Professional Documents
Culture Documents
(Nisio) Stochastic Control Theory (2015)
(Nisio) Stochastic Control Theory (2015)
(Nisio) Stochastic Control Theory (2015)
Makiko Nisio
Stochastic
Control
Theory
Dynamic Programming Principle
Second Edition
Probability Theory and Stochastic Modelling
Volume 72
Editors-in-Chief
Søren Asmussen, Aarhus, Denmark
Peter W. Glynn, Stanford, CA, USA
Thomas Kurtz, Madison, WI, USA
Yves Le Jan, Paris, France
Advisory Board
Joe Gani, Canberra, Australia
Martin Hairer, Coventry, UK
Peter Jagers, Gothenburg, Sweden
Ioannis Karatzas, New York, NY, USA
Frank P. Kelly, Cambridge, UK
Andreas E. Kyprianou, Bath, UK
Bernt Øksendal, Oslo, Norway
George Papanicolaou, Stanford, CA, USA
Etienne Pardoux, Marseille, France
Edwin Perkins, Vancouver, Canada
Halil Mete Soner, Zurich, Switzerland
The Stochastic Modelling and Probability Theory series is a merger and con-
tinuation of Springer’s two well established series Stochastic Modelling and
Applied Probability and Probability and Its Applications series. It publishes
research monographs that make a significant contribution to probability theory or
an applications domain in which advanced probability methods are fundamental.
Books in this series are expected to follow rigorous mathematical standards, while
also displaying the expository quality necessary to make them useful and accessible
to advanced students as well as researchers. The series covers all aspects of modern
probability theory including
• Gaussian processes
• Markov processes
• Random Fields, point processes and random sets
• Random matrices
• Statistical mechanics and random media
• Stochastic analysis
as well as applications that include (but are not restricted to):
123
Makiko Nisio (emeritus)
Kobe University
Kobe, Japan
Osaka Electro–Communication University
Osaka, Japan
First edition published in the series ISI Lecture Notes, No 9, by MacMillan India Limited publishers,
Delhi, c Makiko Nisio, 1981
v
vi Preface
to the upper and lower Isaacs equations. The link between stochastic control and
differential game is given via the risk sensitive control. Chapter 5 is a review on
stochastic evolution equations on Hilbert spaces, in particular stochastic parabolic
equations with colored Wiener noises. Basic definitions and results and Itô’s
formula are presented. Chapter 6 is concerned with control problems for Zakai
equations. We again construct semigroups associated with the value functions.
The dynamic programming principle and viscosity solutions of Hamilton–Jacobi–
Bellman equations on Hilbert spaces are treated by using results obtained in the
previous chapters. We show the connection between controlled Zakai equations and
control of partially observable diffusions.
This book was planned as a new edition of Stochastic Control Theory, ISI
Lecture Notes 9 (1981) following F. Delbaen’s recommendation. I would like to
acknowledge his recommendation together with valuable advice during preparation
of the manuscript.
The author is greatly indebted to W. H. Fleming, who read carefully the
manuscript and offered many valuable comments and suggestions, especially for
Chap. 4, which led to a much improved version. Many thoughtful helps and
encouragements had been given by experts and friends, particularly F. Asakura,
Y. Fujita, H. Nagai, and T. Uratani who helped to improve the book at various stages.
H. Morimoto assisted in writing the manuscript by carefully reading it and making
valuable comments, especially on mathematical economics.
vii
Contents
ix
x Contents
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 245
Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 249
Notations
RdC D fx 2 Rd I x i > 0; i D 1; : : : ; d g
Sr D fx 2 Rd I jxj rg
Rd ˝ Rm denotes the set of d m matrices
Aij D .i; j / entry of A 2 Rd ˝ Rm
A> D transpose of A
tr A D trace of A 2 Rd ˝ Rd
jAj2 D tr.A> A/ D tr.AA> / for A 2 Rd ˝ Rm
Id D d -dimensional unit matrix
S d D the set of symmetric d d matrices
SC d
D fA 2 S d I A is nonnegative definite}
SCC D fA 2 S d I A is positive definite}
d
xiii
xiv Notations
xv
Chapter 1
Stochastic Differential Equations
This section collects basic definitions and results on stochastic processes. Proofs of
the results are mainly referred to standard and easily accessible books.
1. Measurable maps
Let ˝ be a non-empty set and F a -field of ˝. We call the pair (˝; F ) a
measurable space. Let (˝; F ) and .˝ 0 ; F 0 / be measurable spaces and X W ˝ 7!
˝ 0 be a given map. For a -field G . F /, X is called G=F 0 -measurable (or G-
measurable in short), if X 1 .B/ 2 G for any B 2 F 0 . Sometimes X is called
measurable, when G D F .
2. Random variables
Let .˝; F ; P / be a complete probability space, namely .˝; F / is a measurable
space, P a probability measure on F , and N WD fA 2 F I P .A/ D 0g satisfies the
condition
A 2 N; B A H) B 2 N : (1.1)
(1) Let .U; d / be a Polish space, i.e., U is a separable complete metric space with
a metric d , and let B.U / be the -field generated by the open subsets of U .
X W ˝ 7! U is called a U -valued random variable if X is F =B.U /-measurable.
We identify two U -valued random variables X and Y , whenever there is
N 2 N , such that
EŒX I B D 0; if P .B/ D 0;
1.1 Review of Stochastic Processes 3
For 2 F , E. jG/ is called the conditional probability of given G, and denoted
by P . jG/.
Definition 1.2. Let X be a d -dimensional random variable and G be a sub--field
of F . We say X and G are mutually independent, if
i.e.,
We say that X and Y are mutually independent if X and .Y / are mutually
independent.
4. Regular conditional probability
Since P . jG/ is defined uniquely up to sets of P -measure zero (possibly
depending on ), P .jG/.!/ may not be a probability on F when ! is fixed. We
next introduce a regular conditional probability.
Definition 1.3. p W F ˝ 7! Œ0; 1 is called a regular conditional probability given
G, if
(a) p.; !/ is a probability measure on .˝; F /, 8! 2 ˝,
(b) p.A; / is G-measurable, 8A 2 FR,
(c) 8A 2 F , 8B 2 G, P .A \ B/ D B p.A; !/ dP .!/.
Concerning the existence of regular conditional probability we have;
Theorem 1.1. Suppose that ˝ is a Polish space and F D B.˝/. Then there exists
one and only one regular conditional probability given G, i.e., if both of p and pO are
regular conditional probabilities, then
p.A; / D p.A;
O /; 8A 2 F ; P -a.s.
5. Convergence theorems
Let Xn ; n D 1; 2; : : : and X be d -dimensional random variables.
Definition 1.4. (a) Xn converges to X in probability .Xn ! X in prob), if
limn!1 P .jXn X j > "/ D 0, 8" > 0;
(b) Xn converges to X almost surely .Xn ! X P -a.s./, if limn!1 jXn X j D 0
P -a.s.;
(c) Xn converges to X in p-th mean .Xn ! X in Lp /, if limn!1 EjXn X jp D 0;
(d) Xn converges to X in law .Xn ! X in law), if limn!1 Ef .Xn / D Ef .X /,
8f 2 Cb .Rd /.
For the law convergence, X and Xn ; n D 1; 2; : : : may be defined on different
probability spaces.
Proposition 1.1. The following relations hold:
(i) Xn ! X P -a.s. ) Xn ! X in prob ) Xn ! X in law.
(ii) If Xn ! X in prob, then there is a subsequence which converges P -a.s.
Now we state the basic theorems of the passage to the limit under the sign of
conditional expectation.
Let Xn be an integrable real random variable, n D 1; 2; : : : , and G be a sub--
field of F .
Fatou’s Lemma
Let Xn 0; n D 1; 2; : : : . Then
Convergence Theorem
1. Basic definitions
(1) Let .˝; F ; P / be a complete probability space and T a positive constant.
A family of sub--fields .Ft / WD .Ft I t 2 Œ0; T / is called a filtration,
if Fs Ft F for 0 s < t, and F0 contains all T P -null sets.
If Ft is right continuous, namely Ft D Ft C WD s>t Fs ; 8t 2 Œ0; T /, then
the filtration.Ft / is said to satisfy the usual condition.
If .Ft / is a filtration, then .Ft C / satisfies the usual condition.
Definition 1.5. We call a quadruple .˝; F ; .Ft /; P / a filtered probability
space when .Ft / satisfies the usual condition.
(2) A stochastic process X D .X.t/I t 2 Œ0; T / is a family of (d -dimensional)
random variables, defined on .˝; F ; P /. We identify two stochastic processes
X and Y , whenever there is a P -null set N such that
We call Y is a modification of X , if
For details, refer to [LS01], Section 1.3, [KS91], Section 1.1.2 and [IW81],
1.5.
3. Continuous martingales
(1) Let X be a continuous .Ft /-adapted real process.
1.1 Review of Stochastic Processes 7
Then, as N ! 1.
X
jX.tiNC1 / X.tiN /j2 ! hX i.t/ in prob (1.9)
i
Then
We have the following maximal inequality, for the quadratic variational process.
Burkholder–Davis–Gundy Inequality
Let X be a square integrable continuous .Ft /-martingale, with X.0/ D 0. Then, for
p 2 .0; 1/ there exist positive constants cp and Cp , such that
h i h i
cp E sup jX.s/j2p EŒhX i.t/p Cp E sup jX.s/j2p : (1.10)
0st 0st
In detail, c1 D 14 ; C1 D 1,
hX i WD .hX i.t/; t 2 Œ0; T / does not depend on the choice of n , and is called the
quadratic variational process of the local martingale X .
By Mc .Œ0; T ; .Ft // and Mcloc .Œ0; T ; .Ft //, we denote the set of continuous .Ft /-
martingales and continuous .Ft /-local martingales, respectively.
We put
M2c .Œ0; T ; .Ft // D fX 2 Mc .Œ0; T ; .Ft //I EjX.t/j2 < 1; 8t 2 Œ0; T g:
(2) Definition 1.9. By a continuous Markov process with the transition probability
function p./, we mean a continuous .Ft /-adapted process X , with
1
A˚.; x/ D lim .E x ˚. C h; X. C h// ˚.; x//;
h!0 h
@˚
A˚.; x/ D G ˚.; /.x/ C .; x/:
@
When X is given by a stochastic differential equation, we can compute A by
using Itô’s formula (see (2.178)).
Let L0 ..Ft // denote the set of .Ft /-progressively measurable real process and
8 n Z T o
ˆ
ˆ
< L .Œ0; T ˝; .Ft // D ˚ 2 L ..Ft //I E
2 0
j˚.s/j2 ds < 1 ;
Z T
0
ˆ 2 n o (1.11)
:̂ L.Ft / .Œ0; T / D ˚ 2 L ..Ft //I P
0
j˚.s/j2 ds < 1 D 1 :
0
RT
Setting k˚k2T D EŒ 0 j˚.s/j2 ds; L2 .Œ0; T ˝; .Ft // becomes a Hilbert space,
with the norm k kT . Rt
Let us define I.˚/.t; !/ D 0 ˚.s; !/ d W .s; !/ in three steps.
Step 1. ˚ is a simple process, say
f0 .!/; for s D 0;
˚.s; !/ D
fi .!/; for s 2 .ti ; ti C1 ; i D 0; : : : ; n:
X
n
I.˚/.t; !/ D fi .!/.W .ti C1 ^ t; !/ W .ti ^ t; !//
i D0
X
j1
D fi .!/.W .tiC1 ; !/W .ti ; !//Cfj .!/.W .t; !/W .tj ; !//;
iD0
for t 2 Œtj ; tj C1 :
and
Z t
hI.˚/i.t/ D ˚.s/2 ds:
0
Z t
I.˚/.t/ D lim I.˚n /.t/ DW ˚.s/ d W .s/:
n!1 0
Consequently, we have
Z t
I.˚/ 2 M2c .Œ0; T ; .Ft // with hI.˚/i.t/ D ˚.s/2 ds: (1.13)
0
For any two .Ft /-stopping times and , with 0 , P -a.s., we put
Z t
˚.s/ d W .s/ D I.˚/. / I.˚/./: (1.14)
R m ^
Since 0 ˚.s/ d W .s/ 2 M2c .Œ0; T ; .Ft //; m D 1; 2; : : : , the Itô integral
1.1 Review of Stochastic Processes 13
Rt
0 ˚.s/ d W .s/ is defined as a continuous .Ft /-local martingale (see [IW81],
pp. 45–22, [KS98], pp. 129–141, [LS01], pp. 92–112, for details).
Replacing real one-dimensional processes by m-dimensional ones, we define
L2 .Œ0; T ˝; .Ft /I Rm / and L2.Ft / .Œ0; T I Rm / in the same way as in (1.11).
R i process. For ˚ D .˚1 ; : : : ; ˚m / 2
i i i
Let W be an m-dimensional .Ft /-Wiener
L .Œ0; T ˝; .Ft /I R /, we define ˚ .s/ d W .s/ by
2 m
Z t m Z
X t
˚ i .s/ d W .s/ D ˚ji .s/ d W j .s/:
0 j D1 0
where
X0 is an F0 - measurable d -dimensional random variable,
b./ 2 L1 .Œ0; T ˝; .Ft /I Rd /,
R˚./ 2 L .Œ0; T ˝I R ˝ R /,
2 d m
Itô’s Formula
R t^
e 0 c.s/ ds F .t ^ ; X.t ^ // F .0; X0 /
Z t^ Rs n
D e 0 c.h/ dh @t F .s; X.s// C b.s/ @x F .s; X.s//
0
1 o
C tr a.s/@xx F .s; X.s// c.s/F .s; X.s// ds
2
Z t^ Rs
C e 0 c.h/ dh @x F .s; X.s//˚.s/ d W .s/; 8t; (1.18)
0
@F X @F 1 X @2 F
d d
dF .t; X.t// D dt C dX i C dX i dX j C o.dt/
@t i D1
@xi 2 j D1
@xi @xj
Suppose that
(a) ˚./ and b./ are bounded,
(b) a.t; !/ is uniformly positive definite, i.e., there is a positive constant 0 such
that
Let
b W Œ0; T Rd ˝ 7! Rd
and
˛ W Œ0; T Rd ˝ 7! Rd ˝ Rm
jb.t; x; !/b.t; y; !/jCj˛.t; x; !/˛.t; y; !/j ljx yj; 8x; y; 8.t; !/:
8t 2 Œ; T ; i D 1; : : : ; d; P -a.s.
O
P .X.t/ D X.t/; 8t 2 Œ; T / D 1 (1.22)
P Rt Rt
The solution is given by Z.t/ D 1 1
. 0 B.s/ ds/n D exp. 0 B.s/ ds/ and
Rt
nD0 nŠ
its inverse matrix Z 1 .t/ is exp. 0 B.s/ ds/.
18 1 Stochastic Differential Equations
Let .˝; F ; .Ft /; P / be a filtered probability space. Here we are concerned with
transformations of probability distribution. So we write down the probability P
together with .Ft /.
Consider an exponential .Ft ; P /-martingale
Z t Z
1 t
q.t/ D exp ˚.s/ d W .s/ j˚.s/j2 ds ; t 2 Œ0; T ; (1.33)
0 2 0
Q. / D EŒZ ; 8 2 FT : (1.35)
E Q Œ D EŒZ
D EŒE.ZjFt /
h E.Z jF / i
t s
D E Zs
Zs
h E.Zt jFs / i
D E E.ZjFs /
Zs
h E.Z jF / i E.Zt jFs /
t s
DE Z since is Fs -measurable
Zs Zs
h E.Z jF / i
t s
D EQ ; (1.38)
Zs
which yields (1.36). t
u
Now we state a basic theorem.
Theorem 1.3 (Girsanov’s Theorem). Let W be an m-dimensional .Ft ; P /-
Wiener process and ˚ 2 L2 .Œ0; T ˝; .Ft /I Rm /. Suppose that q.t/ given by
(1.33) is an exponential .Ft ; P /-martingale. Then WO .t/ D .WO 1 .t/; : : : ; WO m .t//,
given by
Z t
O
W .t/ D W .t/
i i
˚ i .s/ ds; i D 1; ; m (1.39)
0
where EjX0 j2 < 1 and b.; x; / and ˛.; x; / are .Ft /-progressively measurable
processes satisfying condition .a2 /. Assume that ˛ is bounded.
Let W Œ0; T / Rd ˝ 7! Rm be a measurable map such that, for any
x 2 R ; .; x; / is .Ft /-progressively measurable. Further assume that satisfies
d
and
.t/
2nd term in the RHS c1 C M.t/ (1.45)
1 C ".t/
Z Z
M.t/ 1 t
dM.s/ t
d hM i.s/
D C " :
1 C "M.t/ 1C" 0 .1 C "M.s//2 0 .1 C "M.s//3
Now take the expectation of both sites and apply (1.43) to the 3rd term of RHS.
Letting here " ! 0 yields EM.t/ D 1, from which the proposition follows thanks
to Example 1.2. t
u
Example 1.4 (Transformation of the drift term). Let W be a d -dimensional
.Ft ; P /-Wiener process and b W Œ0; T Rd 7! Rd a Borel function satisfying
and
Z t Z
1 t
q.t/ D exp ˚.s/ d W .s/ j˚.s/j2 ds
0 2 0
8 Z t Z
< 1 t
exp b.s; .s// d W .s/ jb.s; .s//j2 ds ; t;
D 2 (1.52)
:
1; t < ;
Firstly we assume .a1 / and .a2 /, i.e., we have an SDE with Lipschitz continuous
coefficients. By Theorem 1.2, the unique solution X;X .t/ depends only on .; X /
and .W .s/ W ./; s 2 Œ; t/ (this solution is called a strong solution). Put
Namely, the solution is a Markov process with the transition probability function
p./.
For the transition operator T t .x/ D E x .X.t//, Itô’s formula gives the
expression for the backward evolution operator A. For ˚ 2 C 12 .Œ0; 1/ Rd /,
where a D ˛˛ > .
When we drop the continuity condition .a1 / and .a2 /, we can hardly expect the
existence of strong solutions. This leads us to the definition of the notion of weak
solution.
Definition 1.12 (Weak solution). Suppose that XQ and WQ are given on a filtered
Q FQ ; .FQ t /; PQ /. .XQ ; WQ / is called a weak solution of (1.55)–
probability space .˝;
(1.56), if
Q
(a) X./ is a d -dimensional random variable, having the same probability distri-
bution as X ,
(b) WQ is an m-dimensional .FQ t ; PQ /-Wiener process,
(c) XQ is a d -dimensional Q
R t continuous .FtR/-adapted process,
Q Q
(d) X .t/ D X./ C b.s; X .s// ds C ˛.s; XQ .s// d WQ .s/, t, PQ -a.s.
Q t
24 1 Stochastic Differential Equations
Then, the SDE (1.55)–(1.56) admits a weak solution provided EjX j4 < 1.
Proposition 1.7 ([SV79], Theorem 7.2.1). Let b W Œ0; 1/ Rd 7! Rd and ˛ W
Œ0; 1/ Rd 7! Rd ˝ Rm be bounded and Borel measurable. Suppose that ˛.t; x/ is
uniformly continuous w.r.t. x, uniformly in t and ˛˛ > is uniformly parabolic. Then
the SDE (1.55)–(1.56) admits the unique weak solution.
In this subsection, we will overview the terminal value problems for SDEs, called
backward SDE, introduced by Pardoux and Peng [PP90].
Let W be a d -dimensional Wiener process defined on .˝; F ; .FtW /; P /. Let h W
Œ0; T R1 Rd ˝ 7! R1 be measurable and satisfy the conditions
.a1 / h.; y; z; / 2 L2 .Œ0; T ˝; .FtW /I R1 /; 8y; z
and
.a2 / 9l > 0 such that 8t; !,
Now we have;
Theorem 1.4. Under the conditions .a1 / and .a2 / , problem (1.62)–(1.63) admits
a unique solution.
(see [ElKPQ97] and [YZ99], Chapter 7 for details).
Example 1.5 (Backward SDE associated with a quasi-linear parabolic equation).
Let F W Œ0; T Rd R1 Rd 7! R1 be a continuous function satisfying
and
where coefficients b and ˛ are deterministic and satisfy .a1 / and .a2 / with d D m.
Now consider a backward SDE
(
d Y .t/ D F .t; .t/; Y .t/; Z.t// dt C Z.t/ d W .t/; t 2 Œ0; T /;
(1.68)
Y .T / D g..T //;
@t v.t; x/ C Gt v.t; x/ F .t; x; v.t; x/; ˛.t; x/> @x v.t; x/> / D 0 on Œ0; T / Rd ;
(1.69)
26 1 Stochastic Differential Equations
Let us consider a contract to sell a contingent claim with payoff G and maturity T ,
at time 0. The problem is to price this contingent claims. Suppose that the agent
invests his own money in a bond and stocks whose prices evolve according to SDEs
and does not want to run any risk of losing money. The lowest price satisfying this
condition is called the selling price (refer [ElKQ95, ElKPQ97]).
In Sect. 1.3.1, we will formulate the problem and, by using the dynamics of
bond and stocks, we study the selling price and portfolio strategies in Sects. 1.3.2
and 1.3.3.
1.3.1 Formulation
Let us consider a financial market consisting of one bond (riskless asset) and d
stocks (risky assets). We suppose that the price of the bond is given by
S 0 .t/ D e rt ; t 0; (1.72)
where r is a positive constant, and the price of i -th stock evolves according to the
SDE
1.3 Asset Pricing Problems 27
8
< dS i .t/ D S i .t/.i .t; S.t// dt C X i .t; S.t// d W j .t//;
d
ˆ
j t > 0;
(1.73)
j D1
:̂
S i .0/ D s i > 0 .i D 1; : : : ; d /
Let us consider the contract between the agent and a quest at time 0, such that the
agent pays G .2 L2 .˝; FtW I Œ0; 1/// to the quest at the maturity T . The problem
is what is the price of G at 0. .t/I Œ0; T ˝ 7! Rm is called a portfolio strategy,
./
when ./ S./ 2 L2 .Œ0; T ˝; .FtW /I R1 /. By Xx , we denote the solution of
(1.74)–(1.75). Since the agent does not want to run any risk of losing money, he
chooses a portfolio strategy ./ such that
We are going to study the selling price by using a backward SDE. We assume that
the following conditions are satisfied:
(a) W Œ0; T Rd 7! Rd ˝ Rd , W Œ0; T Rd 7! Rd are bounded and
uniformly Lipschitz continuous w.r.t. x,
(b) .t; x/ is invertible and .t; x/1 is bounded.
28 1 Stochastic Differential Equations
Then (1.73) has the unique strong solutions S.t/ D .S 1 .t/; : : : ; S d .t// and
Z t
S .t/ D s exp
i i
i .; S.// d W ./
0
Z t
1
C i .; S.// j i .; S.//j2 ds > 0: (1.78)
0 2
8
ˆ
ˆ d Y .t/ D .rY .t/ C Z.t/> .t; S.t//1 ..t; S.t// r1d //dt
<
C Z.t/> d W .t/; t 2 Œ0; T /; (1.84)
ˆ
:̂
Y .T / D g.S.T //
has the unique solution .Y; Z/ by Theorem 1.4, and comparing (1.82) with (1.84),
we conclude that x G D Y .0/ and
X
d
..t; S.t//1 /i Z j .t/=S i .t/;
j
O i .t/ D i D 1; : : : ; d; (1.85)
j D1
In order to get an explicit formula for x G and O ./, we study the Cauchy problem
for a parabolic equation related to (1.84).
Using Proposition 1.8 and S i .t/ > 0; i D 1; : : : ; d , we have
8
ˆ
ˆ 1 X ij
d X d
ˆ
ˆ C i j
C x i @i v rv D 0;
ˆ
< @t v.t; x/ a .t; x/x x @ij v r
2 i;j D1 i D1
(1.86)
ˆ
ˆ .t; x/ 2 Œ0; T / RdC ;
ˆ
ˆ
:̂
v.T; x/ D g.x/; x 2 RdC
where a D > .
If (1.86) admits a classical solution, then we have
(
v.0; S.0// D x G ;
(1.87)
@i v.t; S.t// D O i .t/; i D 1; : : : ; d:
Since V .t/ WD v.t; S.t// satisfies (1.82)–(1.83), V .t/ provides the selling price at
time t. Accordingly V ./ is called the selling price process. Applying Itô’s formula
to v./, we have
Example 1.6 (Constant coefficient markets). Let ./ and ./ be constants. Then
(1.86) reduces to
8
< @ v C 1 X aij x j x j @ v C r X x i @ v rv D 0;
d d
ˆ
t ij i on Œ0; T / RdC ;
2 ij D1 i D1
:̂
v.T; x/ D g.x/; x 2 RdC :
(1.89)
Referring to (1.88), we have
!
X
d 1
.t/ D ./ exp
i i
ji .W i .t/ W .// C r j i j2 .t /
j
(1.90)
j D1
2
and
Z
r.T t /
v.t; x/ D e g..T t; x; z//pd .T t; z/ d z; (1.91)
Rd
where pd .s; z/ D .2 s/ 2 exp jzj
d 2
2s
and D . 1 ; : : : ; d / with i .s; x; y/ D
x i exp.y i C .r 12 j i j2 /s/, i D 1; : : : ; d . In particular, for d D 1,
Z p
1
2
v.t; x/ D r.T t / g x exp T t z C r .T t/ p1 .1; z/ d z
1 2
(1.92)
holds. Suppose
g.x/ D .x K/ _ 0; (1.93)
where
n x 1 o p
˙ D log C r ˙ 2 .T t/ = T t :
K 2
Equation (1.94) is the famous Black–Scholes formula [BS73].
Chapter 2
Optimal Control for Diffusion Processes
2.1 Introduction
This section is devoted to formulating the time horizon stochastic control and
analyze basic notions. We introduce control processes, payoffs and value functions
in Sect. 2.1.1, and investigate their properties in Sect. 2.1.2.
Before we formulate the stochastic control problem, we give a typical example,
called the linear quadratic control.
Example 2.1 (Linear quadratic (LQ) control). Consider a d -dimensional stochastic
system with an external random force ./,
hZ T i
J.; xI .// D E x .X.t/> M.t/X.t/ C .t/> N.t/.t// dt C jX.T /j2
2.1.1 Formulations
We are going to formulate the finite time horizon stochastic control problem.
Let .˝; F ; .Ft /; P / be a filtered probability space and W an m-dimensional
.Ft /-Wiener process. We call .˝; F ; .Ft /; P; W / a reference probability system.
Let T > 0 be given. A -compact and convex subset of Rq is called a control
region, where q is a positive integer.
Definition 2.1. The -valued .Ft /-progressively measurable process ..t/,
t 2 Œ0; T / is called a control process, if ./ 2 L1 .Œ0; T ˝; .Ft /I /, namely,
there is a compact set ./ . /, such that .t; !/ 2 ./ , for almost all .t; !/.
The 6-tuple A D .˝; F ; .Ft /; P; W; .// is called an admissible control.
We denote by A the set of all admissible controls.
Let b W Œ0; T Rd 7! Rd and ˛ W Œ0; T Rd 7! Rd ˝ Rm be given
and satisfy the following condition:
.b1 /
and
When A.2 A/ is applied, the stochastic system evolves according to the SDE
(more precisely, controlled SDE)
dX.t/ D b.t; X.t/; .t// dt C ˛.t; X.t/; .t// d W .t/; .0 / < t T;
(2.3)
with the initial condition
f and are called the running cost and the terminal cost, respectively, and is
the discount rate.
For the response X D XAx , the cost function on time interval Œs; t . Œ; T / is
given by
C.t; s; AI /
Z t Z
D expf .h; X.h/; .h// dhgf .; X./; .// d
s s
Z t
C expf .h; X.h/; .h// dhg.X.t//: (2.5)
s
When the response X stops at time t, we define the payoff (or criterion) by
is called the value function. If A 2 A gives the infimum (or supremum) of the RHS,
A is called an optimal control.
34 2 Optimal Control for Diffusion Processes
Thus we are concerned with the characterization of the value function and an
optimal control.
Let us introduce different classes of admissible controls: Brownian adapted
controls and feedback controls.
1. Let A D .˝; F ; .Ft /; P; W; .// be an admissible control. A is called
Brownian adapted, if ./ is .FtW /-progressively measurable. AW denotes
the set of all Brownian adapted controls. For a fixed reference probability
system .˝; F ; .Ft /; P; W /, W denotes the set of all .FtW /-progressively
measurable control processes. Since the payoff is calculated in terms of the joint
probability distribution of .W ./; .//, for any given a reference probability
system .˝; F ; .Ft /; P; W /, it holds that
2. We control a system by using the data on the system, in the customary manners.
A -valued Borel function O ./, defined on Œ0; T Rd , is called a Markovian
policy, if it is bounded. When we apply O ./, the system evolves according to the
SDE
We can hardly expect that a strong solution will exist for (2.8), but a weak
solution does exist under mild conditions (see Sect. 1.3.3). Hence there exist
W and X , on an appropriate filtered probability space .˝ ; F ; .Ft /; P /,
such that (2.8) holds. Putting .s/ D O .s; X .s//, we have A D
.˝ ; F ; .Ft /; P ; W ; .// 2 A, but not in AW . We call A an admissible
control associated with the Markovian policy O ./.
h i
E x sup jX A .s/ X A .t1 /j2p Kp .1 C jxj2p /.t2 t1 /p ; 8A 2 A:
t1 st2
(2.10)
(iii) For any x; y 2 Rd and 2 Œ0; T ,
h i
E sup jXAx .s/ Xy
A
.s/j2 K1 jx yj2 ; 8A 2 A: (2.11)
sT
O O
Z.t/ D X A .t/ X A .t/; 4b.t/ D b.t; X A .t/; .t// b.t; X A .t/; O .t//;
Now Gronwall’s inequality and (2.14) lead to the left inequality in (2.13).
Now using the estimation
N
E m.jj/ 2
a2 C EŒm.jj/
N 2
N 1 .a/
I jj > m
a2 C MN 2 P .jj > m
N 1 .a//
Ejj2
a2 C MN 2 ;
N 1 .a//2
.m
then we can choose a subsequence n0 such that, for any .; x/ 2 Œ0; T Rd ,
A 0
lim sup jX xn .s/ XAx .s/j D 0 P -a.s.
n0 !1 sT
and
h i
lim E x
0
sup jX An0 .s/ X A .s/j2 D 0:
n !1 sT
Next we discuss continuity properties of the payoff and the value function,
replying on Propositions 2.1 and 2.2.
Theorem 2.1. Suppose .b1 /–.b5 / hold. Then J.t; ; x; AI / and v.t; ; xI / have
the following properties:
(i) There is a constant KN such that, for any 0 < t T and x 2 Rd ,
N C jxj2 /;
jJ.t; ; x; AI /j K.1 8A 2 A;
N C jxj2 /:
jv.t; ; xI /j K.1 (2.15)
2.1 Introduction 37
and
and
We have
I1 .x; y/
hZ t i
E jf .s; X.s; x/; .s// f .s; X.s; y/; .s//j dsI ˝x \ ˝y \ ˝Q x
hZ t i
O
C KE .2 C jX.s; x/j2 C jX.s; y/j2 / dsI ˝xc [ ˝yc [ ˝Q xc (2.25)
and
n o
˝Q D ! 2 ˝I sup jX.s; 1 / X.s; 2 /j < ı0 :
2 sT
whenever 2 1 < ı2 .
Now (ii) the proof of follows from (2.29) and (2.30).
Since (iii) the proof of mimics of (ii), the result is established. t
u
40 2 Optimal Control for Diffusion Processes
Theorem 2.1 asserts that, if the terminal cost function has quadratic growth,
then the value function also has quadratic growth. Put
n .x/ o
CQ D 2 C.Rd /I Q .x/ WD 2 C b .R d
/ : (2.31)
1 C jxj2
and
Q
kV t 0kCQ K; 8 t: (2.36)
Proof. (i) and (ii) are clear from the definition of payoff.
(iii) From the inequality j.x/ .x/j k kCQ .1 C jxj2 / it follows that
jJ.t; ; x; AI / J.t; ; x; AI /j k kCQ E x 1 C sup jX.t/j2 :
t T
that is,
Before we embark upon the proof, let us consider constant controls. Set A D
.˝; F ; .Ft /; P; W; /, where 2 stands for a constant control process: .t; !/ D
A
, for all t and !. Then its response Xy is the strong solution of (2.3), measurable
42 2 Optimal Control for Diffusion Processes
w.r.t. .W ./ W ./; y; /, and the payoff J.t; ; y; A I / depends only on
t; ; y; and . Denote the payoff by J.t; ; y; I /. We will seek an optimal
.2 / by using the measurable selection theorem (see [SV79], Lemma 12.1.7
and Theorem 12.1.10).
Since J.t; ; y; I / is continuous in .y; / by (2.11) and (2.13), the compact-
ness of implies that, for y 2 Rd , the set
t .yI / 2 y ; 8y 2 Rd (2.43)
Y
j
tk tkC1 D ti ti C1 ti C1 ti C2 tj tj C1 ; (2.46)
kDi
Y
j
VtD
i tj C1
.x/ tk tkC1 .x/ (2.47)
kDi
Y
j
VtiDtj C1 .x/ tk tkC1 .x/: (2.48)
kDi
2.2 Dynamic Programming Principle (DPP) 43
Put X A ./ D XtAi x ./. Since X A .tj / and .tj / are Ftj -measurable and W . C tj /
W .tj / is independent of Ftj ,
Y
j
J.tj C1 ; ti ; x; AI / J.tj ; ti ; x; AI tj tj C1 / tk tkC1 .x/: (2.50)
kDi
Y
j
.ti / D ti ti C1 xI tk tkC1 ; (2.51)
kDi C1
Y
j
.ti C1 / D ti C1 ti C2 X.ti C1 /I tk tkC1 : (2.53)
kDi C2
Y
j
J.tj C1 ; ti ; x; A I / D tk tkC1 .x/; (2.54)
kDi
m1
Y
VtlDtn .x/ D tk tk1 .VtmDtn /.x/
kDl
We again assume D 0.
(i) We have
C.T; 0; ./I /
X
pC1 Z tk
D . D tk / f .s; X .s/; .s// ds C C.T; tk ; ./I / ; (2.58)
kD1 0
where tpC1 D T .
From the definition of ./, it follows that
and
X
pC1 h nZ tk oi
D E0x . Dtk / f .s; X .s/; .s// ds CE.C.T; tk ; ./I /jFtWp /
kD1 0
hZ i
D E0x f .s; X .s/; .s// ds C V DT .X . // :
0
X
pC1
Q .t/ D . D tk /..t/.t < tk / C .t/.t tk //:
kD1
46 2 Optimal Control for Diffusion Processes
X
pC1 h Z tk i
D E0x . D tk / f .s; X.s/; .s// ds C E.C.T; tk ; ./I /jFtWk /
kD1 0
D RHS of (2.57);
and
Then Nl;m ./ provides a switching control at Dm , and, as m ! 1, Nl;m ./ approaches
Nl ./ uniformly on Œ0; T , P -a.s.
Third, we can choose .ln ; mn / so that mn n and
Set
[
As D AD D set of all switching controls.
D
VD1n .VD n
1 2
/ VD1n .V1 2 /: (2.70)
V 2 V 1 .V1 2 / (2.72)
D
Next we shall show the converse of inequality (2.72). Since V 2nCm .x/ D
D D D
V 1nCm .V1 nCm
2
/ V 1nCm .VD m
1 2
/.x/, we obtain, letting n ! 1,
V 2 .x/ V 1 .VD m
1 2
/.x/ .m D 1; 2; : : : /
J.1 ; ; x; AI VD m
1 2
/; 8A 2 A: (2.73)
c2 .1 C jxj4 /; (2.74)
lim VD m
1 2
.X.1 // D V1 2 .X.1 // P -a.s.; (2.75)
m!1
follows from (2.73) and (2.76). Taking the infimum of the RHS of (2.77) over
A 2 A, we obtain the converse of inequality (2.72).
This completes the proof of (2.68) when is convex and compact.
Step 2. Let be convex and -compact.
We can take a sequence of convex and compact sets n ; n D 1; 2; such that
[
n nC1 ; n D 1; 2; : : : and n D : (2.78)
n
We denote by An the set of all admissible controls with control region n . Then
[
An AnC1 and An D A; (2.79)
n
where
VnCm m
1 .V1 2 /.x/; m; n D 1; 2; : : : : (2.84)
50 2 Optimal Control for Diffusion Processes
Letting n ! 1, we get
Thus, the converse of inequality (2.83) follows from (2.85) and (2.86).
This completes the proof of Theorem 2.4. t
u
In this section, we will comment on the value function in the case where one restricts
the control processes to the class of Brownian adapted ones.
We denote the set of all admissible controls, A D .˝; F ; .Ft /; P; W; .// where
./ is .FtW /-progressively measurable, by AW . Put
Then we have:
Proposition 2.5.
Proof. Since we can easily see that (2.82) is also valid for VWt , we may assume
for the proof that is convex and compact. Now the proposition follows from
Remark 2.1 and Corollary 2.2. t
u
For A 2 AW ; J.t; ; x; AI / can be calculated by means of the joint probability
distribution .W; .//. So, we fix a reference probability system .˝; F ; .Ft /; P; W /
and identify A with its control process ./.
For a stopping time , we again denote the payoff by J./:
hZ Z s
J. ; ; x; ./I / DE x expf .; X./; .// dgf .s; X.s/; .s// ds
Z i
C expf .; X./; .// dg .X. // ; (2.89)
where X D X ./ .
2.2 Dynamic Programming Principle (DPP) 51
p n Z t 12
T 2 2 E0x K.1 C jX.s/j4 / ds .use .b4 //
0
n
2 2 c1 .1 C jxj2 / (2.92)
by (2.15), and the RHS of (2.93) is integrable, the dominated convergence theorem
yields
and
X
p nZ tk
J.T; 0; x; ./I / D E0x
"
. D tk / f .s; X " .s/; " .s// ds
kD1 0
Z T o
C f .s; X " .s/; " .s// ds C .X " .T // (2.102)
tk
Since any stopping time can be approached by finitely-many valued ones, (2.100)
and (2.103) complete the proof. t
u
2.2 Dynamic Programming Principle (DPP) 53
Example 2.2 (DPP with hitting time). Assume f ./ D 0 and ./ D 0. Let O be
an open set. For ./ 2 W ; X ./ and ./ denote its response and the hitting
time of O by X ./ respectively. Then DP property yields that the value function
v.; x/ D VT .x/ satisfies
Referring to Theorems 2.1 and 2.4 we summarize the basic properties of the
semigroup .V t ; t/; ; t 2 Œ0; T .
Proposition 2.7. V t W CQ ! CQ satisfies the followings:
(i) V D identity, 8 2 Œ0; T ,
(ii) Semigroup property
(iii) Monotonicity
H) V t V t ;
H t .x/ D J.t; ; x; I /; 8x 2 Rd : (2.107)
Then .H t I t/ is a semigroup satisfying properties (i)–(v) of Proposition 2.7
and
8 2 CQ :
V t H t ; (2.108)
8 2 CQ ;
U t H t ; 8 2 ; 8 t; (2.109)
then
U t V t ; 8 2 CQ ; 8 t: (2.110)
DDf 2 CQ \ Cp2 .Rd /I @x and @xx satisfy the polynomial growth conditiong:
Then
1
G D tr.a.; x; /@xx / C b.; x; / @x
2
is the generator of X , with domain D, and
G .x/ .; x; / .x/ C f .; x; /
is the generator of .H t ; t/. Put
G .x/ WD inf .G .x/ .; x; / .x/ C f .; x; //:
2
with constants K > 0 and p 1. Then, for any positive " and R0 , there is ı"R0 > 0
such that for any ./ 2 W ,
h i
sup E x sup jh.s; X ./ .s/; .s// h.; x; .s//j < " (2.113)
jxjR0 st
1
h.t; x; / D tr.a.t; x; /@xx .x// C b.t; x; / @x .x/
2
.t; x; / .x/ C f .t; x; /; 2 D:
we have
Z t
inf E x h.; x; .s// ds D .t / inf h.; x; /: (2.117)
./2 W 2
Inserting (2.117) into (2.114) and referring to Proposition 2.5, we complete the
proof. u
t
Proposition 2.8 asserts that if the value function is smooth, then it satisfies the
Cauchy problem for the (nonlinear) parabolic equation;
@V
.t; x/ C H.t; x; @xx V .t; x/; @x V .t; x/; V .t; x// D 0; 8.t; x/ 2 Œ0; T / Rd
@t
(2.118)
with the lateral boundary condition
H.t; x; A; P; v/
1
D sup tr.a.t; x; /A/ b.t; x; / p C .t; x; /v f .t; x; / : (2.120)
2 2
t D 0.t / : (2.121)
(
O
d X.t/ D b.XO .t/; O .t// dt C ˛.XO .t/; O .t// d W .t/; t > 0;
(2.123)
XO .0/ D x;
1
G .x/ D lim .Vt .x/ .x//
t !0 t
1
D inf tr.a.x; /@xx / C b.x; / @x .x; / C f .x; /
2 2
(2.126)
for 2 D.
58 2 Optimal Control for Diffusion Processes
Here we assume the following conditions .b1 /0 and .b2 /0 weaker than .b1 /–.b5 /:
.b1 /0
with a constant K.
.b2 /0
O C jxj2 C j j2 /;
jf .t; x; /j C j.x/j K.1 (2.129)
O
with a constant K.
Here we admit a slightly bigger class of control processes. Specifically let
.˝; F ; .Ft /; P; W / be a reference probability system. We take . .t/; t 2 Œ0; T / to
be a -valued .Ft /-progressively measurable process, satisfying
2.3 Verification Theorems and Optimal Controls 59
hZ T i
E j .t/jn ds < 1 for n D 1; 2; : : : : (2.130)
0
dX.t/ D b.t; X.t/; .t// dt C ˛.t; X.t/; .t// d W .t/; t 2 Œ; T ; (2.131)
X./ D x: (2.132)
respectively.
Remark 2.2.
Proof. That v .; x/ v.; x/ is clear. For the opposite inequality, we fix A D
.˝; F ; .Ft /; P; W; .// arbitrarily and claim that
Let N .t/ WD E x Œsup st jX.s/XN .s/j2 . From (2.127) and the Burkholder–
Davis–Gundy inequality, we easily deduce that
Z t
N .t/ cE fN .s/ C .1 C j .s/j2 /.j .s/j N /g ds ; (2.137)
@V
.t; x/ C H.t; x; @xx V .t; x/; @x V .t; x/; V .t; x// D 0; 8.t; x/ 2 Œ0; T / Rd
@t
(2.1180)
with the lateral boundary condition
(ii) Let .t; x/ be given. Suppose that there exists AO D .˝;
O FO ; .FO t /; PO ; WO ; O .//
such that
DGsO .s/ !.s; XO .s// .s; XO .s/; O .s//!.s; XO .s// C f .s; XO .s/; O .s//;
For the case of optimal Markovian policies, we deduce from Theorem 2.6 and
Proposition 1.6 the following
Corollary 2.3. In addition to .b1 /0 and .b2 /0 , we assume that
(a) is convex and compact,
(b) d D m,
(c) b./ is bounded,
(d) ˛./ is bounded, symmetric and uniformly parabolic.
Suppose that the HJB equation (2.1180)–(2.1190) has a classical solution, satisfying
the growth condition (2.138). Then the maximum selector of H./ provides an
optimal Markovian policy.
Regarding the classical solution of HJB equation, we recall the following result.
Theorem 2.7 ([FS06], p. 163). Let D 0 and assume;
(a) is compact and d D m,
(b) ˛./ is uniformly parabolic and independent of . Moreover, ˛./ 2 C 12 .Œ0; T
Rd / and ˛./; ˛./1 , and @x ˛./ are bounded,
Q x/ C ˛.t; x/.t; x; /, where b./
(c) b.t; x; / D b.t; Q 2 C 12 .Œ0; T Rd / and
Q is bounded, and ./ and @x ./ are in Cb .Œ0; T Rd /,
@x b./
(d) f ./ and @x f ./ are in Cp .Œ0; T Rd /,
(e) ./ 2 C 3 .Rd / \ Cp1 .Rd /.
Then the HJB equation (2.1180)–(2.1190) has a unique classical solution.
62 2 Optimal Control for Diffusion Processes
We are concerned with explicit formulations of optimal controls for three simple
models. The first two examples are related to Gaussian diffusions effected by linear
control processes. The third one is a stochastic control with state constraint.
Example 2.4 (Linear Gaussian quadratic regulator). Let D Rq and suppose
that matrix-valued continuous functions of suitable sizes A; B; ; M; N and D are
given on Œ0; T . For an admissible control A D .˝; F ; .Ft /; P; W; .//, we have
the d -dimensional SDE
J.T; ; x; A /
hZ T i
DE x .X.s/> M.s/X.s/ C .s/> N.s/.s// ds C X.T /> DX.T / :
(2.145)
t 2 Œ0; T /; x 2 Rd ; (2.146)
dP
0D .t/ C M.t/ C A.t/> P .t/ C P .t/A.t/
dt
P .t/B.t/N.t/1 B.t/> P .t/; t 2 Œ0; T / (2.150)
with
P .T / D D: (2.151)
Hence, from (2.148) and (2.149) it follows that an optimal Markovian policy is given
by N.t/1 B.t/> P .t/x.
More details for LQ problems are given in [YZ99], Chapter 6.
Example 2.5 (1-dimensional bang-bang control). Let D Œ1; 1. Suppose that
g 2 Cp1 .R1 / is even, g.0/ D 0 and, g 0 .x/ 0, on Œ0; 1/.
Let us consider the following time-homogeneous simple model. For A D
.˝; F ; .Ft /; P; W; .//, its response X A evolves according to the SDE
X.0/ D x .2 R1 /; (2.153)
The (formal) HJB equation for the value function v.t; x/ D infA2A j.t; x; A/ reads
8
ˆ
ˆ @v .t; x/ D 1 @xx v.t; x/ C inf . @x v.t; x//
ˆ
ˆ
< @t 2 j j1
1 (2.155)
ˆ
ˆ D @xx v.t; x/ j@x v.t; x/j; t > 0; x 2 R1 ;
ˆ 2
:̂
v.0; x/ D g.x/; x 2 R1 :
and
8
< 1; x > 0;
sgn.@x v.t; x// D sgn x D 0; x D 0; (2.157)
:
1; x < 0:
@!
.t; x/ D @xx !.t; x/; t > 0; x>0 (2.158)
@t
with the boundary and initial conditions
8
ˆ 1
< @x !.t; 0/ C !.t; 0/ D 0; t > 0;
2 (2.159)
x
:̂ !.0; x/ D e 2 g
x
; x > 0:
2
and
x
lim u.t; x/ D g : (2.161)
t !0 2
By Theorem 2.6, v./ is the value function and the minimum selector O .t; x/ D
sgn.@x v.t; x// D sgn x gives an optimal Markovian policy. Its response X
evolves according to SDE
Since (2.163) admits a unique strong solution, X is the diffusion with generator
1 d2 2jxj
2 dx 2 .sgn x/ dx and e
d
gives the density of the corresponding invariant
probability measure. From (2.156) it follows that for any R > 0, there is a constant
CR > 0 such that
ˇ Z 1 ˇ
ˇ ˇ t
sup ˇv.t; x/ e 2jyj g.y/ dy ˇ CR e 2 ; 8t > 1: (2.164)
jxjR 1
By A
we denote the exit time of X A from O, that is
(
inffs > 0I X A .s/ … Og;
A
D (2.166)
1; if f g is empty:
hZ A i
e s f .X A .s/; .s// ds C e g.X A .
A
J.x; A/ D Ex A
// ; x 2 OI
0
(2.167)
here, when A D 1; e g.X A . A // stands for 0.
A
When ./ is constant 2 , its response X A is the diffusion with the generator
G given by
1
G u D tr.a.x; /@xx u/ C b.x; / @x u (2.168)
2
66 2 Optimal Control for Diffusion Processes
then
From (i) and (2.172), it follows that u.x0 / D v.x0 / and A gives an optimal
control for the initial state x0 .
Regarding the classical solution of (2.170), we can refer to [E83], [E98], [Mo10],
Sect. 5.3.
Example 2.6 (Production planning problem). Consider the production planning for
one commodity, in the presence of random demand, fluctuating according to the
SDE
dZ.t/ D b dt d W; (2.173)
where the positive constant b represents the expected demand rate and and W
denote a positive constant and a real Wiener process. The firm adjusts its production
2.3 Verification Theorems and Optimal Controls 67
X.t/ > 0 (resp. < 0) means a surplus (resp. backlog) of product. Since the inventory
state cannot exceed the buffer size K and the firm fixes the backlog size K.< 0/,
because a new policy is needed for a large amount of backlog, we give the payoff
by
hZ A i
J.x; A/ D Ex e s f .X A .s/; .s// ds C e A g.X A . A
// ; (2.174)
0
1 2 00
u .x/ bu0 .x/ u.x/ C inf . u0 .x/ C 2 / C p.x _ 0/ D 0; 8x 2 .K; K/;
2 2
(2.176)
with the boundary conditions
Since there exists a unique classical solution of (2.176), the verification theorem
asserts that the solution is equal to the value function and an optimal Markovian
policy is given by the minimum selector;
8
ˆ
ˆ 0 on fx 2 OI u0 .x/ 0g;
ˆ
< 0
O .x/ D u .x/ on fx 2 OI u0 .x/ 2 .2K; 0/g;
ˆ
ˆ 2
:̂
K on fx 2 OI u0 .x/ 2Kg:
The prices of the bond and of the i -th asset are given by
(
dS 0 .t/ D S 0 .t/r.t; X.t// dt; t > 0;
(2.178)
S 0 .0/ D s 0 > 0;
the bond. The agent wants to maximize the expected utility from the terminal wealth
by choosing a suitable . i ./; i D 1; : : : ; m/.
As an application of our previous results, we study the problem by using a DPP
argument. We note that the martingale approach is also powerful (refer to [HP81]
and [HP83]).
2.4.1 Formulations
Let
W Œ0; T Rd 7! Rm ˝ Rd Cm ;
b W Œ0; T Rd 7! Rd ;
g W Œ0; T Rd 7! Rm ;
r W Œ0; T Rd 7! Œ0; K0
1
Gt D C b.t; x/ @x ; (2.180)
2
if X.0/ Dconstant.
Since S 0 and S i evolve according to (2.178) and (2.179), respectively, we have
Z t
S 0 .t/ D s 0 exp r.s; X.s// ds (2.181)
0
and
nZ t 1 Z t o
S i .t/ D s i exp g.s; X.s//i j.s; X.s//i j2 ds C .s; X.s//i d WQ .s/ :
0 2 0
(2.182)
g.t; X.t// and .t; X.t// are called the mean return process and the volatility
process, respectively.
We assume
Q
./> .; X.// 2 L1 .Œ0; T ˝; .FtW /I R1 ˝ Rd Cm /: (2.184)
dZ ./ .t/
Z ./ .t/
X
m
D i
.t/.g.t; x.t//i dt C .t; X.t//i d WQ .t// C 0
.t/r.t; X.t// dt
i D1
D.r.t; X.t// C .t/.g.t; X.t// r.t; X.t//1m // dt C .t/> .t; X.t// d WQ .t/
(2.185)
xı
power utility function: U.x/ D ; x>0 with ı < 1; ¤ 0; (2.188)
ı
and
The agent’s objective is to maximize the expected utility from the terminal wealth,
EŒU.Z ./ .T //, by choosing an appropriate investment strategy.
2.4 Optimal Investment Models 71
zı Z t
U.Z ./
.t// D M ./
.t/ exp ı .s; X.t/; .s// ds (2.190)
ı 0
where
Z t Z
1 t
M ./
.t/ D exp ı .s/> .s; X.s// d WQ .s/ jı .s/> .s; X.s//j2 ds
0 2 0
(2.191)
and
1ı
.s; x; .s// D j .s/> .s; x/j2 C .s/ .s; x/ C r.s; x/: (2.192)
2
Q
P ./
DM ./
.T / ı P on FTW : (2.193)
./
The following (i)–(iii) are valid with respect to P
Q
(i) WQ ./
given by (2.194) is an .FtW /-Wiener process, i.e.,
Z t
WQ ./
.t/ D WQ .t/ ı .s; X.s//> .s/ ds; t 2 Œ0; T : (2.194)
0
./ ./
where E denotes the expectation w.r.t. P .
72 2 Optimal Control for Diffusion Processes
Y .t/ D x .2 Rd /; (2.198)
where B is the first d -dimensional component of B.Q We define the payoff j./ and
the value function u./ by
h1 Z T i
j.t; x; .// D Etx exp ı .s; Y .s/; .s// ds (2.199)
ı t
and
1
v.t; x/ D log.ıu.t; x//: (2.201)
ı
Calculating formally, we obtain
@v 1
0D C . v C ıj@x vj2 C 2b.t; x/ @x v C 2r.t; x//
@t 2
n 1ı o
C sup j > N .t; x/j2 C > ..t; x/ C ı .t;
N x/@x / : (2.202)
2Rm 2
1
˘O .t; x/ D fa.t; x/1 ..t; x/ C ı N .t; x/@x .t; x//g; (2.203)
1ı
we obtain
@v 1
0D C v C H.t; x; @x v.t; x// on Œ0; T / Rd ; (2.204)
@t 2
with the lateral boundary condition
v.T; x/ D 0; x 2 Rd ; (2.205)
2.4 Optimal Investment Models 73
where
ın 2 ı o
H.t; x; p/ D jpj C p.N > a1 N /.t; x/p >
2 1ı
ı
C b.t; x/ p C f.> a1 N /.t; x/p > C p.N > a1 /.t; x/g
2.1 ı/
1
C r.t; x/ C .> a1 /.t; x/: (2.206)
2.1 ı/
For the existence of optimal strategy, we need to take a broader class than A . Put
Q
A2 D f ./I ./> .; X.// 2 L2 .Œ0; T ˝; .FtW /I R1 ˝ Rm g: (2.207)
and
Q
./ 2 L1 .Œ0; T ˝; .FtW /I Rm / if ./ 2 A : (2.209)
lim Z n ./
.T / D Z ./
.T / P -a.s. (2.210)
n!1
(ii)
On the other hand, for " > 0, there is a positive constant k D k" , such that
Px sup jX.s/j k < ": (2.213)
0sT
74 2 Optimal Control for Diffusion Processes
Put
where
infft < T I jX.t/j kg;
k D
T if f g D empty:
Replacing ./ by l ./ and l" ./, we define l and l" , respectively. Compute
Px .jl" j > "/ D Px .jl" j > "I k D T / C Px .jl" j > "I k < T /:
(2.216)
Since l .s/ D l" .s/, for all s.< T /, for k D T , one has that,
and
Z T Z T
lim > Q
n .s/ .s; X.s// d W .s/ D .s/> .s; X.s// d WQ .s/ P -a.s.
n!1 0 0
(2.220)
2.4 Optimal Investment Models 75
where
1
uQ .t; x/ D exp.ı vQ .t; x//: (2.223)
ı
Ex ./ Œ ./
.T / Ex ./ Œ ./
.0/ D uQ .0; x/ (2.227)
and
ExO ./ Œ O ./ .T / D ExO ./ Œ O ./ .0/ D uQ .0; x/; (2.228)
we have (2.224).
Now, using Lemma 2.2 we conclude that
b.t; x/ D b0 C b1 x; .t; x/ D ;
g.t; x/ D g0 C g1 x; r.t; x/ D r;
and
(
dS i .t/ D S i .t/..g0i C g1i X.t// dt C i d WQ .t//; t 2 .0; T ;
(2.230)
S i .0/ D s i > 0 .i D l; : : : ; m/;
respectively.
The bond price S 0 .t/ is given by
S 0 .t/ D s 0 e rt : (2.231)
1 >
v.t; x/ D x Q.t/x C R.t/ x C S.t/; (2.232)
2
2.4 Optimal Investment Models 77
where
ı ı
K0 D ıId C N > a1 N ; K 1 D b1 C N > a1 g1 I
1ı 1ı
8
< 0 D dR .t/ C .K1 C K0 Q.t//> R.t/ C Q.t/K2 C K3 ; t 2 Œ0; T /;
dt
:
R.T / D 0;
(2.234)
where
ı 1
K 2 D b0 C N > a1 .g0 r1m /; K3 D g > a1 .g0 r1m /I
1ı 1ı 1
and
8
ˆ
ˆ dS 1 1
ˆ0 D
ˆ .t/ C trQ.t/ C R.t/> K0 R.t/
ˆ
< dt 2 2
1
ˆ
ˆ C K2 R.t/ C r C .g0 r1m /> a1 .g0 r1m /; t 2 Œ0; T /;
ˆ
ˆ 1 ı
:̂
S.T / D 0:
(2.235)
If the Riccati equation (2.233) has a solution, then we can solve (2.234) and (2.235),
and v./ given by (2.232) becomes a solution of (2.204)–(2.205) with linearly
growing @x v./.
Hence
O .t/ WD a1 g0 r1m C g1 X.t/ C ı @ N x v.t; X.t// (2.236)
gives an optimal investment strategy (refer to [KN02, W71, Wo68] for Riccati
equations).
Example 2.8 (Uniformly elliptic volatility model). We consider a random volatility
model presented in [Ph02].
78 2 Optimal Control for Diffusion Processes
Let d D m D 1. Suppose that 2 .1; 1/ and " > 0 are given. We assume the
following forms:
p p p
b.t; x/ D b0 C b1 x; .t; x/ D . c 2 x 2 C "; 1 2 c 2 x 2 C "/;
g.t; x/ D g0 C g1 x; r.t; x/ D r > 0
and
p
dS.t/ DS.t/ .g0 C g1 X.t// dt C c 2 X.t/2 C " d W .t/
p p
C 1 2 c 2 X.t/2 C " dB.t/ ; (2.238)
holds by (2.204). Since [Ph02], Theorem 4.1 provides a solution v./ 2 C 12 .Œ0; T /
R1 / \ C.Œ0; T R1 / with @x
@v
satisfying the linear growth condition, we conclude
that
1 g r C g X.t/ @v
0 1
O .t/ D p p C ı .t; X.t//
c 2 X.t/2 C " c 2 X.t/2 C " @x
Abstract The theory of viscosity solutions was originated by M.G. Crandall and
P.L. Lions in the early 80s for the Hamilton–Jacobi equations and later P.L.
Lions developed it for the HJB equations (Lions, J Commun PDE 8:1101–1134,
1983; Acta Math 16:243–278, 1988; Viscosity solutions of fully nonlinear second-
order equations and optimal stochastic control in infinite dimensions. Part II.
Optimal control of Zakai equation. In: Da Prato, Tubaro (eds) Stochastic partial
differential equations and applications II. Lecture notes in mathematics, vol 1390.
Springer, Berlin/Heidelberg, 1989, pp 147–170, 1989; J Funct Anal 86:1–18, 1989).
In Chap. 2, we have seen the relation between the value function and the HJB
equations. If the value function is smooth, then it provides the classical solution
of the HJB equations. Unfortunately, when the diffusion coefficient is degenerate,
smoothness does not necessarily hold, even for a simple case, and the HJB equations
may in general have no classical equation, either. However, the theory of viscosity
solutions gives a powerful tools for studing stochastic control problems. Regarding
the viscosity solutions for the HJB equations, we claim only continuity for a
solution, not necessarily differentiability. Thus, it has been shown that under mild
conditions the value function is the unique viscosity solution of the HJB equation.
We will revisit this fact in terms of semigroups in Sect. 3.1.3.
This chapter is organized as follows. In Sects. 3.1 and 3.2, we recall some basic
results on viscosity solutions for (nonlinear) parabolic equations for later use. In
Sect. 3.3 we consider stochastic optimal control-stopping problems in a framework
similar to that of finite time horizon controls.
3.1 Formulations
Here we are concerned with the Cauchy problem for the (non-linear) parabolic
equation
@t U.t; x/ C F .t; x; @xx U.t; x/; @x U.t; x/; U.t; x// D 0; .t; x/ 2 Œ0; T / Rd ;
(3.1)
with the lateral boundary condition
PN 2C W .t; x/ D f.q; p; A/ 2 R1 Rd S d I
9.tn ; xn / 2 .0; T / Rd and .qn ; pn ; An / 2 P 2C W .tn ; xn /
such that lim .tn ; xn ; W .tn ; xn /; qn ; pn ; An / D .t; x; W .t; x/; q; p; A/g:
n!1
(3.5)
3.1 Formulations 81
.q; p; A/ 2 P 2˙ W .t; x/
” .q C @t f .t; x/; p C @x f .t; x/; A C @xx f .t; x// 2 P 2˙ .W C f /.t; x/:
(3.7)
P 2C U.t; x/ 3 .q; p; A/
8
ˆ
ˆ q D @t U.t; x/ C c; where c D 0 for t ¤ 0, c 0 for t D 0;
<
” p D @x U.t; x/;
ˆ
:̂
A D @xx U.t; x/ C B; with B 2 SC d
:
Suppose that U./ is a classical solution. Then the ellipticity condition .d2 /
implies
@t U.t; x/ C F .t; x; @xx U.t; x/; @x U.t; x/; U.t; x// D 0;
because .@t U.t; x/; @x U.t; x/; @xx U.t; x// 2 P 2C U.t; x/ \ P 2 U.t; x/.
This completes the proof. t
u
with K > 0 and a non-negative integer p. Let u./ be the function figuring in
Proposition 3.3 (b). Then for any 0 < l < m < 1, we can take uQ 2 C 12 .Œ0; T /
Rd / \ C.Œ0; T Rd /, such that
and
Moreover, uQ ./ satisfies (3.10) and W ./ uQ ./ attains its global maximum at .t; x/.
Proof. Let , W Œ0; 1/ 7! Œ0; 1 be smooth functions, decreasing and increasing,
respectively, and satisfying
8 h i
< 1; 2 0; l C m ;
./ D 2 (3.15)
:
0; 2 Œm; 1/;
and
8
ˆ
< 0; 2 Œ0; l;
./ D hl C m (3.16)
:̂ 1; 2 ;1 :
2
84 3 Viscosity Solutions for HJB Equations
We define uQ ./ by
@t u.t; x/ C F .t; x; @xx u.t; x/; @x u.t; x/; U.t; x// 0: (3.19)
@t u.t; x/ C F .t; x; @xx u.t; x/; @x u.t; x/; U.t; x// 0: (3.20)
(iii) If U./ satisfies the polynomial growth condition (3.11), then we may assume
that u./ in .b/ satisfies
and
[
˙2p D ˙2p .k/: (3.23)
k>0
U 2 C.Œ0; T Rd / N C jxj2 /:
with jU.t; x/j K.1 (3.25)
N
Firstly we will show that U./ is a viscosity subsolution of (3.24). Let u 2 ˙2 .K/.
Suppose that .tN; x/
N 2 Œ0; T / Rd is a global maximizer of U./ u./ and
U.tN; x/
N D u.tN; x/.
N Fix a arbitrary reference probability system .˝; F ; .Ft /; P; W /.
Then Proposition 2.5 and DPP yield
u.tN; x/
N D U.tN; x/
N
hZ t Rs Rt i
D inf EtNxN e tN d
f .s; X.s/; .s// ds C e tN d
U.t; X.t//
./2 W tN
hZ t Rs Rt i
inf EtNxN e tN d
f .s; X.s/; .s// ds C e tN d
u.t; X.t// ;
./2 W tN
(3.26)
0 @t u.tN; x/
N C GtNu.tN; x/;
N (3.28)
1
lim fT Cs . C s'/.x/ .x/g D G .x/ C '.x/;
s!0 s
8.; x/ 2 Œ0; T / Rd : (3.30)
Q
kU.t; /kCQ kTt T Tt T 0kCQ C kTt T 0kCQ K.kkCQ C 1/; (3.31)
we have
Q
jU.t; x/j K.kkCQ C 1/.1 C jxj /; 8t; x:
2
(3.32)
Now let us show that U./ is a viscosity subsolution of (3.24). Let u./ 2 ˙2 .
Suppose that U./ u./ attains its global maximum at .tN; x/
N 2 Œ0; T / Rd and
N N D u.tN; x/.
U.T ; x/ N Then
0 D U.tN; x/
N u.tN; x/ N u.tN; x/
N D TtNT .x/ N
D TtN tNCs U.tN C s; /.x/
N u.tN; x/
N
TtN tNCs u.tN C s; /.x/
N u.tN; x/
N .by U u 0/: (3.33)
we have
1
lim e.s; x/ D 0 (3.34)
s!0 s
uniformly on Rd .
3.1 Formulations 87
Q
kTtN tNCs u.tN C s; / TtN tNCs .u.tN; / C s@t u.tN; //kCQ Kke.s; /kCQ : (3.35)
1
lim jTtN tNCs u.tN C s; /.x/ TtN tNCs .u.tN; / C s@t u.tN; //.x/j D 0 (3.36)
s!0 s
1
lim fV Cs . C s'/.x/ .x/g
s!0 s
uniformly in ./ 2 W .
Further, Proposition 2.8 yields
1
lim .V Cs .x/ .x// D G .x/; 8x 2 Rd : (3.40)
s!0 s
This section deals mainly with the comparison principle between a viscosity
subsolution U./ and a viscosity supersolution V ./ of Eqs. (3.1)–(3.2).
The comparison principle asserts that
Let us recall Ishii’s Lemma first, which states the maximum principle for elliptic
differential equations. This result provides a fundamental tool for the theory of
viscosity solutions.
Definition 3.3. 1. Let u 2 US C.RN /. J 2C u.x/ is defined as the superdifferential
of u./ at x,
n
J 2C u.x/ D .p; A/ 2 RN S N I
JN 2C u.x/ Df.p; A/ 2 RN S N I
9xn and .pn ; An / 2 J 2C u.xn / such that
lim .xn ; u.xn /; pn ; An / D .x; u.x/; p; A/g: (3.42)
n!1
J 2 v.x/ D J 2C .v/.x/
3.2 Uniqueness of Viscosity Solutions 89
and
respectively.
Ishii’s Lemma
and
1 X1 0
C kAk I2N A C "A2 ; (3.44)
" 0 X2
where
See [CIL92], Appendix, [Ko04], p. 71, [Mo10], Theorem 4.4.6 for a proof.
For parabolic differential equations, Crandall and Ishii showed a similar fact, by
using Ishii’s lemma.
.q; p; A/ 2 P 2C ui .t; x/
and jxj C jui .t; x/j C jpj C jAj M H) q c.M /: (3.46)
Then, for any " 2 .0; 1/ and any local maximizer .tN; xN 1 ; xN 2 /.2 .0; T / Rd Rd / of
u1 .t; x1 / C u2 .t; x2 / .t; x1 ; x2 /, there exist qi 2 R1 and Xi 2 S d ; i D 1; 2, for
which
q1 C q2 D @t .tN; xN 1 ; xN 2 /; (3.48)
1 X1 0
C kAk I2d A C "A2 ; (3.49)
" 0 X2
@x1 x1 .tN; xN 1 ; xN 2 / @x1 x2 .tN; xN 1 ; xN 2 /
where A D :
@x2 x1 .tN; xN 1 ; xN 2 / @x2 x2 .tN; xN 1 ; xN 2 /
c
.t; x1 ; x2 / D jx1 x2 j2 C ˇ.T t/ C .jx1 j2 C jx2 j2 /; (3.51)
2 2
. C c/Id Id
where A D :
Id . C c/Id
Set
c ˇ
uQ i .t; xi / D ui .t; xi / jxi j2 .T t/ (3.53)
2 2
and
.xi ; x2 / D jx1 x2 j2 : (3.54)
2
Define !i 2 US C.R1 Rd / by
1
'.t1 ; x1 ; t2 ; x2 / D jx1 x2 j2 C jt1 t2 j2 (3.58)
2 2ı
and
respectively. Then ˚ attains its global maximum at some point z.ı/ D .t1 .ı/; x1 .ı/;
t2 .ı/; x2 .ı//.
Since ˚.0; 0; 0; 0/ D 0, (3.52) and (3.56) yield
and
1 Z1 0
C 2 I2d O
.1 C 2"/A; (3.63)
" 0 Z2
Id Id
with AO D .
Id Id
Step 3. Limit of z.ı/ as ı ! 0
We can easily see that !i ./ also satisfies (3.46), with a different c.M /. Hence,
from (3.60) and (3.46) it follows that there is a positive constant c0 such that, for
any ı > 0,
8
ˆ 1
< @t1 '.z.ı// D .t1 .ı/ t2 .ı// < c0 ;
ı (3.64)
:̂ @t '.z.ı// D 1 .t1 .ı/ t2 .ı// < c0 ;
2
ı
that is
1
jt1 .ı/ t2 .ı/j < c0 ; 8ı > 0: (3.65)
ı
1
lim .t1 .ıj / t2 .ıj // DW q .jqj c0 /; (3.70)
j !1 ıj
and
D 0; i D 0; i D 1; 2: (3.72)
!1 .; 1 / C !2 .; 2 / j1 2 j2 : (3.73)
2
However, (3.73) contradicts (3.57) if .; i ; 2 / ¤ .0; 0; 0/. Thus (3.72) is valid.
Collecting the above results, we see that
(
.q; 0; Z1 ."// 2 PN 2C !1 .0; 0/;
(3.74)
.q; 0; Z2 ."// 2 PN 2C !2 .0; 0/:
Structural Condition
Id 0 A 0 Id Id
3 3 ; (3.75)
0 Id 0 B Id Id
94 3 Viscosity Solutions for HJB Equations
˛ W Œ0; T Rd 7! Rd ˝ Rm ;
b W Œ0; T Rd 7! Rd ;
W Œ0; T Rd 7! Œ0; c0 ;
f W Œ0; T Rd 7! R1 (3.77)
F .t; x; A; p; z/
1
D sup tr.a.t; x; /A/ b.t; x; / p C .t; x; /z f .t; x; / (3.78)
2 2
F .t; x; A; p; z/
n 1
D inf sup tr.a.t; x; I ; II /A/
I 2 I II 2 II 2
o
b.t; x; I ; II / p C .t; x; I ; II /z f .t; x; I ; II / : (3.79)
X
m
jk k j2 c1 jx yj2 ; 8: (3.80)
kD1
3.2 Uniqueness of Viscosity Solutions 95
Observing that
X
m >
k Id Id k
3 .by (3.75)/
k Id Id k
kD1
X
m
D 3 jk k j2 3c1 jx yj2 .by (3.80)/;
kD1
O p;
jF .t; x; A; p; z/ F .t; x; A; O zO/j
O C jp pj
l0 fjA Aj O C jz zOjg; 8t; x; (3.81)
Step 1. Preparation
Let us suppose the contrary, namely there is .; z/ 2 .0; T / Rd , such that
For > 0, we put 0 D 1 and
U .t; x/ WD U.t; x/ ; .t; x/ 2 Œ0; T Rd : (3.85)
t
Then U 2 US C.Œ0; T Rd / and, for .t; x/ 2 .0; T / Rd ,
.q; p; A/ 2 P 2C U.t; x/ ” q C 2 ; p; A 2 P 2C U .t; x/:
t
Thus, putting q D q C t2
, condition .d1 / yields
and
For ˇ 2 .0; 4
T
/ and > 1, we put
'.t; x; y/ D jx yj2 C ˇ.T t/ . 0/ (3.91)
2
and
Then there exists a global maximizer of ˚, say .tNˇ" ; xN ˇ" ; yNˇ" /, because (3.83)
leads to
uniformly in t.
3.2 Uniqueness of Viscosity Solutions 97
UQ .tN; x/
N VQ .tN; y/
N ˚.tN; x;
N y/
N ˚.; z; z/ 4 (3.93)
and
we have
K
jxj
N 2 C jyj
N2 DW C 2 : (3.95)
"
On the other hand, since
0 ˚.tN; x; N ˚.tN; x;
N y/ N D VQ .tN; y/
N x/ N C VQ .tN; x/
N jxN yj
N 2; (3.96)
2
(3.83) and (3.95) yield
N 2 V .tN; y/
jxN yj N C V .tN; x/
N C ".jxj
N 2 C jyj
N 2 / 3K: (3.97)
2
jU.t; x/ .x/j C jV .t; y/ .y/j < ; 8x; y 2 SC : (3.99)
2T
Then for t > t1 ,
N y/
˚.t; x; N < .x/
N .y/
N C < .x/
N .y/:
N (3.100)
2T t
Since ./ is uniformly continuous on SC , (3.97) provides a large 0 ."/, such that
j.x/
N .y/j
N < for > 0 ."/: (3.101)
98 3 Viscosity Solutions for HJB Equations
Finally, we consider the limit of the maximizer. Fix ˇ; ", and . By (3.95), (3.97),
(3.98), and (3.102), we can take n ; n D 1; 2 : : : , tending to 1, so that
and
.q; p; A/ 2 PN 2C UQ .t; x/
” .q; p C 2"x; A C 2"Id / 2 PN 2C U .t; x/; (3.106)
jF .t; x; A; p; U .t; x// F .t; x; A C 2"Id ; p C 2"x; U .t; x//j
p
2l0 ". d C jxj/ .by (3.81)/; (3.107)
and
F .t; x; A; p; UQ .t; x// F .t; x; A; p; U .t; x// .by .d1 //: (3.108)
O p;
Similarly, for .q; O 2 PN 2 VQ .t; y/,
O A/
p
O p;
qO C F .t; y; A; O VQ .t; y// 2l0 ". d C jyj/: (3.110)
3.2 Uniqueness of Viscosity Solutions 99
Now we apply Theorem 3.4 to ˚.t; x; y/. For tN D tNˇ" ; xN D xN ˇ" , and yN D
yNˇ" , we take q; qO 2 R1 and A; AO 2 S d such that
N A/ 2 PN 2C UQ .tN; x/;
.q; .xN y/; N
O .xN y/;
.q; N A/O 2 PN 2 VQ .tN; y/;
N
q qO D ˇ;
Id 0 A 0 Id Id
3 3 :
0 Id 0 AO Id Id
and
p
qO C F .tN; y; O .xN y/;
N A; N VQ .tN; y/
N 2l0 ". d C jyj//:
N (3.112)
LHS of (3.113)
In Sect. 3.3, we control not only the dynamics of stochastic system, but also the
terminal time of its evolution. Accordingly, here an admissible system consists a
pair of admissible control A D .˝; F ; .Ft /; P; W; .// and an .Ft /-stopping time
. By A, we denote the set of all admissible systems. We will deal with this problem
in the same framework as the control problem.
Firstly we formulate control-stopping problems in Sect. 3.3.1. In Sects. 3.3.2
and 3.3.3, we study the DPP and the viscosity solutions for the HJB equations via
semigroup arguments respectively. Section 3.3.4 deals with the American option
price problem as an example.
3.3.1 Formulations
We assume that all coefficients ˛; b; and f are time independent and that
conditions .b1 /–.b5 / in Sect. 2.1 are satisfied. .A; / is called an admissible system
if A D .˝; F ; .Ft /; P; W; .// is an admissible control and is a Œ0; T -valued
.Ft /-stopping time. A denotes the set of all admissible systems.
For .A; / 2 A, we have the following SDE (3.116) for the response X A and the
payoff j./ in the control-stopping problem:
and
C.s; AI /
Z s Z
D expf .X A .h/; .h// dhgf .X A ./; .// d
0 0
Z s
C expf .X A .h/; .h// dhg.X A.s//:
0
Thus, we aim to analyze the value function and find an optimal admissible system.
For simplicity, we assume that T is an integer.
3.3 HJB Equations for Control-Stopping Problems 101
Q C jxj2 /;
sup jv.t; xI /j K.1 (3.119)
0t T
sup jn .y/ .y/j < " for n > N."; R/:
y2SR
Hence
h i
Ex sup jn .X A .s// .X A .s//j
0sT
h i
" C 2KEx 1 C sup jX A .s/j2 I sup jX A .s/j R
0sT 0sT
102 3 Viscosity Solutions for HJB Equations
s q
2 Ex Œsup0sT jX A .s/j2
" C c1 Ex 1 C sup jX A .s/j2
0sT R
1 C jxj3
"c2 (by (2.9)) (3.125)
R
with constants c1 and c2 independent
p
of A.
1C R
Taking R so that “c2 R < "”, we conclude that, for n > N."; R/,
1
sup jv.t; xI n / v.t; xI /j < 2" for jxj < R 6 :
0t T
3.3.2 DPP
Before we start the proof, we prepare two lemmas. First of all, we introduce the
mapping jN W CQ ! CQ , according to [S08], (2.6), [N81], 2.4.1 [N78]. Put 4 D 2N
and fix a reference probability system .˝; F ; .Ft /; P; W /.
Let us define N and jN W CQ ! CQ by
and
Proof. We divide the proof into three steps. Fix a reference probability system
.˝; F ; .Ft /; P; W /. By SW we denote the set of all .FtW /-stopping times with
values in Œ0; T .
Step 1. Construction of ./ on the reference probability system
For 2 CQ , J.4; y; I / is continuous w.r.t. .y; / 2 Rd . Since is
compact, there exists a minimum selector O W Rd ! , such that
Define ./ by
Step 2. Construction of 2 SW
Put
Then D.0/ D Rd and the closed set D.l/ is decreasing. Thinking of D.k l/ as
the stop region at time l4, we define by
.!/ D minfi 4I X .i 4; !/ 2 D.k i /g
D minfi 4I .X .i 4; !// D jNki .X .i 4; !//g: (3.138)
Since jN0 D , is an .FtW /-stopping time taking values in f0; 4; : : : ; k4g.
Step 3. . ./;
/2 W
SW satisfies (3.132). Note that the triviality of F0W
yields
Px . D 0/ D 1 or 0: (3.139)
For x 2 D.k/, we have Px . D 0/ D 1 and
X
k
DW Jl : (3.142)
lD1
3.3 HJB Equations for Control-Stopping Problems 105
Observing that . D k4/ D . .k 1/4/c 2 F.k1/4
W
and
Z k4 Z s
Ex expf ./ dgf .X .s/; O .X ..k 1/4/I // ds
.k1/4 .k1/4
Z k4
C expf ./ dg.X .k4//jF.k1/4
W
.k1/4
DN .X ..k 1/4//; (3.143)
where ./ denotes .X ./; O .X ..k 1/4/I // for simplicity, we have
Therefore,
J D Jk C Jk1 C C J1
D Ex ŒC.4; ./; jNk1 /I
4
D N .jNk1 /.x/ D jNk .x/; (3.146)
because x … D.k/ and Px . 4/ D 1.
Now (3.146) and (3.140) complete the proof of the lemma. t
u
Now we are in the position to prove Proposition 3.6.
Proof. First we will show that, for .A; / 2 AN ,
and
Theorem 3.7. Let be convex and -compact. Suppose .b1 /–.b5 / holds. Then for
2 CQ ,
for 0 s, t s C t T , x 2 Rd .
Proof. We divide the proof into three steps.
Step 1. Let be convex and compact. Let t and s be binary, say, t D i 2p and
s D j 2p .
Since vN .; yI / is decreasing to v.; yI / as N ! 1, Proposition 3.6 and
the monotonicity property of vN ./ yield
and, as N ! 1,
lim C.t ^ ; ./I vm .s; I // D C.t ^ ; ./I v.s; // P -a.s. (3.169)
m!1
and
jC.t ^ ; ./I vm .s; I /j c1 1 C sup jX.t/j2
0t T
lim j.t; x; ../; /I vm .s; I // D j.t; x; ../; /; v.s; I //; (3.170)
m!1
Step 2. Let t and s be given. Suppose that binary tm and sn approach t and
s, respectively. Setting n D v.sn ; I / in Proposition 3.5 and using (3.119)
and (3.121), we get
Now
S
Let AN denote the set of all .A; / with control region N . Then A D N AN
by Definition 2.1. Set
From DPP for vN and (3.175) we obtain (3.164) by using the same arguments as
in Step 2 of the proof of Theorem 2.4.
This completes the proof. t
u
H) Vt Vt ; 8t 2 Œ0; T I
(v)
Vt and Vt Ht ; 8t 2 Œ0; T ; 8 2 ;
Ht .x/ D Ex C.t; I /;
(vi) Maximality
Suppose that .Ut I t 2 Œ0; T / is a semigroup on CQ , satisfying (i)–(v). Then
Ut Vt ; 8t 2 Œ0; T ; 8 2 CQ :
O
G.x/ D 0 ^ inf h .x/; (3.179)
2
where h is the generator of Ht . Then for any R > 0,
ˇ1 ˇ
ˇ O ˇ
lim sup ˇ .Vt .x/ .x// G.x/ ˇD0 (3.180)
t !0 jxjR t
Proof. We use the same arguments as in Proposition 2.8. Fix a reference probability
system .˝; F ; .Ft /; P; W /. For simplicity, we put
On the other hand, Lemma 2.1 asserts that, for " and R > 0, there is t0 D t0 ."; R/
such that
whenever s < t0 .
Plugging (3.182) into the RHS of (3.181), we have
O
D .G.x/ "/t for x 2 SR ; t < t0 : (3.183)
Taking the infimum of the LHS of (3.183) over ../; / 2 W SW and dividing
both sides by t, we obtain
1 O
.Vt .x/ .x// G.x/ > "; 8x 2 SR ; t < t0 : (3.184)
t
For the converse inequality, we notice that
Vt .x/ .x/
Z t Z s
0 ^ inf Ex expf .X./; / dgh .X.s// ds
2 0 0
hZ t Z s i
0 ^ inf Ex expf .X./; / dgh .x/ ds C "t (by (3.182))
2 0 0
0 ^ inf h .x/t C "t C sup jh .x/jt 2 c0 (by .b5 /): (3.186)
2 2
The HJB equation associated with the control-stopping problem reads and is
considered
where
1
H.x; Q; p; z/ D sup tr.a.x; /Q/ b.x; / p C .x; /z f .x; / :
2 2
Proposition 3.9. Suppose that the coefficients ˛; b; , and f are bounded, continu-
ous, and Lipschitz continuous w.r.t. x uniformly on . Then the value function is the
unique bounded continuous viscosity solution of (3.187)–(3.188), for 2 Cb .Rd /.
Proof mimics Theorem 3.3, because V t WD Vt is a 2 parameter semigroup
satisfying (3.30).
Let us apply previous results to the HJB equation arising from the problem of
American option price. Following [Ma00], we consider a time-homogeneous one-
dimensional model.
The asset price ./ evolves according to the linear SDE
where and are positive constants and W is a real Wiener process. Let T >
0 denote the maturity time. For a Lipschitz continuous function g, the American
option gives an asset holder the right to get the amount of money g.. // with the
discount rate at his/her chosen stopping time . T /. Thus the asset holder wants
to maximize its expectation by choosing a suitable stopping time.
First we formulate the problem and recall its HJB equation. Put
WC
S D f I is an.Ft /-stopping time valued in Œ0; T g (3.190)
3.3 HJB Equations for Control-Stopping Problems 113
and
where
Hence, u./ is continuous on Œ0; T .0; 1/ and (3.187) yields that, when u./ is
smooth,
2 @2 u @u
0 D @t u C 0 ^ x 2 2 x u on Œ0; T / .0; 1/ (3.194)
2 @x @x
with the lateral boundary condition
Next we revisit the problem, by changing the space variable x to y D log x. Put
2 @2 2 @
GD C :
2 @x 2 2 @x
114 3 Viscosity Solutions for HJB Equations
Since
E x Œe . /
g.. // D E log x Œe . /
f .. // (3.198)
holds, putting
we have
with
v.T; y/ D g.e y /; y 2 R1 :
and
Let .; y/ be given. ./ denotes the solution of (3.197) with ./ D y. Noticing
that v f and v.T; / D f ./, and using (3.202) and (3.203), we can take an optimal
stopping time in S such that
D inffs 2 Œ; T I v.s; .s// D f ..s//g: (3.204)
3.3 HJB Equations for Control-Stopping Problems 115
Indeed, the Itô–Krylov formula together with (3.202) and (3.203) yields
^
Ey Œe . R /
v.
^ R ; .
^ R / D v.; y/; (3.205)
where R denotes the exit time of ./ from ŒR; R. Since v./ is bounded, the
bounded convergence theorem yields
/
Ey Œe . v.
; .
// D v.; y/;
4.1 Formulations
In this section we introduce some basic concepts for stochastic differential games
and formulate problems. When players I and II choose Wiener-adapted control
processes y WD y./ and z WD z./ respectively, the response X evolves according to
SDE and the cost functional is given by the usual form:
Z t
C.t; ; x; y; zI / D f .X.s/; y.s/; z.s// ds C .X.t//
for y; z with X./ D x: Its expectation is called the payoff. Now player I wants to
maximize the payoff and player II wants to minimize it.
In Sect. 4.1.1, we define control processes and strategies, based on the concepts
in [FS06], XI. 4, and study properties of responses. In Sect. 4.1.2 we formulate
two kinds of stochastic differential games and investigate lower- and upper value
functions.
8
< YW 1
0 Œ; t D L .Œ; t; .Ft /I Y/; Y0 D YW
W
0 Œ0; T ;
(4.1)
: ZW Œ; t D L1 .Œ; t; .F W /I Z/; Z0 D ZW
0 t 0 Œ0; T :
under P .jFW /, we freeze W in (4.2) and .y.t/; t 2 Œ; T / can be regarded as an
WC
element of L1 .Œ0; T ; .Ft /I Y/.
Denote by YSC the set of all switching processes of Y0 , namely y.t/ D y.ti /; t 2
Œti ; ti C1 /; i D 0; : : : ; p with some 0 D t0 < t1 < < tp < tpC1 D T . For Y Y0
and t > 0; Yt denotes the set of restriction of y 2 Y to Œ0; t.
Next we introduce admissible controls, by referring to [FS06], X1.4. Let Y be a
subset of Y0 satisfying
(a) Y YSC ;
(b) For y1 ; y2 2 Y and 2 .0; T /, the concatenation y given by
is in Y;
WC
(c) Under P .jFW /; .y.t/; t 2 Œ; T / can be regarded as an element of YT
(D
C
the set given by YT with W replaced by W ).
Clearly YSC and Y0 satisfy (a)–(c) and
YSC Y Y0 : (4.3)
and
dX.t/ D b.X.t/; y.t/; z.t// dt C˛.X.t/; y.t/; z.t// d W .t/; t 2 .0; T ; (4.6)
X.0/ D x .2 Rd /: (4.7)
We always assume that the coefficients ˛; b, and f .D running cost) are bounded
and Lipschitz continuous, say
(d)
and
j˛.x; y; z/ ˛.x;
Q y;
Q zQ/j C jb.x; y; z/ b.x;
Q y;
Q zQ/j C jf .x; y; z/ f .x;
Q y;
Q zQ/j
l0 .jx xj
Q C jy yj
Q C jz zQj/:
Proposition 4.1. The SDE (4.6)–(4.7) admits a unique .FtW /-adapted solution.
Moreover, the following estimates hold:
h i
Ex sup jX.s/j2p Kp .1 C jxj2p /; p 1; (4.8)
0sT
h i
Ex sup jX.s/ X./j2p Kp jt jp ; p 1; (4.9)
st
h i
E sup jXx1 .s/ Xx2 .s/j2 K1 jx1 x2 j2 ; (4.10)
0sT
We are concerned with two kinds of stochastic differential games, lower games and
upper games. In the lower game, the player I chooses an admissible control y first
and y is known when the player II chooses an element of Z. Hence player II chooses
an admissible strategy from Z . When y 2 Y and 2 Z are chosen, the response
X y evolves according to (4.6) with z D .y/. Thus, when game stops at t, the
payoff is given by
hZ t i
J.t; x; y; I / D Ex f .X.s/; y.s/; .y/.s// ds C .X.t// (4.12)
0
The upper game is defined in the similar way and the .Z; Y /-upper value V C is
given by
C
The Elliott–Kalton upper value. is denoted by VEK . However, it seems to be unfair
that only the stronger player chooses a strategy. Now following [FH11] we reduce
the class of strategies to a smaller one which eliminates the information advantage.
We will treat mainly the lower game for a while.
Definition 4.5. 2 ZEK is called a strictly progressively measurable strategy, if for
each 2 YEK the equations
have a solution, namely and have fixed points. Zs denotes the set of strictly
progressively measurable strategies.
Let 2 Zs and 2 YEK be given. Then, it is immediate that
Since (4.4) yields that, for s 2 Œ0; t1 /, .y/.s/ D .y/.0/ D constant independent of
y.DW zQ /, we see that
holds by Definition 4.2. Hence, zk .t1 / D .yk /.t1 /.k D 1; 2; : : :/ does not depend
on k, and (4.21) holds for j D 1.
If (4.21) holds for j , then the same arguments lead to
and
zj C2 D zj C1 on Œ0; tj C2 / (4.24)
and
yj C2 D yj C1 on Œ0; tj C1 /; (4.25)
Indeed, for 2 YEK , we put y0 D .z0 /. Then z0 .y0 / D z0 yields that .y0 ; z0 /
is a solution. If .Oy; zO / is a solution, then z0 .Oy/ D zO shows that zO D z0 . Hence
yO D .z0 / D y0 .
Regarding the solution of (4.15) and related topics in deterministic differential
games, [FKSh10], Sect. 4 presents interesting results.
The case when the equality holds in (4.17), i.e.,
Consequently,
and so
Before we end Sect. 4.1, let us list some basic properties of payoff, by using
Proposition 4.1.1.
Proposition 4.2. Let 2 Cbu .Rd /.
(i) Bound
(iv) Contractiveness
4.2 DPP
We have already emphasized the importance of DPP for stochastic control problems
in Chap. 2. Here we will consider DPP (sometimes we say minimax principle) for
stochastic differential games. In this section, we deal with stochastic games where
the player with information advantageous uses Elliott–Kalton strategies and the
other player uses switching controls. Hence, it is dealing with an unfair game. By
using semi-discretization arguments, introduced in [N88] and [FSo89], we firstly
consider D-lower and D-upper value functions in Sect. 4.2.1. These cases admit the
discrete time DPP. Taking finer and finer divisions of Œ0; T , we obtain the DPP for
the limit games in Sect. 4.2.2 (see Theorem 4.2).
Here we are concerned with DPP for vD and uD . From now on, we mainly consider
vD , because uD can be treated by the same arguments.
126 4 Stochastic Differential Games
(v) Contractiveness
Note. Let k < p and 2 Œtk ; tkC1 / be given. By D we denote the division 0 <
tkC1 < tkC2 < < tn < tnC1 D T on Œ0; T .
4.2 DPP 127
We define vD in the same way as vD . Then Theorem 4.1 leads to the DPP
whenever x1 ; x2 2 j .j D 1; 2; : : : /.
We fix xj 2 j arbitrarily and put
1
X
y"t .x/ D yO t .xj /j .x/: (4.48)
j D1
X
p1
" .z/.t/ D Œti ;ti C1 / .t/y";ti C1 ti ;i C1 .X.ti // C Œtp ;T .t/y"; tp ; .X.tp //
i D0
(4.49)
where
pC1 D ; p D . tp /pC1 ; : : : ;
l D .tlC1 tl /lC1 ; : : : ; 0 D .t1 /1 ; (4.50)
and X is the response for ." ; z/ constructed by the following stepwise procedure.
Let x be the initial state. Fix z 2 Z0 arbitrarily and define " by
.X.s/; s t1 / denotes the response for ." ; z/ with X.0/ D x. Next we define
." .z/.s/; s 2 Œt1 ; t2 // by
" .z/.s/ D " .z/.t1 / D y";t2 t1 ;2 .X.t1 //: (4.52)
Thus, we have " .z/ on Œ0; t2 / and its response .X.s/; s 2 Œ0; t2 /. Repeating this
procedure, we obtain " .z/ on Œ0; tp / and its response .X.s/; s 2 Œ0; tp /. Finally,
we put
" .z/.s/ D " .z/.tp / D y"; tp ; .X.tp //; s 2 Œtp ; T : (4.53)
Step 4. Construct a nearly optimal " 2 ZEK by using arguments similar to those
in Steps 1 and 2.
Let " > 0 be given. By (4.36), we can take d" > 0, such that
1 X
X l
zt .y; x/.s; w/ D i .x/Yj .y/z"t ij .s; w/: (4.62)
i D1 j D1
.X.s/; s 2 Œ0; t1 / denotes the response for .y; " / with X.0/ D x. Then we put
Observing that " 2 ZEK and letting " ! 0 in (4.69) we conclude (4.59).
Finally we show the opposite of inequality (4.59) by using (4.57). Since "
of (4.57) is in Y D , Example 4.1 says that, for any 2 ZEK , there exist yO 2 YD and
zO 2 Z0 such that
Therefore,
which concludes the opposite inequality in (4.59). This completes the proof of
theorem. u
t
4.2 DPP 131
Then
Q
jvD .t; xI / vD .t; xI /j < 2"; 8x; if tpC1 tp < "˚ ; (4.74)
and
Q
vD .t; xI / D .t1 / .tp tp1 /. tp /.tpC1 /˚.x/: (4.76)
Put
and
Q
jvD .t; xI / v .t; x/j k. tp /.tpC1 /˚ ˚kC < ";
Q
v.t; x/ D v.Dn / .t; xI /; vQ .t; x/ D v.Dn / .t; xI /;
and Dnk D Dn [ DQ k .
Then for all n; k D 1; 2; : : : ,
and
uniformly in n.
Proof of Lemma. Let X be the response for y 2 Y0 and z 2 Z0 . Then (4.8)
implies
(
1; for jyj " ;
" .y/ D (4.89)
0; for jyj " C 1:
Then we have
by Lemma 4.2. Thus, (4.93) and (4.94) give the opposite inequality.
This completes the proof of the theorem. t
u
To conclude Sect. 4.2, we introduce the upper value function. Under the same
conditions as in Proposition 4.4, the Dn -upper value function uDn .I / is decreasing
uniformly on any bounded set of Œ0; T Rd , as n ! 1. Its limit u.I /, called the
upper value function, is independent of Dn ; n D 1; 2; : : : and the DPP holds:
for 0 t, s t C s T .
Using (4.380), (4.58), and Theorem 4.1, we have
Hence,
When the two value functions coincide, we use the name value function.
In this section, we study the dynamics of lower- and upper value functions, via the
semigroup formulation of DPP. Referring to Sect. 2.2, we introduce semigroups on
Cbu .Rd / related to DPP for lower (resp. upper) value function in Sect. 4.3.1. Their
generators lead to a (nonlinear) parabolic equation, called the lower (resp. upper)
Isaacs equation. In Sect. 4.3.2, we prove that the lower value function is the unique
viscosity solution of the lower Isaacs equation, by using the same arguments as in
Sect. 3.1.3. Similar results are valid for the upper value function.
4.3 Isaacs Equations 135
(ii) Monotonicity
H) Vt Vt ; 8t:
(iii) Contractiveness
kVt Vt kC k kC ; 8t:
(iv) Continuity
Fix y 2 Y arbitrarily. From the definition of I.t; x; yI / (see (4.41)) the maps
y
It I Cbu .Rd / 7! Cbu .Rd / given by
y
It .x/ D I.t; x; yI /; t 2 Œ0; T ; (4.99)
form a semigroup, which formulates the DPP for the value function of stochastic
control (cf. Sect. 2.2). Hence, putting
1
G.y; z/ .x/ D tr.a.x; y; z/@xx .x// C b.x; y; z/ @x .x/ C f .x; y; z/
2
(4.100)
for 2 Cb2 .Rd /, we have
Proposition 4.6. For any y 2 Y,
(i) Semigroup
y
I0 Didentity map,
y y y
It Cs D It .Iys / D Iys .It /; 0 t; s t C s T:
136 4 Stochastic Differential Games
(ii) Monotonicity
y y
H) It It ; 8t:
(iii) Contractiveness
y y
kIt It kC k kC ; 8t:
(iv) Continuity
y y y
lim kIt C I kC lim kIt kC D 0; 8;
t !0 t !0
(v) Generator
For 2 Cb2 .Rd /,
1 y
lim .It .x/ .x// D inf G.y; z/ .x/:
t !0 t z2Z
By using the same arguments as in Sect. 2.2.5, the following proposition is also
immediate.
Proposition 4.7. Let Q denote the set of all continuous, monotone, and contractive
semigroup .Qt / WD .Qt I t 2 Œ0; T / satisfying
y
Qt It ; 8 2 Cbu .Rd / (4.101)
for any t and y. Then .Vt / is the minimal element of Q, that is .Vt / 2 Q and for any
.Qt / 2 Q,
1
G .x/ D lim .Vt .x/ .x// D sup inf G.y; z/ .x/ .WD G .x//: (4.103)
t !0 t y2Y z2Z
8x 2 SR ; y 2 Y0 ; 2 ZEK : (4.105)
X
p Z tj C1
c.t/ inf sup E G.y.tj /; .y/.s// .x/ ds C "t: (4.106)
2ZEK y2YD tj
j D0
O
G.y; .y// D inf G.y; z/ .x/: (4.107)
z2Z
Define 2 ZEK by
O
.y/.s/ D .y.tj // for s 2 Œtj ; tj C1 /; j D 0; 1; : : : ; n: (4.108)
RHS of (4.106)
X
p Z tj C1
sup E G.y.tj /; .y/.s// .x/ ds C "t (4.109)
y2YD j D0 tj
and
Z tj C1
G.y.tj /; .y/.s// .x/ ds D inf G.y.tj /; z/ .x/.tj C1 tj /
tj z2Z
G .x/.tj C1 tj /: (4.110)
Step 3. Conclusion
Suppose that Dn ; n D 1; 2; : : : satisfy the conditions of Proposition 4.4. Then
vDn .I / is increasing to v.I / uniformly on any bounded subset of Œ0; T Rd ,
and so (4.111) yields
1
lim sup .v.t; xI / .x// G .x/: (4.112)
t !0 t
1
lim inf .Vt .x/ .x// G .x/: (4.114)
t !0 t
In the case when the lower value function v.I / is smooth, Theorem 4.3 says that
v.I / satisfies
Now we intend to prove that the lower value function satisfies (4.115)–(4.116) in
the viscosity sense. Define the Hamiltonian H W Rd S d Rd 7! R1 by
1
H.x; A; p/ D inf sup tr.a.x; y; z/A/ b.x; y; z/ p f .x; y; z/ ; (4.117)
y2Y z2Z 2
Theorem 4.4. Let .St ; t 2 Œ0; T / be a continuous semigroup on Cbu .Rd / with
monotone and contraction properties. Suppose that for any 2 Cb2 .Rd / and 2
Cbu .Rd /,
1
lim .St . C t/ .x// D G .x/ C .x/; 8x: (4.119)
t !0
t
Then, for g 2 Cbu .Rd /; .St g; t 2 Œ0; T / is the unique viscosity solution of the
lower Isaacs equation (4.118) with the initial condition v.0; x/ D g.x/.
Proof. We can mimic the proof of Theorem 3.3. Put U.t; x/ D St g.x/. First let us
show that
Taking the 12 " -net ftj ; j D 1; : : : ; pg of Œ0; T and noticing that U.tj ; / is in
Cbu .Rd /, we can choose a constant ı" > 0, so that
0 D U.tO; x/
O .tO; x/
O
D S U.tO ; /.x/
O .tO; x/
O
S .tO ; /.x/
O .tO; x/:
O (4.124)
140 4 Stochastic Differential Games
On the other hand, the contractiveness of S together with (4.123) imply that
O @t .tO; x/
0 G.tO; x/ O
we deduce (4.119) from (4.103) and (4.127). This completes the proof. t
u
Finally, we consider the upper value function, in the same way. Then the
following two results are easy to verify.
Proposition 4.8. For 2 Cbu .Rd /, the upper value function u.t; xI / is the unique
bounded viscosity solution of the upper Isaacs equation
8
< @t u.t; x/ inf sup G.y; z/u.t; x/ D 0; t 2 .0; T ; x 2 Rd ;
z2Z y2Y (4.129)
: u.0; x/ D .x/; x 2 Rd :
4.4 Risk Sensitive Stochastic Controls and Differential Games 141
1
D sup inf tr.a.x; y; z/A/ C b.x; y; z/ p C f .x; y; z/ (4.130)
y2Y z2Z 2
for x 2 Rd ; A 2 S d ; p 2 Rd . Then,
This section is concerned with a link between stochastic controls and differential
games, via risk sensitive stochastic controls. Here we want to minimize the
exponential-of-integral risk sensitive criterion
1 Z T
E exp f .X ./ .t/; .t// dt ;
" 0
where the response X ./ evolves according to an SDE with small noise
(see (4.132)). Applying the logarithmic transformation to the value function, we
obtain the upper Isaacs equation, the corresponding stochastic differential games
of which are considered in Sect. 4.4.1. Section 4.4.2 deals with differential games
obtained as the limit of these stochastic differential games when " tends to 0. In
Sect. 4.4.3, we mention a control problem with infinite time horizon.
˛ W Rd 7! Rd ˝ Rd ;
b W Rd 7! Rd ;
f W Rd 7! R1
142 4 Stochastic Differential Games
1
are in Cbu .Rd / and Cbu
2
.Rd / uniformly on , say
where a D ˛˛ > .
Let " 2 .0; 1/ be given. For a control process ./ 2 W
, its response evolves
according to the SDE
p
dX.t/ D b.X.t/; .t// dt C "˛.X.t/; .t// d W .t/ (4.132)
Since a controller wants to minimize the payoff, the value function is defined by
Next we explain how v" is related to the upper value function of stochastic
differential game, by performing the logarithmic transformation of v" .
4.4 Risk Sensitive Stochastic Controls and Differential Games 143
1 > 1
p ap D sup p > ˛y jyj2 (4.140)
2 y2Rd 2
with
where
"
G" .y; / .x/ D tr.a.x; /@xx .x//
2
1
C @x .x/ .b.x; / C ˛.x; /y/ C f .x; / jyj2 : (4.143)
2
Since, by (4.139), j@x V " .t; x/j M with some constant M D M" , the
function @x V " .t; x/> ˛.x; r/y 12 jyj2 attains its maximum at yO t;x; 2 SMK0 .
Consequently (4.141) is recast as
for N MK0 .
Proposition 4.11. Let U 2 Cb .Œ0; T Rd / be a viscosity solution of (4.141N )–
(4.142) for some large N . Then U coincides with the upper value function of the
stochastic differential game and U D V " holds.
Proof. Since (4.141N ) is the upper Isaacs equation, U is equal to the upper value
function of an stochastic differential game between player I and the controller
(player II) with control regions SN and , respectively. Now the proposition follows
from the uniqueness of the viscosity solution. t
u
144 4 Stochastic Differential Games
with
V .0; x/ D 0 on Rd : (4.145)
S
Put YR D L1 .Œ0; T I SR /; Y0 D R>0 YR and D L1 .Œ0; T I /. For y 2
YR and ./ 2 , the response is given by the solution of the differential equation
dX
.t/ D b.X.t/; .t// C ˛.X.t/; .t//y.t/; t 2 .0; T ; (4.146)
dt
with initial condition
X.0/ D x .2 Rd /:
and
and
Z t
1
J.t; xI y; .// kf kC t jy.s/j2 ds; (4.151)
2 0
Indeed, (4.149) shows that the sub- and super differential of vR .t; / are in SM1 .
Hence for jpj M1 and R > M1 k˛kC it holds that
1 1
sup p > ˛.x; /y jyj2 D sup p > ˛.x; /y jyj2 ; (4.156)
y2Rd 2 jyjR 2
1. Formulation
Let B and W be mutually independent d - and m-dimensional Wiener processes.
S 0 and S i .i D l; : : : ; m/ denote the price processes of bond and i -th asset,
respectively, and evolve according to the SDE
(
dS 0 .t/ D rS 0 .t/ dt; t 2 .0; T ;
(4.157)
S 0 .0/ D s 0 .> 0/;
and
8 r
ˆ " X i
m
< dS i .t/ D S i .t/ g i .X.t// dt C k .X.t// d W k .t/ ; t 2 .0; T ;
1C"
kD1
:̂
S i .0/ D s i > 0;
(4.158)
where X is the d -dimensional factor process, described by the SDE
p
dX.t/ D b.X.t// dt C " dB.t/: (4.159)
and
@u" " " r
0D C u C b.x/ @x u" u"
@t 2 "
n 1 1 o
C sup .x/ j > .x/j2 u" ; t 2 .0; T ; x 2 Rd ;
2Rm " 2
(4.163)
with a positive constant c, independent of t; x and ". Thus, (4.163) admits a classical
solution with bounded @x u" , where the bound depends on ".
3. Logarithmic transformation and small noise limit
Take the logarithmic transformation, v" D " log. 1" u" /. Then
1
>
rt v" .t; x/ t sup r C .x/ j .x/j2 : (4.164)
;x 2
On the other hand, the Isaacs equation associated with the differential game,
8
ˆ @v
ˆ
ˆ 0 D C b.x/ @x v C r
ˆ
ˆ @t
ˆ
ˆ 1
< 1 >
C inf jyj2 y @x v C sup .x/ j .x/j2 ;
y 2 j jc 2 (4.166)
ˆ
ˆ
ˆ
ˆ t 2 .0; T ; x 2 R d
;
ˆ
ˆ
:̂
v.0; x/ D 0; x 2 Rd
Finally we sketch a model problem with infinite time horizon, given in [FMcE95],
Sect. 7.
For a control process ./ 2 W , its response X 2 X ./ evolves according to
the SDE
r
"1
dX.t/ D b.X.t/; .t// dt C d W .t/; (4.167)
2
with a positive constant . The payoff is given by the long-run expected rate
1 1 Z T
J ..// D " lim inf log Ex exp
"
f .X ./ .s/; .s// ds ; (4.168)
T !1 T " 0
where " 2 .0; 1/ is called the noise intensity and the control region is convex and
compact.
4.4 Risk Sensitive Stochastic Controls and Differential Games 149
" 1
"
D W " .x/ C j@x W " .x/j2
42 42
C min.b.x; / @x W " .x/ C f .x; //; x 2 Rd : (4.169)
2
holds.
3. Using (4.140), Eq. (4.169) is recast of an ergodic Isaacs equation:
"
"
D W " .x/ C max .y @x W " .x/ 2 jyj2 /
42 y2Rd
Let us take the upper value function, v WD v" , in much the same way as in
Sect. 4.1.2. Then
"
@t v C v v
42
C min.b.x; / @x v C f .x; //
2
v.0; x/ D 0; x 2 Rd : (4.176)
7. Let us consider the lower value function u" for the stochastic differential game
given by (4.173)–(4.174). Then u" satisfies (4.175)–(4.176), by changing v to u" .
8. Combining 5, 6, and 7, we see that the ergodic differential game related to the
small noise limit has the value.
For various applications of ergodic type stochastic controls, we can refer to
[Br06, BP99, FSh99, FSh00, FSh02, HNSh10] and [HS10].
Chapter 5
Stochastic Parabolic Equations
where L and M are second-order elliptic and first-order differential operators and
WQ is a colored Wiener process (see Example 5.1).
These equations are generalization of finite-dimensional SDEs and appear in the
study of random phenomena in natural sciences and the unnormalized conditional
probability of finite-dimensional diffusion processes (see Sect. 5.5), related to
filtering equations derived in Fujisaki et al. (Osaka J Math 9:19–40, 1972) and
Kushner (J Differ Equ 3:179–190, 1967).
In Sect. 5.1 we collect basic definitions and results for Hilbert space-valued pro-
cesses; in particular, for continuous martingales, quadratic variations and correlation
operators are treated. Stochastic integrals are introduced in Sect. 5.2. Section 5.3
is devoted to the study of stochastic parabolic equations from the viewpoint of
Hilbert space-valued SDEs, following Rozovskii (Stochastic evolution systems.
Kluwer Academic, Dordrecht/Boston, 1990). By using the results presented, we
also consider a semilinear stochastic parabolic equation with Lipschitz nonlinearity
in Sect. 5.3.4. Section 5.4 deals with Itô’s formula and in Sect. 5.5 Zakai equations
related to filtering problems are given.
5.1 Preliminaries
This section gives preliminaries for SDEs in Hilbert spaces. Here Hilbert space
always means separable real one. We first collect basic definitions and results for
stochastic processes taking values in Hilbert spaces. Proofs can be founded in the
accessible books (cf. [DaPZ92, M88, R90]).
For a Hilbert space H; k k WD k kH and . ; / WD . ; /H denote its norm and
inner product, respectively. H is equipped with a usual Borel field B.H/, generated
by the open sets. H denotes the conjugate space, namely the set of all continuous
linear functional on H. h 2 H can be identified with h 2 H, such that
B.H/ is nothing but the smallest -field which makes h measurable for any
h 2 H .
X
n
lim X X i ei D 0 P -a.e. (5.2)
n!1
i D1
Expectation of X
X
n
EX D EX i ei : (5.3)
i D1
Conditional Expectation
for GQ G F .
X
n
kM.t/k2 D lim M i .t/2 ; (5.14)
n!1
i D1
uniformly in t, P -a.s.
Hence kM./k2 is a continuous .Ft /-submartingale. Now the Doob–Meyer
decomposition asserts that there exists a unique continuous increasing .Ft /-adapted
process hM i./, such that
kdM.t/k2 D d hM i.t/:
Theorem 5.1.
X
n
hM i.t/ D lim hM i i.t/; uniformly in t; P -a.s.; (5.16)
n!1
i D1
Conversely suppose that ./ is continuous, bounded variation and .Ft /-adapted,
and moreover .M./; N.// ./ 2 Mc .Œ0; T ; .Ft /I R1 /.
Then
where M h .t/ D .M.t/; h/ and M g .t/ D .M.t/; g/. Hence, qM .t/ becomes a
symmetric, nonnegative nuclear (trace class) operator on H with tr qM .t/ D 1.
Definition 5.7. qM is called the correlation operator of M (refer to [R90],
Ch. 2.1.13 for details). Generally, in our context a symmetric nonnegative nuclear
operator on H is called a covariance operator.
Example 5.1 (Colored Wiener process). M 2 M2c .Œ0; T ; .Ft /I H/ is called a
colored .Ft /-Wiener process with the covariance operator Q, if the following two
conditions are satisfied.
(a) hM i.t/ D t tr Q,
(b) qM .t/ D .tr Q/1 Q.
M can be expressed in terms of real Wiener processes in the following way.
Since we can take the ONB .eQi ; i D 1; 2; : : : / so that
X
l p
M.t/ D i ˇ i .t/eQi :
i Dl
This section deals with Hilbert space valued stochastic integrals. We are mainly
concerned with stochastic integrals driven by colored Wiener processes. First we
treat a broad class. For an H-valued martingale M and a progressively measurable
integrand ˚ whose values R are Hilbert–Schmidt operators from H into Y, we define
the stochastic integral, ˚.s/ dM.s/, and list some of its basic properties in
Sect. 5.2.1, following [R90]. Section 5.2.2 is devoted to the Burkholder–Davis–
Gundy inequality for H-valued martingales.
First we define stochastic integral in the case H D R1 . Let 2 M2c .Œ0; T ; .Ft /I R1 /
be given. By L2 .hi P; .Ft /I Y/, we denote the set of all .Ft /-progressively
measurable Y-process , satisfying
Z T
E k.s/k2Y d hi.s/ < 1: (5.31)
0
I .t/ D .t/:
5.2 Stochastic Integrals 161
It is clear that
Z t
hI i.t/ D k.s/k2Y d hi.s/: (5.34)
0
Next we suppose that M 2 M2c .Œ0; T ; .Ft /I H/ is given. L2 .HI Y/ denotes the
Hilbert space of Hilbert–Schmidt operators, equipped with the norm
X 12
kkL2 .HIY/ D kei k2Y : (5.35)
i
By L2 .hM i P; .Ft /I L2 .HI Y// we denote the set of L2 .HI Y/-valued processes ˚,
such that .˚./h; y/Y is .Ft /-progressively measurable for any .h; y/ 2 H Y, and
Z T
E k˚.s/k2L2 .HIY/ d hM i.s/ < 1: (5.36)
0
Rt
We define the stochastic integral I˚ .t/ WD 0 ˚.s/ dM.s/ as an element of
M2c .Œ0; T ; .Ft /I Y/, following [R90], Ch. 2.2.2. We still omit the subscript H
R t . /H and kikH . For M .t/ D .M.t/; ei /,2cthe stochastic integral Ji .t/ WD
i
in
0 ˚.s/ei dM .s/ is given as an element of M .Œ0; T ; .Ft /I Y/ by (5.33). Define
˘np ˚ 2 L2 .hM i P; .Ft /I L2 .HI Y// by
(
˚.s/ei ; i D n; n C 1; : : : ; p;
˘np ˚.s/ei D
0; otherwise:
Pp
Since k i Dn Ji .t/k2Y is a real submartingale, we have
h X
p
2i h X
p
2i
E sup Ji .t/ 4E Ji .T /
0t T Y Y
i Dn i Dn
Z T 1
D 4E k˘np ˚.s/qM2 .s/k2L2 .HIY/ d hM i.s/
0
Z T
4E k˘np ˚.s/k2L2 .HIY/ d hM i.s/ ! 0 as n; p ! 1:
0
We will define the stochastic integral of ˚ as a local martingale, in the same way as
in Step 3 of Sect. 1.1.3. Namely, put
8 n Z t o
<
inf t < T I k˚.s/k2L2 .HIY/ d hM i.s/ N ;
N D (5.39)
: 0
T; if f g D empty
and
X Z t
hI˚ i.t/ D i k˚.s/ei k2Y ds
i 0
Z t
D .˚.s/Q; ˚.s//L2 .HIY/ ds; 8t; P -a.s.; (5.43)
0
Z t
hI˚ ; I i.t/ D .˚.s/Q; .s//L2 .HIY/ ds; 8t; P -a.s.: (5.44)
0
P; .Ft /I L .HI R // and define the stochastic integral I .t/ WD 0 ..s/; dM.s//
2 1
by (5.37), namely
Z t XZ t
..s/; dM.s// D ..s/; ei / dM i .s/: (5.45)
0 i 0
and
h ˇZ t ˇ2 i Z T
ˇ ˇ
E sup ˇ ..s/; dM.s//ˇ 4E k.s/k2 d hM i.s/: (5.47)
0t T 0 0
Rt
Example 5.4 (Martingale part of kM.t/k2 ). Put M # .t/ D 0 .M.s/; dM.s//. Then
M # is in Mc .Œ0; T ; .Ft /I R1 / and 2M # .t/ gives the martingale part of kM.t/k2 ,
namely
Indeed, for the exit time N of kM.t/k from Œ0; N ; M i .t/ WD .M.t/; ei / satisfies
Z t^ N
M i .t ^ N/
2
D2 M i .s/ dM i .s/ C hM i i.t ^ N /; 8t; P -a.s. (5.49)
0
kM.t ^ N /k
2
D 2M # .t ^ N/ C hM i.t ^ N /: (5.50)
5
EhM i.T /: (5.51)
2
Thus the monotone convergence theorem for (5.50) implies that EjM #.t/j <
1 and (5.48).
Note that if M # 2 M2c .Œ0; T ; .Ft /I R1 /, then
Z t
hM #i.t/ D .M.s/qM .s/; M.s// d hM i.s/: (5.52)
0
Therefore,
We have already stated a martingale inequality in Theorem 5.2. Now we are going
to prove the Burkholder–Davis–Gundy inequality for H-valued martingales.
Theorem 5.3. Let M 2 M2c .Œ0; T ; .Ft /I H/ and put
Then we have
1
EŒM .t/2 EhM i.t/ EŒM .t/2 ; (5.55)
4
cp EŒM .t/2p EhM i.t/p Cp EŒM .t/2p (5.56)
where
8 2p 2p p
ˆ
ˆ
< cp D p.2p 1/ ; Cp D .4p/p ; if p > 1;
2p 1
p (5.57)
ˆ
:̂ cp D p ; Cp D p 2p ; if p 2 .0; 1/:
16
Proof. Since (5.55) is clear from Theorem 5.2, we will prove (5.56), by using
arguments similar to those in the real martingale cases.
5.2 Stochastic Integrals 165
Let N be the exit time of kM.t/k from Œ0; N . By considering M.t ^ N / instead
of M.t/, we assume that M is bounded. We divide the proof into four steps.
Step 1. Compute cp for p > 1
Let " > 0 be given. Put X.t/ D kM.t/k2 . Applying Itô’s formula to (5.48) and
using (5.52), we obtain
and
X Z t p1
N.t/ D N .t/ei D
i
hM i.s/ 2 dM.s/: (5.61)
i 0
and
Thus
1 p1
EhN i.t/ D EkN.t/k2 4.EM .t/2p / p .EhM i.t/p / p (5.65)
by (5.61), we get
1p
kM.t/k 2N .t/hM i.t/ 2 ; (5.67)
Let us put
Z t
N i .t/ D ." C M .s//p1 dM i .s/: (5.72)
0
and
1
kN.t/k ." C M .t//p : (5.75)
p
nX
d X
d
d u.t; x; !/ D @i aij .t; x; !/@j u.t; x; !/ C b i .t; x; !/u.t; x; !/
i D1 j D1
o
C .t; x; !/u.t; x; !/ C z.t; x; !/ dt
X p nXd
j
C k gk .t; x; !/@j u.t; x; !/
k j D1
o
C „k .t; x; !/u.t; x; !/ C fk .t; x; !/ d W k .t/: (5.78)
We will consider (5.78) from the viewpoint of Hilbert space-valued SDE, following
[P79], [P93], [R90], Chs. 3 and 4. In Sect. 5.3.1, we introduce concepts, notations,
168 5 Stochastic Parabolic Equations
and basic results needed for later sections. We study the Cauchy problem for (5.78)
and properties of solutions in Sects. 5.3.2 and 5.3.3. By changing z.t; x; !/ to
z.t; x; u.t; x/; @x u.t; x/; !/, we will consider a semilinear stochastic parabolic
equation with Wiener noise in Sect. 5.3.4, based on results obtained in Sect. 5.3.3.
5.3.1 Preliminaries
In this subsection, we list some notations and basic results we need later.
Let X and H be Hilbert spaces. By X Õ H, we mean that the embedding; X 7! H
is continuous and dense. We use the identification H D H (D conjugate space of
Q is called normal, if X Õ H D H Õ X
H). The triple .X; H; X/ Q and there is a
constant N such that
hy; xi.X;X/
Q D lim .yn ; x/H : (5.80)
n!1
The RHS of (5.80) does not depend on the choice of .yn /, and
jhy; xi.X;X/
Q j N kykXQ kxkX ; Q
8y 2 X; x 2 X: (5.81)
Qek D k ek ; k D 1; 2; : : : : (5.82)
Put
X
L2Q .HI H/ D f 2 L.HI H/I k k2Q WD k k ek k2H < 1; g (5.83)
k
1
i.e., 2 L2Q .HI H/ means that Q 2 is a Hilbert-Schmidt operator.
Q and G.t; !/ 2 L.XI L2 .HI H//, consider the SDE
For given A.t; !/ 2 L.XI X/ Q
.A2 / Coerciveness
There are constants N > 0 and N 2 R1 , such that
N
CkG.t/k2Q kk Q Ckk
N 8 2 X; 8t;
2
2hA.t/; i.X;X/
Q X H; P -a.s.:
Q
.A3 / Z 2 L2 .ŒT0 ; T ˝; .Ft /I X/.
.A4 / 2 L .ŒT0 ; T ˝; .Ft /I L2Q .HI H//.
2
.u.t/; /H D.u0 ; /H
Z t
C hA.s/u.s/ C Z.s/; i.X;X/
Q ds
T0
Z t
C ..G.s/u.s/ C .s// d WQ .s/; /H ; 8t; P -a.s. (5.86)
T0
Rt
we confirm that M.t/ WD T0 G.s/u.s/ d WQ .s/; t 2 ŒT0 ; T is a continuous local
martingale with
Z t
hM i.t/ D kG.s/u.s/k2Q ds:
T0
Theorem 5.4 (Theorem 3.1.4 in [R90], p. 90). Under conditions .A0 /–.A5 /, the
SDE (5.84)–(5.85) has the unique solution u./, such that
(ii) Estimate
h Z T i
E sup ku.t/k2H C ku.t/k2X dt
T0 t T T0
h Z T i
NE ku0 k2H C .kZ.s/k2XQ C k .s/k2Q / ds ; (5.88)
T0
N and .
where N is a constant depending only on K; , N
From now on, we are mainly concerned with L2 .Rd / as H and introduce the
following notations:
H0 WD L2 .Rd / with the usual inner product . ; / and the norm k k.
Hp WD f 2 H0 ; the generalized derivatives of order p belong to H0 g,
endowed with the inner product
p p
.; /Hp D ..1 / 2 ; .1 /2 / (5.89)
p
and the norm kkHp D k.1 / 2 k, where 1 is the identity map and is the
d -dimensional Laplacian operator.
5.3 Stochastic Parabolic Equations with Colored Wiener Noises 171
For example, the triple (H2 ; H1 ; H0 ) is normal and is used later in Proposi-
tion 5.4.
1
H1 WD fI is a Borel function such that .1 / 2 2 H0 g;
endowed with the inner product and the norm such that
1 1
.; /H1 D ..1 / 2 ; .1 / 2 / (5.90)
and
1
kkH1 D k.1 / 2 k
D sup.j.; /jI 2 H1 with k kH1 D 1/: (5.91)
Let us come back to (5.78). Here we will study equation (5.78) in the Hilbert space
framework. But we firstly introduce the notion of generalized solution from the
viewpoint of parabolic equations (cf. [R90], p. 130).
Let
.u.t/; / D .u0 ; /
d Z t X
X d
aij .s; /@j u.s/ C b i .s; /u.s/; @i ds
i D1 T0 j D1
172 5 Stochastic Parabolic Equations
Z t
C ...s; /u.s/; / C hz.s; /; i/ ds
T0
Xp Z t X
d
j
C k gk .s; /@j u.s/C„k .s; /u.s/Cfk .s; /; d W k .s/;
k T0 jD1
.B2 / Superparabolicity
X
j
^.t; x/ WD 2aij .t; x/ k gki .t; x/gk .t; x/
ij D1;:::;d
k
X
d X
d
A.t; !/ D @i aij .t; x; !/@j C b i .t; x; !/ C .t; x; !/; (5.98)
i D1 i D1
X
d
Gk .t; !/ D gki .t; x; !/@j C „k .t; x; !/; (5.99)
j D1
and
for 2 H1 .
Thus, (5.78) can be recast as the following SDE on .H1 ; H0 ; H1 /:
Since CK1 .Rd / is dense in H1 and since (5.96) and (B2 ) imply (A1 ) and (A2 ),
Theorem 5.4 leads to the following
Theorem 5.5. Under conditions (B1 )–(B3 ), Eq. (5.78) has a unique generalized
solution u./ 2 L2 .ŒT0 ; T ˝; .Ft /I H1 /\L2 .˝I C.ŒT0 ; T I H0 // with u.T0 / D u0 .
Moreover, (i) and (ii) hold:
(i) Energy equality
For any (Ft )-stopping time taking values in ŒT0 ; T ,
8t 2 ŒT0 ; T ; P -a.s.:
(ii) Estimate
h Z T i
E sup ku.s/k2 C jjju.s/jjj2 ds
T0 S T T0
h Z T i
NE ku0 k2 C .kz.s/k2 C k .s/k2Q / ds ; (5.103)
T0
Further (B2 ) and (B4 ) yield the coerciveness on (H2 ; H1 ; H0 ), namely there is a
constant 0 .2 R1 / depending only on K1 ; 0 , and , such that for any 2 H2 ,
ˇˇˇX ˇˇˇ2 X X j
ˇˇˇ ˇˇˇ
ˇˇˇ gki @i ˇˇˇ D gki @i ; .1 / gk @j :
i i j
For any " > 0, there is a constant K" , independent of , such that
We apply the same calculations to other terms in the LHS of (5.106). Then letting
" small, depending only on 0 and K1 , and putting all estimates together, we
conclude (5.106) for 2 H3 .
For 2 H2 , we can take n 2 H3 , so that limn!1 I n I D 0. Hence, (B4 )
and (5.105) yield
v.T0 / D u0 : (5.109)
Then Theorem 5.4 yields the unique solution of problem (5.108)–(5.109) in the
space L2 .ŒT0 ; T ˝; .Ft /I H2 / \ L2 .˝I C.ŒT0 ; T I H1 //. Hence, for any 2 H2 ,
..v.t/; // D..u0 ; //
Z t
C hA.s/v.s/ C z.s/; i.H0 H2 / ds
T0
Xp Z t
C k ..Gk .t/v.t/ C fk .t/; // d W k .t/; t 2 .T0 ; T :
k T0
(5.110)
176 5 Stochastic Parabolic Equations
Consequently, v./ coincides with the unique generalized solution of (5.78) with
u.T0 / D u0 .
This completes the proof.
For the dependence on the time parameter, we have;
Proposition 5.5. Suppose z D 0 and fk D 0; k D 1; 2; : : : . Let u./ be the unique
generalized solution of (5.78) with u.T0 / D u0 .
(i) Let u0 2 L2 .˝; FT0 I H1 / and ; t 2 ŒT0 ; T . Under .B1 /, .B2 / and .B4 /,
h i
E sup ku.s/ u./k2 N1 jt jEjjju0 jjj2 ; (5.111)
st
Then
h i
E sup jjju.s/ u./jjj2 N2 .t /EIu0 I2 ; (5.112)
st
Further, z./ D A./u./ and ./ D G./u./ are in L2 .Œ; T ˝; .Ft /I H1 / and
L2 .Œ; T ˝; .Ft /I L2Q .H0 I H0 //, respectively. (B1 ) yields
where A.t/ and Gk .t/ are given by (5.98) and (5.99), respectively.
The definition of generalized solution for (5.116) is given in the same way
as (5.94), by replacing hz.s; /; i with hz.s; ; u.s/; @x u.s//; i.
Let us introduce the SDE on .H1 ; H0 ; H1 / associated with (5.116), whose
solution provides the generalized solution of (5.116) (see (5.118)).
Using (b) and (c), we can define Z W Œ0; T H1 ˝ 7! H0 by
Hence, for 2 L2 .Œ0; T ˝; .Ft /I H1 /, .t/ WD Z.t; .t// is defined as an element
of L2 .Œ0; T ˝; .Ft /I H0 / by (b) and (c), and
Z T Z T
E kZ.t; 1 .t// Z.t; 2 .t//k dt l0 E 2
jjj1 .t/ 2 .t/jjj2 dt
0 0
that is
nX
d X
d
d u.t/ D @i aij .t; x/@j u.t/ C b i .t; x/u.t/
i D1 j D1
o
C .t; x/u.t/ C z.t; x; u.t; x/; @x u.t; x// dt
X p nXd
j
o
C k gk .t; x/@j u.t/ C „k .t; x/u.t/ C fk .t; x/ d W k .t/;
k i D1
X
n1
.C0 .t T0 //j 1
n .t/ NEjjju0 jjj2
j D1
.j 1/Š
Z X
n1
t
.C0 .t s//j 1
CN EŒ2kz0 .s/k2 C jjj .s/jjj2Q ds
T0 j D1
.j 1/Š
.C0 .t T0 //n
C Ejjju0 jjj2 ; (5.126)
nŠ
where C0 D 2N l0 .
Step 2. Convergence of un ./
Put !1 D u1 and !nC1 D unC1 un ; n D 1; 2; : : :. Then !nC1 satisfies
8
ˆ
ˆ d!nC1 .t/ D .A.t/!nC1 .t/ C Z.t; un .t// Z.t; un1 .t/// dt
<
C G.t/!nC1 .t/ d WQ .t/; t 2 .T0 ; T ; (5.127)
ˆ
:̂
!nC1 .T0 / D 0:
we have
.C0 .t T0 //nC1
nC1 .t/ sup 1 .s/: (5.129)
.n C 1/Š T0 st
Therefore,
hX Z T X i
E sup jjj!n .s/jjj C I!n .s/I ds < 1: (5.130)
n T0 sT T0 n
P Pn
Since supT0 sT n jjj! P n .s/jjj < 1, P -a.s., we deduce that un ./ D kD1 !k ./
converges to u./ WD k !k ./ in C.ŒT0 ; T I H1 /, P -a.s.
On the other hand, un ./ converges to some uQ ./ in L1 .ŒT0 ; T I H2 /, P -a.s.,
by (5.130). Therefore, un ./ converges to uQ ./ in L1 .ŒT0 ; T I H1 /, P -a.s., and so
u.t/ D uQ .t/, a.e. t 2 ŒT0 ; T , P -a.s. By using (5.126) and the Fatou’s lemma, we
conclude that
p Z T
l0 I IE jjjun .s/ u.s/jjj ds: (5.133)
T0
!.T0 / D 0; (5.136)
.t/ D 0; t 2 ŒT0 ; T ;
nX
d X
d
d u.t; x/ D @i aij .x/@j u.t; x/
i D1 j D1
X
d o
C sup b i .x; /@i u.t; x/ C .x; /u.t; x/ C z.x; / dt
2 i D1
m X
X d
C gki .x/@i u.t; x/ C „k .x/u.t; x/ C fk .x/ d W k .t/;
kD1 i D1
t 2 .0; T ; x 2 Rd ; (5.139)
where 2 H1 .
We assume that
(a) is convex and compact,
(b) All coefficients are time-homogeneous and bounded continuous non-random
functions,
(c) aij 2 Cb2 .Rd /; gki ; „k 2 Cb1 .Rd /; ij D 1; : : : ; d; k D 1; : : : ; m,
5.4 Itô’s Formula 183
(d) Superellipticity
X
m
j
2aij .x/ gki .x/gk .x/
ij D1;:::;d
kD1
X
d
z.x; u; v/ D sup b i .x; /vi C .x; /u C z.x; / :
2 i D1
where 2 L2 .Œ0; T ˝; .Ft /I H/, X0 2 L2 .˝; F0 I H/, M 2 M2c .Œ0; T ; .Ft /I H/.
In Sect. 5.4.2, we will show that, for F 2 C 12 .Œ0; T H/ satisfying local
boundedness condition (F0 ), Itô’s formula holds:
F .t ^ ; X.t ^ // D F .0; X0 /
Z t^ n o
C @s F .s; X.s// C ..s/; DF .s; X.s///H ds
0
Z t^
C .DF .s; X.s//; dM.s//H
0
Z t^
1
C tr.D 2 F .s; X.s//qM .s// d hM i.s/; 8t; P -a.s.
2 0
(5.142)
184 5 Stochastic Parabolic Equations
for any (Ft )-stopping time taking values in Œ0; T . In the differential form,
5.4.1 Preliminaries
This subsection lists basic definitions and results that will be needed in the sequel.
We state linear- and bilinear maps, Hilbert–Schmidt operators, nuclear operators
and first and second derivatives.
Proofs are given in accessible books, cf. [La83, Y80].
1. Linear- and bilinear maps
Let X; Y, and Z be Hilbert spaces. L.YI Z/ denotes the Banach space of
continuous linear mappings from Y into Z, endowed with the norm k kL.YIZ/ given
by
X
N
˚D
j
Qj .
j D adjoint of i /: (5.145)
j D1
X X
N
tr.˚/ WD .˚ei ; ei / D . Qj ; j /L2 .H/ : (5.146)
i j D1
kF .x C h/ F .x/ DF .x/hkZ
lim D 0: (5.149)
khkH !0 khkH
Formally,
Q D D 2 F .x/.h;
D 2 F .x/.h; h/ Q h/; 8h; hQ 2 H
Taylor Formula
We denote by C 12 .Œ0; T H/ (or Cu12 .Œ0; T H/) the set of F 2 C.Œ0; T H/ such
that @t F , DF , and D 2 F are continuous (or uniformly continuous) on Œ0; T H.
For F 2 C 12 .Œ0; T H/, we assume the local boundedness condition;
.F0 / jF ./j, j@t F ./j, kDF ./kH and kD 2 F ./kL.HIH/ are bounded on any bounded
subset of Œ0; T H.
5.4 Itô’s Formula 187
Theorem 5.6. Assume .F0 /. For X given by (5.141), Itô’s formula (5.142) holds.
Proof. Put
Q
and, for ! 2 ˝,
jF .t; /j C j@t F .t; /j C kDF .t; /kH C kD 2 F .t; /kL.H;H/ K;
8.t; / 2 Œ0; T H (5.154)
and kX0 kH K, P -a.s. During the proof, we omit the suffix H in k kH and . ; /H ,
for simplicity.
From Taylor’s formula (5.151) it follows that
Put
From the continuity of @s F and the bounded convergence theorem, it is clear that
Z T
lim j@s F .s; X.NN .s/// @s F .s; X.s//j ds D 0 (5.157)
N !1 0
188 5 Stochastic Parabolic Equations
we have
Z T
lim E j.DF .N .s/; X.N .s///DF .s; X.s//; .s//j ds D 0: (5.159)
N !1 0
Step 3. Put
Z t
4M .; t/ D M.t/ M./ and I.; t/ D .s/ ds:
and
ˇ X ˇ Z t
ˇ ˇ N
Eˇ J2 .tj ; tj C1 /ˇ K2 E k.s/k2 ds: (5.164)
tj C1 t 0
5.4 Itô’s Formula 189
1 1
K.Eh˘n? M i.t// 2 .EhM i.t// 2 ! 0;
as n ! 1; uniformly w.r.t. the divisions DN : (5.165)
X
n X
D D 2 F .tj ; X.tj //.ek ; ep /.4M .tj ; tj C1 /; ek /.4M .tj ; tj C1 /; ep /
k;pD1 tj C1 t
n Z
X t
! D 2 F .s; X.s//.ek ; ep /q k;p .s/ d hM i.s/; P -a.s. as N ! 1:
k;pD1 0
(5.166)
X Z t
lim J4 .tj ; tj C1 / D tr.D 2 F .s; X.s//qM .s// d hM i.s/ P -a.s.
N !1 0
tj C1 t
(5.167)
Putting (5.157), (5.159)–(5.161), (5.163), (5.164), and (5.167) together, we
complete the proof.
Example 5.7 (Energy equality). For F ./ D kk2H , Itô’s formula gives the energy
equality;
with
Xp
WQ .s/ D k ˇ k .s/ek
k
190 5 Stochastic Parabolic Equations
1X o
C k D 2 F .t; X.t//. k .t/; k .t// dt
2
k
Xp
C k .DF .t; X.t//; k .t// dˇ
k
.t/: (5.169)
k
In this subsection, we always assume the conditions (B1 ), .B2 /, .B4 /, and (B5 ) are
satisfied. Since the solution u./ of (5.102) with the initial state u0 2 L2 .˝; FT0 I H1 /
belongs to L2 .ŒT0 ; T ˝; .Ft /I H2 /\L2 .˝I C.ŒT0 ; T I H1 //, Itô’s formula is valid
for F satisfying condition .F0 /, by using
Z t
.s/ WD A.s/u.s/ C z.s/ and M.t/ WD .G.s/u.s/ C .s// d WQ .s/:
T0
Remark 5.2. The following facts are obvious: under .F0 / and .F1 /,
(i) There is a constant KDF , such that
(
jjjDF .t; /jjj KDF .1 C jjjjjj/; 8.t; / 2 Œ0; T H1 ;
(5.171)
kDF .t; /k KDF .1 C kk/; 8.t; / 2 Œ0; T H0 :
then
Z T
lim E jjjDF .t; n .t// DF .t; .t//jjj2 dt D 0:
n!1 0
Since D 2 F .t; / D B0 D 2 FQ .t; B0 /B0 , we conclude that (F0 ) and .F1 / hold.
Theorem 5.7. Assume that conditions .B1 /; .B2 /; .B4 / and .B5 / are satisfied.
Suppose that F 2 C 12 .ŒT0 ; T H0 / satisfies .F0 / and .F1 /. Then Itô’s formula
holds for the solution u./ of problem (5.102)–(5.93).
dF .t; u.t//
n
D @t F .t; u.t// C hA.t/u.t/ C z.t/; DF .t; u.t//i
1 1 1
o
C tr ..G.t/u.t/ C .t//Q 2 / D 2 F .t; u.t//.G.t/u.t/ C .t//Q 2 / dt
2
C ..G.t/u.t/ C .t// d WQ .t/; DF .t; u.t/// (5.172)
Proof. By taking the exit time of u./ from a ball of H0 , we may assume that the
boundedness condition (5.154) holds for H D H0 . We divide the proof into two
cases.
Case 1. ku0 k C0 (D constant), P -a.s.
192 5 Stochastic Parabolic Equations
By un ./, we denote the solution of (5.102) with the initial state un0 . Then un 2
L2 .ŒT0 ; T ˝; .Ft /I H2 / \ L2 .˝I C.ŒT0 ; T I H1 // and
h Z T i
E sup kun .t/ u.t/k C 2
jjjun .t/ u.t/jjj2 dt
t T0
NEkun0 u0 k ; 2
n D 1; 2; : : : (5.174)
and
and
Z T
lim j@s F .s; un .s// @s F .s; u.s//j ds D 0; P -a.s. and in L1 .˝/:
n!1 T
0
(5.178)
Step 2. For DF terms, we have
Z T
E jhA.s/.un .s/ u.s//; DF .s; un .s//ij ds
T0
Z T
KKDF E jjjun .s/ u.s/jjj.1 C jjjun .s/jjj/ ds
T0
Z T 12 Z T 12
2
KKDF E jjjun .s/ u.s/jjj ds E .1 C jjjun .s/jjj/2 ds
T0 T0
! 0 as n ! 1; (5.179)
5.4 Itô’s Formula 193
thanks to (5.174) and (5.103). Similarly, for any " > 0, we have
Z T
E jhA.s/u.s/; DF .s; un .s// DF .s; u.s//ij ds
T0
Z T
KE jjju.s/jjjmDF .jjjun .s/ u.s/jjj/ ds
T0
Z T
KE jjju.s/jjj." C C" jjjun .s/ u.s/jjj/ ds; (5.180)
T0
Step 3. For the stochastic integral terms, (5.154) and the dominated convergence
theorem yield
hZ T
E .Gk .s/.un .s/ u.s//; DF .s; un .s///2 ds
T0
Z T i
C .Gk .s/u.s/; DF .s; un .s// DF .s; u.s///2 ds ! 0 as n ! 1;
T0
(5.182)
X hZ T i
k E jIk n .s/j C jJk n .s/j ds ! 0 as n ! 1; (5.185)
k T0
(ii) Let u./ and uQ ./ be the solutions of (5.100) with initial states u0 and uQ 0
respectively. Then
h Z T i
E sup ku.s/Qu.s/k2 C ku.s/Qu.s/k2 ds NEku0 Qu0 k2 (5.189)
T0 sT T0
N N and K1 of (5.104).
where the constant N depends only on ;
1
Proof. Put A0 D 1 and B0 D .1 / 2 for simplicity. Referring to [GŚ00],
Lemma 3.3, we have for 2 H3
with a constant c3 .
196 5 Stochastic Parabolic Equations
Put
Z t
M.t/ D ..G.s/A0 v.s/ C .s// d WQ .s/; v.s//:
T0
This section is devoted to certain stochastic parabolic equations, called the Zakai
equations, related to filtering problems. Let X be a diffusion described by an
SDE. X cannot be measured directly. Only a partial measurement of X can be
obtained by means of another process Y affected by an observation noise. Hence,
we need to estimate X.t/, by using the information of Y up to time t. This is the
filtering problem (cf. [BC09]). Hence the problem is how to compute the conditional
probability of X.t/, given FtY , by using results obtained in Sects. 5.3 and 5.4.
In Sect. 5.5.1 we are concerned with partially observable controlled diffusions
and in Sect. 5.5.2 we study an explicit solution of the Zakai equation for conditional
Gaussion process.
dX.t/ D ˛.t; X.t/; .t// dB.t/ C .t; X.t/; .t// d W .t/ C b.t; X.t/; .t// dt;
198 5 Stochastic Parabolic Equations
where B and W are mutually independent Wiener processes and ./ is a control
process. As mentioned above we suppose that X cannot be measured directly. Only
a partial measurement of X can be obtained by means of another process Y , given
by
Z t
Y .t/ D h.s; X.s/; .s// ds C W .t/:
0
˛ W Œ0; T Rd 7! Rd ˝ Rd0 ;
W Œ0; T Rd 7! Rd ˝ Rm ;
b W Œ0; T Rd 7! Rd ;
and
h W Œ0; T Rd 7! Rm
jz.t; x; / z.t; x;
Q /j ljx xj;
Q 8t; (5.201)
and
jb.t; x; / b.t; x;
Q /j lR jx xj
Q for jxj; jxj
Q R; 8t; I
System Equation
8
ˆ
ˆ dX.t/ D ˛.t; X.t/; .t; Y // dB.t/ C .t; X.t/; .t; Y // d W .t/
<
C b.t; X.t/; .t; Y // dt; t 2 .0; T ; (5.203)
ˆ
:̂
X.0/ D X0 .2 L2 .˝; F0 I Rd //;
and
Observation Equation
(
d Y .t/ D h.t; X.t/; .t; Y // dt C d W .t/; t 2 .0; T ;
(5.204)
Y .0/ D 0:
Now we seek a solution .X; Y / and then compute P .X.t/ 2 jFtY /. So the next
step is to introduce the dynamics of the conditional probability density process,
according to [R90], V.
Let ˇ and ˇQ be d0 - and m-dimensional Wiener processes, defined on .˝; F ; P /,
and 0 be a random variable with the same probability distribution as X0 . We assume
Q and 0 are mutually independent. Set
that ˇ; ˇ,
;ˇ; Q̌
Ft D Ft 0 D right continuous and complete -field
Q
generated by f0 ; ˇ.s/; ˇ.s/; s tg:
where
and
(
Q
d.t/ D d ˇ.t/; t 2 .0; T ;
(5.206)
.0/ D 0:
By conditions .d1 /–.d3 /, the auxiliary SDEs (5.205)–(5.206) admit a unique strong
solution .; /.
Applying Girsanov’s transformation, we obtain
Proposition 5.8. The SDE (5.203)–(5.204) has a weak solution. Moreover, the
probability distribution of a weak solution is unique.
Proof. Put
Z t Z
1 t
M.t/ D exp Q
h.s; .s/; .s; // d ˇ jh.s; .s/; .s; //j2 ds :
0 2 0
(5.207)
Then by Proposition 1.5 M is an exponentialRmartingale and PQ WD M.T / ı P gives
a probability on FT . It is easy that PQ . / D M.t/ dP for 2 Ft . Put
Z t
Q
WQ .t/ D ˇ.t/ Q ds;
h.s; .s/; .s; ˇ// t 2 Œ0; T : (5.208)
0
Then Girsanov’s theorem shows that ˇ and WQ are mutually independent .Ft /-
Wiener processes under PQ . Since .; / satisfies (5.203)–(5.204) with ˇ and WQ
on .˝; F ; .Ft /; PQ /; .; / gives a weak solution of (5.203)–(5.204). Moreover, the
distribution of the weak solution is unique, because (5.205)–(5.206) has a unique
strong solution.
Q ..t//jFt /, where EQ D the expectation w.r.t. PQ . Since
Next we study E.F
E.M.t/jFt / 2 .0; 1/ P -a.s., we can use Bayes formula (Proposition 1.3)
Q ..t//jFt / D E.M.t/F ..t//jF
E.F
t /
; P -a.s.; (5.209)
E.M.t/jFt /
Q x/ 0;
q.t; 8.t; x/; P -a.s. (5.210)
5.5 Zakai Equations 201
Proposition 5.9.
(i) Besides .d1 /–.d3 /, we assume that ˛ and are in C 1 .Rd / w.r.t. x. Suppose
that 0 has the probability density p0 2 H0 . If the density-valued process qQ is in
L2 .Œ0; T I H1 / \ C.Œ0; T I H0 /, P -a.s., then qQ is a generalized solution of the
following stochastic linear parabolic equation (5.212)–(5.213).
(ii) If ˛˛ > is uniformly positive definite, then the generalized solution is unique.
X
m
Q D L.t;/ .t/ q.t/ .t/ q.t/
.t;/
d q.t/ Q dt C lk Q dk .t/; t 2 .0; T ;
kD1
(5.212)
with the initial condition
Q
q.0/ D p0 ; (5.213)
where
1
L .t/.x/ D tr..˛˛ > C > /.t; x; /@xx .x// C b.t; x; / @x .x/;
2
(5.214)
X
d
lk .t/.x/ D ki .t; x; /@i .x/ C hk .t; x; /.x/: (5.215)
i D1
Z t m Z t
X
.s;/
D L.s;/
Q
.s/F; q.s/ ds C Q
lk .s/F; q.s/ dk .s/:
0 kD1 0
(5.216)
and X.0/ has probability density p0 2 H0 . The price of the k-th asset is given by
(
dS k .t/ D S k .t/.hk .X.t// dt C d W k .t//; t > 0;
(5.217)
S k .0/ D s0k > 0; k D 1; : : : ; m;
.S 1 ;:::;S m / .Y 1 ;:::;Y m /
and Ft D Ft . Hence Proposition 5.9 yields the following stochastic
linear parabolic equation for its density-valued process q.t/:
8
ˆ X d X d X d
ˆ
ˆ D ij
i
ij
ˆ
ˆ dq.t/ @i a .x/@ j q.t/ b .x/ @j a .x/ q.t/ dt
ˆ
< i D1 j D1 j D1
Xm
ˆ 1
ˆ
ˆ C h .x/
k k
q.t/ d W .t/; t > 0;
ˆ
ˆ 2
:̂ kD1
q.0/ D p0 :
X
m
.t /
d u.t; x/ D A.t / .t/u.t; x/ dt C Gk .t/u.t; x/ d W k .t/; (5.218)
kD1
where
X
d X
d
A .t/.x/ D @i aij .t; x; /@j .x/ C b i .t; x; /.x/ C .t; x; /.x/;
i D1 j D1
X
d
j
Gk .t/.x/ D gk .t; x; /@j .x/ C „k .t; x; /.x/:
j D1
Hence (5.212) is the Zakai equation, because .t/ is a Wiener process on .˝; F ; P /.
Let us return to (5.209). For the normalization factor, N.t/ WD E.M.t/jFt /,
observing that
X
m
dN.t/ D E M.t/hk .t; .t/; .t; //jFt dk .t/;
kD1
where
Q D E.h.t;
h.t/ Q
.t/; .t; //jFt /:
Q
because (5.209) yields dN.t/ D N.t/h.t/d.t/.
8
ˆ
ˆ dX.t/ D .r0 .t; Y / C r1 .t; Y /X.t// dt
<
C ˛.t; Y / dB C .t; Y / d W; t 2 .0; T ; (5.220)
ˆ
:̂
X.0/ D X0 .N.0; v0 /-distributed random variable/:
(
d Y .t/ D h.t; Y /X.t/ dt C d W .t/; t 2 .0; T ;
(5.221)
Y .0/ D 0:
and
(
Q
d.t/ D d ˇ.t/; t 2 .0; T ;
(5.225)
.0/ D 0:
Since the system (5.224)–(5.225) has the unique strong solution .; / and .0 ; ˇ; /
are mutually independent, (5.224) shows that the conditional probability of
..ti /; i D 1; : : : ; n/ given Ft , is Gaussian for any t and 0 t1 < t2 < < tn t.
Note. This property is called the conditional Gaussian property.
Now we define a new probability PQ by
PQ D M.T / ı P on FT ; (5.226)
Q
2. Conditional probability P..t/jF
t /
According to [LS01], Th. 12.6, the conditional Gaussian property is also valid
under PQ .
We see that PQ ..t/ 2 jFt / is Gaussian distribution with mean m.t/ D
Q
E..t/jF
Q >
t / and variance V .t/ D E...t/ m.t//..t/ m.t// jFt /. Since .; /
satisfies (5.220)–(5.221) with ˇ and WQ on .˝; F ; .Ft /; PQ /, filtering theory asserts
that
n Z t o
WN .t/ WD 1 .t/ Q
h.s; /E..s/jF
s / ds (5.229)
0
is an .Ft /-Wiener process and .V .t/; m.t// formally satisfies the following equa-
tions:
Riccati Equation
8
ˆ
ˆ dV
ˆ
ˆ .t/ D r1 .t; /V .t/ C V .t/r1 .t; />
ˆ
ˆ dt
ˆ
ˆ
ˆ
ˆ C .˛.t; /˛.t; /> C .t; /.t; /> /
ˆ
ˆ
<
ˆ
ˆ 2 ..t; / C V .t/h.t; /> /..t; / C V .t/h.t; /> /> ;
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ t 2 .0; T ;
ˆ
ˆ
ˆ
:̂
V .0/ D v0 :
(5.230)
Filtering Equation
8
ˆ
ˆ d m.t/ D .r0 .t; / C r1 .t; /m.t// dt
<
C ..t; / C V .t/h.t; /> / d WN .t/; t 2 .0; T ; (5.231)
ˆ
:̂
m.0/ D 0:
d
When (5.230) admits a unique global solution V ./ in C.Œ0; T I SCC / for any
path of , (5.231) yields that for F 2 Cb .R /,
d
Z
Q ..t//jFt / D
E.F F .x/g.t; x/ ds; (5.232)
Rd
5.5 Zakai Equations 207
where g is an (Ft )-progressively measurable process given by the Gaussian density
1
1
g.t; x/ D ..2 /d det V .t// 2 exp .x m.t//V 1 .t/.x m.t//> : (5.233)
2
3. Density-valued process q./
Now Proposition 5.9 shows that q./ is the unique generalized solution of (5.222).
Refer to [Be92], 4 and 6 for the solution of Zakai equations.
Chapter 6
Optimal Controls for Zakai Equations
Ft D FtW and D W
:
6.1.1 Preliminaries
nX
d X
d
d u.t; x/ D @i aij .x; .t//@j u.t; x/ C b i .x; .t//u.t; x/
i D1 j D1
o
C .x; .t//u.t; x/ dt
m X
X d
j
C gk .x; .t//@j u.t; x/ C „k .x; .t//u.t; x/ d W k .t/:
kD1 j D1
(6.1)
j j
z WD aij ; bi ; @i aij ; @i b i ; ; gk ; @j gk ; and „k ;
.i; j D 1; : : : ; d; k D 1; : : : ; m/
In this subsection, we will show that the solution depends continuously on the initial
state, the time parameter and the control process. Here we deal with Eq. (6.1) in the
Hilbert space framework, by putting
X
d X
d
A./ .t/.x/ D @i aij .x; .t//@j .x/Cb i .x; .t//.x/ C.x; .t//.x/;
i D1 jD1
./
X
d
j
Gk .t/.x/ D gk .x; .t//@j .x/ C „k .x; .t//.x/;
j D1
and
G ./ .t/ D .G1 ./.t/; : : : ; Gm
./
.t//:
(B10 ) and (B20 ) ensure that the following coercivity condition is satisfied.
212 6 Optimal Controls for Zakai Equations
There are two constants, N > 0 and N 2 R1 , such that, for any ./,
N
2hA./.t/; iC3kG ./ .t/k2m jjjjjj2
C kk
N 2
; 8t; P -a.s.; 8 2 H1
(6.6)
and
N ,
with a constant N2 depending only on ; N and K.
Proof. For simplicity of notation, we omit ./ and .T0 ; u0 /. ci ; i D 1; 2; : : : will
denote constants independent of t; ./ and u0 .
From the energy equality
it follows that
6.1 Controlled Zakai Equations 213
by (6.6). Let be the exit time of ku.t/k from [0; n]. Then (6.12) yields
n
Z t^ n
Eku.t ^ n /k4 Eku0 k4 2E N
.ku.s/k 2
jjju.s/jjj2 C ku.s/k
N 4
/ ds:
T0
(6.13)
Therefore,
N T0 /
Eku.t ^ n /k
4
e 2.t Eku0 k4 (6.14)
and
Z n
E N 1 e 2T
ku.s/k2 jjju.s/jjj2 ds .2/ N
Eku0 k4 : (6.15)
T0
1 hZ t^ n i
J.t ^ n/ C c2 E ku.s/k2 jjju.s/jjj2 ds ; (6.16)
8 T0
Finally, we prove (6.11) in the same way. Suppose that Ejjju0 jjj4 < 1. Then u./
has continuous H1 -paths. Let n be the exit time of jjju.t/jjj from Œ0; n, and put
2
.t/ D jjju.t ^ n /jjj : (6.18)
The energy equality (cf. Theorem 5.4) gives the dynamics of .t/ and we obtain
X
m
C 4.t/ ..Gk .t/u.t/; u.t/// d W k .t/
kD1
and
Z n
E N 1 e 2T
.s/Iu.s/I2 ds .2/ N
Ejjju0 jjj4 : (6.21)
T0
We estimate the stochastic integral term in the same way as (6.16), namely
h ˇZ X
m ˇi
ˇ ˇ
E sup ˇ .s/ ..Gk .s/u.s/; u.s/// d W k .s/ˇ c5 Ejjju0 jjj4 : (6.22)
T0 t ^ n T0 kD1
Proposition 6.3.
N N and K, such that, for any ; t
(i) There is a constant N3 , depending only on ;
and ./,
h i
ET0 u0 sup ku./ .s/ u./ ./k2 N3 jt jEjjju0 jjj2 ; (6.23)
st
h i
ET0 u0 sup ku./ .s/ u./ ./k2 N3 jt jEku0 k2 : (6.24)
st
6.1 Controlled Zakai Equations 215
(ii) Let u0 D .2 H0 /. For " > 0, there is ."; / > 0, such that for any t1 ; t2
and ./,
h i
ET0 sup ku./ .t/ u./ ./k2 < " (6.25)
;t 2Œt1 ;t2
sup ku.t/ u./k2 2 sup ku.t/ u.t1 /k2 C 2 sup ku./ u.t1 /k2 ;
;t 2Œt1 ;t2 t 2Œt1 ;t2 2Œt1 ;t2
Q 2< "
k k : (6.29)
12N1
Since for the response uQ .t/ with initial state Q (6.8) yields
h i "
E sup ku.t/ uQ .t/k2 < ; (6.30)
T0 t T 12
we obtain
h i
E sup ku.t/ u./k2
;t 2Œt1 ;t2
h i h i
6E sup ku.t/ uQ .t/k2 C 3E sup kQu.t/ uQ ./k2
T0 t T ;t 2Œt1 ;t2
"
<" if t2 t1 < N ; Q .DW ."; //:
6
This gives (ii).
Since (iii) is clear by (ii), we have completed the proof.
216 6 Optimal Controls for Zakai Equations
(ii) For 2 H0 and " > 0, there is ı."; / > 0, such that
h Z T i
Q ./
ET0 Q sup ku ./
.t/u .t/k C
2
jjju./ .t/uQ ./ .t/jjj2 dt < " (6.32)
T0 t T T0
Q 2<
for any Q 2 H0 with k k "
48N1 whenever EŒ1./ Q ./14 < ı."; /.
Q
Proof. For simplicity, we put u D u./ ; uQ D u./ and v D u uQ . Then we have
d v.t/ D A./ v.t/ dt C G ./ .t/v.t/ d W .t/ C Z.t/ dt C .t/ d W .t/; t 2 .T0 ; T
(6.33)
with
v.T0 / D 0;
where
and
and
Z T h i
E j.t/ Q .t/j2 jjjQu.t/jjj2 dt E sup jjjQu.t/jjj2 1./ Q ./12
T0 T0 t T
h i 12 1
E sup jjjQu.t/jjj4 .E1./ O ./14 / 2
T0 t T
Hence,
" "
kQ # k2 < Q 2<
for k k : (6.37)
12N1 48N1
./ Q
./
uT Q .t/ uT Q .t/
0 0
Thus, putting ı."; / WD "2 .2N4 jjj # jjj2 /2 we complete the proof.
where the integral part and F .u./ .t// denote a running cost and a terminal cost,
respectively. We want to minimize (or maximize) the expectation of the cost, by
taking a suitable ./. This optimization is a generalization of optimal control
218 6 Optimal Controls for Zakai Equations
problem for partially observable diffusions, see Example 6.1. For the terminal cost,
we introduce a Banach space C # (see (6.40)). The basic properties of the terminal
costs and running costs are studied in Sects. 6.2.2 and 6.2.3, respectively. Putting
these results together, we state properties of the value function in Sect. 6.2.4.
6.2.1 Preliminaries
F ./
F # ./ D for 2 H0 ; (6.39)
1 C kk
and define C # by
We write
jF ./j p
1 _ N0 kF k# ; (6.44)
1 C kk
with the constant N0 of (6.4), and, for " > 0, there is ı1 ."I F / > 0 independent of
t; T0 and ./, such that
6.2 Formulations of Control Problems 219
ˇ F ./ Q ˇˇ
F ./
ˇ
ˇ ˇ<" (6.45)
1 C kk 1 C kk Q
Q < ı1 ."; F /.
whenever k k
Proof. ci denotes a constant depending only on ; N ;
N K, and l. By (6.8), F is
./
continuous. For simplicity, we put u.t/ D uT0 .t/. Equation (6.44) follows from the
inequality
jF .u.t//j 1 C ku.t/k
kF k# : (6.46)
1 C kk 1 C kk
./
Putting uQ .t/ D uT Q .t/ and v.t/ D u.t/ uQ .t/, let us show (6.45). We have
0
ˇ 1 C ku.t/k 1 C kQu.t/k ˇˇ
ˇ #
ˇF .u.t// F # .Qu.t// ˇ
1 C kk Q
1 C kk
1 C ku.t/k ˇ 1 C ku.t/k 1 C kQu.t/k ˇ
ˇ ˇ
mF # .kv.t/k/ C kF k# ˇ ˇ
1 C kk 1 C kk Q
1 C kk
WD J1 C kF k# J2 : (6.47)
Since the modulus m./ satisfies that, for any " > 0, there is C" > 0, such that
m.a/ " C C" a; a 0, we have
Q 2 /:
.EJ1 /2 c1 ."2 C C"2 k k (6.49)
kv.t/k Q
.1 C kQu.t/k/k k
J2 C : (6.50)
1 C kk Q
.1 C kk/.1 C kk/
Q
EJ2 c2 k k: (6.51)
Let us study how the terminal cost function, .t; T0 ; ; ./I F / WD ET0 F .u./.t//,
depends on T0 ; , and ./.
2. Dependence on time
Proposition 6.7. For 2 H0 and " > 0, there is ı3 ."; I F / > 0, such that, for
t1 ; t2 2 ŒT0 ; T ,
h i
ET0 sup jF .u./.s2 // F .u./.s1 //j < "; 8./ 2 ; (6.53)
s1 ;s2 2Œt1 ;t2
we have
h i
ET0 sup jF .u.s// F .u.t1 //j
t1 st2
3. Dependence on ./
Proposition 6.8.
(i) For positive constants " and n, there is ı4 ."; nI F / > 0, such that
6.2 Formulations of Control Problems 221
h i
Q
ET0 sup jF .u./.t// F .u./ .t//j < " (6.56)
T0 t T
Q < ı.
for any ./ 2 , whenever k k
Next we fix 2 H1 , so that k k < ı and k k < n. Then
./ Q
sup jF .uT0 .t// F .uT Q .t//j
t 0
./ ./
sup jF .uT0 .t// F .uT0 .t//j
t
./ Q
./
C sup jF .uT0 .t// F .uT0 .t//j
t
Q
./ Q
./
C sup jF .uT0 .t// F .uT Q .t//j
t 0
1 "
ı5 ."; nI F / D 1 ^ ı2 ; n C 1I F
2 3
and
"
ı6 ."; I F / D ı4 ; jjj jjjI F :
3
222 6 Optimal Controls for Zakai Equations
Put
When a control process ./ 2 is taken, the running cost R on ŒT0 ; t is given by
Z t
./
R.t; T0 ; ; .// D r.uT0 .s/; .s// ds: (6.62)
T0
Let us list several properties of the running cost function, by using results
obtained in Sects. 5.4 and 6.1. By Ki , we denote a constant depending only on
N ;
; N K; l; lr and r0 .
1. Bound
h i Z T
E sup jR.t; T0 ; ; .//j r0 T C lr ET0 ku./ .s/k ds
T0 t <T T0
K1 .1 C kk /: (6.63)
3. Dependence on time
Z t p
ET0 jr.u./ .s/; .s//j ds K3 .1 C kk / t : (6.65)
Q
Proof. For simplicity, we put u.t/ D u./ .t/; uQ .t/ D u./ .t/ and v.t/ D u.t/ uQ .t/.
Then v satisfies
6.2 Formulations of Control Problems 223
8
ˆ
ˆ d v.t/ D .A./ .t/v.t/ C Z.t// dt
<
C .G ./ .t/v.t/ C .t// d W .t/; t 2 .T0 ; T ;
ˆ
:̂
v.T0 / D 0;
where
(
Z.t/ D .A./ .t/ AQ ./ .t//Qu.t/;
.t/ D .G ./ .t/ G Q ./ .t//Qu.t/:
and
we obtain (6.66).
ER.t;T0 ;;.//
5. R ./ WD
#
1Ckk
is bounded and Lipschitz continuous uniformly in
.t; T0 ; .//
Indeed, (6.63) and (6.64) imply
and
and
From (6.57) and (6.66), we deduce that, for " > 0 and 2 H0 , there is a positive
constant ıC ."; I F / such that
h i
E sup jC.t; T0 ; ; ./I F / C.t; T0 ; ; Q ./I F /j < " (6.70)
T0 t T
whenever E1./ Q ./12 < ıC ."; I F /. Now we want to minimize the payoff by
choosing a suitable control process from .
Definition 6.1. The minimum payoff
From the results in Sects. 6.2.2 and 6.2.3, the following properties of payoff and
the value function are immediate.
Proposition 6.9. Let F 2 C # be given.
(i) V .t; T0 ; I F / 2 C # .
(ii) V .t; T0 ; I F / D V .t T0 ; 0; I F /.
(iii) Dependence on the initial state
There are modulus functions, mn .I F /; n D 1; 2; : : : , such that
8
ˆ
ˆ jJ.t; T0 ; 1 ; ./I F / J.t; T0 ; 2 ; ./I F /j mn .k1 2 kI F /;
ˆ
ˆ
ˆ
< 8t; T0 ; ./;
ˆ
ˆ jV .t; T0 ; 1 I F / V .t; T0 ; 2 I F /j mn .k1 2 kI F /;
ˆ
ˆ
:̂
8t; T0 ;
(6.73)
whenever k1 k; k2 k n,
(iv) Dependence on time
For 2 H0 , there is a modulus function m.I ; F /, such that
8
ˆ
ˆ jJ.t1 ; T0 ; ; ./I F / J.t2 ; T0 ; ; ./I F /j m.jt1 t2 jI ; F /;
<
8./ 2 ;
ˆ
:̂
jV .t1 ; T0 ; I F / V .t2 ; T0 ; I F /j < m.jt1 t2 jI ; F /:
(6.74)
As we said in Chap. 2, the DPP is a useful tool for analyzing control problems for
finite-dimensional processes governed by SDEs. In this section, we focus on DPP
for Hilbert space-valued processes governed by Zakai equations. We will formulate
DPP via the semigroup constructed by means of the value function, in the same way
as in Sect. 2.2. First, we show the DPP for the discrete-time case in Sect. 6.3.1. Then
we deal with the continuous-time DPP in Sect. 6.3.2, by using a time discretization
technique. We study properties of the semigroup and its generator in Sects. 6.3.3
and 6.3.4, respectively.
Now we have;
Remark 6.2.
Q
D DQ H) VDt F ./ VDt F ./; 8 2 H0
and
F FQ H) VDt F VDt FQ :
6.3 Dynamic Programming Principle (DPP) 227
is non-empty and compact. Hence there is a minimum selector O tF (Borel mapping;
H0 7! ) such that O tF ./ 2 ; 8 2 H0 . Thus
holds.
Theorem 6.2. For 0 i j p,
Y
j
VtiD;tj C1 F ./ D ti ti C1 tj tj C1 F ./ DW tk tkC1 F ./; 8 2 H0 : (6.84)
kDi
Outline of Proof. Since we can apply the same arguments as in Theorem 2.2, we
only sketch the proof.
Let ./ 2 D be given. Since (W .s/ W .tj /; s tj ) is independent from Ftj ,
(u .t/I t 2 Œtj ; tj C1 /) is the response for the constant control .tj / with initial
./
state u./ .tj /, under the conditional probability P .jFtj /. Hence, by freezing .tj /
and u./ .tj / we obtain
Y
j
J.tj C1 ; ti ; ; ./I F / tk tkC1 F ./: (6.86)
kDi
D
Taking the infimum over we obtain
Y
j
VtiDtj C1 F ./ tk tkC1 F ./: (6.87)
kDi
228 6 Optimal Controls for Zakai Equations
uO .t/ denotes its response with uO .ti / D . Repeating this procedure, we construct
O ./ 2 D and its response uO ./. Then we have
Y
j
J.tj C1 ; ti ; ; O ./I F / D tk tkC1 F ./: (6.89)
kDi
Since (6.89) yields the opposite inequality for (6.87), we have completed the
proof.
Corollary 6.1. Discrete-time DPP
Let D D .t1 ; : : : ; tp / be a division of Œ0; T . For ti tl tj ,
We need to establish the semigroup property of fV t I tg, which formulates the
DPP.
6.3 Dynamic Programming Principle (DPP) 229
Then
For the opposite inequality, we fix ./ 2 arbitrarily and take n0 ./ 2 Dn0 ,
so that
8
< lim E1n0 ./ ./12 D 0;
n0 !1
(6.96)
: lim n0 .t/ D .t/; a.e. on Œ0; T ˝:
0 n !1
Observing that
D
J.t; ; ; n0 ./I F / V t n F ./ lim VDt n F ./;
0
n!1
we have
VD n
1 3
F ./ D VD n
1 2
.VD n
2 3
F /./
VD n
1 2
.V2 3 F /./ (6.99)
lim VD m
2 3
F .u./ .2 // D V2 3 F .u./ .2 // P -a.s.
m!1
and
jVD m
2 3
F .u./ .2 //j c.1 C kF k# /.1 C ku./ .2 /k/ P -a.s.
Collecting previous results, we can easily deduce that .Vt / WD .Vt ; t 2 Œ0; T / is
a monotone semigroup on C # , satisfying
(i) 0 -contractiveness
kVt F Vt FQ k# e 0 t kF FQ k# (6.104)
p
where 0 D 1 C N1 with the constant N1 of (6.8).
(ii) t-continuity
For any 2 Œ0; T /,
St F Vt F; 8F 2 C # ; 8t 2 Œ0; T ; (6.108)
that is, .Vt / is the maximal element of ˙, called the envelope of ..Jt /; 2 /.
Proof. Let t be given. Take a sequence of divisions, Dn ; n D 1; 2; : : : ; so that,
the monotonicity of Jt yields
tj 1 tj .tj t F /./ inf Jtj tj 1 .St tj F /./
2
1X 2
m
L F ./ D hA ; DF ./i C hD F ./Gk ; Gk i; 2 : (6.112)
2
kD1
and
./
I.s/ WD L.s/ F .u0 .s// L.s/ F ./: (6.114)
We are going to show that, for " > 0, there is ."; / > 0, independent of ./, such
that
h i
E sup jI.s/j < " for t < ."; /: (6.115)
0st
6.3 Dynamic Programming Principle (DPP) 233
./
We divide the proof of (6.115) into four steps. For simplicity, we put u.t/ D u0 .t/
and ci denotes a constant independent of t; and ./.
Step 1. There is 1 ."; / > 0, independent of ./, such that
h i
E sup jhA./.u.s/ /; DF .u.s//ij < "; for t < 1 ."; /: (6.116)
0st
Q C jjjQ jjj
E sup jjju.s/ uQ .s/jjj C E sup jjjQu.s/ jjj
0sT 0st
p p p
Q
. N1 C 1/Q" C N2 t II; (6.118)
k D 1; 2; : : : ; m; (6.120)
LHS of (6.120)
h i
./ .s/
E sup jjjD 2 F .u.s//Gk .s/u.s/jjj kGk .s/.u.s/ /k
0st
h i
K0 K 2 E sup ku.s/k ku.s/ k
0st
12 12
K0 K 2 E sup ku.s/k2 E sup ku.s/ k2
0st 0st
p
c1 kk jjjjjj t (by (5.111));
k D 1; 2; : : : ; m; (6.121)
Thus, (6.115) and (6.123) yield that for any " > 0, there is ."; / > 0, such
that
h i
E sup jL.s/ F .u.s// C r.u.s/; .s// L.s/ F ./ r.; .s//j < "; 8./
0st
(6.124)
whenever t < ."; /.
Finally, we note that
Z t
inf E .L.s/ F ./ C r.; .s/// ds
./2 0
Z t
E inf .L F ./ C r.; // ds
0 2
Vt F ./ F ./
hZ t i
D inf E r.u.s/; .s// ds C F .u.t// F ./
./2 0
hZ t i
D inf E .r.u.s/; r.s// C L.s/ F .u.s/// ds
./2 0
n hZ t i
D inf E .r.u.s/; .s// C L.s/ F .u.s// r.; .s// L.s/ F .// ds
./2 0
Z t o
CE .r.; .s// C L.s/ F .// ds :
0
Now Theorem 6.5 shows that U.t; / WD VT t F ./ has the backward dynamics
This section is devoted to the study of the HJB equations introduced at the end of
Sect. 6.3. Let us consider HJB equation on H0
@t U.t; /
1X 2
m
C inf hA ; DU.t; /i C hD U.t; /Gk ; Gk i C r.; / D 0;
2 2
kD1
2 H0 ; t 2 .0; T / (6.127)
When the value function is smooth, the HJB equation provides the backward
evolution equation by Theorem 6.5. However we can hardly expect that the value
function will be regular. Here we will consider viscosity solution of the HJB
equation according to [L89] and [GŚ00].
In Sect. 6.4.1, we give the definition of the viscosity solution. Since we treat
the equation on H0 , the relation between the value function and the viscosity
solution has already been stated in [GŚ00]. Thus we will sketch the proof of the
assertion that the value function becomes a viscosity solution in Sect. 6.4.2 and
the uniqueness theorem in Sect. 6.4.3. Applying these results, we revisit control
problems for partially observable diffusions in Example 6.1.
6.4.1 Definitions
Put
n o
D ı./ 2 C 1 .0; T /I lim ı.t/ D lim ı.t/ D 1 and ı D inf ı.t/ > 0 :
t !0 t !T t
(6.129)
Elements of Cbu12
..0; T / H1 / is called test functions. For a test function ./ and
2 , we put
6.4 Viscosity Solutions of HJB Equations 237
1
L .t; / D hA ; D.t; /i C hD 2 .t; /G ; G im : (6.130)
2
P
where h ; im D m kD1 h k ; k i .
0 O
O C ı .t / kk
0 @t .tO; / O 2
2
n o
C inf L .tO; /O C 1 ı.tO/L kk
O 2 C r.;
O / (6.131)
2 2
holds and
holds and
O
I.t/ WD EŒU.t; u .t// U.tO; /:
Z tn
I.t/ E @s .s; u .s// C hA u .s/; D.s; u .s//i
tO
1
C hD 2 .s; u .s//G u .s/; G u .s/im
2
1 0
C ı .s/ku .s/k2 C ı.s/hA u .s/; u .s/i
2
1 o
C ı.s/kG u .s/k2m ds (6.138)
2
by Itô’s formula (see Sect. 5.4).
Next we calculate the expression inside f g in the RHS of (6.138). Take an
integer n0 , so that
h 1i
jı.s/ ı.tO/j C jı 0 .s/ ı 0 .tO/j 1 for s 2 tO; tO C : (6.139)
n0
6.4 Viscosity Solutions of HJB Equations 239
Z t
O O
O
I.t/ c3 .t t /.1 C kk / ı E
2
jjju .s/jjj2 ds: (6.141)
2 tO
Noting that
Z t
E jr.u .s/; /j ds .t tO/c4 1 C E sup ku .s/k/
tO tOsT
O
c5 .t tO/.1 C kk/; (6.142)
Put t D tO C 1
n
.n D n0 ; n0 C 1; : : : /. Then there is tn 2 ŒtO; tO C n1 , such that
O 2 D 0;
lim EŒku .tn0 / k (6.146)
n0 !1
R1
Since the LHS D 0 Ejjjun ./ .tO C n /jjj2 d, there is n 2 ŒtO; tO C n1 , such that
O 2 D 0:
lim EŒkun0 ./ .n0 / k (6.150)
n0 !1
hZ t
0 inf E r.u./ .s/; .s// ds
./2 tO
i
O C 1 .ı.t/ku./ .t/k2 ı.tO/kk
C .t; u./ .t// .tO; / O 2/
2
Z t
ı 0 .s/ ./
D inf E r.u./ .s/; .s// C @t .s; u./.s// C ku .s/k2
./2 tO 2
1
C L.s/ .s; u./ .s// C ı.s/L.s/ ku./ .s/k2 ds: (6.151)
2
6.4 Viscosity Solutions of HJB Equations 241
whenever t tO < O
1 ."; /.
Indeed, for "Q > 0, we take 2 H2 , so that jjjO jjj2 < ".
Q Let v./ denote the
./ O
response for ./, with v .t / D . Then, the inequality
O
whenever t tO < 2 ."; /.
We treat L .s; u .s// and other terms of (6.151) in the same way as the
.s/ ./
proof of Theorem 6.5. Then (6.151) shows that the subsolution inequality holds.
Since we can establish the supersolution inequality by the same arguments, this
completes the proof.
To show the uniqueness of the viscosity solution, we state the comparison theorem
[GŚ00], Th. 6.1. Denote by C the set of all U 2 C.Œ0; T H0 / such that
(a)
jU.t; /j
lim sup D 0;
kk!1 t kk2
Comparison Theorem
Thus, (6.155) and (6.156) show that U.t; / 2 Cu .H1 / uniformly on Œ0; T ,
namely there is a modulus m./, such that
Q m.k k
jU.t; / U.t; /j Q / 8t: (6.157)
Next we consider time continuity. Let u./ be the response for ./, with u.0/ D
. Since U.t; / D V0 T t F ./, we see that, for " > 0,
and
(
d Y .t/ D h.X.t/; .t; Y // dt C d W .t/; t 2 .T0 ; T ;
(6.161)
Y .T0 / D 0;
where
X
d hX
d X
d i
A D @i K ij .x; /@j b i .x; / @j K ij .x; / ;
i D1 i D1 j D1
X
d X
d
Gk D ki .x; /@i C h .x; /
k
@i ki .x; / ; k D 1; : : : ; m:
i D1 i D1
244 6 Optimal Controls for Zakai Equations
where q ./ ./ is the solution of (6.163), and J.T0 ; X0 I / depends on T0 , and .
Further, the value function V .T0 ; / WD inf2 J.T0 ; X0 I / is nothing but the
infimum of the RHS of (6.164) over ./ 2 ˇ . Since the terminal cost F . / WD
.f; / is in C.H1 /, Theorem 6.7 says that the value function V ./ is the unique
viscosity solution of the HJB equation
n 1X 2
m o
@t V .t; / C inf hA ; DV .t; /i C hD V .t; /Gk ; Gk i C .l.; /; /
2 2
kD1
D 0; t 2 .0; T /; 2 H1 ;
[A03] R.A. Adams, Sobolev Spaces, 2nd edn. (Academic, Amsterdam/Boston, 2003)
[BC09] A. Bain, D. Crisan, Fundamentals of Stochastic Filtering (Springer, New York/
London, 2009)
[Be52] R. Bellman, On the theory of dynamic programming. Proc. Nat. Sci. U.S.A. 38,
716–719 (1952)
[Be57] R. Bellman, Dynamic Programming (Princeton University Press, Princeton, 1957)
[Be75] V.E. Benes, Composition and invariance methods for solvinig some stochastic
control problems. Adv. Appl. Prob. 7, 299–329 (1975)
[BSW80] V.E. Venes, L.A. Shepp, H.S. Witsenhaussen, Some solvable stochastic control
problems. Stochastics 4, 39–83 (1980)
[Be92] A. Bensoussan, Stochastic Control of Partially Observable Systems (Cambridge
University Press, Cambridge/New York, 1992)
[BN90] A. Bensoussan, M. Nisio, Nonlinear semigroup arising in the control of diffusions
with partial observation. Stoch. Stoch. Rep. 30, 1–45 (1990)
[BP99] T.R. Bieleckii, S.R. Pliska, Risk sensitive dynamic asset management. Appl. Math.
Optim. 39, 337–366 (1999)
[BS73] F. Black, M. Scholes, The pricing of options and corporate liabilities. J. Polit. Econ.
81, 637–659 (1973)
[BM07] R. Buckdahn, J. Ma, Pathwise stochastic control problems and stochastic HJB
equations. SIAM J. Control Optim. 45, 2224–2256 (2007)
[Br06] S. Brendle, Portfolio selection under incomplete information. Stoch. Proc. Appl.
116, 701–723 (2006)
[CPY09] M.H. Chsng, T. Pang, J. Yong, Optimal stopping problem for stochastic differential
equations with random cofficients. SIAM J. Control Optim. 48, 941–971 (2009)
[CI90] M.G. Crandall, H. Ishii, The maximum principle for semicontinuous functions.
Differ. Integral Equ. 3, 1001–1014 (1990)
[CIL92] M.G. Crandall, H. Ishii, P.L. Lions, A user’s guide to viscosity solutions. Bull. AMS
NS 27, 1–67 (1992)
[DaPZ92] G. Da Prato, J. Zabczyk, Stochastic Equations in Infinite Dimenmensions
(Cambridge University Press, Cambridge, 1992)
[ElKPQ97] N. El Karoui, S. Peng, M.C. Quenez, Backward stochastic differential equations in
finance. Math. Finance 7, 1–71 (1997)
[ElKQ95] N. El Karoui, M.C. Quenez, Dynamic programming and pricing of contingent
claims in an incomplete market. SIAM. J. Control Optim. 33, 29–66 (1950)
[La83] S. Lang, Real Analysis, 2nd edn. (Addison-Wesley, New York, 1983)
[L83] P.L. Lions, Optimal control of diffusion processes and Hamilton-Jacobi-Bellman
equations I. J. Commun. PDE. 8, 1101–1134 (1983)
[L88] P.L. Lions, Viscosity solutions of fully nonlinear second-order equations and
optimal stochastic control in infinite dimensions. Part I. The case of bounded
stochastic evolution. Acta Math. 161, 243–278 (1988)
[L89] P.L. Lions, Viscosity solutions of fully nonlinear second-order equations and
optimal stochastic control in infinite dimensions. Part II. Optimal control of Zakai
equation, in Stochastic Partial Differential Equations and Applications II, ed.
by G. Da Prato, L. Tubaro. Lecture Notes in Mathematics, vol. 1390 (Springer,
Berlin/Heidelberg, 1989), pp. 147–170. Part III. Uniqueness of viscosity solutions
for general second order equations. J. Funct. Anal. 86, 1–18 (1989)
[LN83] P.L. Lions, M. Nisio, A uniqueness result for the semigroup associated with HJB
equations. Proc. Jpn. Acad. 58, 273–276 (1983)
[LS01] R.S. Liptser, A.N. Shiryayev, Statistics of Random Processes I, II, 2nd edn.
(Springer, New York/Berlin, 2001)
[Ma00] C. Martini, American option prices as unique viscosity solutions to degenerate HJB
equations, Rapport de rech, INRIA, 2000
[Me71] R.C. Merton, Optimal consumption and portfolio rules in continuous time. J. Econ.
Theory 3, 373–413 (1971)
[Me73] R.C. Merton, Theory of rational option pricing. Bell J. Econ. Manag. Sci. 4,
141–183 (1973)
[M88] M. Metivier, Stochastic Partial Differential Equations in Infinite Dimensional
Spaces (Scuola Superiore, Pisa, 1988)
[My66] P.A. Meyer, Probability and Potentials (Blaisdell, Waltham, 1966)
[Mo10] H. Morimoto, Stochastic Control and Mathematical Modeling, Applications in
Economics (Cambridge University Press, Cambridge/New York, 2010)
[Na03] H. Nagai, Optimal strategies for risk sensitive portfolio optimization problems for
general factor models. SIAM J. Control Optim. 41, 1779–1800 (2003)
[N76] M. Nisio, On a nonlinear semigroup attached to stochastic optimal control. Pull.
RIMS 12, 513–537 (1976)
[N78] M. Nisio, On nonlinear semigroups for Markov processes associated with optimal
stopping. Appl. Math. Optim. 4, 143–169 (1978)
[N81] M. Nisio, Lecture on Stochastic Control Theory. ISI Lecture Notes, vol. 9 (McMil-
lan India, Delhi 1981)
[N88] M. Nisio, Stochastic differential games and viscosity solutions of Isaacs equations.
Nagoya Math. J. 110, 163–184 (1988)
[P79] E. Pardoux, Stochastic partial differential equations and filtering of diffusion
processes. Stochastics 3, 127–167 (1979)
[P93] E. Pardoux, Stochastic partial differential equations, a review. Bull. Sc. Math. 117,
29–47 (1993)
[PP90] E. Pardoux, S. Peng, Adapted solution of backward stoshastic equation. Syst.
Control Lett. 14, 55–61 (1990)
[Pe92] S. Peng, Stochastic Hamilton-Jacobi-Bellman equations. SIAM J. Control Optim.
30, 284–304 (1992)
[Ph02] H. Pham, Smooth solutions to optimal investment models with stochastic volatilities
and portfolio constraints. Appl. Math. Optim. 46, 55–78 (2002)
[R90] B.L. Rozovskii, Stochastic Evolution Systems (Kluwer Academic,
Dordrecht/Boston, 1990)
[S08] A.N. Shiryayev, Optimal Stopping Rules, 2nd edn. (Springer, Berlin/Heidelberg
2008)
[St11] L. Stettner, Penalty method for finite horizon stopping problems. SIAM J. Control
Optim. 49, 1078–1099 (2011)
248 References
A D
adapted, 6 density-valued process, 200, 201
admissible control, 32, 118 differential games, 144
admissible strategy, 118 discrete-time DPP, 41, 102
admissible system, 100 discrete-time DP property, 44
almost surely, 4 DP property, 51
American option price, 112 DPP, 47, 107, 132, 228
approximation, 46, 106
E
B Elliott–Kalton strategy, 118
backward evolution operator, 10 Elliott–Kalton upper value, 121
backward stochastic differential equations, 24 Elliott-Kalton lower value, 120
Bayes formula, 18 energy equality, 170
Black–Scholes formula, 30 expectation, 2, 154
Brownian adapted control, 34, 50 exponential martingale, 14
Burkholder–Davis–Gundy inequality, 8, 164
F
C factor processes, 68
classical solution of HJB, 61 filtered probability space, 5
colored Wiener process, 159
Comparison Principle, 95
conditional expectation, 2, 155
conditional probability, 3 G
continuous martingale, 6, 156 generalized solution, 171
control process, 32 generator, 54, 110, 136, 232
control-stopping problem, 100 Girsanov transformation, 18
controlled Zakai equation, 211 Girsanov’s theorem, 19
convergence, 4
correlation operator, 159
cost function, 33 H
covariance operator, 159 HJB equation, 56, 60, 94, 112, 236
Crandall–Ishii Lemma, 89 H-process, 156
criterion, 33 H-random variable, 154
M
Markov process, 10 T
Markovian policy, 34 Taylor formula, 186
mean, 2 time-homogeneous, 57
mean return process, 69 transition operator, 10
mean value theorem, 186 transition probability, 9
O U
optimal control, 33 unique viscosity solution, 242
optimal investment problems, 68 uniqueness, 99
unnormalized conditional expectation, 200
P upper Isaacs equation, 140
parabolic differentials, 80 upper value, 121
parabolic equation, 80 upper value function, 134
payoff, 33 utility function, 70
portfolio strategies, 26
power utility function, 70
probability space, 2 V
progressively measurable, 6 value function, 33, 134, 224
verification theorem, 60
viscosity solution, 81, 112, 139, 140, 237
Q viscosity solution of the HJB equation, 85, 236
quadratic variational process, 7, 157, 158 viscosity subsolution, 81, 84, 237
viscosity supersolution, 81, 84, 237
R volatility process, 69
random variable, 2
reference probability system, 32
regular conditional probability, 3 W
response, 33 weak solution, 23
risk sensitive stochastic controls, 141 Wiener process, 9
S Z
saddle point property, 123 Zakai equation, 203, 204