SECOND-ORDER HAMILTON-JACOBI
EQUATIONS
IN INFINITEDIMENSIONS*
PIERMARCO CANNARSA AND GIUSEPPE DA PRATO$
Abstrnet. Somesecond-order Hamilton-Jacobiequationsconnectedtostochasticoptimal control prob-lemsfor infinite-dimensional systemsdrivenbyawhite noisearestudied.Adirectmethodtoproveexistence and uniqueness ofmildsolutionsisdeveloped.Then thissolution isidentifiedasthe valuefunctionof the related stochastic controlproblem,andafeedback formula for optimal controlsisderived.
Keywords. Hamilton-Jacobi equations,stochastic optimal control, dynamicprogramming, viscosity solutions,whitenoise,infinitedimensions
AMS(MOS)subjectclassifications. 49C10, 49A60,93E20
1. Introduction. Second-order Hamilton-Jacobi equations in infinite dimensions have been studiedbyseveral authorsin connection withthe stochasticoptimalcontrol of distributed parameter systems; see
[Lecture Notes
in Mathematics, Vol. 1390,Springer-Verlag, Berlin,
1989]
and the references quoted therein. Most of the workson this subject concern systems governed by stochastic partial differential equations drivenby aHilbertspace-valuedWienerprocess. Inthis paper,wefocus our attention onthe case of stochastic systems that are drivenby awhite noise. Forsuch problems
fewer results are available in the literature.
In order to explain the context we have in mind, let X be a separable Hilbert
space, and consider the problemofminimizing:
{fT"
12
}
(1.1)
J(t,x;
z)=
[g(y(s))+-glz(s)l
ds+ch(y(T))over all controls z
Mw(t,
T; X)
satisfyingIz(s)l
<-R almost surely for all s t,T].
Here e,
R,
and Tare given positive numbers and g,b:X-
Rarebounded uniformlycontinuous functions. In
(1.1),
y is the mild solution of the stochastic differential equationdy(s)=(Ay(s)+F(y(s))+z(s)) dt+/-[ dW(s),
t<-_s<=T,
(1.2)
(t)
x,where W is a cylindrical
Wiener
process(or
white noise) on a probability space(l),
o,
P). Moreover,
M(t,
T;
X)
denotes the space of the X-valuedprocessesx(s)
that are adapted to W and satisfyIx(s)
ds<.
As
iswell known,the dynamicprogramming approachtoproblem(1.1), (1.2)
consistsofstudyingthe value
function
V,
defined as(1.3)
V(t,x)=inf{J(t,x;z)’zM(t,T;X),lz(s)l<=Ra.s.
Vs[t,T]}.
ReceivedbytheeditorsDecember 4, 1989;acceptedfor publication(inrevisedform)April25, 1990. DipartimentodiMatematica,Universit/diPisa,ViaF.Buonarroti,2,56127 Pisa, Italy. This research was partially supported by the Italian National Project M.P.I. Equazioni Differenziali e Calcolo delle Variazioni.
$Scuola Normale Superiore, Piazza deiCavalieri, 7,56126 Pisa,Italy.
Thisresearchwaspartiallysupported bythe Italian National ProjectM.P.I.Equazionidi Evoluzione eApplicazioni Fisico-Matematiche.
474
The function
u
(t, x)
V
T- t,x)
isrelatedtothe Hamilton-Jacobi-Bellman equationTr(u,x)+(Ax+F(x),u,)-H(ux)+g(x)
in]0,T[xX,
(1.4)
at 2 whereu(O, x)
4(x),
Ipl
ifIpl
R,
H(p)
R
2Rip
I--
ifIPl
->-R.The maingoalof thispaperis todevelopadirectmethod ofsolution forequation
(1.4).
By
"directmethod"wemeanamethod that makes no use of the control theoretic interpretation ofproblem(1.4).
More
precisely, we will solve the above problem as an initial value problem for a semilinear parabolic equation.Then,
after havingidentified the direct solution of
(1.4)
as the value function(1.3),
we can transfer information from apartial differential equation context into a variational setting, byderiving afeedback formula for optimalcontrols.
Wenow explain the main ideas of our method.
As
in finite dimensions, we firstconsiderthe linearproblem
Tr(ux)+(Ax, u)
in]0,T[xX,
(1.6)
at 2u(O,x)=(x)
whose solution can berepresented by the probabilistic formula
(1.7)
u(t,x)=
d e’ax+x/-
e(’-)adW(s)
=:(Ttb)(x).
Indeed,
whenA
is self-adjoint, strictlynegative andA
-
is nuclear, it is shown in[7]
that
(1.7)
isthe unique classical solution ofproblem(1.6).
This resultisrecalled and improved in 3 of this paper, by proving the uniform convergence of some finite-dimensional approximations of(1.7).
Then,
we define a mild solution of(1.4)
as asolution ofthe integral equation(.a
u(,-=
r,+
r,_((F,
u(s, .)-g(u(s,
.+g
s.
We
solve(1.8)
by fixed-point argumentsin aspaceof functions which are C inx for>
0and satisfya suitable blowup condition inzero,(see
4).
Inorder to applythe above result to problem
(1.1), (1.2)
we have to identifythe functionu(T-t, x)
as the value functionV(t, x).
This could be done by standard verification techniques, computing theIt
differentialdu(T-t,y(t)),
if u were suciently smooth and the covariance ofWhad finite trace.Toovercome thisdiculty,westudyasuitablefinite-dimensionalapproximation of
(1.4),
for whichthesmoothnessof the solution
u
iswell known.Wethenapplythe Itg formulatou
and passto the limit as n.
To makethisprocedure rigoro,us,it is essential toshow thatu
convergesto
u,
uniformlyon the bounded setsof[r,
T]xX
for all 0<<
T
(see
Theorem4.5).
The techniques of this paper could be easily arrangedto studythe equation Ou e
Tr(Quxx)+(Ax+F(x),ux)-H(u)+g(x)
in]0,T[X,
(1.9)
Ot 2u(O, x)= 4(x),
where
Q
is a self-adjoint positive nuclear operator in X. This problem is related to the optimal control ofa system driven by a "genuine" Wiener process. Unlike(1.4),
for which no other result seems available in the literature, equation
(1.9)
has been consideredbyseveral authors.In 1],
problem(1.9)
is studied withF 0and assuming g and&
to be convex. In[9],
the case ofA
=0 is treated by the theory ofabstract Wienerspaces. Notethat,eventhoughthe equation considered in[9]
looks like(1.4),
it is indeed equivalentto
(1.9).
The general equation(1.9)
is solved in[5]
by using a probabilistic formula like(1.7).
In
this case,however,
we do notget C regularity, butonly differentiabilityin some special directions related to
Q.
Weconclude this Introductionby recallingthat several resultsonviscosity solutions are availabletodayfor Hamilton-Jacobi equations in infinite dimensions
(see
[4]
for first-orderequations). Second-order equations havebeen treatedbyLions in10]-[ 12].
Inthistheory,the existence of solutions to
(1.9),
forweaklycontinuousdata,isusuallyobtained by variational methods based on the representation formula
(1.3).
2. Preliminaries. LetXbeaseparableHilbert space, with norm
I.
ForanyR>
0,we set
BR
{X6 X;
Ix[
R}.
For any x,y X we denote by x(R)y the operator defined by
x(R) y z
(y, z)x.
Let Y
be another Hilbertspace.We
denote byCb(X, Y)
the Banach space of all boundeduniformlycontinuous mappings4:
X- Yendowed with the sup normI1" 11o.
Likewise,
Cb
h(X,
Y),
h 0, 1, 2,..,
endowed with the natural normI1"
I1,
is the setof all the mappings
b"
X- Y which are h times Fr6chet differentiable and such that the kth derivativeb
(k is uniformly continuous and bounded for all k-<h.Moreover,
we set
Chb(X,
R)=
C(X).
For any
&
Cb(X)
we denote by w6 a continuity modulus ofb, i.e., continuous functionw6:[0,
ee[-[0,
oe[
satisfyingw+(0)=0
and such that[b(x)-&(y)]
-<o+(Ix-yl)
for allx, y
X. It is well known that any functionb
Cb(X)
possesses a concave continuity modulus.Lip
(X,
Y)
is the space of all Lipschitz continuous andbounded functions fromXto
Y,
endowed with the norm"bl’
sup{
’b(x)
b(y)’
}
[x_y
;x,yX;xCy
/11 11o,
Throughoutthe wholepaperwefix acompleteorthonormal system in
X,
denotedby
{ek}u.
We define the projection 17, ofX
onto the span of{el,
e2," "’,e,}
asfollows:
(2.1)
1-In
e
(R)e
Vn
N.k=l
Indeed, setw(t) sup{[b(x) 4’(Y)I;
Ix-
Yl
=<t}.Then o) is a nondecreasing subadditive continuitymodulus forb. Sowecancheck that theconcaveenvelopeofw has the required properties.
Now,
let{Ck}
be asequence of positive real numbers. Then there exists auniqueself-adjoint operator
A
in X such thatAek
=--ckek.As
is wellknown, A
is denselydefined and closed with domain
D(A)
x eX"
Y
ak(X,
ek)
2<
ook=l
Moreover,
sinceA
is negative,A
generates an analytic semigroup eta in X and(2.2)
etax
,
e-tk(x,
ek)
k=l
for all x X.
Consider now a complete probability space
{f,
,
P}
and a sequence{/3k}
of standard one-dimensional Brownian motions, mutually independent.We
denote byWn(t)
the n-dimensional Brownian motion given by(2.3)
W"(t)=
flk(t)ek
k=l
for all ->0. We set
(2.4)
(2.5)
W(t)=
ek e-"kt-)dk(s),
k-=lIo
WA(t)
k
ek e-(t-s)dflk(S).
=1We
note thatW(t)
is the stochastic convolutionIo
WnA( t)
e(’-’)AdW’(s).
In general,
(2.5)
is meaningless since the series in theright-handsidemaynotconverge.The following proposition shows that it becomes meaningful,under some restrictions
onthe sequence
{Ck}.
PROPOSITION 2.1.
Assume
that(2.6)
2
1k=l
Then,
theseries in(2.5)
converges inL2(),
if,P;
H)
for
all >-O.Moreover,
Wa(
t)
isa Gaussianprocess with mean zero andcovariance operatorQt
given by1 e
(2.7)
Qtx
E
(x, ek)ek,
xX.
k=l
2Cek
Proof
Forall 0, we have[Io
E
V
e-"(’-)d(s)
e -2"as
Z
k=l =1 k=l
which is finite in view of
(2.6). Therefore,
the series in(2.5)
converges in X for all>
0 almost surely to a GaussianprocessWa(t).
In orderto prove(2.7)
it sufficestoremark
that,
for allx,
yX,
we haveIo
W(
t), x)(
W(
t),
y)
2
(x, e)(y, e)
e-(’-’ds.k=l
Remark 2.2. If
(2.6)
is fulfilled, we write the stochastic convolution(2.5)
asfollows:
WA( t)
e(’-s)AdW(s),
where
W(t)=
k=l
ekk(t)
is usually interpreted as a white noise.In
ordertoshowthatWA(t)
has continuous trajectories, we will strengthen(2.6),
assuming
2o’-1
(2.8)
Y
ce
<
oofor some cr
]0,
1/2[.
We set1 e-2kt
q(t)--k=
20k
We notethat(2.8)
yields(2.9)
q(t)-Mt
20.for all 0 and someconstant
M
0.PROPOSITION 2.3.
Assume
(2.8).
en
WA(t)
hasa-Hldercontinuous trajectoriesfor
alla]0, [.
Proo
Since{k)
are independent, we havelWa(t)-
WA(S)I===,
e-"(t-o)d(p)
+
:, e-%-)d(p) -2
Z
e-"(’-) dill(p) e-"(’-)dill(p) =1=Z
e-2"0++
e-2"p+-2
=1 =1 =1for all s 0.
Now,
by changingthe variable p with+
s 2p, we obtain(2.10)
glWA(t
WA(S)12=q(t)+q(s)+2[q(7) -q()]
for all tsO.
Next,
notethat,sinceq(t)
q(s)
q(t-
s),
(2.9)
yields qeC2([0,
m[). Therefore,
there exists C
>
0 such thatFrom
(2.10)
it follows that1
(
(1
c(
+-
.
Since
W(t)- W(s)
is a Gaussianprocess,W(t)-
W(s)lm
C’l-x
m for all t, s 0 and a suitable constantC’.
The Kolmogorov testyields the conclusion.PROPOSITION 2.4.
Assume
(2.8).
en,
for
allT> O,
Proof
We usethe factorization method as in[6].
SetY(s)
2
ek(S- r)
e-k(s-r)dk(r),
k=l
for all s _->0.
Then,
by astraightforward computation,(2.13)
WA()
sino
(t- s)
-
e(’-g(s)s,
(2.14)
W(t)
sin(c-s)
-
e(’-Y(s)
ds,
where
An
AIIn.
Therefore,
(2.15)
wherew(t)
w2(t)
n.(t)
+
c.(t),
sinBn(t)
|(t--s)-I[e(’-)A--e(t-s)A"]y(s)
ds,
.o
io
Cn(t)
sinro(t-s)
-1e(t-’A"[Y(s)
Yn(s)]
ds.Wewill estimate
B(t)
andC(t)
separately. After some computations we obtain(2.16)
l
g(s)l
2=2
e-2ak(s-r)(s
r)
-
drNko
2
1_2=:kl
k=l
J0
k=l kfor some constant
ko>
0. SinceY(s)
is a Gaussianprocess,(2.16)
impliesthat,
for allmN,
(2.17)
lY(s)lk
for some constant
k
>
0.Now,
by H61der’s inequalityand(2.13)
it follows that supWA(t)l
2mS2m(-l)/(2m-1)ds
1
g(s)l
mds.OtT 0
Moreover,
we conclude that supIBn(t)
2ms2m(-l)/(2m-1)]]esA--eSAl2m/(2m-1)
dskONtT
g(s)l
mds.0
Now,
since the semigroup e"
is analytic,lie
-
e’ll
0for all s>
0 as n.
Thus,
in view of
(2.17),
the dominated convergence theorem yields(2.18)
lim
(o,rsup
IB.(,)I
m)
=0provided that
(1-)2m/(2m-1)<l.
We will now estimateC(t):
we haveg(s)-
g(s)
=
e-(’-(s-
r)
-
drk=n+l
N
ko
_"k=n+lk
for some constant
kln>O
such thatlimn kin--0.
SinceY(s)-
Y,(s)
is a Gaussian process,(2.19)
implies that,for all mN,
(2.20)
l
Y(s)-
Yn(s)l
2mcm(kln)
for some constant Cm
>
0.NOW,
by H61der’s inequalitywe conclude that(
sup[C,(t)[
2m)
(si7)2m(ff0T
)2m-1
S2m(-l)/(2m-1) ds kOtT g()-gn(s)l
md.Hence
(2.21)
,-lim
(,
0__<,__<[C,(t)l
2m)
=0,which, along with
(2.18),
gives the conclusion, lq3. Linearparabolic equations.
In
this section we studythe linear problem:Ou e
Tr(u)+(ax,
u)
in[0, T]X,
(3.1)
Ot 2u(O, x)
4(x),
where e is a given positive number and
h
C(X).
Here
A:D(A)=X-X
is a self-adjoint negative operator in X satisfying,for all kN,
(3.2)
Ae
-ce,with
ce
>
0. In(3.1)
thesubscript x represents Fr6chet partial derivative with respectto x andTr denotes thetrace; i.e.,
Tr
(u)
2
(Uxe, e).
k=l
The following result, provedin
[7],
statesthat,for any4
C(X),
problem(3.1)
has aunique classical solution givenby
(3.3)
u(t, x)
g(ch(e’Ax
+x/- Wa(t)))=:
Td)(x),
where
WA(t)
istheprocessdefined in(2.5).
The method of[7],
based onaprocedure ofGalerkintype,usesthefollowingsequenceofsemigroupswhichisprovedtoconvergepointwise to the semigroup defined in
(3.3):
(3.4)
(qb(etAI-[nXq"r
W(t)))=:
(TTqb)(x)
for all
b
C,(X). In
thispaperwe improve the result of[7],
by showingthat(T’;4)(x)
converges to
(Ttch)(x)
uniformly. We will use this convergence result in 5. In thefollowing we denoteby
etaQf
thebounded operator defined by2ak
e-tk(3.5)
e,AQ
-
ek -2,,k ek, k N.1-e
PROPOSaqON 3.1.
Assume
(3.2)
and(2.6). Then,
for
anyr]0, T[,
(i) lim
T’
6)(x)
(Tt4)(x),
(3.6)
(ii)
lim(T’6)x(x)
(T,6)x(x)
(3.7)
(3.8)
(3.9)
(3.0)
whereuniformlyon the boundedsets
of
r,T]
X.Moreover,
thefunction
u
:[0,
oo[
X--->R, u(
t,x)
Tt6
)(x)
is continuous.
Furthermore,
u(
t,C(X)
for
all> O, u(., x)
C(]O,
oo[)
for
all xD(A)
andu,,(t, x) =v/7
(e*AQ-[
WA(t)cb(etAx
+v/7 WA(t))),
U(t,
X)=
e(etAQ-[
W(t)(R)etAQ-[ WA(t)cb(etAx
+v/7
W(t))),
lu(
t,x)l
<-,/7
p(t)ll,/,
IIo,
ITr
[u(t,
x}]l--<
(t}
IIo,
(3.11)
/92(t)
Z
20ek
e-2tck
1 e-2t%
k
Finally,
(3.1)
.isfulfilled
for
all xD(A)
and>
O.Proof
First,we note thatformulas(3.7)
and(3.8)
arederived in[7].
Wewillgivea detailed proof ofthe uniform convergence in
(3.6).
The proofs of the remaining statements will be only sketched for the reader’s convenience.From
(3.3)
and(3.4)
we obtainI(
T,cb)(x)-(
T74,)(x)[
<-g149(etAx
q-x/7Wa(t))-
p(etAI-[nx
+x/7
W.(t))[
(3.12)
_--<o
(le
tAx e’al-l,x[
+,/7
WA(t)
WA(t)I
<--oo(le’Ax-e’ArI.xl+,/7 l
Wa(t)--where to, is aconcave continuitymodulus for qS.
Now,
since e’A is compact for>
0,(3.13)
lim[etax etaIInX[
--0uniformly for
(t,
x)
e r,T]
xBR,
R
>
0.Moreover,
(3.14)
gl
Wa(t
W"A(t)l
<--4-
,/l
Wa(t)
WA(t)I 2.
Thus, (3.6)(i)
follows from(3.13)
and(3.14),
in light of Proposition 2.1. Equation(3.6)(ii)
followsbyasimilar argument,using formula(3.7)
forux(t, x).
Next,
wehave[(Ux(t.x)
e)12=ee_2,%(
2ak
)2
fO
2
1_e-2,% g
e-2(’-*%
dk(S)Ch(e’Ax+
WA(t))
(3.15)
<=
ee_2,.(
2a
)
2Io’
22
1 e -2’% ge-2’-
)%d/3g(s)
b
Iio
and
(3.9)
followstaking the sum overk. Similarly,(3.8)
yields!
I(ux(
t,x)e e}
e-2’I
\1--e-2t%]
12
-2(t--s)akd[k(S
dp(e’Ax+
W(t))
8e-2tak 2ok 1_----2,% I111o,
which in turn implies(3.10).
(3.17)
where
Remark 3.2. Under the strongerJcondition
(2.8),
we have2
(20lke--tak)
2e
-2tak --t‘L
(t)<_-sup
qt)<_-- q),
k1--(3.18)
L=4max e_2,. 0 1Hence,
estimate(3.10)
and(2.9)
yield(3.19)
ITr
[lgxx( t,X)]
t2_2c
4. Semilinear parabolic equations. Let T>0 and consider theCauchy problem
Ou
--=-Tr(uxx)+(Ax+F(x),ux)-H(ux)+g(x)
in[0, T]xX,
(4.1)
Ot 2u(O,x)=6(x),
wherewe assume the following, in addition to
(3.2)
and(2.8),
(i)
6,gCb(X),
(4.2)
(ii)
H Lip(X),
H(0)=
0,(iii)
FCb(X; X).
Obviously, the requirement
H(0)-0
implies no loss ofgenerality, as we can replaceg by
g-H(0).
We will solveproblem(4.1)
in the Banach space5;
{v
Cb([O,
T]X)"
v)C(]0,
T]X;
X),
vx
B(]O, T]
xX; X)},
i[vll,.-sup{lv(t,x)l/ltl-vx(t,x)l
(t, x)@ ]0, r] xX},
where er is defined in
(2.8)
andB(]0, T]
xX;
X)
denotes the space of all boundedX-valued functions defined in
]0,
T]
xX.DEFINITION 4.1.
A
function uZ
is called amild solution of problem(4.1)
ifu is asolution of the integral equation(4.3)
u(t,’)=Tt6+
Tt_,((F, ux(s,.))-g(u(s,.))+g)as
Vt[O,T],
where
T,
is the semigroup defined in(3.3).
LZMMA 4.2.
Assume
(2.8)
andletq" ]0, T]
x X-R
be such that(i) qt
Cb([%
T]xX)
for
all’6]0,
T],
(ii)
It
1-’O(t,
x)[
<-K
for
all(t, x)
]0, T]
xX
andsome constantK>
O. Set(4.4)
f(t, )=
Tt_s(tP(s, ))
dsVt
]0,
T].
Then
f
2,.Proof
Step
1.f
Cb([O, T]
xX).
Fix e>
0and let-
]0,
T]
such that K"
=<
ere.Let
(t,x), (t’, x’)
[0, T]xX.
We
shall consider two cases separately.Case 1. t,
t’ [0, r].
Then wehave,
obviously,[f(t,x)-f(t’,x’)[<=
ITt_,(sl-4,(s,.))(x)ls
*-1ds+
IT,_s(s-4,(s,.))(x’)]s
-
ds<=
2K s-1ds <- 2e.Case 2.
It/2, T[, t’ [r, T].
Then we haveIf(,x)-f(’,x’)l<-2e/
r,_,(O(s,
.))(x)
ds-r,,_,(4,(s,
.))(x’)
dsr/2 r/2
In view of
(ii),
standard continuity properties of the integral in the right-hand side imply that there exists 8>
0suchthat,
ifIt
t’
+
Ix
x’
<
a,
then_(4,(s,"
))(x)
ds- ,,_(q,(s,.))(x’)
ds<
.
To conclude the reasoning, set 6’=min
{6,
r/2}.
Then,
from the above analysis itfollows that
]f(
t,x)
f(
t’,
x’)l
<=
3e provided]t-t’l+lx-x’]<6.
This proves thatf6
Cb([0, T]
X).
Step
2.t-fx
(t,
x)
isbounded on]0, T]
X. First, notethat, by(3.9), (2.8),
and Remark3.2, we obtainCoK
(4.5)
](T,_sq,(s,.
))
<- 0<s<
t,--(t__s)l-O-sl-Cr,
where
Co V’eLM,
L and M being defined in(3.17)
and(2.9),
respectively.Hence,
fx(t, x)
exists for all t>0 and x eX.
Moreover,
by(ii),
(4.6)
It’-fx(t,x)l<-t’-Co
K(t-s)-’s
-’ds. On the otherhand,(4.6)
yields the conclusion ofStep
2 since22(-)
(4.7)
(t-
s)-’s
-Ids
t2-’fl
(o
",o’)
_<-2-1,
where/3
is the Euler beta function.Step
3.fx C
([
to,T]
xX;
X)
forallto
e]0,
r[.
Fixto
e]0,
r[
andto
--<
_-<t’
_-<T,
x, x’
eX. ThenIfx(t,x)-fx(t’,x’)l
(4.8)
[(
T,_.,(s,.
))x(X)-(
r,,_,q,(s,.
))x(X’)]
as
"(
Tt,_,,lp(s,))x(X)
dsMoreover,
recalling(4.6),
we obtain(4.9)
"
T,,_,b(s,
))(x)
ds <--CoK
(t’--s)-ls
-
ds <-Co--
to
(t’- t)
.
O"
On the otherhand, for all r/>0there exists
-
]0, to/2]
such that(4.10)
[(:r,_q(s,
.))(x)-(T,,_sO(S,
.))x(X’)]
as
----<27+
[(r,_,O(s,
"))x(X)-(t,_,O(s,
.))(x’)]
dsIndeed,
it suffices to take"
so thatFinally, the conclusion follows from
(4.8)-(4.10)
and the uniform continuity ofthemapping
(s,t,x)->(Tt_sO(S,’))x(X)
on{(s,t,x):to<=t<-T,
’<-_s<=t-",x6X}.
Theproofof the lemma is thus complete.
THEOREM 4.3.
Assume (3.2), (2.8),
and(4.2).
Then problem(4.1)
has a uniquemild solution.
Proof
Suppose
firstthat Tis sufficientlysmall,
i.e.,1
(4.12)
(1
+
22(1-a)Co)
IIFIIo+
IlHIll
T_
(r
2’
where
Co
x/eLM,
L and M being defined in(3.17)
and(2.9),
respectively. Define a mapF
onE
as follows"(4.13)
(Fv)(t,
.)=
Ttch+
T_.((F,
vx(s, "))-H(vx(S,
"))+g)
ds for all[0, T]. From
Lemma 4.2, it follows thatF"
E-*E.
Moreover,
(4.14)
I(rv)(t,x)-(rz)(t,x)l
IlFll+llslll
Tllv-zll,
(4.15)
tl-l(rV)x(t,
x)-
(rZ)x(t,
x)]
Co(
Ell0
+
gII,)(,
)TII
v-zll,
where/3
isthe Euler beta function. Sincefl(cr,
O’)
=<22(1-)/O
",(4.13)-(4.15)
imply thatF
isacontraction inE
and the conclusionfollowsbythe contraction mapping principle. Finally, condition(4.12)
can beremovedbyafinitenumber of iterations of thepreviousfixed-point argument.
In the sequelwe will considerthe following "finite-dimensional" approximation
of
(4.1):
OUn
(4.16)
Ot ETr(un,xx)+(AHnx+HnF(Hnx)
Unx)-H(unx)+g(IInx)
Un(O X)"--
I(HnX),
which has the integral form
(4.17)
u,(t,
.)
T’4
II,
+
TT_s((FHn,
un,x(S,.))-H(un,x(S,.))+gHn)ds.
THEOREM 4.5.
Assume (3.2),
(2.8),
and(4.2)
and letu andun
be the solutionsof
(4.1)
and(4.16),
respectively.Then,
for
all"
]0, T[,
(i)
lim]u(t,x)- u,(t,x)]
=O,
(4.18)
(ii)
limlu,(t,
x)
u.,,,(t, x)]
0 uniformlyfor
’,T]
and xin boundedsetsof
X.Wefirst prove the following lemma.
LEMMA 4.6.
Assume
(2.8)
and let4’,,:
]0,
T]
X.R,
nN,
be such that(i) On,
qCb([’,T]xX)
forallz]O, T]
(ii)
It’-q(t,x)l=<K, It’-q(t,x)l=<Kforall
(t, x) ]0, T]
Xandsome constant K>0.(iii)
lim,_.oosup{Itl-((t,x)-q,,(t,x))l:
t[0,
T],lxl<-R}=O
for
all R>0.Set
f(
t,T,-sO,)(s,
ds,
Then,
for
allR >
0f(
t,Tt_O)(s,
ds.(4.19)
limIf(
t,x) f(
t,x)l
O,
(4.20)
limtl-lL,
x(t,
x)
-fx(t,
x)l
0uniformly
for
[0,
T]
andIx[-<-
R.Proof
First,we note thatfn,
f
Z
in view ofLemma
4.2.Now,
fix R>
0and let 6[0, T],
]x
_-<R. We haveIf(
t,x)
f,(
t,x)l
<=
T,_f
qt(s, p.(s,.)]
ds(4.21)
+
[T
t_,-T,_,]O(s,
ds We claimthat(4.22)
!im
T,_
T,"__]O(s,
ds 0uniformly for
t[O,
T],
Ixl
<-R.Indeed,
fix>0
and letre]O,
T[
be such that(K/er)
"
<r/.Then,
if0=<t-< r,(4.23)
T,_
TtQ]O(s,
dsOn the other
hand,
ifr<
=<
T,
thent[ T_s
TT_]q(s, ds(4.24)
+
<----27+
T,_-
T’/_,]q(s, ds /2 t-rTt--
T’_s]O(s,"
ds=<4*7+
Tt_
Tt"__,]O(s,"
dsand
(4.22)
follows from(3.6).
Next,
let us show thatio
(4.25)
!iIn
T_.[O(s,
")-O,(s, ")]
as
=0 uniformly for[0,
T],
[x[
_-<R.
Recalling(3.4),
we obtainT_[O(s,
)-,(s, .)](x)
[O(s,
e(t-)alI,x+v/-
Wa(t-s))
(4.26)
-O,(s,
e(’-)aHnx+
W%(t-s))].
Moreover,
Proposition2.4 implies that there exists a random variableC,
suchthat(4.27)
[e(’-s)aHn
x+
W2(t-
s)[
NR+
Cfor n
N,
0NsN NTandIx[
NR.So,
in lightof hypothesis(iii),
lim sup
[O(s,
e(’-s)anx+
W(t-s))
n[x[R,zst(4.28)
-,(s,
e(’-n,x+
w2(t-s))[=o
almostsurelyfor all z
]0,
T].
Hence,
bythe dominatedconvergencetheorem,
for allz
]0,
T].
(4.29)
i
T,[O(s,
")-O(s,
.)](x)
ds =0uniformly for
IxlR
and r]0,
T].
Nowfix >0andchooser sothat(K/)r <
.
Then
(4.30)
T,,[O(s, ")-O(s,-)](x)
dsN2+
T,,[O(s,
")-O(s,
.)](x)
ds and(4.25)
follows from(4.28).
Finally,(4.21), (4.22),
and(4.25)
imply(4.19). Next,
weprove
(4.20).
For
all t]0, T],
x
NR we have-lL(,xl-L,(,xl
-
(rr_s[(s,.l-(s,.l]lxS
(4.31)
+
1-([ Tt-,-
Tt%s])(s,"
dsNow,
fix>
0and let z6]0,
T[
besuch that8CoKt
<
,
whereCo
eLM,
LandM beingdefined in
(3.17)
and(2.9),
respectively.Then,
by(4.5)
and(4.7)
weconcludethat, if 0
T,
-
([r_,-T_,]O)(s,.)d
s2CoK
-Onthe other
hand,
if<
NT,
we obtain, as in(4.24),
-
([
T,_
T_.]O)x(s
ds(4.32)
+
-
([
r,_,
rL,])(s,.
s
So,
by(3.6)(ii),
(4.33)
lim 1-’([ T,_
TT-])xO(S,
ds =0uniformly for
]0,
T]
and[x[
_-<R. Next we prove thatIo
(4.34)
limt-
(T’_[O(s,.)-O,(s,.)])ds
=0uniformly for t
]0, T]
andIxl-<-R.
From
(3.7)
itfollows thatT_[(s,.
)-
o/.(s,.
)])x (x)
(4.35)
x/--
{e(’-)AQ-ll-I,,WA(t--s)[qt(s,e(t-)AII,x+xf-
Wa(t-s))
6,(s,
e’-Arl.x
+,/7
W"(t-
s))]}.
Recalling
(4.28)
and(2.11),
we concludethat, if >_--,
then lim suple(’-)AQ,5
II,WA(t-s)[-b,]
(4.36)
(s,
e(t-)AII,x+x/-
W"a(t-s))[=O
almostsurelyfor
]x]
=<
R, ’/2
=<
s=<
’/2.
Now,
arguingasabove,
we canprove(4.34)
and so(4.20).
Proof of
Theorem 4.5. SetG,(x, p)
(F(1-I,x), p)- H(p)
+
g(II,x),
(4.37)
G(x, p)
(F(rlx), p)- H(p)
+
g(rlx)
for all
x,
peX and(4.38)
(Fv)(t,.)=
r;4+
r,_,G(
.,
v(s,.))
dsVve.
From Lemma
4.2itfollows thatF
maps into;.
Arguing asintheproofof Theorem 4.3 itfollows that(4.39)
IIr.-
r,z[I,
<-11
-
zll,.
provided that T satisfies
(4.12).
Therefore,
F,
has a unique fixed-point u,, which is the unique mild solution of(4.16).
Moreover,
(4.40)
[[u,-
F
,"
(0)11,.
-<_21-"(Tllgllo
+
b
11o),
where
F,"
denotesthe/x-iterate
ofF,.
We claimthat forall/x
N and all R>
0,(4.41)
limF(0)(t,
x)=
F"(0)(t,
x)
uniformly for
[-, T]
andIx[-<_
R. Infact,
(4.41)
is truefor/x
1,in view of(3.6)(i).
Now,
suppose that(4.41)
holdsfor/x
N. Thenthe functions(4.42)
q(t,x)=G(x,F"(O))x(t,x),
,(t,x)=G(x,F(O))x(t,x)
satisfy the assumptions of
Lemma
4.6. Consequently, by(4.19)
(4.43)
limFn+l(0)(t,
x)=
F"+i(0)(t,
x)
uniformly for
[0, T],
]x]
_--<R. Therefore(4.41)
holds for all kt N.Finally,toprove
(4.18)(i)
note thatfor all,
n N]U(t,X)--U,,(t,X)
(4.44)
<-_lu(t,x)-F(O)(t,x)l+lF’(O)(t,x)-F.(o)(t,x)l+lF.(o)(t,x)-u.(t,x)l
--<- 22-"(
Yllg
[[o +
Fix
r/>0
andlet/x,
Nbe such that2=-,(TIIgll0+ I111o)<
7. Then, (4.44)yields(4.45)
lu(
t,x)
un(
t,x)
_-<2r/+
IF",
(0)(
t,x)
F, (0)(
t,x)l.
Now,
(4.18)(ii)
can be easily derived by minor modifications ofthe above argument(using
(3.6)(ii)
instead of(3.6)(i)
and(4.20)
instead of(4.19)).
Therefore,
the proofis complete.
5. Application tostochasticoptimal control. Let
{,
,
P}
beacompleteprobability space and{ilk}
asequenceof standard one-dimensional Brownian motions, mutuallyindependent. For anys 0 let be the -algebragenerated by
{k(S)"
k 1, 2,’’’0s
t}.
LetM(t,
T;
X)
denote the space of the X-valued processes x such thatx(s)
in-measurable
for all s T andConsidera stochastic systemgoverned by the state equation
(5.1)
y(s)=e(-’x+
e(’-r[F(y(r))+z(r)]dr+
W(t,s),
stO,
where x
X, A
is a self-adjoint operator satisfying(3.2)
and(2.8),
F Lip(X, X),
2
z M
w(t, T; X),
andW(t, s)
is defined by(5.2)
W(t,
s)
k=l
Equation
(5.1)
canberegardedasthe"mild"form of the stochastic differential equationdy(s)={Ay(s)+F(y(s))+z(s)}ds+dW(s),
tsT,
(.3)
y(t)=x,
where
W(t)
is acylindrical Wiener process(see
Remark2.2).
Wenowprove the existence of solutionsto
(5.3)
aswellas aGalerkin approxima-tion result.POPOSTION 5.1.
Assume (3.2),
(2.8)
and let FLip(X,
X).
en,
for
all zM(t,
T;
X),
equation(5.1)
hasaunique solutiony(.;
t,x,
z),
whichis continuouswithprobabilityone.
2
Proof
LetA= {v Mw(t, T;X):
on
A
as follows"(5.4)
a(v)(s)=e(S-t)Ax+
e(X-r)A[F(v(r))+z(r)]
dr+
WA(t,S),
tNSFrom
Proposition 2.4 it follows thatWa(t," )@
m.
Hence,
1"A A.Moreover,
is acontraction provided that
T-t<I/[[FII
and the conclusion follows by standardfixed-point arguments.
POPOSON 5.2.
Assume (3.2),
(2.8)
andletF Lip(X,
X).
Lety,,(.
t,x, z)
be the solutionof
dy,,(s)=
{AH,y,,(s)+H,F(H,y,,(s))+H,z(s)}
ds+
dW"(s),
(5.5)
y,,(t)=Hx,
tNsNT,
wherez
MZw(t,
T; X)
andWn(t)
isdefined
in(2.3).
Then,
(5.6)
n-limg(t__<_=rsup
[y(s;
t,x,
z)
y,n(s;
t,x,
z)[)
Ofor
all xX.
Proof
LetA
bethe space definedinthe previousproofand define amapAnn
onA
as follows:An(V)(s)=
e(s-t)Al-Inx
+
e(s-r)arln[F(I-[nV(r))+
l-Inz(r)]
dr+x/
I-[nWA(t, s),
t<_s<_T.
Then,
hn
is acontraction
inA
uniformly with respect to nN,
provided that T-t<
1/[I
F[]I.
Moreover;
Proposition2.4 implies thatlim(
sup[An(v)(s)-A(v)(s)l)=O
for all v
M(t,
T; X).
The conclusion then follows by the contraction mappingtheorem depending on a parameter.
We
will now study thefollowing stochastic optimal control problem.Given R
>
0, minimize the costfunctional(5.7)
J(t,x; z)=
g(y(s;
t,x,z))+glz(s)l
ds+(y(T;
t,x,z))
over all controls z
M(t,
T;
X)
satisfying[z(s)[R
almost surely for alls[t,
T].
The valuefunction
ofproblem(5.7)
is given by(5.8)
V(t,x)=inf{L(t,x;
z)"
zM(t,
T;
X),[z(s)[R}.
The corresponding Hamilton-Jacobi-Bellman equation reads asfollows:
(5.9)
--+-Tr(vxx)+(Ax+F(x), vx)-H(v)+g(x)=O
Ot 2 in[0,
T]
xX,
v(T,x)-(x),
where H is defined by(5.10)
H(p)
-IP
ifIpl
R,
R
RIp
I----f-
if[pl-->
e.
From Theorem 4.3 we obtain the result below.
THEOREM 5.3.
Assume
(2.1),
(2.8), (4.2)(i)
andletF Lip(X,
X).
Thenproblem(5.9)
has a unique mildsolution, which coincides with the valuefunction
V.
Moreover,
for
any t,x)
[0, T]
xX,
thereexistsanoptimalcontrolfor
problem(5.7).
Furthermore,
any optimal control
z*
is related to the corresponding optimalstatey*
by thefeedback
formula
(5.11)
z*(s)=-h(
OVOx
(s,y*(s))),
t<s<T,=
where(5.12)
pif
lpl=R,
h(p)= pR
-1
iflP[>=R"
Proof
First, we note that the existence and uniqueness ofa mild solution v to(5.9)
follows fromTheorem 4.3,sincethe functionH
defined in(5.10)
fulfills(4.2)(ii).
Let
us show that vV.
We
claim that v satisfies the dynamic programmingprinciple below: for any
]0, T[,
x X and zMw(t,
T;
X)
such that]z(s)l
<=
R
almost surely, we have
(5.13)
{Iz(s) +
v(s, y(s;
t,x,
z))[:-x(v(s,
y(s;
t,z,
z))-n)}
dsg(y,(s;
t,x,
z))+-lz(s)l
d+6(y(T;
t,x, z))
where
l’(a)=
0 if a<=
0 andx(a)=
a if a>_-0.Indeed,
letun
be the solution of problem(4.16)
withH
given by(5.10)
and setVn(t,
X)
U,( T-
t,x).
We
claimthatv, isregular. Toshow this fact let’(t,
xl,,
x,)
be defined as
(t,
xl,...,x,)=v,(t,xe+...+xne,),
for all(t,x,...,x,)
]0, T[
Rn.
Then sr is a classical solution of the problem-+
A-
E
Ea,x,-(F(x,
el+...
+
x,en),
ei)]Ot i=1 OXi
el+...n
t-OXn
en)
+
g(xe +.
+
x,en)
O,
(
T,
x,.
.,
Xn)
(Xlel
+"+
X,en).
So,
we can usetheIt6
formulatodifferentiatevn(s,
y,n(s))
wherey,n(s)
y,,(s;
t,x,z).
Thus,
we obtaindn(S,
y,(s))=--(s,
y,n(S))ds+(dY,n(S),
Un,(S,
7,n(S))>+
Tr(u,(s,
yn(S)))as.
Now,
recall(4.16)
and(5.5),
integrate on[t, T]
and take expectation to obtainVn(t,x)___
c_
{ll-Inz(S)+Vn,x(S,
y,n(S))12-X(lVn,x(S,
y,n(S))l-e}
dsg(y,(;
t,x,z+ylnz(s)l
d+(y,,(T; t,x,z)
By
Proposition 5.2 and Theorem4.5, we obtain(5.13)
inthe limit as nee.
Next,
wenotethat thefollowing inequalityholds:
(5.14)
Iz(s)+v(s,y(s;
t,x,z))12-X(V(s,y(s;
t,x,z))-R)>-O.
Thus,
from(5.13)
and(5.14)
it follows thatv(t,x)<=
V(t,x).
To
prove the reverseinequality, let us consider the closed-loop equationy(s)=e(S-t)Ax+
e(’-r)A[F(y(r))-h(vx(r,y(r)))]
dr+x/
WA(t,s),
(5.15)
T>s_->t->0,
which can be solved by the Schauder fixed-pointtheorem
(see,
e.g.,[13,
Cor.2.3]).
Indeed,
from(2.8)
it follows that e’a is compact for >0. Lety*
be amild solutionof
(5.15).
Taking(5.16)
z(s)
:-h(vx(s,
y*(s))),
we have the equality in
(5.14),
and sov(t,x)>= V(t,x)
for all t<T.
Moreover,
thechoice
(5.16)
providesanoptimal control at(t, x).
Finally,thefeedbackformula(5.11)
follows from
(5.13)
and the fact thatv(t, x)--
V(t,
x).
Example 5.4.
Let
XL2(0,
7r)
and defineD(A)
H2(0,
7r)
f3H(0,
7r),
02x
(5.17)
ax
:-
Vx
D(A),
F(x(:
=f(x(,
g(x
(x(,
:,
4,(x)
((
,
where
f
Cb(R),
o,1
Cb(R).
By
inspection,F,
g, and4
fulfill the hypotheses ofTheorem 5.3.
As
for operatorA, (3.2)
is satisfied by k,
and so(2.8)
holds trueforany cr
]0,
1/2[.
Therefore,
the results of this sectionapplytothefollowingstochasticoptimal control problem.
Minimize
(5.18)
J(t,x; z)= V
a(y(s,
))+-lz(s,
c)12
d6ds+
fi(y(T,))
d
over all controls z
M2w([t,
T];
L(0,
r))
satisfyingIz(s,
:)12
d:-<_
R2almost surely
for all s
It,
T],
where the state y is subject tody(s,
()=
-xy(s,
()+
f(y(s,
())+z(s, )
ds+x/
dW(s),
(5.19)
y(s,
0)
--y(s,
7r)
:0,s[t, T],
y(
t,x( ).
Remark 5.5. Let us consider the same problem as in Example 5.4 for an
N-dimensional parabolic state equation, i.e., taking
X--L2([0,
7r]
N)
andAx
Ax,
with Dirichlet orNeumann
boundary conditions.Then,
Theorem 5.3 does not apply. Infact,
we can show thatq(t)=
tl-N/Zo(t)
(see
[7])
and(2.6)
is notsatisfied.However,
if we consider the iterated Laplace operator
Ax=(-1)’-(-A)"x,
(with Dirichletboundary
conditions),
wehaveq(t)
-N/zmO(l)
and Theorem 5.3appliesifN<
2m. Remark 5.6. Using Theorem 5.3 and the variational technique of[8]
we can characterize the value functions of deterministic optimal control problems as limits,as e$ 0, ofthe mild solutions
u
of(4.1).
For example,considerthefollowing problem.Given R
>
0,minimize(5.20)
J(t,x; z)=
g(y(s;t,x,z))+-g[z(s)l
2ds+dp(y(T;
t,x,z))
over all controls zL2(t,
T; X)
satisfyingIz(s)l
<=
R.Here y(.;
t,x,z)
is the mild solution of the state equationy’(s)
Ay(s)
+
V(y(s))
+
z(s),
<-_s<-_T,
y(t)=x.
Define the value function ofproblem
(5.20)
asV(t, x)
inf{J(t,
x;z):
zLZ(t,
T;
H),
]z(s)]-<
R}.
Then, for all
(t, x)
[0,
T]
xH,
we can show that(5.21)
lu(T-t,x)- V(t,x)l<=
wheretog(respectively,
to)
denotesaconcave modulus ofcontinuityfor g(respectively,4)
andC
x/q(T)
q(t)
eREFERENCES
[1] V. BARBUANDG.DAPRATO,Hamilton-Jacobi Equations inHilbertSpaces,Pitman, Boston, 1982. [2] P.CANNARSAANDG. DAPRATO, The vanishing viscosity methodin infinitedimensions, Atti Accad.
Naz.Lincei,toappear.
[3]
,
AsemigroupapproachtoKolmogoroffequationsinHilbert spaces,Appl. Math.Lett.,toappear.[4] M.G. CRANDALLANDP.L.LIONS,Hamilton-Jacobiequations ininfinitedimensions.PartIV,preprint.
[5] G.DAPRATO, SomeresultsonBellmanequation inHilbert spaces,SIAM J.Control Optim., 23(1985),
pp. 61-71.
[6] G. DA PRATO, S.KWAPIEN, ANDJ. ZABCZYK, Regularityofsolutionsoflinearstochastic equations inHilbertspaces,Stochastics,23 (1987),pp.1-23.
[7] G. DA PRATO AND J. ZABCZYK, Smoothing properties oftransition semigroups in Hilbert spaces, Stochastics,toappear.
[8] W. H. FLEMING,TheCauchy problemforanonlinearfirstorder partialdifferentialequation,J.Differential Equations, 5 (1969),pp. 515-530.
[9] T. HAV,RNEANU, Existencefor the dynamicprogramming equation ofcontroldiffusionprocesses in Hilbertspace,NonlinearAnal. Theory MethodsAppl., 9(1985),pp. 619-629.
[10] P. L. LIONS, Viscosity solutionsoffullynonlinearsecond-orderequationsand optimal stochastic control in infinitedimensions. Part I: The caseofbounded stochastic evolutions, Acta Math., 161 (1988), pp. 243-278.
[11]
,
ViscositySolutionsofFullyNonlinearSecond-Order Equations and Optimal Stochastic Control inInfiniteDimensions.Part II: Optimal ControlofZakai’sEquation,Lecture Notesin Mathematics, Vol. 1390,Springer-Verlag, Berlin,1989.[12]
,
Viscosity solutions offully nonlinearsecond-order equationsand optimalstochasticcontrol ininfinite dimensions. Part III: uniqueness ofviscosity solutionsforgeneralsecond-order equations,
J.Funct. Anal.,86(1989),pp. 1-18.
13] A.PAZY, SemigroupsofLinearOperatorsand ApplicationstoPartialDifferentialEquations, Springer-Verlag, New York,1983.