ)’ - (BC)’,
with Hdefined by (3.4),
then u EC(fl),
andulll
has aunique
continuous extension
to a
which is also aviscosity
solutionof (HJ)’ - (BC)’.
It may
happen
that some ofF, E
andFo
are empty in this theorem.Proposition
1.3 or Theorem 3.1guarantees
the existence of aviscosity
solutionof
(HJ)’ - (BC)’.
Thus we see, from this theorem,that under
the abovehypotheses
(HJ)’ - (BC)’
has aunique viscosity
solution inC (0).
The
assumption
of Theorem 4.1 ismotivated by
thefollowing
observation.Let T > 0, and
set li = (0, T) x fi.
Let E = =(0,
andro = {T}
x Q.Clearly E
and r are,respectively,
a closed and an open subsets of 9H andro
is an open subset of F. Letf and g
be as above andsatisfy (Al).
Let A E R, and definef :
x A - R andg :
x A -by
=
and g(t, x, a) = (-1, g(x, a)),
for(t,
x,a)
E x A.The associated Hamiltonian
given by
for t,
s E R and x, p and thecorresponding
Bellmanequation
is:On the other
hand,
if(A3)
and(A5)
hold for z E afi, then(A2), (A3),
(A4),
with A =Fo,
and(A5) hold, respectively,
for z EE, z
EF, z
Efo
andz E
fBfo.
This observation and Theorem 4.1imply
that if(A3)
and(A5)
holdfor z E ao and h is a
given
continuous functionona fi,
then theproblem
in
on
has a
unique viscosity
solution in where Notethat this
problem
for u isequivalent
to theproblem
in
on
for where Also note
u C .r.
that the
following
theorem and Lemmas 4.1 and 4.2assert
that the solution of thisproblem
does notdepend
on the values of h onFo
andequals
to h on ~.Thus our result covers the
initial-boundary
valueproblem
for Hamilton - Jacobiequations.
THEOREM 4.2. Assume the
hypotheses of
Theorem 4.1. Let u and vbe, respectively, viscosity
sub- andsupersolutions of (HJ)’ - (BC)’,
with Hdefined by (3.4).
Thenon
Moreover, (ull1
and(vIOBfo,O)*
are,respectively, viscosity
sub-and
supersolutions of (HJ)’ - (BC)’.
For the time
being,
we assume Theorem 4.2 is correct and prove Theorem 4.1.PROOF. Let u be a
viscosity
solution of(HJ)’ - (BC)’. By
virtue of Theorem4.2 we have
on
while on
ii, by
definition. Thereforeon
Thus is continuous
on n
and is aviscosity
solution of(HJ)’
-
(BC)’, by
Theorem 4.2. Theuniqueness
followsimmediately
from Theorem2.1. D
In what follows we
always
assume thehypotheses
of Theorem 4.2 and will prove the theorem.LEMMA 4.1. One has
for
PROOF. We
begin by observing
thatwhere ,
Fix z E
By translation,
we may assume z = 0. Sinceby assumptions (A2)
and(A5)
there is an open convex cone K and an e E(0,1)
such thatand
Choose finite sequences and so that
and Set and for
Note that .... -- for
By
thecontinuity
of g,we see
that e (x)
EK,
forx
in aneighbourhood
of 0. SetCc =
x
E B(0, E) }, E,
=cnE, CE
andKa
=cneEe.
ThenCE , Ee
andKe
are nonemptyclosed convex cones.
Replacing
e > 0by
a smaller one, we may assume thatand
for
Define d :
R N -+ 0, oo ) by
= dist( x, E~ ) ,
for x EJR N.
It iswell-known that d e and
for ~ 1
where
P ~ x)
denotes the nearestpoint
ofE,
from x.(One
way to seethese,
forinstance,
consists inobserving
thatand then
applying general
results(e.g.,
H. Br6zis[3, Prop. 2.11])
for convexfunctions to
d2 ).
Note also that d ispositively homogeneous
ofdegree
1.We claim that
for
In view of
(4.4), by continuity,
it isenough
to prove(4.6)
for x EK~ .
Tothis
end,
recall that and set thenAssume x E
Kc
and that~P(x)~ > ,8lxl.
ThenAs and we find that , The
homogeneity
of dyields that
. ~
Thus we have
proved (4.6).
Let x E
E~
and p ECa satisfy Ipl
= 1.Let q
EB ( o,1 ) .
Since p + eq EE,
and
P(x) ~ (P(x) - x)
= 0, we haveTaking q
= with x EE~ ,
in thisinequality,
we find thatfor and
Now fix M > 0, so that
Let a > 0, and select a constant C > 0 so that
Let p E
E~,
and set for withto be fixed later on. Since we have for
and
hence, by (4.4),
By
the same reason, we havefor
Therefore, using (4.6),
we see that if B >2 max 1!2, ~ } ,
then w > h onB (0, e)
n an and w > u* onaB(0, e) nn. Using (4.7)
and(4.5),
we calculatethat
for Fix
and set
no
= n nB(0, e)°
andho (x)
=w (x),
for x E(9f]o.
Then z,v satisfiesin
on
in the classical sense, while u* is a
viscosity
subsolution of(HJ)’ - (BC)’,
with00
andho
inplace
of n andh.
Asimple
argument which leads to a contradictionnow shows that u *
on 00 and,
inparticular, u*(0)
+ a +B d (- p).
Sending
p - 0, whilekeeping
p EE~,
and then a10,
we conclude(4.2).
D LEMMA 4.2. One hasfor
Our
proof
is similar to the above one, and weonly give
its outline here.OUTLINE OF PROOF. We may assume z = 0. Fix e > 0. We set and
By assumption (A2),
we can choose an e > 0 so thatand
for
Let and
B > 0,
and defineby
There is a
positive
constantBa , independent
of p, such that if B >Ba ,
then w
h onB(0,e)n80,
z,v v*on 8B(0,e)nn
and 0in in the
classical
sense. Thus weconclude,
as in the aboveproof,
that w v* on
B (0, e) n a,
and henceh(0) v* (0).
DIn what
follows,
we writeand
We note that u u * and v > v * on an and that û = u * and v = v * in fi.
LEMMA 4.3. Assume
u* ( x ~ > - C’ for
xE fi
and some constant C > 0.Then, for
each z E r, there is an open convex cone K and a constant e > 0 such thatand
for
PROOF. Let z E r. We may assume z = 0.
By assumption (A3),
there isan open convex cone K and a constant e E
(0,1)
such thatG(0)
nK ~ ~
andfor We select finite sequences
and so that
and Set
and for
Also,
set and defineand
Replacing e
> 0by
a new one, we may assume that forand
for
In view of
(4.11),
it isenough
to show that(4.9)
holds for K =Feo.
Tothis
end,
we will provethat,
for any r E(0,e],
for where
and
Note that
(4.9),
with K =Fz,
is a direct consequence of(4.13). Also,
notice thatLet r E