Presentación 2 Webconference

July 27, 2007 10:5 Proceedings Trim Size: 9.75in x 6.
5in Chap8-Contents
xi
Copyright © 2007. World Scientific. All rights reserved. May not be reproduced in any form without permission from the publisher, except fair uses permitted under U.S. or
Part Four: Dynamics and Optimal Control of Partial

Differential Equations
Optimal Control of Quasilinear Elliptic Obstacle Problems . . . . . . . . . . . . . . . . . . . 263

Qihong Chen and Yuquan Ye
Controllability of a Nonlinear Degenerate Parabolic System

with Bilinear Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
Ping Lin, Hang Gao, and Xu Liu
Near-Optimal Controls to Infinite Dimensional Linear-Quadratic

Optimal Control Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
Liping Pan and Qihong Chen
Shape Optimization of Stationary Navier–Stokes Equation . . . . . . . . . . . . . . . . . . . 323

Gengsheng Wang, Lijuan Wang, and Donghui Yang
Solution Map of Strongly Nonlinear Impulsive Evolution Inclusions . . . . . . . . . . .338

Xiaoling Xiang, Yunfei Peng, and Wei Wei
Study on Repairable Series System with Two Components — A Semigroup

Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
Hong Xu and Dinghua Shi
Finite Dimensional Reduction of Global Dynamics and Lattice Dynamics

of a Damped Nonlinear Wave Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
Yuncheng You
Recent Progress on Nonlinear Wave Equations via KAM Theory . . . . . . . . . . . . . . 387

Xiaoping Yuan
applicable copyright law.
EBSCO Publishing : eBook Academic Collection (EBSCOhost) - printed on 9/11/2017 7:08 PM via UNIVERSIDAD
NACIONAL ABIERTA Y A DISTANCIA - UNAD
AN: 235971 ; Yong, J., Tang, Shanjian, Li, Xunjing.; Control Theory And Related Topics: In Memory Of
Professor Xunjing Li
Account: ns145102
This page intentionally left blank

Account: ns145102
July 27, 2007
9:43
Proceedings Trim Size: 9.75in x 6.5in
Copyright © 2007. World Scientific. All rights reserved. May not be reproduced in any form without permission from the publisher, except fair uses
Chap5-ConferencePhoto
permitted under U.S. or applicable copyright law.
From left: back row — Zengjing Chen, Jianxiong Huang, Liping Pan, Qihong Chen, Gengsheng Wang, Hongwei Lou, Hanzhong Wu
and Xiaobo Bao. Middle row — Hang Gao, Yuan Zhou, Kangsheng Liu, Chunjiang Qian, Xu Zhang, Xun Li, Yashan Xu and Lei Wang.
Front row — Xiaoling Xiang, Shanjian Tang, Shige Peng, Shuping Chen, Yuncheng You, Jin Ma, Xunyu Zhou, Wei Lin and Ying Hu.
EBSCO Publishing : eBook Academic Collection (EBSCOhost) - printed on 9/11/2017 7:08 PM via UNIVERSIDAD NACIONAL ABIERTA Y A DISTANCIA - UNAD
AN: 235971 ; Yong, J., Tang, Shanjian, Li, Xunjing.; Control Theory And Related Topics: In Memory Of Professor Xunjing Li
Account: ns145102

Account: ns145102
July 27, 2007 11:33 Proceedings Trim Size: 9.75in x 6.5in Chap6-Participants
LIST OF PARTICIPANTS
Xiaobo Bao
School of Mathematical Sciences, Fudan University, Shanghai, China
032018054@fudan.edu.cn
Qihong Chen
Department of Applied Mathematics,
Shanghai University of Finance and Economics, Shanghai, China
chenqih@yahoo.com
Shuping Chen
Department of Mathematics, Guizhou University, Guizhou, and
Department of Mathematics, Zhejiang University, Hangzhou, China
chensp@gzu.edu.cn
Zengjing Chen
Department of Mathematics, Shandong University, Jinan, China
chenqih@online.sh.cn
Hang Gao
School of Mathematics & Statistics,
Northeast Normal University, Changchun, China
hangg@nenu.edu.cn
Ying Hu
IRMAR, Université Rennes 1, Rennes, France
ying.hu@univ-rennes1.fr
Jianxiong Huang
Shanghai College of Electric Power, Shanghai, China
13311716836@133sh.com
Xun Li
National University of Singapore, Singapore
matlx@nus.edu.sg
xv
Account: ns145102
xvi
Wei Lin
Department of Electrical Engineering and Computer Science,
Case Western Reserve University, Ohio, USA
linwei@nonlinear.cwru.edu
Daobai Liu
Fudan Univeristy, Shanghai, China
dbliu@fudan.edu.cn
Kangsheng Liu
Zhejiang University, Hangzhou, China
ksliu@zju.edu.cn
Hongwei Lou
School of Mathematical Sciences, Fudan Univeristy, Shanghai, China
hwlou@fudan.edu.cn
Jin Ma
Department of Mathematics, Purdue University, Indiana, USA
majin@math.purdue.edu
Liping Pan
lppan@fudan.edu.cn
Shige Peng
School of Mathematical Sciences, Fudan Univeristy, Shanghai, and School
of Mathematics and System Science, Shandong University, Jinan, China
peng@sdu.edu.cn
Chunjiang Qian
Department of Electrical and Computer Engineering,
University of Texas at San Antonio, San Antonio, Texas, USA
cqian@utsa.edu, Chunjiang.Qian@utsa.edu
Shanjian Tang
sjtang@fudan.edu.cn
Gengsheng Wang
Department of Mathematics, Wuhan University, Wuhan, Hubei, China
wanggs@public.wh.hb.cn
Account: ns145102
xvii
Jianbo Wang
jbwang@fudan.edu.cn
Lei Wang
Fudan Univeristy, Shanghai, China
Hanzhong Wu
hzwu@fudan.edu.cn
Xiaoling Xiang
Department of Mathematics, Guizhou University,
Guiyang 550025, Guizhou, China
xxl3621070@yahoo.com.cn
Yashan Xu
Yuncheng You
Department of Mathematics,
University of South Florida,Tampa, Florida, USA
you@math.usf.edu
Xiaoping Yuan
xpyuan@fudan.edu.cn
Xu Zhang
School of Mathematics, Sichuan University, Chengdu, China
xu.zhang@uam.es
Hailang Zhou
Shanghai Bank, China
Xun Yu Zhou
Department of Systems Engineering and Engineering Management,
The Chinese University of Hong Kong, Shatin, Hong Kong, China
xyzhou@se.cuhk.edu.hk
Account: ns145102
xviii
Yuan Zhou
mayzhou@fudan.edu.cn
Daxun Zhu
madxzhu@fudan.edu.cn
Huimin Zhu
hmzhu@fudan.edu.cn
Liang Zhu
Shangwei Zhu
Shanxi Finance & Economics University, Taiyuan, China
fdswzhu@126.com
Account: ns145102
July 27, 2007 11:26 Proceedings Trim Size: 9.75in x 6.5in Chap7-Contributors
LIST OF CONTRIBUTORS
Xiaobo Bao
School of Mathematical Sciences, Fudan University, Shanghai, China
Qihong Chen
chenqih@yahoo.com
Shuping Chen
Department of Mathematics, Guizhou University, Guizhou, and
Department of Mathematics, Zhejiang University, Hangzhou, China
chensp@gzu.edu.cn
Zengjing Chen
Department of Mathematics, Shandong University, Jinan, China
chenqih@online.sh.cn
Matt Davison
University of Western Ontario London, ON, Canada
mdavison@uwo.ca
Marco Fuhrman
Dipartimento di Matematica, Politecnico di Milano, Milano, Italy
marco.fuhrman@polimi.it
Hang Gao
School of Mathematics & Statistics,
hangg@nenu.edu.cn
Ying Hu
IRMAR, Université Rennes 1, Rennes, France
ying.hu@univ-rennes1.fr
xix
Account: ns145102
xx
Hanqing Jin
hqjin@se.cuhk.edu.hk
Hao Lei
Ping Lin
Graduate Student, School of Mathematics & Statistics,
llinp258@nenu.edu.cn
Wei Lin
Kangsheng Liu
ksliu@zju.edu.cn
Xu Liu
Graduate Student, Department of Mathematics, College of Science,
Zhejiang University, Hangzhou and School of Mathematics & Statistics,
Hongwei Lou
hwlou@fudan.edu.cn
Jin Ma
majin@math.purdue.edu
Liping Pan
lppan@fudan.edu.cn
Account: ns145102
xxi
Shige Peng
School of Mathematical Sciences, Fudan Univeristy, Shanghai, and
School of Mathematics and System Science, Shandong University, Jinan, China
peng@sdu.edu.cn
Yunfei Peng
Department of Mathematics, Guizhou University, Guiyang, Guizhou, China
pengyf0803@126.com
Jason Polendo
Jason.Polendo@utsa.edu
Chunjiang Qian
cqian@utsa.edu, Chunjiang.Qian@utsa.edu
Mark Reesor
University of Western Ontario London, ON, Canada
mreesor@uwo.ca
Dinghua Shi
Department of Mathematics, Shanghai University, Shanghai, China
Shanjian Tang
sjtang@fudan.edu.cn
Gianmario Tessitore
Dipartimento di Matematica, Università di Parma, Parma, Italy
gianmario.tessitore@unipr.it
Gengsheng Wang
wanggs@public.wh.hb.cn
Lijuan Wang
hepeijie@public.wh.hb.cn
Account: ns145102
xxii
Jianfeng Wei
Wei Wei
Department of Mathematics, Guizhou University, Guiyang, Guizhou, China
wei66@yahoo.com
Xiaoling Xiang
Department of Mathematics, Guizhou University,
Guiyang 550025, Guizhou, China
xxl3621070@yahoo.com.cn
Chao Xu
yuxin@zju.edu.cn
Hong Xu
Department of Mathematics, Shanghai University, Shanghai, China
Mingyu Xu
School of Mathematics and System Science,
Shandong University, Jinan, China
xvmingyu@hotmail.com
Yashan Xu
Bo Yang
Department of Mathematics and Statistics,
Texas Tech University, Lubbock, Texas, USA
bo.yang@ttu.edu
Donghui Yang
dongfyang@yahoo.com.cn
Yuquan Ye
yyq@mail.shufe.edu.cn
Account: ns145102
xxiii
Jiongmin Yong
Department of Mathematics, University of Central Florida, Orlando, Florida, USA,
and School of Mathematical Sciences, Fudan University, Shanghai, China
jyong@mail.ucf.edu
Yuncheng You
Department of Mathematics, University of South Florida,Tampa, Florida, USA
you@math.usf.edu
Xin Yu
yuxin@zju.edu.cn
Yuhua Yu
yyu@math.purdue.edu
Xiaoping Yuan
xpyuan@fudan.edu.cn
Xu Zhang
School of Mathematics, Sichuan University, Chengdu, China
xu.zhang@uam.es
Ying Zhang
Department of Mathematics and Statistics Acadia University, Canada
ying.zhang@acadiau.ca
Xun Yu Zhou
xyzhou@se.cuhk.edu.hk
Shangwei Zhu
Shanxi Finance & Economics University, Taiyuan, China
fdswzhu@126.com
Account: ns145102

Account: ns145102
July 27, 2007 11:34 Proceedings Trim Size: 9.75in x 6.5in Chap9-PartOneTitle
PART ONE
XUNJING LI’S ACADEMIC LIFE

Account: ns145102
July 27, 2007 11:34 Proceedings Trim Size: 9.75in x 6.5in Chap9-PartOneTitle
2
Account: ns145102
July 27, 2007 11:42 Proceedings Trim Size: 9.75in x 6.5in Chap10-Tribute
A TRIBUTE IN MEMORY OF PROFESSOR XUNJING LI ON HIS

SEVENTIETH BIRTHDAY
JIN MA
Purdue University,
West Lafayette, IN 47907-1395, USA
E-mail: majin@math.purdue.edu
YUNCHENG YOU
University of South Florida,
Tampa, FL 33620-5700, USA
E-mail: you@math.usf.edu
Professor Xunjing Li passed away in February, 2003, at age 68, in Shanghai,

where he had lived for about half a century, almost equivalent to all his professional
life. Professor Li was known to many of his students as “Lao Ban”, which means
“Boss” in English. This is not only because of his absolute authority sensed by his
students, but most importantly is because of his vision in finding new directions
of research, his rigorous attitude towards each and every detail in the research
work, and his role as a mentor to many young mathematicians he had fostered. He
has been greatly missed by all his students, colleagues, collaborators, especially in
the event such as the international conferences on stochastic control, mathematical
finance, and/or backward stochastic differential equations, like the one held this
year (2005) in Fudan University.
Professor Xunjing Li was born in Qingdao, Shandong Province, China, on June
13, 1935. He came to Shanghai in 1956, at age of 21, as a graduate student in the
Department of Mathematics, Fudan University, after receiving a Bachelor degree
of Mathematics from Shandong University. He spent three years for his graduate
study, under the supervision of the renowned mathematician Professor Jiangong
Chen, in the area of approximation theory of functions. He started his teaching
and research career in 1959, when he became an assistant professor in the Depart-
ment of Mathematics, Fudan University. He was promoted to Lecturer in 1962, to

Associated Professor in 1980a , and to Full Professor in 1984. In 1985, he became
a nationally appointed Doctoral Supervisor, and was named as a Distinguished
a From 1966 to 1976, China was in the period of “Cultural Revolution”. All academic titles in
universities were abolished during that period
Account: ns145102
4
Professor of Fudan University in 1997. Professor Li retired in 2001.

During the forty-five year span of his professional life, Professor Xunjing Li grew
from a student to a researcher and an educator, and made ineffaceable contributions
to the advances in control theory and related fields in China. He was best known
for his works on Maximum Principle for optimal controls of infinite dimensional
control systems, and is one of the most prominent pioneers in the area of the optimal
control theory of distributed parameter systems in China. Professor Li’s another
major contribution to the Chinese and the international control community was,
arguably, the fact that he almost single-handedly found and fostered a stochastic
control group in Fudan University, which has produced a flock of scholars who later
become influential players in the area across China, Europe and North America.
Professor Li’s research achievements can be chronologically summarized accord-
ing to the following main periods.
1. 1959—1976. Almost as soon as Professor Li finished his graduate study

and started work as an assistant lecturer in the Department of Mathematics of
Fudan University, he switched his research direction from Function Theory to Or-
dinary Differential Equations. This was partly due to a call from the department
to expand the academic areas, and as a young faculty member Professor Li enthu-
siastically answered. Following the leadership of Professor Fulin Jin, Professor Li
spent tremendous amount of time in the teaching and research in this new area.
In 1962, when he was only 27 years old, he co-authored the text book “Ordinary
Differential Equations” with Professor Fulin Jin. The book has been widely used as
a main text/reference book by researchers and students in China for many years.
In the meantime, as a natural extension of the theory of ordinary differential equa-
tions, Professor Li began to explore the area of control theory in early 1960’s. While
working on the subject of absolute stability of (finite dimensional) dynamic systems,
he participated in another important service activity to the Chinese control theory
community: this time was the translation of the celebrated monograph “Mathe-
matical Theory of Optimal Processes” by L. S. Pontryagin et al. In the middle of
1960’s, when Professor Li’s research activities just started to take off, the whole
direction of China took an unfortunate turn. During the ten-year period of 1966-
1976, the normal education and scientific research were strongly discouraged and
even interrupted due to the “Cultural Revolution”. Taking the only opportunity
in late 1960’s to apply his expertise in optimal control theory, Professor Li turned
to industrial and applied mathematics. He participated in several projects asso-
ciated with Shanghai Petroleum Refinery Factory, as well as Jin-Shan Petroleum

Chemical Cooperation. These experiences later became an important factor for his
perspectives towards control theory throughout his research.
Account: ns145102
5
2. 1977—1987. The year 1977 marked a resurrection and a new era of Chi-
nese education system, as well as a turning point in Professor Li’s research career.
Starting from that year, many traditional teaching and research activities began
to be restored, including the national college entrance examination, a long tradi-
tion in China for universities to enroll freshman students, at both undergraduate
and graduate levelb . However, the scar in the scholastic activities and scientific re-
search caused by the ten-year turmoil period was clearly visible, the research results
achieved by many scholars like Professor Li but shelved for a decade became a little
out-of-date. Inspired by the new spring in the scientific field, he started to ponder
over the new germing points of his research. After a careful survey of articles and
evaluating his strength, he decided to attack the infinite dimensional optimal control
problems. The first break through came out in 1978, in a joint work with Professor
Yunlong Yao, then an assistant professor. For a time optimal control problem of
infinite dimensional linear systems, they realized, in general the attainable sets is
not necessarily convex (a substantial difference from the finite dimensional case),
but they discovered that its closure must be. Such an observation, together with the
separation theorem for convex sets in infinite dimensional spaces, lead to a proof of
maximum principle of time optimal control for infinite dimensional linear systems.
Their work was published in the top journal in China, Scientia Sinica (“Science in
China”), and was later presented in the 8th International Federation of Automa-
tion Conference (IFAC), Kyoto, Japan. While this might be considered usual by
today’s standard, but back then when China was just opened up, it was indeed a
highly recognizable event. The subsequent several years then witnessed a series of
research accomplishments by Professor Li and his group, including the second au-
thor. The vector-valued measures in the infinite dimensional optimal control theory
was investigated in depth, and the Pontryagin’s maximum principle was extended
to various cases of general semi-linear evolutionary distributed-parameter systems.
Among many other results, the one that involves terminal constraints is particularly
worth mentioning. It was known that in the finite-dimensional case, the maximum
principle requires only the differentiability of the coefficients, provided the the ter-
minal constraint set is closed and convex. But there exist counter-examples showing
that this is no longer the case in general for the infinite dimensional systems. As
a consequence, seeking the proper conditions under which the maximum principle
remains valid became a long-standing challenging problem. In 1985, Li and Yao
successfully resolved the problem with rigor and elegance. They proved that, for
the general semi-linear evolutionary distributed-parameter systems, if the terminal
constraints satisfy the finite co-dimension condition, then the maximum principle
holds. This result was highly recognized by the international control community,
and was later regarded as the foundation of the “Fudan School” research on the
infinite-dimensional optimal control theory.
b Both authors of this article were the beneficiaries of this new policy, though at different levels.
Account: ns145102
6
3. 1987—2001. Since 1985, a group of new researchers, mostly the new gen-
eration of graduate students after the class of 1977, started to join Professor Li’s
research group. By 1988, some new Ph.D’s from Europe and US returned to China
and became the important new addition. Among others, most notable were the
Post-doctor Shige Peng (France) and Associate Professor Jiongmin Yong (USA).
In 1989, Professor Li, in collaboration with Jiongmin Yong, further extended the
maximum principle to the general semilinear evolutionary distributed systems with
mixed initial-terminal constraints, by using Ekeland’s Variational Principle and an
improved “spike” variational method. This ignited another wave of activities, and
a variety of infinite dimensional versions of the maximum principle were knocked
down. It is commonly recognized that finite-dimensional optimal control theory
has three milestones: the maximum principle by L. S. Pontryagin, the method of
dynamic programming by R. Bellman, and the linear quadratic optimal control
theory by R. E. Kalman. Professor Li’s research covered essentially all the areas,
although the main focus in his earlier years was on the representation of the Pon-
tryagin maximum principle in the infinite dimensional spaces. Many of Professor
Li’s works and thoughts, along with many results obtained by the control theory
research group of Fudan University (a.k.a. “Fudan School”), can be found in the
book “Optimal Control Theory for Infinite Dimensional Systems”, co-authored by
Xunjing Li and Jiongmin Yong, and published by Birkhäuser in 1995. The book
summarized quite exhaustively the latest results in the optimal control theory of
nonlinear, deterministic, infinite dimensional systems up to that point, from the
perspectives of the aforementioned three milestones. It was very well commented
by researchers in the field of control theory.
Although for the most part of his professional life, Professor Li considered him-
self a “deterministic person”, he was nevertheless in essence the main reason of the
existence of several research groups, including the stochastic control group and later
the mathematical finance group, in Fudan University. Since 1985, especially after
he visited several universities in USA, he had a vision that Fudan had to develop
the research direction on stochastic control. He began by organizing a stochastic
control seminar, and directing several of his graduate students, including the first
authorc , to study and to explore new problems in that area. With Professor Li’s
cultivation, the scope of the research on stochastic control was quickly expanded to
most of the subjects in the field. In his late years, Professor Li personally involved
in many research projects on stochastic control theory. Collaborating with Shup-
ing Chen, Ying Hu, Shige Peng, Shanjian Tang, Jionming Yong, Xunyu Zhou, and
others, he worked on various problems in linear quadratic control problems, Max-
imum Principle for stochastic control systems with partial observations, and with
jumps. Apart from these works, Professor Li also made important contributions
in many other areas such as multi-player differential games, infinite dimensional
c The first author later went on to complete his Ph.D dissertation, on singular stochastic control
problems, at University of Minnesota.
Account: ns145102
7
linear quadratic unbounded optimal control, and optimal control of elliptic partial
differential equations.
As a closely related subject to his research, Professor Li also had far-reaching
perspectives toward infinite-dimensional dynamical systems and applications, an
area that has been rapidly and richly expanding since 1980’s. In 1982 and 1983, he
invited three leading mathematicians in this area, Professors Jack Hale, George Sell,
and Shui-Nee Chow, to visit Fudan University. Inspired by these successful visits
and encouraged by Professor Li himself, Xiaobiao Lin (now a professor at the NCSU,
Raleigh) and the second author later finished their doctoral dissertations at Brown
university and at the University of Minnesota, respectively, on topics of infinite-
dimensional dynamical systems and then become active and prolific researchers in
this area.
In addition to his mathematical research, Professor Li’s life-long pedagogical

achievements constitute another highlight of his life. Professor Li supervised four
postdoctors, eleven Ph.D students, thirteen Master students, as well as many junior
faculty members. Professor Li was famously known of being strict to his students
(this in effect earned him the name “Lao Ban”, as we mentioned before). In fact,
looking back, almost all his former advisees could tell some anecdotes where he
or she learned lessons, sometimes embarrassing, from Professor Li. However, this
might exactly be one of the main reasons that many of them became successful later
on, when they became professors, researchers, supervisors, and principal investiga-
tors themselves. As one of the main figures in dynamic system and control theory
in Fudan University and in China, Professor Li showed tremendous leadership by
not only encouraging young faculty in his research group to boldly explore new
areas, but also guiding his graduate students in their studies and investigations in
areas unfamiliar to himself. This philosophy of Professor Li was the key for success
in many cases with his graduate students. It was because of these efforts that the
research directions of the Fudan (control theory) group expanded and developed
progressively, from distributed parameter control systems in 1970’s and 1980’s, to
stochastic control theory in 1980’s and 1990’s, and to mathematical finance through-
out the 1990’s and continuing through the 2000’s. It would not be exaggerating to
say that without Professor Li, the Fudan University would not have a control group
of a history like what it is seen today.
This year when we cherish the memory of Professor Xunjing Li on his seventieth
birthday, we all feel that his adamant scholastic spirit and his rigorous scientific
approach have more or less become a part of our own professional life in conducting
research and in educating younger generations of graduate students. We are proud

to be a part of “Fudan School”, and glad to see that the named of the group is
being carried on by many more talented and dedicated mathematicians year after
year, and hopefully for years to come. We believe that this is what Professor Li
would be pleased to see as well.
Account: ns145102
8
Acknowledgments
Another important obituary that describes Professor Xunjing Li’s role in the
development of control theory and related fields in China is the preface (in Chinese)
of his collection of articles1 . Our tribute can be considered an adapted version of
that article, and we owe our sincere gratitude to the authors of that article—Shuping
Chen, Shige Peng, and Jiongmin Yong, for their effort of collecting all the historical
information regarding Professor Li’s professional life, which is indeed the basis of
this tribute.
References
1. Xunjing Li, Selection of Mathematical Papers by Xunjing Li, Fudan University Press,
2003.
Account: ns145102
July 27, 2007 15:2 Proceedings Trim Size: 9.75in x 6.5in Chap11-PublicationsLi-ModifiedYong
PUBLICATIONS OF XUNJING LI
I. Books
1. Ordinary Differential Equations, Shanghai Science and Technology Press,

1962, 2nd ed., 1984, (with Fulin Jin et al., in Chinese).
2. Control Theory Applied to Computers, Fudan University press, 1988 (with

Laixiang Sun and Yougen Chen, in Chinese).
3. Theory of Ordinary Differential Equations for Optimal Control Systems,

Higher Education Press, 1989 (with Xueming Zhang and Zuhao Chen, in
Chinese).
4. Optimal Control Theory for Infinite Dimensional Systems, Birkhäuser, 1995

(with Jiongmin Yong).
5. Basics of Control Theory, Higher Education Press, 2002 (with Jiongmin

Yong and Yuan Zhou, in Chinese).
6. Collection of Mathematical Papers, Fudan University press, 2003 (forwarded

in Chinese by Shuping Chen, Shige Peng and Jiongmin Yong).
II. Edited Conference Proceedings
1. Control Theory of Distributed Parameter Systems and Applications, Lecture

Notes in Control and Information Sciences, vol. 159, Springer-Verlag, 1991
(with Jiongmin Yong).
2. Control of Distributed Parameter and Stochastic Systems, Kluwer Academic

Publishers, 1999 (with Shuping Chen, Jiongmin Yong and Xun Yu Zhou).
III. Research Papers
1. Stability of second-order linear ordinary differential equations with periodical

coefficients, Collection of Students’ Papers, Shandong University, 2 (1956),
1–4 (in Chinese).
Account: ns145102
10
2. Four-point common circles and circles of invariance under bilinear transfor-

mation, Collection of Students’ Papers, Shandong University, 2 (1957), 5–9
(in Chinese).
3. An extension of Nikolisky’s theorem, Journal of Fudan University (Natural

Science) no. 2 (1957), 264–269 (in Chinese, with an English summary).
4. Cesàro summability in Banach space, Acta Mathematica Sinica, 10 (1960),

41–54 (in Chinese); English Translation in Chinese Mathematics, 2 (1960),
40–52.
5. Uniform approximation of a function by the Cesàro means of its Faber series,

Journal of Fudan University (Natural Science), 5 (1960), no. 2, 159–166 (in
Chinese, with an English summary).
6. On the absolute stability of the indirect control systems Journal of Fudan

University (Natural Science), 7 (1962), no. 1, 25–34 (with Huimin Xie and
Junben Chen; in Chinese, with an English summary).
7. On the absolute stability of systems with time lags, Acta Mathematica

Sinica, 13 (1963), 558–573 (in Chinese); English Translation in Chinese
Mathematics, 4 (1963), 609–626.
8. Time optimal control of linear systems, Journal of Fudan University

(Natural Science), 9 (1964), no. 4, 501–512 (with Huimin Xie; in Chinese,
with a Russian summary).
9. Stability and time optimal control for automatic regulated systems,

Collection of Papers, Research Institute of Mathematics, Fudan University,
(1964), 95–111 (with Huimin Xie et al.; in Chinese).
10. Mathematical principle for electro-magnetic power-driven compressors (I) —

resonance analysis in the case of free loads, Journal of Fudan University
(Natural Science), 10 (1965), nos. 3–4, 151–154 (in Chinese).
11. Period-analysis approach to time series, Selection of Long-Term Forecast

Methods, (Eastern China 1973 meeting on precipitation forecast in the flood
season), Eastern China Meteorological Observatory Center, (1973), 71–73

(in Chinese).
12. Derived dynamic equation and feed-forward control for controlled plants —
an application of electronic digital computers in the control of atmospheric
Account: ns145102
11
distillation, Acta Math. Appl. Sinica, 1 (1976), no. 2, 33–45 (in Chinese).
13. Analytical design of control equations, Journal of Fudan University (Natu-

ral Science), 17 (1978), no. 2, 38–48 (in Chinese).
14. On the analytic design of the optimal regulator for the systems with time
lags, Journal of Fudan University (Natural Science), 19 (1980), no. 2, 189–
195 (in Chinese, with an English summary).
15. An analysis of the mathematical model to the furnace of thermal dissocia-

tion for generating ethene, Journal of Fudan University (Natural Science),
19 (1980), no. 1, 115–116 (in Chinese).
16. Time optimal boundary control for systems governed by parabolic equations,
Chin. Ann. Math., 1 (1980), nos. 3–4, 453–458 (in Chinese, with an English
summary).
17. Time optimal control of distributed parameter systems, Scientia Sinica, 24

(1981), 455–465 (with Yao Yunlong).
18. On the evolution equation in Banach space, Chin. Ann. Math., 2 (1981), no.
4, 479–489 (in Chinese, with an English summary).
19. On applications of vector measure to the optimal control theory for dis-
tributed parameter systems, Chin. Ann. Math., 3 (1982), no. 5, 655–662 (in
Chinese, with an English summary).
20. Methods and applications of control theory Nature Magazine, 5 (1982), 435–
438 (in Chinese).
21. On the stability of nonlinear control systems with time-lag, Control Theory
and Applications, 1 (1984), no. 3, 117–123 (in Chinese, with English sum-
mary).
22. Vector-valued measure and the necessary conditions for the optimal control
problems of linear systems, Journal of Mathematical Research and Exposi-
tion, 4 (1984), no. 4, 51–56.
23. Maximum principle of distributed parameter systems with time-lags,

Springer Lecture Notes in Control & Information Science, vol. 75 (1985),
410–427 (with Yunlong Yao).
Account: ns145102
12
24. Margin of stability for the optimal regulator of distributed parameter sys-
tems Control Theory and Applications, 3 (1986), no. 1, 76–82 (in Chinese,
with an English summary).
25. N-person differential games governed by infinite dimensional systems, J. Op-

tim. Theory & Appl., 50 (1986), 431–450.
26. Maximum principle of optimal periodic control for functional differential

systems, J. Optim. Theory & Appl., 50 (1986), 421–429.
27. Maximum principle of optimal controls for functional differential systems,

J. Optim. Theory & Appl., 54 (1987), 335–360 (with S. N. Chow).
28. On optimal control of functional differential systems, Springer Lecture Notes

in Control & Information Sciences, vol. 102 (1987), 112–119 (with Fulin Jin).
29. Optimal control for infinite dimensional systems, Springer Lecture Notes in
Control & Information Sciences, vol. 159 (1991), 96–105.
30. Maximum principle for optimal control of nonlinear generalized systems —

finite dimensional case, Acta Automatica Sinica, 17 (1991), 17–23 (with
Shige Peng; in Chinese).
31. Necessary conditions for optimal control of distributed parameter systems,

SIAM J. Control & Optim., 29 (1991), 895–906 (with Jiongmin Yong).
32. Dynamical model of neural network (I), Acta Biophysica Sinica, vol. 8, no.
2 (1992), 339–345 (with Fanji Gu and Jiong Ruan).
33. Dynamical model of neural network (II), Acta Biophysica Sinica, vol. 8, no.
3 (1992), 412–418 (with Fanji Gu and Jiong Ruan).
34. A realistic model of neural networks, J. of Electronics, vol. 9, no. 4 (1992),

289–295 (with Fanji Gu and Jiong Ruan).
35. Progress in control theory (distributed parameter systems), 1990 Yearbook

of Natural Science, 3.24–3.26, Shanghai Translation Publishing Company,
Shanghai, 1992 (in Chinese).
36. Optimal control theory for infinite dimensional systems, Progress in Natural
Science, 2 (1992), 104–112 (with Jiongmin Yong; in Chinese).
Account: ns145102
13
37. Maximum principle for optimal control problem of nonlinear generalized sys-
tems — infinite dimensional case, Acta Math. Appl. Sinica, 15 (1992),
no. 1, 99–104 (with Ying Hu and Shige Peng; in Chinese with an English
summary).
38. A linear quadratic optimal control problem with disturbances — an algebraic

Riccati equation and differential games approach, Appl. Math. Optim., 30
(1994), 267–305 (with Shuping Chen, Shige Peng and Jiongmin Yong).
39. Necessary conditions for optimal control of stochastic systems with random
jumps, SIAM J. Control & Optim., 32 (1994), 1447–1475 (with Shanjian
Tang).
40. Maximum principle for optimal control of distributed parameter stochas-

tic system with random jumps, Lecture Notes in Pure & Appl. Math., 152
(1994), 867–890 (with Shanjian Tang).
41. The effect of small time delays in the feedback on boundary stabilization,
Science in China (Series A), 36 (1993), 1435–1443 (with Kangsheng Liu).
42. General necessary conditions for partially observed optimal stochastic con-
trols, J. Appl. Prob., 32 (1995), 1118–1137 (with Shanjian Tang).
43. Stochastic verification theorems within the framework of viscosity solutions,

SIAM J. Control & Optim., 35 (1997), 243–253 (with Jiongmin Yong and
Xunyu Zhou).
44. Tracking control for nonlinear affine systems, J. Math. Control & Informa-
tion, 14 (1997), 307–318 (with K. L. Teo and W. Q. Liu).
45. Stochastic linear quadratic regulator with indefinite control weight costs,
SIAM J. Control & Optim., 36 (1998), 1685–1702 (with Shuping Chen and
Xunyu Zhou).
46. Minimum period control problem for infinite dimensional system, Chin. Ann.
Math. (Ser. B), 19 (1998), no. 1, 113–128 (with Liping Pan).
47. The equivalence between two types of exponential stabilities, Chinese Science
Bulletin, 43 (Chinese Series, 1998), no. 16, 1787–1788; 43 (English Series,
1998), no. 18, 1583–1584 (with Hanzhong Wu).
Account: ns145102
14
48. Linear quadratic problem with unbounded control in Hilbert spaces, Chinese
Science Bulletin, 43 (1998), no. 20, 1712–1717 (with Hanzhong Wu).
49. Synthesis of upper-triangular non-linear systems with marginally unstable

free dynamics using state-dependent saturation, Int. J. Control, 72 (1999),
1078–1086 (with Wei Lin).
50. Research into unexpected events: a key project of mathematical finance,

Chinese Science, vol. 51, no. 2 6–9, March, 1999 (with Zhiyuan Shen and
Jiongmin Yong; in Chinese).
51. Social responsibility of scientists, Chinese Science, vol. 52, no. 1, 24–25,
January, 2000 (with Zhiyuan Shen; in Chinese).
52. A linear quadratic problem with unbounded control in Hilbert spaces, Differ-
ential Integral Equations, 13 (2000), no. 4–6, 529–566 (with Hanzhong Wu).
53. Necessary conditions for optimal control of elliptic systems, J. Austral. Math.
Soc. (Ser. B), 41 (2000), 542–567 (with Hang Gao).
IV. Conference Proceedings Papers
1. On optimal control for distributed parameter systems, Proc. of 8th IFAC

World Congress, Kyoto, Japan, 1981, 207–212 (with Yunlong Yao).
2. Time optimal control of distributed Parameter systems (in Chinese),

Proc. National Exchange Meeting on Control Theory and Applied Mathe-
matics, Science Press, Beijing, 1981, 210–211 (with Yunlong Yao).
3. Some problems on optimal control theory and computer control, Proc.

National Exchange Meeting on Control Theory and Applied Mathe-
matics, Science Press, Beijing, 1981, 230–232 (in Chinese).
4. Vector-valued measure and the necessary conditions for the optimal con-
trol problems of linear systems, Proc. IFAC 3rd Symposium on Control of
Distributed Parameter Systems, Toulouse, France, 1982, 503–506.
5. On the stability of nonlinear control systems with time-lags, Proc. 1983

Beijing Symposium on Differential Geometry and Differential Equations,
Science Press, Beijing, China, 1986, 477–480.
6. Bounded real lemma and stability of the time lag nonlinear control systems,
Proc. 9th IFAC World Congress, Budapest, Hungary, 1984, 67–71.
Account: ns145102
15
7. Quadratic optimal control for generalized linear stochastic control sys-

tems, Preprint of First China-Japan Symposium on System Control & Its
Appl., Hangzhou, China, 1989 (with Ying Hu and Shige Peng).
8. Robust design of regulator for controlled systems via LQ differential games,

Preprint of 2nd Japan-China Joint Symposium on System control Theory
& its Appl., Osaka, Japan, 1990 (with Shuping Chen).
9. Optimal control for a class of distributed parameter systems, Proc. 29th

CDC, Honolulu, USA, 1990, 2319–2320 (with Jiongmin Yong).
10. Optimality conditions for controls of infinite dimensional systems, Proc.

1st World Congress on Nonlinear Analysis, Tampa, Florida, USA, August,
1992.
11. Some Advances in the Theory of Optimal Control, Proc. National Annual
Meeting of Control Theory and Applications, 1992 (in Chinese).
12. On stochastic linear controlled systems, SIAM Conference on Control and

Appl., Minneapolis, USA, Sept. 1992 (with Shuping Chen, Shige Peng, and
Jiongmin Yong).
13. Works on Control Science of Fudan University, Proc. Annual Meeting of

Shanghai Society of Automation, 1992 (with Jiongmin Yong; in Chinese).
14. Contributions to the Theory of Optimal Stochastic Controls, Differential

Equations and Control Theory (Wuhan, 1994), 169-175, Lecture Notes in
Pure & Appl. Math., vol. 176 (1996), Marcel Dekker (with Shanjian Tang).
15. Necessary conditions for optimal control of infinite dimensional systems,

Proc. 1994 Hong Kong International Workshop on New Directions of Con-
trol and Manufacturing, Hong Kong, 1994, 214–221.
16. A Class of Non-conventional Stochastic Linear Quadratic Regulators, Proc.

2nd Asia Control Conference, 1997, 1-317-320 (with Shuping Chen and
Xunyu Zhou).
17. Necessary Conditions for Optimal Controls of Infinite Dimensional Systems,

Proc. 2nd Asia Control Conference, 1997, 1429–432.
Account: ns145102
16
18. Global Stabilization of Nonlinear Systems with Marginally Unstable Free

Dynamics by Small Controls, Proc. of 1998 American Control Conference,
Philadelphia, 1998, 303–307 (with Wei Lin).
19. The Linear quadratic Optimal Control in Hilbert Spaces with Unbounded
Controls, Proc. 14th World Congress of IFAC, vol. F, 1999, 121–124.
20. Optimal Control Theory: from Finite Dimensions to Infinite Dimensions,

Control of Distributed Parameter and Stochastic Systems (Proc. of the
IFIP WG7.2 International Conference on Control of Distributed Parame-
ter and Stochastic Systems, June 19–22, 1998), (eds.: Shuping Chen, Xun-
jing Li, Jiongmin Yong and Xun Yu Zhou), Kluwer Academic Publishers,
Boston, 1999, 85–94.
Account: ns145102
July 27, 2007 12:2 Proceedings Trim Size: 9.75in x 6.5in Chap12-StudentsLi-ModifiedYong
STUDENTS AND POST-DOCTORS ADVISED BY XUNJING LI
I. Post-doctors.
1. Shige Peng
Final research report: General Stochastic Maximum Principle, Backward
Stochastic Differential Equations, and Application of Singular Perturbation
in Annealing Simulation and Neutral Networks, (March, 1988—November,
1989).
2. Daode Gao
Final research report: Estimation of Potential Taxation, (June, 1996—June,
1998).
3. Aiguo Kong
Final research report: Research into Capital Structure of Financial Markets,
(December, 1996—July, 1998).
4. Jiang Yu
Final research report: Existence of Discrete Breather for a Class of Infinite
Dimensional Coupled Oscillators, (June, 1999–June, 2001).
II. Doctors.
1. Shuping Chen Riccati Equation: a New Approach, and Its Applications,

Ph.D. thesis, 1985.
2. Xunyu Zhou, Maximum Principle, Dynamic Programming and Their Rela-

tionship in Optimal Control Theory, Ph.D. thesis, July, 1989.
3. Ying Hu, Maximum Principle for Optimal Control of Stochastic Systems,

Ph.D. thesis, July, 1989.
4. Liping Pan, Optimal Control of Distributed Parameter Systems with Time

Lags, and Infinite Dimensional Leader-Follower Differential Games, Ph.D.
thesis, March, 1991.
17
Account: ns145102
18
5. Kangsheng Liu, Stabilization and Local Boundary Control of Distributed

Parameter Systems, Ph.D. thesis, December, 1991.
6. Jianxiong Huang, Bifurcation and Chaos of Homoclinic Orbits in Infinite

Dimensional Dynamical Systems, Ph.D. thesis, July, 1992.
7. Shanjian Tang, Optimal Control in Hilbert Space of Stochastic Systems with

Random Jumps, Ph.D. thesis, January, 1993.
8. Pingjian Zhang, Differential Riccati Equations with Uncertain Coefficients

and Their Applications in H ∞ -Optimization and L-Q Games, Ph.D. thesis,
1994.
9. Hang Gao, Domain Optimization and Optimal Control of Systems Governed

by Elliptic Equations, Ph.D. thesis, March, 1996.
10. Hanzhong Wu, The Linear Quadratic Optimal Control Problem with
Unbounded Control in Hilbert Spaces, Ph.D. thesis, July, 1998.
11. Qihong Chen, Indirect Obstacle Optimal Control Problem for Variational
Inequalities, Ph.D. thesis, July, 1999.
12. Shangwei Zhu, Two Problems in Optimal Control Theory and Applications,
Ph.D. thesis, July, 2005.
III. Masters.
1. Yuncheng You, Optimal Control of Linear Systems in Abstract Spaces with

Indefinite Quadratic Criteria, Master thesis, January, 1981.
2. Guozhu Gao, Uniform Asymptotic Stability of Functional Differential

Equations of Neutral Type, Master thesis, July, 1981.
3. Jin Ma, On Infinite-Time State Estimation, Master thesis, July, 1984.
4. Yinping Wang, Convergence of Self-Tuned Regulators, Master thesis, July,

1985.
5. Huiheng Zheng, Stability of Nonlinear Control Systems with Time Lags,

Master thesis, July, 1986.
Account: ns145102
19
6. Yan Qin, Singular Perturbation Approach to Computation of Optimal

Control Problems, Master thesis, July, 1987.
7. Yuan Zhou, Singular Quadratic Optimal Control of Infinite Dimensional

Linear Systems, Master thesis, July, 1987.
8. Hong Xu, Necessary Conditions for Optimal Control of Distributed Param-

eter Systems, Master thesis, July, 1988.
9. Chunjiang Qian, Positive Real Lemma and Linear Quadratic Optimal

Control of Time Delayed Systems, Master thesis, July, 1994.
10. Guanghui Li, The Positive Real Lemma and Absolute Stability for Neutral
Differential-Difference Equation, Master thesis, July, 1994.
11. Qi Zhou, High-Order Necessary Conditions for Singular Optimal Control

of Distributed Parameter Systems, Master thesis, March, 1995.
12. Shihong Wang, Averaging of Hamilton-Jacobi Equation in Infinite Dimen-

sions and Application, Master thesis, July, 1997.
13. Zuoyi Zhou, H∞ Control Problem for Nonlinear Systems, Master thesis,
July, 1997.
Account: ns145102

Account: ns145102
July 27, 2007 12:5 Proceedings Trim Size: 9.75in x 6.5in Chap13-PartTwoTitle
PART TWO
STOCHASTIC CONTROL,
MATHEMATICAL FINANCE,
AND BACKWARD STOCHASTIC
DIFFERENTIAL EQUATIONS
21
Account: ns145102
July 27, 2007 12:5 Proceedings Trim Size: 9.75in x 6.5in Chap13-PartTwoTitle

Account: ns145102
July 31, 2007 14:52 Proceedings Trim Size: 9.75in x 6.5in Chap14-BaoTang-WithProofs
AXIOMATIC CHARACTERISTICS FOR SOLUTIONS OF

REFLECTED BACKWARD STOCHASTIC DIFFERENTIAL
EQUATIONS∗
XIAOBO BAO
Institute of Mathematics,
School of Mathematical Sciences,
Fudan University, Shanghai 200433, China
& Key Laboratory of Mathematics for Nonlinear Sciences,
(Fudan University), Ministry of Education, China
SHANJIAN TANG
Department of Finance and Control Sciences,
Fudan University, Shanghai 200433, China,
& Key Laboratory of Mathematics for Nonlinear Sciences,
(Fudan University), Ministry of Education, China
E-mail: sjtang@fudan.edu.cn
In this paper, we introduce the notion of an {Ft , 0 ≤ t ≤ T }-consistent dynamic operator

with a floor in terms of four axioms. We show that an {Ft , 0 ≤ t ≤ T }-consistent dynamic
operator {Es,t , 0 ≤ s ≤ t ≤ T } with a continuous upper-bounded floor {St , 0 ≤ t ≤ T },
is necessarily represented by the solutions of a backward stochastic differential equation
reflected upwards on the floor {St , 0 ≤ t ≤ T }, if it is E µ -super-dominated for some
µ > 0 and if it has the non-increasing and floor-above-invariant property of forward
translation.
Keywords: American contingent claims, reflected BSDE, dynamic nonlinear evalua-

tion, filtration-consistent nonlinear expectation, nonlinear Doob-Meyer decomposition.
2000 Mathematics Subject Classification: Primary 60H10; Secondary 60H30,

60A05, 49N90, 91B30, 91B24.
1. Introduction
Let {Bt , 0 ≤ t ≤ T } be a d-dimensional standard Brownian motion defined on

a probability space (Ω, F, P ). Let {Ft , 0 ≤ t ≤ T } be the natural filtration of
∗ Thiswork is partially supported by the NSF of China under Grant No. 10325101 (distinguished
youth foundation), the National Basic Research Program of China (973 program) with Grant
No. 2007CB814904, and the Chang Jiang Scholars Program.
23
Account: ns145102
24
{Bt , t ∈ [0, T ]}, augmented by all P -null sets of F. Set

L2 (FT ) := {ξ : ξ is an FT -measurable random variable s.t. E|ξ|2 < +∞},

L2F (0, T ; Rm) := φ : φ is Rm -valued and {Ft , 0 ≤ t ≤ T }-adapted s.t.
Z T
E |ϕt |2 dt < ∞ ,
0
n o
H2 := ϕ ∈ L2F (0, T ; R) : ϕ is predictable ,
n o
S2 := ϕ ∈ H2 : ϕ is a continuous process s.t. E max |ϕt |2 < +∞ .
0≤t≤T
2 2 2
For ξ ∈ L (FT ), denote by L (FT ; ξ) the subset of L (FT ) whose elements are not
less than ξ. We shall use L2+ (FT ) to stand for L2 (FT ; 0).
Consider the following reflected backward stochastic differential equation (RB-
SDE):
 Z T Z T


 Ys = ξ + g(r, Yr , Zr )dr + KT − Ks − hZr , dBr i, 0 ≤ s ≤ T;
s sZ (1)
T


 Ys ≥ Ss , a.s.0 ≤ s ≤ T ; K0 = 0 and (Yr − Sr )dKr = 0.
0
Here, the three objects are previously given: a terminal value ξ, a random field
g : Ω × [0, T ] × R × Rd −→ R, and a continuous progressively measurable real-valued
random process {St , 0 ≤ t ≤ T }. Assume that
(C1) ξ ∈ L2 (FT ).
(C2) g(·, y, z) ∈ H2 for (y, z) ∈ R × Rd .
(C3) |g(t, y, z) − g(t, y 0 , z 0 )| ≤ α(|y − y 0 | + |z − z 0 |) a.s. with y, y 0 ∈ R and
z, z ∈ Rd for some positive constant α. And
0
(C4) S + ∈ S2 and ST ≤ ξ a.s..

The solution to RBSDE (1) is a triple {(Yt , Zt , Kt ), 0 ≤ t ≤ T } of {Ft , 0 ≤
t ≤ T }-progressively measurable processes taking values in R × Rd × R+ such that
(i) Z ∈ H2 , and (ii) {Kt , 0 ≤ t ≤ T } is continuous and increasing. In view of El
Karoui et al., 5 we know that there exists unique solution {(Yt , Zt , Kt ), 0 ≤ t ≤ T }
of RBSDE (1) if the four conditions (C1)–(C4) are satisfied. Define
r;g,S
Et,T [ξ] := Yt , ∀ ξ ∈ L2 (FT ; ST )
and
r,g r;g,0
Et,T [ξ] := Et,T [ξ], ∀ ξ ∈ L2+ (FT ).
Here, the superscript r indicates that the underlying operators are generated by
a reflected BSDE, and the superscripts (g, S) specify the generator and the obsta-
cle of the underlying RBSDE. El Karoui and Quenez 8 argued that the operators
r;g,S
{Es,t , 0 ≤ s ≤ t ≤ T } introduce a nonlinear pricing system for square-integrable
American contingent claims, possessing the following properties:
Account: ns145102
25
(A1) Monotonicity. Es,t [Y ] ≥ Es,t [Y 0 ] if Y, Y 0 ∈ L2 (Ft ; St ) such that Y ≥ Y 0 .

(A2) Time consistency. Er,s [Es,t [Y ]] = Er,t [Y ] if r ≤ s ≤ t ≤ T and Y ∈
L2 (Ft ; St ).
A financial theoretically interesting problem is the converse one: Is a dynamic
operator that possesses similar properties like (A1) and (A2), necessarily repre-
sented by a RBSDE? That is, we are concerned with the axiomatic characteristics
for a RBSDE.
Throughout the paper, we make the following assumption (denoted by (H0)):
(H0) The predictable process S is continuous, S + ∈ S2 , and there is a positive

constant C such that ess sup0≤t≤T St ≤ C a.s..
The process S is given, either as the (lower) obstacle of the underlying RBSDE or
as the floor of the underlying dynamic operator (see Definition (1.1) below). In this
paper, we shall formulate and discuss the above converse problem for RBSDE (1)
with a given general upper-bounded floor S > −∞. More precisely, we introduce
the notion of a dynamic operator with a floor S, and then represent the dynamic
operator by a BSDE reflected upwards on the floor S. In this way, we characterize
on one hand the solutions of RBSDEs in terms of axioms, and on the other hand,
we give a representation for the dynamic operator in terms of RBSDEs (1).
We first introduce the following notion of an {Ft , 0 ≤ t ≤ T }-consistent dynamic
operator with floor S.
Definition 1.1. A time parameterized system of operators
Es,t [·] : L2 (Ft ; St ) −→ L2 (Fs ; Ss ), 0≤s≤t≤T (2)
is called an {Ft , 0 ≤ t ≤ T }-consistent dynamic operator with floor S if it satisfies
the following: Es,t [Y ] is continuous in (s, Y ) ∈ [0, t] × L2 (Ft ; St ) for t ∈ [0, T ], and
furthermore, it satisfies the following four axioms.
(D1) Floor-above strict monotonicity Es,t [Y ] ≥ Es,t [Y 0 ] a.s. for Y and
Y ∈ L2 (Ft ; St ) such that Y ≥ Y 0 a.s.. If Es,t [Y ] > Ss a.s. for any s ∈ [r, t] and
0
some r ∈ [0, t], then Y 0 = Y a.s. if Y 0 ≥ Y a.s. and Er,t [Y 0 ] = Er,t [Y ] a.s..
(D2) Es,t [Y ] = Y a.s. for each Y ∈ L2 (Fs ; C).
(D3) Time consistency Er,s [Es,t [Y ]] = Er,t [Y ] a.s. for Y ∈ L2 (Ft ; C) if
r ≤ s ≤ t ≤ T.
(D4) Zero-one law For each s ≤ t, Es,t [1A Y + C̃] − C̃ = 1A (Es,t [Y + C̃] − C̃),
a.s., ∀A ∈ Fs for any constant C̃ dominating the floor.
The main result of the paper is stated as follows.
Theorem 1.1. Consider an {Ft , 0 ≤ t ≤ T }-consistent dynamic operator {Es,t , 0 ≤

s ≤ t ≤ T } with the floor S satisfying assumption (H0). We make the following
two assumptions (H1) and (H2):
(H1) E µ -super-domination. There is some µ > 0 such that
µ
Et,T [X + Y ] − Et,T [X] ≤ Et,T [Y ], a.s. (3)
Account: ns145102
26
µ
for t ∈ [0, T ], X ∈ L2 (FT ; ST ) and Y ∈ L2+ (FT ). Here, Et,T [Y ] := yt with {yt , 0 ≤
t ≤ T } being the first component of the solution of the following BSDE:
Z T Z T
ys = Y + µ|zr |dr − hzr , dBr i, 0 ≤ s ≤ T. (4)
s s
(H2) The non-increasing and floor-above-invariant property of forward trans-

lation. Et,T [X + Y ] ≤ Et,T [X] + Y , a.s. for X ∈ L2 (FT ; ST ) and Y ∈ L2+ (Ft ).
Furthermore, for X ∈ L2 (FT ; ST ) such that Et,T [X] > St a.s. for t ∈ [0, T ], We
have Et,T [X + Y ] = Et,T [X] + Y for any Y ∈ L2+ (Ft ).
Then, there is a random field g : Ω × [0, T ] × Rd → R such that the following are
satisfied:
(i) g(t, 0) = 0 for a.e. t ∈ [0, T ],
(ii) |g(t, z1 ) − g(t, z2 )| ≤ µ|z1 − z2 |, and
r;g,S
(iii) Es,t (Y ) = Es,t (Y ) for any Y ∈ L2 (Ft ; St ) and s ∈ [t, T ] with t ∈ [0, T ].
For the particular case of the zero floor (S = 0), the last assertion is still true
if the assumption (H1) is replaced by the weaker one (H1)’:
(H1)’ E µ -domination. There is some µ > 0 such that
µ
E0,T [X + Y ] − E0,T [X] ≤ E0,T [Y ], for any X ∈ L2 (FT ; ST ) and Y ∈ L2+ (FT ). (5)
Note that the assumption (H1)’ is much weaker than that of E µ -domination used
by Coquet,et al., 3 in that Y is here restricted within L2+ (FT ) instead of taking
values in the whole space L2 (FT ) like the latter. This difference will complicate our
subsequent arguments.
Remark 1.1. For the general case of the upper bounded floor S, the primal dy-
C
namic operator {Es,t , 0 ≤ s ≤ t ≤ T } deduces the following new one {Es,t ,0 ≤ s ≤
t ≤ T} :
C
Es,t [·] : L2 (Ft ; St − C) −→ L2 (Fs ; Ss − C), 0 ≤ s ≤ t ≤ T (6)
with
C
Es,t [X] := Es,t [X + C] − C, ∀X ∈ L2 (Ft ; St − C). (7)
C
It is easy to prove that Es,t [X] for X ∈ L2 (Ft ; St − C), satisfies all the conditions
(H1), (H2), and (D1)–(D4). Then, it follows immediately that if Theorem 1.1 is
true for the negative floor S − C, it is also true for the general upper bounded floor
S. Hence, it is sufficient to prove Theorem 1.1 for the case of the negative floor.
Even in the simpler case of negative floors, two key points are still worth to
be mentioned here for the proof of Theorem 1.1. One is to make full use of the
non-increasing and floor-above-invariant assumption (H2) of forward translation
and the assumption (H1) of E µ -super-domination for some µ > 0, to extend the
underlying {Ft , 0 ≤ t ≤ T }-consistent dynamic operator from the subset L2 (FT ; ST )
of floor-dominating square-integrable random variables to the whole space L 2 (FT )
Account: ns145102
27
of square-integrable random variables. The extended dynamic operators are shown

to be identified to an {Ft , 0 ≤ t ≤ T }-consistent expectation E. e The generator
g of its BSDE representation given by Theorem 2.3 turns out to be that of the
desired RBSDE. The second key point is the following observation: the process
e
{Et,T [X], 0 ≤ t ≤ T } turns out to be an E-supermartingale for each X ∈ L2 (FT ; ST ).
This fact allows us to apply a nonlinear Doob-Meyer’s decomposition theorem in
Ref. 3, to give the increasing process of {Et,T [X], 0 ≤ t ≤ T } as a solution to some
BSDE reflected upwards on the floor. Eventually, our proof of Theorem 1.1 is both
natural and elegant. To diverse the difficulty in the above arguments, the whole
proof is divided into Sections 2 and 3. The above two key points are exposed in
detail separately in these two sections.
The main result of this work has been announced in Bao and Tang 1 . In Bao
and Tang 2 , we study a dynamic operator with a very general continuous floor
S which may be unbounded from the above. We define a dynamic operator as a
stopping-times parameterized system of operators, and also give the representation
by RBSDEs of a dynamic operator.
The rest of the paper is organized as follows. In Section 2, we consider the case
of the zero floor. We concentrate our attention to show how to extend the family of
dynamic operators defined on the subset L2+ (FT ) of L2 (FT ) to an {Ft , 0 ≤ t ≤ T }-
expectation, which is defined on the whole space L2 (FT ). Section 3 is devoted to the
case of the negative floor. Restricting the underlying {Et,T , 0 ≤ t ≤ T }-consistent
dynamic operator to L2+ (FT ), we get an {Ft , 0 ≤ t ≤ T }-expectation by extending
the restriction to L2 (FT ) in the way as shown in the preceding section. The {Ft , 0 ≤
t ≤ T }-expectation gives the generator g of an RBSDE by Theorem 2.3. In addition,
we have to give the amount {KtX , 0 ≤ t ≤ T } to push upwards for {Et,T [X], 0 ≤ t ≤
T } with X ∈ L2 (FT ; ST ), that is, the increasing process in relevant to an RBSDE.
For this purpose, a nonlinear Doob-Meyer’s decomposition theorems is shown how
to be used. The key point is to observe that the process {Et,T [X], 0 ≤ t ≤ T } turns
e
out to be an E-supermartingale for each X ∈ L2 (FT ; ST ).
2. The Case of the Zero Floor

In the case of the zero floor, the properties (D2) and (D4) read
(D2)’ Es,t [Y ] = Y a.s. for Y ∈ L2+ (Fs ) and s ∈ [0, t].
(D4)’ Zero-one law Es,t [1A Y ] = 1A Es,t [Y ], a.s. for s ∈ [0, t], A ∈ Fs , and
Y ∈ L2+ (Ft ).
From (D1), (D2)’, (H1) and (H2), we can give the following:
(H2)’ We have for t ∈ [0, T ], X ∈ L2+ (FT ) and Y ∈ L2+ (Ft ),

Et,T [X + Y ] = Et,T [X] + Y a.s. .
In fact, consider X ∈ L2+ (FT ) and Y ∈ L2+ (Ft ). It follows from (D2)’ that
Et,T [] = for t ∈ [0, T ] and very constant > 0. Therefore, from (D1), we deduce
that Et,T [ + X] ≥ Et,T [] = > 0 a.s. for t ∈ [0, T ]. Consequently, by (H2), we
Account: ns145102
28
conclude that
Et,T [ + X + Y ] = Et,T [ + X] + Y a.s. . (8)
By (H1), we can take the limit → 0, and we have the desired equality.
Remark 2.1. Assume that conditions (C2) and (C3) are satisfied. Moreover,
assume that g(·, ·, 0) ≡ 0. Let (y, z) be the adapted solution to the following BSDE:
Z T Z T
Yt = ξ + g(s, Ys , Zs )ds − hZs , dBs i, 0 ≤ t ≤ T. (9)
t t
with the terminal condition ξ ∈ L2+ (FT ).

Then the triple (y, z, 0) is the adapted
solution of RBSDE (1) with the parameter (Y, g, 0). This shows that the adapted
solution of RBSDE (1) gives the one of the corresponding BSDE (9) for any terminal
value ξ ∈ L2+ (FT ) since the obstacle {St , 0 ≤ t ≤ T } is not active in this case.
The following lemma is immediate and will be used later.
Lemma 2.1. We have

µ
(i) E Et,T [X]p ≤ exp(2−1 p(p − 1)−1 µ2 (T − t))E[X p ] for all µ > 0, t ∈ [0, T ],
p
and X ∈ L (FT ), with p ∈ (1, 2];
µ
−µ
(ii) Et,T [X + Y ] = Y − Et,T [−X] a.s. for all µ > 0, t ∈ [0, T ], X ∈ L2 (FT ), and
2
Y ∈ L (Ft ); And
µ
(iii) E0,T −µ
[·] and E0,T [·] are strongly continuous in L2 (FT ).
Proof. We only prove the first assertion (i). The other two assertions are easy
to see.
µ
For simplicity of notations, set yt := Et,T [X]. Then, by definition, there is unique
2 d
z ∈ LF (0, T ; R ) such that (y, z) is the unique adapted solution of BSDE (4). Using
Itô’s formula, we have
Z T Z T
1
E|yt |p + p(p−1) |ys |p−2 |zs |2 ds = E|X|p +2µp |ys |p/2 |ys |p/2−1 |zs | ds. (10)
2 t t
Since
p 1
µp|ys |p/2 |ys |p/2−1 |zs | ≤ µ2 |ys |p + p(p − 1)|ys |p−2 |zs |2 , (11)
2(p − 1) 2
we have
Z T
p
E|yt |p ≤ E|X|p + µ2 |ys |p ds. (12)
2(p − 1) t
The standard arguments of using Gronwall’s inequality then gives the desired in-
equality.
Remark 2.2. See Coquet et al. 3 for the detailed proof of the first assertion in
the case of p = 2, which is easy and standard.
Account: ns145102
29
In view of Remark 2.1, from Proposition 3.6, page 164, Lemma 3.2 and Propo-
sition 3.7, page 165 of Ref. 12 concerning the properties of the solutions to BSDEs,
we have the following
Theorem 2.1. Assume that conditions (C2) and (C3) are satisfied. Moreover,
r,g
assume that g is independent of y and g(·, ·, 0) ≡ 0. Then, {Es,t ,0 ≤ s ≤ t ≤ T}
satisfies (H1), (D1), (D2)’, (D3), (D4)’ and (H2)’. Therefore, it is an {F t , 0 ≤ t ≤
T }-consistent dynamic operator with the zero floor.
Roughly speaking, Theorem 2.1 asserts that an RBSDE with the obstacle being
zero introduces an {Ft , 0 ≤ t ≤ T }-consistent dynamic operator. In what follows, we
shall consider the converse problem. That is, we shall associate an {Ft , 0 ≤ t ≤ T }-
consistent dynamic operator with the zero floor to a BSDE reflected on the zero
floor. For this purpose, we establish the following six preliminary lemmas. First,
we introduce some notations.
Definition 2.1. For an {Ft , 0 ≤ t ≤ T }-consistent dynamic operator {Es,t [·], 0 ≤

s ≤ t ≤ T }, define the system of operators E[·|Ft ] : L2+ (FT ) −→ L2+ (Ft ), 0 ≤ t < T
by
E[Y |Ft ] := Et,T [Y ], a.s. for Y ∈ L2+ (FT ) (13)
2
and the nonlinear functional E[·] : L (FT ) → R by
E[Y ] := E0,T [Y ] for Y ∈ L2+ (FT ). (14)
The two notations E[·] and E[·|Ft ] behave in L2+ (FT ) exactly like an {Ft , 0 ≤
t ≤ T }-consistent expectation and its conditional {Ft , 0 ≤ t ≤ T }-consistent expec-
tation on Ft . The only difference lies in the domains of variables: the former’s are
L2+ (FT ), while the latter’s are L2 (FT ). This can be seen from the following obvious
lemma.
Lemma 2.2. For an {Ft , 0 ≤ t ≤ T }-consistent dynamic operator {Es,t [·], 0 ≤ s ≤

t ≤ T }, E[·] and E[·|Ft ] have the following properties (E1)’-(E3)’.
(E1)’ E[Y1 ] ≤ E[Y2 ] if Y1 , Y2 ∈ L2+ (FT ) and Y1 ≤ Y2 . Furthermore, if Y1 ≤ Y2 ,
then Y1 = Y2 if E[Y1 ] = E[Y2 ].
(E2)’ E[C] = C for any constant C e > 0. And
(E3)’ For any Y ∈ L2+ (FT ), there is unique η ∈ L2+ (Ft ) such that
E[η1A ] = E[Y 1A ], ∀A ∈ Ft , (15)
which is equal to E[Y |Ft ]. Therefore, E[Y |Ft ] can be viewed as the E-expectation
of Y ∈ L2+ (FT ) conditioned on Ft for any t ∈ [0, T ], though E is in general not an

{Ft , 0 ≤ t ≤ T }-consistent expectation at all.
Lemma 2.3. Assume that E[Y |Ft ] satisfies conditions (H1)’ for µ > 0. Then,
1
|E[ξ1 ] − E[ξ2 ]| ≤ exp( µ2 T )||ξ1 − ξ2 ||L2 , ∀ξ1 , ξ2 ∈ L2+ (FT ). (16)
2
Account: ns145102
30
Therefore, E[·] is strongly continuous in L2 (FT ).
Proof. Using (D1) and (H1)’, we have

E[ξ1 ] − E[ξ2 ] ≤ E[|ξ1 − ξ2 | + ξ2 ] − E[ξ2 ]
≤ E µ [|ξ1 − ξ2 |].
Using Lemma 2.1, we have
E µ [|ξ1 − ξ2 |]2 ≤ exp (µ2 T )E[|ξ1 − ξ2 |2 ]. (17)
Therefore,
1 1
E[ξ1 ]−E[ξ2 ] ≤ exp( µ2 T )(E|ξ1 −ξ2 |2 )1/2 = exp( µ2 T )||ξ1 −ξ2 ||L2 , ∀ξ1 , ξ2 ∈ L2+ (FT ).
2 2
(18)
Identically, we can show
1
E[ξ2 ] − E[ξ1 ] ≤ exp( µ2 T )||ξ1 − ξ2 ||L2 , ∀ξ1 , ξ2 ∈ L2+ (FT ). (19)
2
The proof is then complete.
Lemma 2.4. (E µ -domination) Assume that E[Y |Ft ] satisfies (H1)’ and (H2)’.
Then, we have
E −µ [Y ] ≤ E[X + Y ] − E[X] ≤ E µ [Y ], ∀X, Y ∈ L2+ (FT ). (20)
Proof. In view of (H1)’, it is sufficient to prove the following

E[X + Y ] − E[Y ] ≥ E −µ [Y ], ∀X, Y ∈ L2+ (FT ). (21)
If Y ≤ n a.s. for some integer n, then using (H2)’, (H1)’, and Lemma 2.1, we
have
n − (E[X + Y ] − E[X]) = n + E[X] − E[X + Y ] = E[X + n] − E[X + Y ]
≤ E µ [n − Y ] ≤ n − E −µ [Y ].
Therefore,
E[X + Y ] − E[X] ≥ E −µ [Y ]. (22)
In general, consider Y ∈ L2+ (FT ). Define Yn := Y 1{Y ≤n} . Then, Yn converges
to Y strongly in L2 (FT ). The above arguments show that
E[X + Yn ] − E[X] ≥ E −µ [Yn ].
Then using Lemma 2.3, the desired result (21) follows by passing to the limit n → ∞
in the last inequality.
Remark 2.3. The above proof of Lemma 2.4 is more complicated than that of
Ref. 3 since both assumptions (H1)’ and (H2)’ are weaker.
Account: ns145102
31
ζ
For ζ ∈ L2+ (FT ), define the operator Es,t [·] : L2+ (Ft ) → L2 (Fs ) by
ζ
Es,t [X] := Es,t [X + ζ] − Es,t [ζ], ∀X ∈ L2+ (FT ). (23)
Identically as in the case of an {Ft , 0 ≤ t ≤ T }-expectation (see Ref. 3 for
detailed proof), we can show the following lemma.
Lemma 2.5. Let ζ ∈ L2+ (FT ). If {Es,t [·], 0 ≤ s ≤ t ≤ T } is an {Ft , 0 ≤ t ≤ T }-

consistent dynamic operator defined on L2+ (FT ) and satisfies (H1)’ and (H2)’, then
ζ
the operator {Es,t [·], 0 ≤ s ≤ t ≤ T } is also an {Ft , 0 ≤ t ≤ T }-consistent dynamic
operator defined on L2+ (FT ), and satisfy (H1)’ and (H2)’. The expectation E ζ [X|Ft ]
of X ∈ L2+ (FT ) conditioned on Ft is given by the formula:
E ζ [X|Ft ] = E[X + ζ|Ft ] − E[ζ|Ft ]. (24)
Lemma 2.6. Assume that the two F-consistent dynamic operators E 1 [·] and E 2 [·]
defined on L2+ (FT ) satisfy (H1)’ and (H2)’. If
E 1 [X] ≤ E 2 [X], ∀X ∈ L2+ (FT ),
then for all t,
E 1 [X|Ft ] ≤ E 2 [X|Ft ], a.s. for all X ∈ L2+ (FT ).
Proof. The proof is divided into the two steps.

Step 1. The case of X ≤ n for some positive integer n. Set
η = E 2 [X|Ft ] − E 1 [X|Ft ].
Then
−η1{η≤0} = E 1 [X1{η≤0} |Ft ] − E 2 [X1{η≤0} |Ft ] ≥ 0
and
n ≤ E 1 [−η1{η≤0} + n]
= E 1 [E 1 [X1{η≤0} |Ft ] − E 2 [X1{η≤0} |Ft ] + n]
= E 1 [X1{η≤0} − E 2 [X1{η≤0} |Ft ] + n]
≤ E 2 [X1{η≤0} − E 2 [X1{η≤0} |Ft ] + n]
= E 2 [E 2 [X1{η≤0} − E 2 [X1{η≤0} |Ft ] + n|Ft ]]
= E 2 [E 2 [X1{η≤0} |Ft ] + n − E 2 [X1{η≤0} |Ft ]]
= n.
Thus,
E 1 [−η1{η≤0} + n] = n.
From the floor-above strict monotonicity (D1), we have
−η1{η≤0} = 0.
Account: ns145102
32
That is, P ({η ≤ 0}) = 0, a.s.. Thus,
E 1 [X|Ft ] ≤ E 2 [X|Ft ], a.s..
Step 2. The general case of X ∈ L2+ (FT ).

For X ∈ L2+ (FT ), define the truncation Xn := X1{X≤n} . Then it follows from
Lemma 2.3 that
lim E 1 [Xn ] = E 1 [X].

n→∞
From Step 1, we have
E 1 [Xn |Ft ] ≤ E 2 [Xn |Ft ], n = 1, 2, . . . .
Obviously, it is sufficient to prove the following
lim E 1 [Xn |Ft ] = E 1 [X|Ft ], ∀X ∈ L2+ (FT ) (25)

n→∞
We now prove it by contradiction.

Otherwise, there exists 0 ≤ ≤ 1 and A ∈ Ft such that P (A) > 0 and
E 1 [Xn |Ft ]1A ≤ (E 1 [X|Ft ] − )1A .
Since E 1 [Xn + 1|Ft ] = E 1 [Xn |Ft ] + 1 and E 1 [X + 1|Ft ] = E 1 [X|Ft ] + 1 (by (H2)’),
we have
E 1 [Xn + 1|Ft ]1A ≤ (E 1 [X + 1|Ft ] − )1A .
Using (D1), we have E 1 [E 1 [Xn + 1|Ft ]1A ] ≤ E 1 [(E 1 [X + 1|Ft ] − )1A ]. Then letting
n → ∞, we have
lim E 1 [E 1 [Xn + 1|Ft ]1A ] ≤ lim E 1 [(E 1 [X + 1|Ft ] − )1A ]

n→∞ n→∞
< E 1 [(X + 1)1A ]. (26)
While we have by Lemma 2.1 the following
lim E 1 [E 1 [Xn + 1|Ft ]1A ] = lim E 1 [(Xn + 1)1A ]

n→∞ n→∞
= E 1 [(X + 1)1A ]. (27)
This contradicts (26). Therefore, (25) is true.

Combining Lemmas 2.6 and 2.4, we have
Lemma 2.7. Let {Es,t [·], 0 ≤ s ≤ t ≤ T } be an F-consistent dynamic operator

defined on L2+ (FT ) and satisfy (H1)’ and (H2)’. Then, for each t ≤ T , we have
E −µ [Y |Ft ] ≤ E[X +Y |Ft ]−E[X|Ft ] ≤ E µ [Y |Ft ], a.s. for all X, Y ∈ L2+ (FT ). (28)
Remark 2.4. Lemma 2.7 shows that (H1)’ together with (H2)’ and (D1)–(D4)
implies (H1), as pointed in the introduction.
Account: ns145102
33
In the following, we extend the underlying {Ft , 0 ≤ t ≤ T }-consistent dynamic

operator from the subset L2+ (FT ) of nonnegative square-integrable random variables
to the whole space L2 (FT ) of square-integrable random variables. As a preliminary,
we have the following fact.

defined on L2+ (FT ) and satisfy (H1)’ and (H2)’. Let Xn := X1{X≥−n} and Yn :=
E[Xn + n|Ft ] − n for X ∈ L2 (FT ) and n = 1, 2, . .p.. Then {Yn }∞n=1 is a Cauchy
sequence in L (Ft ) equipped with the norm || · || = E| · |2 . If X ∈ L2+ (FT ), then
2
Xn = X and Yn = E[X|Ft ] for n = 1, 2, . . ..
Proof. For the two positive integers m and n such that m > n, we have
Ym − Yn = E[Xm + m|Ft ] − m − (E[Xn + n|Ft ] − n)

= E[Xm + m|Ft ] − E[Xn + m|Ft ]
= E[Xn + X1{−m≤X≤−n} + m|Ft ] − E[Xn + m|Ft ]. (29)
Thus, from Lemmas 2.7 and 2.1, we have
E(Ym − Yn )2 = E(E[Xn + m|Ft ] − E[Xn + X1{−m≤X≤−n} + m|Ft ])2

≤ E(E µ [−X1{−m≤X≤−n} |Ft ])2
2
≤ eµ (T −t)
E[X 2 1{−m≤X≤−n} ]. (30)
Since X ∈ L2 (FT ), we have
lim E[X 2 1{−m≤X≤−n} ] = 0.

m,n→∞
Therefore,
lim E(Ym − Yn )2 = 0.
n,m→∞
b
Lemma 2.8 shows that E[X|F t ] introduced below is well defined for any X ∈
2
L (FT ).
Definition 2.2. For X ∈ L2 (FT ), denote
Xn := X1{X≥−n} . (31)
We introduce a dynamic operator {Ebs,t [X], X ∈ L2 (FT ); 0 ≤ s ≤ t ≤ T } by

b n |Ft ] := E[Xn + n|Ft ] − n,

E[X n = 1, 2, . . . ; (32)
and
b
E[X|F b
t ] := lim E[Xn |Ft ] (33)
n→∞
for any X ∈ L2 (FT ).
Account: ns145102
34
b
Remark 2.5. It is obvious that if X ∈ L2+ (FT ), we have E[X|F t ] = E[X|Ft ] a.s.
for t ∈ [0, T ].
b t ].
We have the following continuity of the extended operator E[·|F
Theorem 2.2. Let {Es,t [·], 0 ≤ s ≤ t ≤ T } be an F-consistent dynamic operator

defined on L2+ (FT ) and satisfy (H1)’ and (H2)’. For each t ∈ [0, T ], the conditional
b t ] is strongly continuous from L2+ (FT ) to L2 (Ft ). That
expectation operator E[·|F
is, if
lim Yn = Y, strongly in L2 (FT ),
n→∞
then
b n |Ft ] = E[Y
lim E[Y b |Ft ], strongly in L2 (Ft ). (34)
n→∞
Proof. For the two positive integers m and n such that m > n, define
b n |Ft ] − E[Yn 1{Y ≥−m} + m|Ft ] − m
δ1 := E[Y n
and
δ2 := (E[Yn 1{Yn ≥−m} + m|Ft ] − m) − (E[Y 1{Y ≥−m} + m|Ft ] − m).
Using Lemma 2.7, we have
|δ2 | ≤ |E[Y 1{Y ≥−m} + m + |Yn 1{Yn ≥−m} − Y 1{Y ≥−m} ||Ft ]
−E[Y 1{Y ≥−m} + m|Ft ]|
≤ E µ [|Yn 1{Yn ≥−m} − Y 1{Y ≥−m} ||Ft ]
≤ E µ [|Yn 1{Yn ≥−m} − Yn ||Ft ] + E µ [|Yn − Y ||Ft ] + E µ [|Y − Y 1{Y ≥−m} ||Ft ].
Thus, using Lemma 2.1, we have

E|δ2 |2 ≤ 3EE µ [|Yn 1{Yn ≥−m} − Yn ||Ft ]2 + 3EE µ [|Yn − Y ||Ft ]2
+3EE µ [|Y − Y 1{Y ≥−m} ||Ft ]2
2
≤ 3eµ (T −t)
E|Yn 1{Yn ≥−m} − Yn |2
2
+3eµ (T −t)
(E|Yn − Y |2 + E|Y − Y 1{Y ≥−m} |2 )
2 2
≤ 3eµ T
E|Yn 1{Yn <−m} |2 + 3eµ T
(E|Yn − Y |2 + E|Y 1{Y <−m} |2 ).
Letting m → ∞ in the following
b n |Ft ] − E[Y
E|E[Y b |Ft ]|2
b |Ft ]|2 )
≤ 3(E|δ1 |2 + E|δ2 |2 + E|E[Y 1{Y ≥−m} + m|Ft ] − m − E[Y

≤ C1 Eδ12 + E|Yn 1{Yn <−m} |2 + E|Yn − Y |2 + E|Y 1{Y <−m} |2

b 2
+E|E[Y 1{Y ≥−m} + m|Ft ] − m − E[Y |Ft ]| ,
Account: ns145102
35
we have
b n |Ft ] − E[Y
E|E[Y b |Ft ]|2 ≤ C1 E|Yn − Y |2 .
Therefore,
b n |Ft ] = E[Y
lim E[Y b |Ft ].
n→∞
Now consider BSDE (9), which can be viewed as the extremal case of the negative
infinite floor (that is, S = −∞) for RBSDE (1). Under conditions (C1)–(C3), it has
g
a unique adapted solution (Y, Z). In the following, denote Yt by Et,T [ξ] to emphasize
the dependence on the parameter (ξ, g) and the initial and terminal times pair (t, T ).
Coquet et al. 3 introduces the following notion of an {Ft , 0 ≤ t ≤ T }-consistent
expectation.
Definition 2.3. An {Ft , 0 ≤ t ≤ T }-consistent expectation is defined to be a

nonlinear functional E[·] on L2 (FT ) which satisfies the following three axioms (E1)–
(E3).
(E1) E[Y1 ] ≤ E[Y2 ] if Y1 , Y2 ∈ L2 (FT ) and Y1 ≤ Y2 . Furthermore, if Y1 ≤ Y2 ,
then Y1 = Y2 if E[Y1 ] = E[Y2 ].
(E2) E[C]e =C e for any constant C. e And
(E3) The E-expectation (denoted by E[Y |Ft ]) of Y ∈ L2 (FT ) conditioned on
Ft exists uniquely for any t ∈ [0, T ]. That is, there is unique η ∈ L2 (Ft ) such that
E[η1A ] = E[Y 1A ], ∀A ∈ Ft . (35)
The third axiom (E3) shows that an {Ft , 0 ≤ t ≤ T }-consistent expectation E[·]
induces an {Ft , 0 ≤ t ≤ T }-consistent dynamic operator, which will be denoted by
{Es,t [·], 0 ≤ s ≤ t ≤ T }. We shall identify them each other.
Coquet et al. 3 proved the following representation result.
Theorem 2.3. Assume that an {Ft , 0 ≤ t ≤ T }-consistent expectation {Es,t , 0 ≤

s ≤ t ≤ T } satisfies (H1)’ for some positive constant µ and the following assump-
tion:
(H2)’ Translation invariance. Es,t [X + Y ] = Es,t [X] + Y a.s. for s ≤ t, X ∈
L2 (Ft ) and Y ∈ L2 (Fs ).
Then, there is a random field g : Ω × [0, T ] × Rd → R such that
(i) g(t, 0) = 0 for a.e. t ∈ [0, T ] and g(·, z) ∈ L2F (0, T ) for z ∈ Rd ,
(ii) |g(t, z1 ) − g(t, z2 )| ≤ µ|z1 − z2 | for z1 , z2 ∈ Rd and a.e. t ∈ [0, T ], and
g
(iii) E0,T [Y ] = E0,T [Y ] for any Y ∈ L2 (FT ).
Note that an {Ft , 0 ≤ t ≤ T }-consistent expectation {Es,t , 0 ≤ s ≤ t ≤ T } has

the following four properties.
(A1) Strict Monotonicity. Es,t [Y ] ≥ Es,t [Y 0 ] if Y, Y 0 ∈ L2 (Ft ) and Y ≥ Y 0 .
Furthermore, if Y 0 ≥ Y and Es,t [Y 0 ] = Es,t [Y ], then Y 0 = Y .
(A2) Es,t [Y ] = Y a.s. for s ∈ [0, t] and Y ∈ L2 (Fs ).
Account: ns145102
36
(A3) Time consistency. Er,s [Es,t [Y ]] = Er,t [Y ] if r ≤ s ≤ t ≤ T and Y ∈

L2 (Ft ).
(A4) Zero-One law. Es,t [1A Y ] = 1A Es,t [Y ] a.s. for s ≤ t, A ∈ Fs , and Y ∈
L2 (Ft ).
Theorem 2.3 means that an {Ft , 0 ≤ t ≤ T }-consistent expectation always has
a BSDE representation if it is both E µ -dominated for some µ > 0 and constant-
preserving. Note that (H1) is stronger than (H1)’ and however, the latter together
with (H2)’ and (A1)-(A4) implies the former (see Remark 2.4 for further details).
Further, (H2)’ is stronger than (H2). Therefore, Theorem 2.3 is the special case of
S ≡ −∞ for Theorem 1.1.
The above extended dynamic operator Eb can be shown to be identified to an
b
{Ft , 0 ≤ t ≤ T }-consistent expectation E.
Theorem 2.4. Let {Es,t [·], 0 ≤ s ≤ t ≤ T } be an {Ft , 0 ≤ t ≤ T }-consistent dy-

namic operator defined on L2+ (FT ) and satisfy (H1)’ and (H2)’. Then, the dynamic
b t ] defined by (33) satisfies (H1)’, (H2)’, (D1), (D2)’, (D3), and (D4)’.
operator E[·|F
Therefore, it is an {Ft , 0 ≤ t ≤ T }-consistent expectation.
Proof. For X ∈ L2 (FT ) and Y ∈ L2 (FT ), define

Xn := X1{X≥−n} , Yn = Y 1{Y ≥−n} .
It is easy to check all the properties of Monotonicity (D1), Lemma 17, (D2)’, (D3),
(D4)’, (H2)’ and (H1)’ for Xn and Yn . From Theorem 2.2, let n → ∞, monotonicity
(D1), Lemma 17, (D2)’, (D3), (D4)’, (H2)’, and (H1)’ then follow immediately.
If Y ≥ X and E[Yb |Ft ] = E[X|F
b t ], then it follows from Lemma 17 that
b |Ft ] − E[X|F
0 = E[Y b t] ≥ E
−µ
[Y − X|Ft ] ≥ 0.
Thus,
E −µ [Y − X|Ft ] = 0.
Therefore,
Y = X.
This shows the strict monotonicity.
The generator g of the BSDE representation of the {Ft , 0 ≤ t ≤ T }-consistent
b given by Theorem 2.3 turns out to be that of the desired RBSDE.
expectation E[·]
Theorem 2.5. Theorem 1.1 is true in the case of the zero floor S ≡ 0. Moreover,
the assumption (H1) may be weakened to (H1)’.
Proof. Theorem 2.4 shows that E[X], b X ∈ L2 (FT ), defined by (33) is an

{Ft , 0 ≤ t ≤ T }-expectation, and satisfy (H1)’ and (H2)’. From Theorem 2.3,
there exists a function g : Ω × [0, T ] × Rd → R, satisfying (C2) and (C3) and the fol-
b
lowing properties: g(0, z) = 0, |g(t, z)| ≤ µ|z| for a.e. t ∈ [0, T ], and E[X] = E g [X]
2
for any X ∈ L (FT ).
Account: ns145102
37
In version of Remark 2.5 and Lemma 2.1, we have

b
E[X] = E[X] = E g [X] = E r,g [X] for anyX ∈ L2+ (FT ).
Hence, for any A ∈ Ft and X ∈ L2+ (FT ), we have
E[X1A ] = E r,g [X1A ]. (36)

This shows that E[X|Ft ] = E r,g [X|Ft ] for any X ∈ L2+ (FT ). That is, Et,T [X] =
r,g
Et,T [X] for any X ∈ L2+ (FT ).
3. The Case of the Negative Floor

Consider the dynamic nonlinear operator with a negative floor S ≤ 0.
In what follows, we discuss the properties of the solution Y to RBSDE (1) from
a view point of operators. Suppose that the generator g satisfies conditions (C2)
and (C3).
Definition 3.1. Define, for 0 ≤ s ≤ t < ∞ and ξ ∈ L2 (Ft ; St ),

r;g,S
Es,t [ξ] := ys , (37)
where (y, z, K) is the solution of RBSDE (1) with the parameter (ξ, g, S).
Theorem 3.1. Let the random field g satisfy (C2) and (C3). Moreover, assume
that g(t, y, z) does not depend on the variable y and that g(·, 0) = 0. Then if
r;g,S
St ≤ 0 a.s. for t ∈ [0, T ], the dynamic operator {Es,t , 0 ≤ s ≤ t ≤ T } satisfies
Axioms (H1), (H2), and (D1)–(D4).
Proof. Let us prove (H1) first. Consider ξi ∈ L2 (Ft ) for i = 1, 2 such that
ξ1 ≥ ξ2 ≥ ST . Let {(Ys1 , Zs1 , Ks1 ), 0 ≤ s ≤ t} and {(Ys2 , Zs2 , Ks2 ), 0 ≤ s ≤ t} be the
adapted solutions of RBSDE (1) with the values ξ = ξ1 and ξ = ξ2 at the terminal
time t, respectively. Denote by {(ys , zs ), 0 ≤ s ≤ t} the unique adapted solution
of BSDE (4) with the value ξ1 − ξ2 at the terminal time t. Note that the terminal
time is takes to be t, that is T = t in RBSDE (1) and BSDE (4). Applying Itô’s
formula to |(Ys1 − Ys2 − ys )+ |2 , and taking the expectation, we have:
Z t
1 2 + 2
E|(Ys − Ys − ys ) | + E 1{Yr1 −Yr2 ≥yr } |Zr1 − Zr2 − zr |2 dr
s
Z t
≤ 2E (Yr1 − Yr2 − yr )+ [g(r, Zr1 ) − g(r, Zr2 ) − µ|zr |]dr
s
Z t
+ 2E (Yr1 − Yr2 − yr )+ (dKr1 − dKr2 ).

s
Since Yr1 − Yr2 − yr ≤ Yr1 − Sr , we have

Z t Z t
(Yr1 − Yr2 − yr ) +
(dKr1 − dKr2 ) =− (Yr1 − Yr2 − yr )+ dKr2 ≤ 0.
s s
Account: ns145102
38
It follows from the conditions (C2) and (C3) that

Z t
E|(Ys1 − Ys2 − ys )+ |2 + E 1{Yr1 −Yr2 ≥yr } |Zr1 − Zr2 − zr |2 dr
s
Z t
≤ 2E (Yr1 − Yr2 − yr )+ µ[Zr1 − Zr2 − zr |]dr
s
Z t Z t
1 2 2
≤E 1{Yr1 −Yr2 ≥yr } |Zr − Zr − zr | dr + E µ2 [Zr1 − Zr2 − zr |]2 ds.
s s
Hence
Z t
E|(Ys1 − Ys2 + 2
− ys ) | ≤ E µ2 [Zr1 − Zr2 − zr |]2 ds,
s
and from Gronwall’s lemma, (Ys1 − Ys2 − ys )+ = 0, 0 ≤ s ≤ t. Therefore,

r;g,S r;g,S µ
Es,t [ξ1 ] − Es,t [ξ2 ] ≤ Es,t [ξ1 − ξ2 ], ∀ξ2 ∈ L2 (Ft ; St ) and ξ1 − ξ2 ∈ L2+ (Ft ).
Similarly, we can show the following inequality in (H2):
r;g,S r;g,S
Es,t [ξ1 ] ≤ Es,t [ξ2 ] + ξ1 − ξ2 , ∀ξ2 ∈ L2 (Ft ; St ), ξ1 − ξ2 ∈ L2+ (Fs ).
r;g,S
If Es,t [ξ2 ] > Ss a.s. for s ∈ [r, t], in version of the following equality
Z t
r;g,S
[Es,t [ξ2 ] − Ss ]dKs2 = 0, (38)
0
we have
Ks2 ≡ K02 = 0, a.s. ∀s ∈ [r, t]. That is K 2 ≡ 0, a.s..
Thus the solution of the RBSDE with the terminal condition ξ2 ≥ ξ1 is actually
that of a BSDE. It follows from the strict monotonicity of BSDEs that
r;g,S r;g,S
ξ2 = ξ1 , a.s. ⇐⇒ ξ2 ≥ ξ1 , a.s. and Er,t [ξ2 ] = Er,t [ξ1 ], a.s..
8 13
The proof of (D1)–(D4) can be found in El Karoui and Quenez and Xu .
Let us recall the definitions of E-martingales and E-supermartingales for an

{Ft , 0 ≤ t ≤ T }-consistent expectation E, introduced by Coquet et al.. 3
Definition 3.2. Let E[·] be an {Ft , 0 ≤ t ≤ T }-consistent expectation. A square-

integrable process {Mt , 0 ≤ t ≤ T } is called an E-martingale (E-supermartingale,
respectively) if for 0 ≤ s ≤ t ≤ T ,
Ms = E[Mt |Fs ] (Ms ≥ E[Mt |Fs ], respectively).
The following nonlinear Doob-Meyer’s decomposition theorem, which is Theo-
rem 6.3 in page 20 of Coquet et al., 3 will play a key role in the following arguments.
Lemma 3.1. Let E[·] be an {Ft , 0 ≤ t ≤ T }-consistent expectation, and let {Yt , 0 ≤
t ≤ T } be a continuous E-supermartingale such that
E[ sup |Yt |2 ] < ∞. (39)
t∈[0,T ]
Account: ns145102
39
Then there exists an A ∈ L2F (0, T ) such that A is continuous and increasing with
A0 = 0, and such that Y + A is an E-martingale.
Let {Es,t [·], 0 ≤ s ≤ t ≤ T } be an F-consistent dynamic operator with floor

S ≤ 0 and satisfy (H1) and (H2). Under suitable conditions, we can show that it
can be represented as a BSDE reflected upwards on a negative obstacle. For this
purpose, consider its restriction on L2+ (FT ) and denote it as
+
Et,T [X] := Et,T [X], ∀X ∈ L2+ (FT ). (40)
+
It is straightforward to check that Et,T satisfies conditions (H1)’, (H2)’, (D1), (D2)’,
(D3) and (D4).
Proceeding identically as in Definition 2.2, we define an {Ft , 0 ≤ t ≤ T }-
e t ], 0 ≤ t ≤ T } by extending E + from L2+ (FT ) to
consistent expectation {E[·|F t,T
L2 (FT ).

with the negative floor S ≤ 0 and satisfy (H1) and (H2). For X ∈ L2 (FT ; ST ),
(i) we have for t ∈ [0, T ],
e
E[X|F t ] ≤ Et,T [X]. (41)
Furthermore, if Et,T [X] > St a.s. for any t ∈ [r, T ] with some r ≥ 0, then
e
E[X|F t ] = Et,T [X] a.s. for any t ∈ [r, T ].
e
(ii) The process {E[X|Ft ], 0 ≤ t ≤ T } is an E-supermartingale.
Proof. First, we prove assertion (i). Set Xn = X1{X≥−n} . From (H2), we have
lim Et,T [Xn ] = Et,T [X].
n→∞
e t ] and (H1), we have

From the definition of E[·|F
e
E[X|F t ] = lim (E[Xn + n|Ft ] − n)
n→∞
= lim (Et,T [Xn + n] − n)
n→∞
≤ lim Et,T [Xn ] = Et,T [X]. (42)
n→∞
Therefore, we have
e
E[X|F t ] ≤ Et,T [X].
If Et,T [X] > St a.s. for t ∈ [r, T ], then we have from (D1) that
Et,T [Xn ] ≥ Et,T [X] > St a.s. for t ∈ [r, T ].

In view of (H2), we see that the inequality in (42) is actually an equality for t ∈ [r, T ].
Therefore,
e
E[X|F t ] = Et,T [X] a.s. for t ∈ [r, T ].
Account: ns145102
40
We now prove assertion (ii). Set Xt = Et,T [X] for t ∈ [0, T ]. Then from assertion
(i), we have for all 0 ≤ s ≤ t ≤ T
e t |Fs ] ≤ Es,t [Xt ] = Xs .
E[X
e
Therefore, {Et,T [X], 0 ≤ t ≤ T } is an E-supermartingale.
Using Lemmas 3.2 and 3.1, we can show the following.
Theorem 3.2. Let {Es,t [·], 0 ≤ s ≤ t ≤ T } be an F-consistent dynamic operator

+
with the negative floor S ≤ 0 and satisfy (H1) and (H2). Let Et,T [·] denote the
2 e +
restriction of Et,T [·] on L+ (FT ), and E[·|Ft ] be the extension of E [·] given in t,T
Section 2. Then for X ∈ L2 (FT ; ST ), there exists an {Ft , 0 ≤ t ≤ T }-adapted
continuous increasing process {AX X
t , 0 ≤ t ≤ T } with A0 = 0 such that
e + AX |Ft ] − AX for any t ∈ [0, T ].

Et,T [X] = E[X (43)
T t
Proof. Assertion (ii) of Lemma 3.2 shows that {Et,T [X], 0 ≤ t ≤ T } is an

e
E-supermartingale.
In version of (D2) and (H1), we have
St ≤ Et,T [X]
= Et,T [X − ST + ST ]
µ
≤ Et,T [ST ] + Et,T [X − ST ]
µ
≤ Et,T [X − ST ].
Therefore, from the first assertion of Lemma 2.1, we have for some positive constant
C1 ,
E sup |Et,T [X]|2

t∈[0,T ]
µ
≤ E sup Et,T [X − ST ]2 + E sup |St |2
t∈[0,T ] t∈[0,T ]
µ
≤ 2E sup Et,T [|X|]2 + 2E|ST |2 + E sup |St |2
t∈[0,T ] t∈[0,T ]
µ
≤ 2E sup Et,T [|X|]2 + 3E sup |St |2
t∈[0,T ] t∈[0,T ]
≤ C1 E[X 2 ] + E sup |St |2 < ∞.
t∈[0,T ]
In view of Lemma 3.1, there exists a continuous increasing {Ft , 0 ≤ t ≤ T }-

adapted process {AX X X
t , 0 ≤ t ≤ T } with A0 = 0 such that {Et,T [X]+At , 0 ≤ t ≤ T }
e
is an E-martingale. Therefore,
e T,T [X] + AX |Ft ] = Et,T [X] + AX ,

E[E ∀t ∈ [0, T ].
T t
While ET,T [X] = X, the proof is then complete.

In the following, we introduce three lemmas for subsequent arguments.
Identical to the case of an {Ft , 0 ≤ t ≤ T }-consistent expectation, we can show
Account: ns145102
41

with the negative floor S ≤ 0 and satisfy (H1) and (H2). For all 0 ≤ s ≤ t ≤ T ,
X, Y ∈ L2 (Ft ; St ), and B ∈ Fs , we have
Es,t [X1B + Y 1B c ] = Es,t [X]1B + Es,t [Y ]1B c . (44)
Lemma 3.4. Let {Es,t [·], 0 ≤ s ≤ t ≤ T } be an F-consistent dynamic operator with
the negative floor S ≤ 0 and satisfy (H1) and (H2). Let the process {A X t ,0 ≤ t ≤ T}
with AX 0 = 0 be given in Theorem 3.2. For a given X ∈ L 2
(F T ; S T ) and some
r ∈ [0, t), if
Es,T [X](ω) > Ss (ω), s ∈ [r, t], a.s. ω ∈ B, B ∈ Fr ,
then
AX X
s 1B = A t 1B , a.s. for s ∈ [r, t]. (45)
Proof. For > 0, set Yt = Et,T [X]1B + 1B c . Then
Es,t [Yt ] > Ss , s ∈ [r, t].
From Lemma 3.2, Yt satisfies
e t |Fs ].
Es,t [Yt ] = E[Y (46)
Then from Lemma 3.4, we have for ∀s ∈ [r, t],
e t + (AX − AX )1B |Fs ] = E[
E[Y e E[X
e + AX − AX |Ft ] + (AX − AX )|Fs ]1B + 1B c
t s T t t s
e e X X
= E[E[X + AT − As |Ft ]|Fs ]1B + 1B c
e + AX − AX |Fs ]1B + 1B c
= E[X T s
= Es,T [X]1B + 1B c ,
= Ys , a.s. (47)
and
e t |Fs ] = Es,t [Yt ]
E[Y
= Es,t [Et,T [X]]1B + 1B c
= Ys , a.s.. (48)
Since AX X
t 1B − As 1B ≥ 0 , we have from (47) and (48) that
AX X
s 1B = A t 1B , a.s. for ∀s ∈ [r, t].

with the negative floor S ≤ 0 and satisfy (H1) and (H2). Let the process {A X t ,0 ≤
t ≤ T } with AX 0 = 0 be given in Theorem 3.2. For a given X ∈ L 2
(F T ; ST ),
{Et,T [X], 0 ≤ t ≤ T } satisfies the following
Z T
(Et,T [X] − St )dAX
t = 0, a.s. . (49)
0
Account: ns145102
42
Proof. We firstly prove 1{Et,T [X]>St } dAX

t = 0 a.s.. Set
Bn := {ω : Er,T [X](ω) > Sr (ω), ∀[(t − 1/n) ∨ 0] < r < [(t + 1/n) ∧ T ])};
Cn := ∪{C ∈ F(t−1/n)∨0 : C ⊆ Bn }.
Since {Ft , 0 ≤ t ≤ T } is the natural filtration of Bt , augmented by all P -null sets
of F, we have
∪∞
n=1 Cn = {Et,T [X] > St }.
Lemma 3.4 implies that 1Cn dAX

t = 0, a.s. . Therefore,
1{Et,T [X]>St } dAX

t =0 a.s. .
For Et,T [X] − St ≥ 0, we have
Z T Z T
(Et,T [X] − St )dAX
t = (Et,T [X] − St )1{Et,T [X]>St } dAX
t = 0. (50)
0 0
Theorem 3.3. Theorem 1.1 is true in the case of the negative floor S ≤ 0.
Proof. From Theorem 2.3, there exists a function g = g(t, z) : Ω × [0, T ] × Rd
satisfying (C2) and (C3) and g(·, ·, 0) ≡ 0, such that the following holds:
e |Ft ] = E g [Y |Ft ], ∀Y ∈ L2 (FT ), t ∈ [0, T ].
E[Y
Therefore, for X ∈ L2 (FT ; ST ), we have E[Xe + AX |Ft ] = E g [X + AX |Ft ] for t ∈
T T
[0, T ].
From the definition of E g [X + AX g X
T ] and E [X + AT |Ft ], we know that there is
2 d
unique Z ∈ LF (0, T ; R ) such that
Z T Z T
g X X
E [X + AT |Ft ] = X + AT + g(s, Zs )ds − hZs , dBs i, a.s. for any t ∈ [0, T ].
t t
Set
Z T Z T
et := X + g(s, Zs )ds − Zs dBs + AX X
X T − At . (51)
t t
Then from Lemma 3.2, we have
e + AX |Ft ] − AX
Et,T [X] = E[X T t
= E g [X + AX X e
T |Ft ] − At = Xt . (52)
Since Xet = Et,T [X] ≥ St , it follows from Lemma 3.5 that {(X et , Zt , A ), 0 ≤ t ≤ T }
X
t
is the solution of RBSDE (X, g, S). That is,
Et,T [X] = X et = E r;g,S [X], ∀X ∈ L2 (FT ; ST ). (53)

t,T
Remark 3.1. From the proof of Theorem 3.3, {(Et,T [X], Zt , AX t ), 0 ≤ t ≤ T } is

the solution of RBSDE (X, g, S). Therefore, the increasing process {AXt ,0 ≤ t ≤ T}
is unique for a given X ∈ L2 (FT ; ST ).
Account: ns145102
43
Acknowledgments
Both authors thank Professor Shige Peng for his helpful comments.
References
1. X. Bao and S. Tang, Representation via Reflected Backward Stochastic Differential
Equations of a Filtration-Consistent Nonlinear Dynamic Operator with a Floor, sub-
mitted to CRAS.
2. X. Bao and S. Tang, Filtration-consistent dynamic nonlinear evaluation with a general
continuous floor and associated reflected backward stochastic differential equations,
preprint.
3. F. Coquet, Y. Hu, J. Mémin, and S. Peng, Filtration-consistent nonlinear expectations
and related g-expectations, Probability Theory and related Fields, 123, 1–27 (2002).
4. N. El Karoui, S. Peng and M. C. Quenez, Backward stochastic differential equations
in finance, Mathematical Finance, 7, 1–71 (1997).
5. N. El Karoui, C. Kapoudian, E. Pardoux, S. Peng and M. C. Quenez, Reflected
solutions of backward SDE’s, and related obstacle problems for PDE’s, The Annals
of Probability, 25, 702–737 (1997).
6. N. El Karoui and M. C. Quenez, Programmation dynamique et évalution des actifs
contingents en marchés incomplet, C. R. Acad. Sci. Paris, Sér. I, 313, 851–854 (1991).
7. N. El Karoui and M. C. Quenez, Dynamic programming and pricing of contingent
claims in incomplete market, SIAM J. Control Optim., 33, 29–66 (1995).
8. N. El Karoui and M. C. Quenez, Nonlinear pricing theory and backward stochastic
differential equations, in : Financial Mathematics (ed.: W. J. Runggaldier), Lecture
Notes in Mathematics 1656, Springer Verlag, 191-246 (1996).
9. I. Karatzas and S. Shreve, Methods of Mathematical Finance, World Publishing Cor-
poration, Beijing (2004).
10. S. Peng, Backward SDE and related g-expectation, in: N. El Karoui and L. Ma-
zliak (eds.), Backward Stochastic Differential Equations, Pitman Research Notes in
Mathematics 364, 141–159 (1997).
11. S. Peng, Monotonic limit theorem of BSDE and nonlinear decomposition theorem of
Doob-Meyer’s type, Probability Theory and related Fields, 113, 473–499 (1999).
12. S. Peng, Nonlinear expectations, nonlinear evaluations and risk measures, Stochastic
methods in finance, Lecture Notes in Math., 1856, Springer, Berlin, 165–253 (2004).
13. M. Xu, Contributions to Reflected Backward Stochastic Differential Equations: the-
ory, numerical analysis and simulations, thesis, Shandong University, 57–67 (2005).
Account: ns145102
July 31, 2007 14:58 Proceedings Trim Size: 9.75in x 6.5in Chap15-YongJiongmin
A LINEAR QUADRATIC OPTIMAL CONTROL PROBLEM FOR

STOCHASTIC VOLTERRA INTEGRAL EQUATIONS
SHUPING CHEN
Zhejiang University,
Hangzhou 310027, CHINA
E-mail: chensp@gzu.edu.cn
JIONGMIN YONG
University of Central Florida,
Orlando, FL 32816, USA, and
Fudan University, Shanghai 200433, CHINA
E-mail: jyong@mail.ucf.edu
A linear quadratic optimal control problem is considered for a stochastic Volterra integral
equation. As a necessary condition for the optimality, a forward-backward stochastic
Volterra integral equation (FBSVIE, for short) is derived, via a duality principle for
stochastic integral equations.
Keywords: Linear quadratic optimal control problem, forward-backward stochastic

integral equations, duality principle.
AMS Mathematics subject classification: 60H10.
1. Introduction
Let (Ω, F, lF, P) be a complete filtered probability space, on which a d-dimensional
∆
standard Brownian motion W (·) is defined with lF ={Ft }t≥0 being its natural filtra-
tion augmented by all the P-null sets. We consider the following controlled linear
(forward) stochastic Volterra integral equation (FSVIE, for short):
Z th i
X(t) = ϕ(t) + A0 (t, s)X(s) + B0 (t, s)u(s) ds
0
d Z th i (1.1)
X
+ Ai (t, s)X(s) + Bi (t, s)u(s) dWi (s), t ∈ [0, T ],
i=1 0
where X(·) is the state and u(·) is the control, taking values in Euclidean spaces
lRn and lRm of dimensions n and m, respectively, ϕ(·), Ai (· , ·) and Bi (· , ·) are given
44
Account: ns145102
45
deterministic functions of suitable dimensions. The cost functional associated with

(1.1) is given by
Z T h i
J(u(·)) = E h Q(t)X(t), X(t) i +2 h S(t)X(t), u(t) i + h R(t)u(t), u(t) i dt,
0
(1.2)
where Q(·), S(·), and R(·) are matrix-valued deterministic functions of suitable di-
mensions. Our optimal control problem is to minimize cost functional (1.2) subject
to the state equation (1.1) with the control u(·) taken from a space U of admissible
controls. Such a problem is referred to as linear quadratic (LQ, for short) problem
for a stochastic Volterra integral equation.
LQ problem for ordinary differential equations (ODEs, for short) was studied by
Bellman–Glicksberg–Gross in 1958 the first time 5 . Kalman 14 and Letov 16 solved
LQ problem (for ODEs) in terms of linear feedback control and Riccati equation in
1960. Since then, many authors extend the theory to various situations. See Refs. 2,
27, and 29 for some summaries of LQ theory for ODEs; see Refs. 20, 18, and 17 for
LQ theory of infinite-dimensional (deterministic) systems; see Refs. 28, 8–11, for
LQ theory of stochastic differential equations. Some detailed historic remarks on
LQ theory can be found in Ref. 34. For some recently relevant works, see Refs. 30,
1, 32, and 31.
LQ problem for (deterministic) Volterra integral equations was firstly studied in
1967 by Vinokurov 26 . See also Ref. 22. Since then some extensions were developed
(see, for example, Refs. 7, 3, 4, 24, 35, and references cited therein). On the other
hand, stochastic Voterra integral equations were studied by several authors 6,25,23 ,
and some interesting applications were indicated in Ref. 13.
In this paper, we are going to study the LQ problem for stochastic Volterra inte-
gral equations with the state equation of form (1.1) and cost functional (1.2). Due to
the nature of our state equation, one could not use Itô’s formula; all the derivations
have to be carried out under integral(s). In our discussion, we will extend/modify
some results of backward stochastic Volterra integral equations (BSVIEs, for short)
developed in Ref. 33 (see also Ref. 19). By a duality principle for linear stochastic
integral equations, we derive a forward-backward stochastic Volterra integral equa-
tions (FBSVIEs, for short) whose solvability leads to the existence of critical point
of the cost functional. Then under some kind of nonnegativity condition on the cost
functional, an optimal control will exist.
Note that, in our discussion, we do not a priori impose nonnegativity condition
on matrix-valued functions Q(·) and R(·). The case of random coefficients and/or
a term like h GX(T ), X(T ) i appears in the cost functional will have much more
things involved, and we hope to present the relevant results in a forthcoming paper.
The rest of the paper is organized as follows. Some preliminary results concern-
ing the state equation will be presented in Section 2. Section 3 is devoted to a
discussion on a minimization problem of a quadratic functional in a Hilbert space
which will give us some abstract idea of our LQ problem. In Section 4, we present
Account: ns145102
46
some results on linear BSVIEs, among other things, we mainly modify some results
from Ref. 33. In Section 5, we introduce a two-point boundary value problem as a
necessary condition for optimal control of our LQ problem. This leads to a linear
coupled FBSVIE.
2. State Equation and Optimal Control Problem

In this section, we present some preliminaries. To begin with, let us introduce some
spaces. First of all, we let lRn×m be the set of all (n × m) matrices, and S n be the
set of all (n × n) symmetric matrices. For H = lRm , lRm×d , S n , etc., we define
n Z T o

L2F (0, T ; H) = ϕ : (0, T ) × Ω → H ϕ(·) is lF-adapted, E |ϕ(t)|2 dt < ∞ ,
n 0
2 2
CF ([0, T ]; L (Ω; H)) = ϕ(·) ∈ LF (0, T ; H) t 7→ ϕ(t, ·) is continuous
o
from [0, T ] to L2 (Ω; lRn ), sup E|ϕ(t)|2 < ∞ ,
t∈[0,T ]
n

L2F (Ω; C([0, T ]; H)) = ϕ(·) ∈ L2F (0, T ; H) ϕ(·) has continuous paths,
h i o
E sup |ϕ(t)|2 < ∞ .
t∈[0,T ]
Spaces L∞ F (0, T ; H), CF ([0, T ]; L (Ω; H)), and LF (Ω; C([0, T ]; H)) can be defined
∞ ∞
in an obvious way. We point out that any process ϕ(·) ∈ CF ([0, T ]; L2(Ω; H)) is con-
tinuous as a map from [0, T ] to L2 (Ω; H), and does not necessarily have continuous
paths. Next, for any RBanach space Y, we let L2 (0, T ; Y) be the set of all maps ϕ :
T
[0, T ] → Y such that 0 |ϕ(t)|2 dt < ∞. The spaces L∞ (0, T ; Y) and C([0, T ]; Y) can
be defined in a similar way. Note that we may take Y = L2F (0, T ; H), L∞ (0, T ; H),
2 2
L∞F (0, T ; H), LF (Ω; L (0, T ; H)), etc. For example, process Z : [0, T ] × Ω → H
∞
2 2 2
belongs to L (0, T ; LF (0, T ; H)) if it is B([0, T ] ) ⊗ FT -measurable; for almost all
t ∈ [0, T ], Z(t, ·) is lF-adapted; and
Z TZ T
E |Z(t, s)|2 dsdt < ∞.
0 0
We now make the following standing assumptions.

(H1) Suppose the following hold:

 2 2 n×n
 A0 (· , ·) ∈ L (0, T ; L (0, T ; lR


)),
 2 n×n
Ai (· , ·) ∈ L (0, T ; L (0, T ; lR
∞
)), 1 ≤ i ≤ d,
(2.1)
 B0 (· , ·) ∈ L (0, T ; L (0, T ; lR
 2 ∞ n×m
)),


 2 n×m
Bi (· , ·) ∈ L (0, T ; L (0, T ; lR
∞
)), 1 ≤ i ≤ d.
Q(·) ∈ L∞ (0, T ; S n ), S(·) ∈ L∞ (0, T ; lRm×n ), R(·) ∈ L∞ (0, T ; S m ). (2.2)
Account: ns145102
47
From (1.1), we see that the value of Ai (t, s) and Bi (t, s) with 0 ≤ t < s ≤ T are
irrelevant. Hence, in what follows, we make the following convention:
Ai (t, s) = 0, Bi (t, s) = 0, 0 ≤ t < s ≤ T, 0 ≤ i ≤ d. (2.3)
Next, we let
X = L2F (0, T ; lRn ), U = L2F (0, T ; lRm ),
and define A : X → X and B : U → X as follows:


 Z t d Z t
X



 (AX)(t) =
 A0 (t, s)X(s)ds + Ai (t, s)X(s)dWi (s),
0 i=1 0
(2.4)

 Z t d Z t
X



 (Bu)(t) = B 0 (t, s)u(s)ds + Bi (t, s)u(s)dWi (s).
0 i=1 0
The following result is concerned with the well-posedness of the state equation (1.1).
Proposition 2.1. Let (H1) hold. Then operators A : X → X and B : U → X are

bounded and A is quasi-nilpotent, i.e.,
1
lim kAk k k = 0. (2.5)
k→∞
Consequently, (I −A)−1 : X → X is bounded, hence, for any ϕ(·) ∈ X and u(·) ∈ U,

state equation (1.1) admits a unique solution X(·) ∈ X .
Proof. By (2.4), we have that for any X(·) ∈ X , and t ∈ [0, T ], (note (2.3))
n Z t 2 Xd Z t o
2
E (AX)(t) ≤ (d + 1)E A0 (t, s)X(s)ds + Ai (t, s)X(s)|2 ds
0 i=1 0
nh Z T d
X iZ T o
≤ (d + 1)E |A0 (t, s)|2 ds + esssup |Ai (t, s)|2 |X(s)|2 ds .
0 i=1 s∈[0,T ] 0
(2.6)
Thus, by (H1), for some L > 0,
Z T Z T
2

E (AX)(t) dt ≤ LE |X(s)|2 ds. (2.7)
0 0
This implies the boundedness of the operator A. The boundedness of B can be

proved similarly. From (2.7), we see that for any k ≥ 1,

Z t Z tZ s
k 2 k−1 2 2

E (A X)(t) ≤ L
E (A
X)(s) ds ≤ L 2
E (Ak−2 X)(r) drds
0 0 0
Z t k Z t
2 L
= L2 (t − s)E (Ak−2 X)(s) ds ≤ · · · ≤ (t − s)k E|X(s)|2 ds, t ∈ [0, T ].
0 k! 0
Account: ns145102
48
Hence,
Z T k Z T Z t
k
k(Ak X)(·)k2X = E (A X)(t)|2 dt ≤ L (t − s)k E|X(s)|2 dsdt
0 k! 0 0
Z T
Lk k+1 2 Lk T k+1
= (T − s) E|X(s)| ds ≤ kX(·)k2X .
(k + 1)! 0 (k + 1)!
This implies that
Lk T k+1
kAk k ≤ , k ≥ 1. (2.8)
(k + 1)!
Thus, A is quasi-nilpotent.
Using operators A and B, we can write the state equation (1.1) as
X = ϕ + AX + Bu. (2.9)
Since A is quasi-nilpotent, we have the existence and boundedness of (I − A)−1 .
Moreover, one has the expansion:
X
∞
(I − A)−1 = Ak . (2.10)
k=0
Hence, for any ϕ(·) ∈ X and u(·) ∈ U, state equation (1.1) admits a unique solution
X(·) ∈ X given by
X
∞
X = (I − A)−1 (ϕ + Bu) ≡ Ak (ϕ + Bu). (2.11)
k=0
This proves our theorem.

By the above result, we see that under (H1)–(H2), the cost functional J(u(·))
is well-defined on U. Consequently, we can state our optimal control problem as
follows.
Problem (LQ). Minimizing J(u(·)) over u(·) ∈ U.
The following notions are concerned with Problem (LQ).
Definition 2.1. (i) Problem (LQ) is said to be accessible if

inf J(u(·)) > −∞. (2.12)
u(·)∈U
(ii) Problem (LQ) is said to be (uniquely) solvable if there exists a (unique)

ū(·) ∈ U such that
J(ū(·)) = inf J(u(·)). (2.13)

u(·)∈U
Any control ū(·) ∈ U satisfying (2.13) is called an optimal control, and the cor-
responding state process X̄(·) is called an optimal state process. We also call
(X̄(·), ū(·)) an optimal pair.
Account: ns145102
49
Our goals are: (a) Characterize the accessibility and solvability of Problem (LQ),
and (b) in the case that Problem (LQ) is solvable, find or characterize optimal
controls.
3. Quadratic Functionals in Hilbert Spaces

In this section, we briefly look at some properties of quadratic functionals in Hilbert
spaces which will be helpful for us to understand Problem (LQ) from a functional
analysis point of view.
Let H be a Hilbert space and Φ : H → H be a bounded self-adjoint operator.
We denote R(Φ) and N (Φ) to be the range and kernel
of Φ, respectively. Since Φ is
self-adjoint, N (Φ)⊥ = R(Φ) (and we always have Φ R(Φ) ⊆ R(Φ)). Thus, under
the decomposition H = N (Φ) ⊕ R(Φ), we have the following representation for Φ:
!
00
Φ= b , (3.1)
0Φ
where Φ b : R(Φ) → R(Φ) is self-adjoint. Now, we define the pseudo-inverse Φ† by

the following:
!
0 0
Φ =
†
b −1 , (3.2)
0Φ
with domain

D(Φ† ) = N (Φ) + R(Φ) ≡ {u0 + u1 u0 ∈ N (Φ), u1 ∈ R(Φ)} ⊇ R(Φ). (3.3)
From the above, we can easily seen the following facts:
(i) Φ† is (closed, densely defined, and) self-adjoint; R(Φ) is closed if and only if
Φ is bounded.
†
(ii) By the definition of Φ† (see (3.2)), together with (3.3), one has that
ΦΦ† Φ = Φ, Φ† ΦΦ† = Φ† , (Φ† )† = Φ. (3.4)
Thus, by (i), R(Φ† ) is closed since Φ is bounded.
(iii) Although D(Φ† ) is not necessarily closed, the operator ΦΦ† : D(Φ† ) → H
is an orthogonal projection onto R(Φ). Thus, we may naturally extend it, still
denoted it by itself, to D(Φ† ) = H. Hence, ΦΦ† : H → R(Φ) ⊆ H is the orthogonal
projection onto R(Φ). Note that since Φ is bounded, Φ† Φ is an orthogonal pro-
jection from H onto R(Φ† ) = N (Φ† )⊥ = N (Φ)⊥ = R(Φ). Therefore, in fact, we

have
ΦΦ† = Φ† Φ ≡ PR(Φ) ≡ orthogonal projection onto R(Φ). (3.5)
(iv) The map Φ 7→ Φ† is not continuous (which can be seen even from one-
dimensional case).
Account: ns145102
50
Now, let us consider a quadratic functional on H:
J(u) = h Φu, u i +2 h v, u i, u ∈ D(Φ) ⊆ H, (3.6)
where Φ : D(Φ) ⊆ H → H is a self-adjoint linear operator and v ∈ H. The following

result is concerned with the completing square and critical point(s) of the functional
J(·). Note here that we do not assume positive (negative) semi-definite condition
on Φ.
Proposition 3.1. (i) There exists a û ∈ H such that
J(u) = h Φ(u − û), u − û i − h Φû, û i, ∀u ∈ H, (3.7)
if and only if

v ∈ R(Φ) ⊆ D(Φ† ) . (3.8)
(ii) Any û ∈ H satisfies (3.7) if and only if it is a solution of the following

equation:
Φû + v = 0, (3.9)
which is equivalent to the following:
û = −Φ† v + (I − Φ† Φ)e
v, (3.10)
for some ve ∈ H (in particular, û = −Φ† v is a solution).

(iii) When (3.7) holds, it is necessary that
J(u) = h Φ(u − û), u − û i − h Φ† v, v i, ∀u ∈ H. (3.11)
Moreover, û is unique if and only if N (Φ) = {0}.
Proof. (i) For any û ∈ H, on has
J(u) ≡ h Φu, u i +2 h v, u i
(3.12)
= h Φ(u − û), u − û i +2 h Φû + v, u i − h Φû, û i, ∀u ∈ H.
Hence, there exists a û ∈ H such that (3.7) holds if and only if (3.9) holds, which
gives (3.8) (and the first part of (ii)).
Conversely, if (3.8) holds, then there exists a û ∈ H such that (3.9) holds.
Consequently,
h Φ(u − û), u − û i − h Φû, û i = h Φu, u i −2 h Φû, u i + h Φû, û i − h Φû, û i

(3.13)
= h Φu, u i +2 h v, u i = J(u),
proving (3.7).
(ii) We have proved the first part of (ii) (from (3.12)). The second part is
straightforward.
Account: ns145102
51
(iii) For any û ∈ H satisfying (3.7), one must have (3.9). Hence,
h Φ(u − û), u − û i − h Φ† v, v i = h Φu, u i −2 h Φû, u i + h Φû, û i − h Φ† Φû, Φû i

= h Φu, u i +2 h v, u i = J(u),
(3.14)
which proves (3.11). Finally, by (3.9), we see that û is unique if and only if N (Φ) =
{0}.
Note that (3.9) is equivalent to the following:
1
0 = Φû + v ≡ ∇J(û). (3.15)
2
Thus, û is actually a critical point of functional J(·). Hence, Proposition 3.1 char-
acterizes critical points of the quadratic functional J(·). Equations (3.7) and (3.11)
are completion of square for the functional J(·) (although Φ is not necessarily pos-
itive/negative semi-definite).
Next, for any self-adjoint operator Φ, regardless whether it is bounded or un-
bounded, we have the following spectrum decomposition 12
Z
Φ= λdPλ , (3.16)
σ(Φ)
where σ(Φ) ⊆ lR is the spectrum of Φ, (which is

a compact set if Φ is bounded, and
it is unbounded if Φ is unbounded); and {Pλ λ ∈ σ(Φ)} is a family of projection
measures. In the case that
Φ ≥ 0, (3.17)
one has from (3.16) that σ(Φ) ⊆ [0, ∞), and
 Z


 Φα =
 λα dPλ , ∀α ≥ 0,
σ(Φ)
Z (3.18)

 † α α †

 (Φ ) = (Φ ) = λ−α dPλ , ∀α > 0.
σ(Φ)\{0}
Now, we can consider minimization problem for functional J(·).
Proposition 3.2. Let Φ : H → H be bonded and self-adjoint, and v ∈ H.

(i) The following holds:
inf J(u) > −∞, (3.19)
u∈H
if and only if (3.17) holds and

1
v ∈ R(Φ 2 ). (3.20)
In this case,
1
inf J(u) = −|(Φ† ) 2 v|2 . (3.21)
u∈H
Account: ns145102
52
(ii) There exists a û ∈ H such that
J(û) = inf J(u), (3.22)

u∈H
if and only if (3.17) and (3.8) hold; and in this case, all the conclusions in Propo-
sition 3.1 hold.
Proof. We need only to prove (i). First, let (3.19) hold. It is straightforward
that one must have (3.17). Next, we prove (3.20) by contradiction. Suppose (3.20)
does not hold. For any n ≥ 1, let
Z
vn = dPλ v.
1
σ(Φ)∩[ n ,n]
Then vn ∈ R(Φ), and

Z
1
h v, Φ vn i =
†
λd|Pλ v|2 = |(Φ† ) 2 vn |2 → ∞, n → ∞.
1
σ(Φ)∩[ n ,n]
Hence, letting un = −Φ† vn , we obtain

1
J(un ) = h Φun , un i +2 h v, un i = −|(Φ† ) 2 vn |2 → −∞, n → ∞,
contradicting (3.19).
Conversely, if (3.17) and (3.20) hold, then for any u ∈ H, one has
1 1 1 1 1 1 1
J(u) = |Φ 2 u|2 +2 h(Φ† ) 2 v, Φ 2 u i = |Φ 2 u+(Φ† ) 2 v|2 −|(Φ† ) 2 v|2 ≥ −|(Φ† ) 2 v|2 > −∞.
(3.23)
Hence, sufficiency follows.
Finally, from the fact that
1 1
R((Φ† ) 2 ) ⊆ R(Φ 2 ) = R(Φ),
we can always find a sequence un ∈ H so that (note (3.23))

1 1 1 1
J(un ) = |Φ 2 un + (Φ† ) 2 v|2 − |(Φ† ) 2 v|2 → −|(Φ† ) 2 v|2 , n → ∞.
Thus, (3.21) follows.

The above result tells us that the existence of minimum is strictly stronger than
the finiteness of the infimum of the functional J(·), which have been described by
1
conditions (3.8) and (3.20), respectively. Note here that R(Φ) ⊆ R(Φ 2 ) when (3.17)
holds. The following example shows the necessity of condition (3.20) in a concrete
way.
Example 3.1. Let H = `2 . For any u = {ai }∞

i=1 ∈ H, define Φu by
Φu = {β i−1 ai }∞
i=1 ,
Account: ns145102
53
where β ∈ (0, 1). Then Φ : H → H is bounded, self-adjoint, and positive definite

(but not uniformly). Let v = {i−1 }∞
i=1 ∈ H1 . Then v ∈ R(Φ) since v = n→∞
lim Φun
with
1 1 1
un = {1, , ,··· , , 0, 0, · · · } ∈ H.
2β 3β 2 nβ n−1
But, clearly, v ∈
/ R(Φ). Now, consider a quadratic functional
∞
X 2ai
J(u) = h Φu, u i +2 h v, u i = β i−1 a2i + .
i=1
i
Then by letting un as above, we see that
J(−un ) = h Φun , un i −2 h v, un i
n h
2 i
X Xn
1 1
= β i−1 2 2(i−1) − 2 i−1 = − 2 β (i−1)
→ −∞, as n → ∞.
i=1
i β i β i=1
i
This means that
inf J(u) = −∞.
u∈H
An interesting point here is that positive semi-definiteness of Φ does not even ensure
the finiteness of the infimum of J(·).
Now, we return to Problem (LQ). Denote

(
QX = Q(·)X(·), SX = S(·)X(·), ∀X(·) ∈ X ,
(3.24)
Ru = R(·)u(·), ∀u(·) ∈ U.
Then under (H1)–(H2) and with ϕ(·) ∈ L2F (0, T ; lRn ), the cost functional (1.2) can
be written as follows:
J(u) = h QX, X i +2 h SX, u i + h Ru, u i

Q S∗ (I − A)−1 (ϕ + Bu) (I − A)−1 (ϕ + Bu)
=h , i
S R u u

Q S∗ (I − A)−1 (I − A)−1 B ϕ (I − A)−1 (I − A)−1 B ϕ
=h , i
S R 0 I u 0 I u
= h Φ2 u, u i +2 h Φ1 ϕ, u i + h Φ0 ϕ, ϕ i,
(3.25)
where


 Φ = (I − A∗ )−1 Q(I − A)−1 ,
 0 n o
Φ1 = B ∗ (I − A∗ )−1 Q + S (I − A)−1 ,



Φ2 = B ∗ (I − A∗ )−1 Q(I − A)−1 B + S(I − A)−1 B + B ∗ (I − A∗ )−1 S ∗ + R.
(3.26)
Consequently, by Propositions 3.1 and 3.2, we obtain the following abstract result.
Account: ns145102
54
Proposition 3.3. (i) If Problem (LQ) admits an optimal control, then
Φ2 ≥ 0. (3.27)
(ii) When (3.27) holds, Problem (LQ) admits an optimal control if and only if
Φ1 ϕ ∈ R(Φ2 ). (3.28)
In this case, ū is an optimal control if and only if it is a solution of the equation:
Φ2 ū + Φ1 ϕ = 0, (3.29)
which is given by
ū = −Φ†2 Φ1 ϕ + (I − Φ†2 Φ2 )v, (3.30)
for some v ∈ U, with the optimal cost functional
J(ū(·)) = h(Φ0 − Φ∗1 Φ†2 Φ1 )ϕ, ϕ i . (3.31)
Further, if Φ2 is invertible, the optimal control is unique.
Although the above gives necessary and sufficient conditions under which Prob-
lem (LQ) is (uniquely) solvable, the conditions imposed a little too abstract, for
example, it is by no means obvious how one can represent A∗ and B ∗ . Therefore,
we need to make further efforts in our investigation.
4. BSVIEs and Duality Principle

In this section, we recall/modify some results on BSVIEs from Ref. 33, which will
be useful in studying Problem (LQ).
To begin with, let us consider the following BSVIE introduced in Ref. 33:
Z T Z T
Y (t) = ψ(t) + g(t, s, Y (s), Z(s, t))ds − Z(t, s)dW (s), t ∈ [0, T ], (4.1)
t t
n n×d n
where g : [0, T ]2 × lR × lR → lR is a given map. According to Ref. 33, we have
the following definition.
Definition 4.1. A pair of process
(Y (·), Z(· , ·)) ∈ L2F (0, T ; lRn ) × L2 (0, T ; L2F (0, T ; lRn×d ))
is called an adapted solution of (4.1) if for almost all t ∈ [0, T ], almost surely, (4.1)
is satisfied in the Itô sense.
In Ref. 33, it was proved that under the condition that (y, z) 7→ g(t, s, y, z) is
uniformly Lipschitz, there exists a unique adapted solution (Y (·), Z(· , ·)) to (4.1).
Based on this, some further results were established (see Ref. 33).
Account: ns145102
55
Before going further, let us look at a simple situation. For any Y (·) ∈
L2F (0, T ; lRn ) and Z(· , ·) ∈ L2 (0, T ; L2F (0, T ; lRn )), we call
Z T
ψ(t) = Y (t) + Z(t, s)dW (s), t ∈ [0, T ]. (4.2)
t
2
Then ψ(·) ∈ L ((0, T ) × Ω), and ψ(·) is not necessary lF-adapted. Clearly,
(Y (·), Z(· , ·)) is an adapted solution of the following BSVIE:
Z T
Y (t) = ψ(t) − Z(t, s)dW (s), t ∈ [0, T ]. (4.3)
t
On the other hand, for ψ(·), by martingale representation theorem, there exists a
b , ·) ∈ L2 (0, T ; L2 (0, T ; lRn )) such that
unique Z(· F
Z T
ψ(t) = Eψ(t) + b s)dW (s),
Z(t, t ∈ [0, T ]. (4.4)
0
If we define
Z t
∆
Yb (t) = Eψ(t) + b s)dW (s),
Z(t, t ∈ [0, T ], (4.5)
0
then (Yb (·), Z(·

b , ·)) is also an adapted solution of (5.14). Hence, we have
Z Th i
b
Y (t) − Y (t) = b s) − Z(t, s) dW (s),
Z(t, t ∈ [0, T ]. (4.6)
t
By taking conditional expectation, we see that (note (4.5) and Eψ(t) = EY (t))
Z t
b
Y (t) = Y (t) = EY (t) + b s)dW (s),
Z(t, t ∈ [0, T ]. (4.7)
0
Hence, we must have
b s),
Z(t, s) = Z(t, s ∈ [t, T ]. (4.8)
In such a sense, the adapted solution to (5.14) is unique. But, in general, (4.8)
might not be true for s ∈ [0, t]. Hence, we might not have (comparing with (4.5)
and (4.7))
Z t
Y (t) = EY (t) + Z(t, s)dW (s), t ∈ [0, T ]. (4.9)
0
To further convince ourselves, let us look at the following example.
Example 4.1. Example 4.2. Take d = 1, Y (t) = W (t), and Z(t, s) ≡ 2. Then
Z T
ψ(t) = Y (t) + Z(t, s)dW (s) = 2W (T ) − W (t), t ∈ [0, T ],

t
b s) = 2 −
and (trivially) (5.14) holds. On the other hand, (4.5) holds with Z(t,
I[0,t] (s). Clearly, (4.8) holds, but
b s),
Z(t, s) 6= Z(t, s ∈ [0, t],
Account: ns145102
56
and
Z T
Y (t) = W (t) 6= 2W (T ) − W (t) = EY (t) + Z(t, s)dW (s), t ∈ [0, T ]. (4.10)
t
The above shows the following: (i) The uniqueness of adapted solution
(Y (·), Z(· , ·)) to BSVIE (4.1) does not give the uniqueness of Z(t, s) for s ∈ [0, t],
and (ii) in general, we do not have (4.9) for any adapted solution (Y (·), Z(· , ·)).
However, from Ref. 33, we know that relation (4.9) played an important role in
establishing duality relation. Hence, the following notion will be useful.
Definition 4.2. A pair of process

(Y (·), Z(· , ·)) ∈ L2F (0, T ; lRn ) × L2 (0, T ; L2F (0, T ; lRn×d ))
is called an adapted M-solution of (4.1) if (4.1) is satisfied in the Itô sense, and (4.9)
holds.
In the above, “M” in “M-solution” stands for “Martingale Representation”.

Now, we state the following result concerning the existence and uniqueness of
adapted M-solutions of BSVIE (4.1).
Proposition 4.1. Suppose that (y, z) 7→ g(t, s, y, z) is uniformly Lipschitz continu-

ous. Then for any ψ(·) ∈ L2 ((0, T ) × Ω), (4.1) admits a unique adapted M-solution
(Y (·), Z(· , ·)).
Proof. Let M be the set of all processes (y(·), z(· , ·)) ∈ L2F (0, T ; lRn ) ×
L2 (0, T ; L2F (0, T ; lRn×d )) satisfying
Z t
y(t) = Ey(t) + z(t, s)dW (s), t ∈ [0, T ].
0
Then M is a (nontrivial) closed subspace of L2F (0, T ; lRn )×L2 (0, T ; L2F (0, T ; lRn×d )).
Now, for any (y(·), z(· , ·)) ∈ M, let
Z T
ˆ = ψ(t) +
ψ(t) g(t, s, y(s), z(s, t))ds, t ∈ [0, T ].
t
By martingale representation theorem, there exists a unique Z(· , ·) ∈
L2 (0, T ; L2F (0, T ; lRn×d )) such that
Z T
ˆ = E ψ(t)
ψ(t) ˆ + Z(t, s)dW (s), t ∈ [0, T ].
0
Next, we define
Z t
ˆ +
Y (t) = E ψ(t) Z(t, s)dW (s), t ∈ [0, T ].
0
Then (Y (·), Z(· , ·)) ∈ M and
Z T Z T
Y (t) = ψ(t) + g(t, s, y(s), z(s, t))ds − Z(t, s)dW (s), t ∈ [0, T ].
t t
Account: ns145102
57
Hence, we obtain a map Φ : M → M, (y(·), z(· , ·)) 7→ (Y (·), Z(· , ·)). Then by
a contraction mapping argument similar to that in Ref. 33, we can show that Φ
admits a unique fixed point, which is the adapted M-solution that we are looking
for.
Note that unlike adapted solution, the uniqueness of adapted M-solution really
gives the uniqueness of Z(· , ·) over [0, T ] × [0, T ], due to the additional requirement
(4.9).
In Example 4.1, we see that although (Y (·), Z(· , ·)) is merely an adapted solution
b , ·)) is the adapted M-solution
and (4.9) fails, we can modify Z(· , ·) so that (Y (·), Z(·
(therefore (4.7) holds).
Sometime, we might need some other restriction other than (4.9). Inspired by
the above, we have the following result.
Proposition 4.2. Let (y, z) 7→ g(t, s, y, z) be uniformly Lipschitz continuous. Then

for any ψ(·) ∈ L2 ((0, T )×Ω) and ϕ(·) ∈ L2F (0, T ; lRn ), (4.1) admits a unique adapted
solution (Y (·), Z(· , ·)) satisfying
Z t
ϕ(t) = Eϕ(t) + Z(t, s)dW (s), t ∈ [0, T ]. (4.11)
0
Proof. First of all, by martingale representation theorem, we can find ζ(· , ·) ∈

L2 (0, T ; L2F (0, T ; lRn )) such that
Z t
ϕ(t) = Eϕ(t) + ζ(t, s)dW (s), t ∈ [0, T ]. (4.12)
0
Note that since ϕ(·) is lF-adapted, ζ(t, s) = 0 for s ∈ [t, T ]. Next, consider the
following BSVIE:
Z T Z T
Y (t) = ψ(t) + g(t, s, Y (s), ζ(s, t))ds − Z(t, s)dW (s), t ∈ [0, T ]. (4.13)
t t
By Proposition 4.1, the above admits a unique adapted solution (Y (·), Z(· , ·)). Note
that in (4.13), only the values Z(t, s) with 0 ≤ s ≤ t ≤ T are used and by changing
the values Z(t, s) for 0 ≤ s ≤ t ≤ T only, (4.13) remains unchanged. On the other
hand, in the drift term, only ζ(t, s) with 0 ≤ s ≤ t ≤ T are used. Hence, it we
redefine
Z(t, s) = ζ(t, s), 0 ≤ s ≤ t ≤ T,
then (Y (·), Z(· , ·)) is still an adapted solution of (4.1) and (4.11) holds.
Note of course that the adapted solution (Y (·), Z(· , ·)) obtained in Proposi-
tion 4.2 is not an adapted M-solution. Actually, (4.9) is replaced by (4.11). Also,
we see that in the proof of Proposition 4.2, we have taken the advantage that the
drift does not depending on Z(t, s) (only depending on Z(s, t)). In the case that
the drift also depends on Z(t, s), we will still have a unique adapted M-solution.
But, it is not clear if Proposition 4.2 holds.
Account: ns145102
58
Now, we strengthen the standing assumptions (H1) for the coefficients of the
state equation as follows.
Ai (· , ·) ∈ L∞ ((0, T )2 ; lRn×n ), Bi (· , ·) ∈ L∞ ((0, T )2 ; lRn×m ), 0 ≤ i ≤ d.

(4.14)
The following result gives representations for A∗ and B ∗ in terms of BSVIEs.
Theorem 4.1. Suppose (H3) holds. Let A and B be defined by (2.4). Then for any
ψ(·) ∈ L2F (0, T ; lRn ) and v(·) ∈ L2F (0, T ; lRm ),
(A∗ ψ)(t) = η(t), t ∈ [0, T ], a.s. , (4.15)
and
(B ∗ v)(t) = λ(t), t ∈ [0, T ], a.s. , (4.16)
where (η(·), ζ(· , ·)) is the unique adapted M-solution to the following BSVIE:
Z T h d
X i Z T
η(t) = A0 (s, t)T ψ(s)+ Ai (s, t)T ζi (s, t) ds− ζ(t, s)dW (s), t ∈ [0, T ],
t i=1 t
(4.17)
satisfying
Z t
ψ(t) = Eψ(t) + ζ(t, s)dW (s), t ∈ [0, T ], (4.18)
0
and (λ(·), µ(· , ·)) is the unique adapted M-solution to the following BSVIE:
Z T h d
X i Z T
T T
λ(t) = B0 (s, t) v(s)+ Bi (s, t) µi (s, t) ds− µ(t, s)dW (s), t ∈ [0, T ],
t i=1 t
(4.19)
satisfying
Z t
v(t) = Ev(t) + µ(t, s)dW (s), t ∈ [0, T ]. (4.20)
0
Proof. By Proposition 4.2, BSVIE (4.17) admits a unique adapted solu-

tion (η(·), ζ(· , ·)) ∈ L2F (0, T ; lRn ) × L2 (0, T ; L2F (0, T ; lRn )) satisfying (4.18). Conse-
Account: ns145102
59
quently, for any X(·) ∈ L2F (0, T ; lRn ), one has

Z T Z T
E h(A∗ ψ)(t), X(t) i dt ≡ E h ψ(t), (AX)(t) i dt
0 0
Z T n Z t d Z
X t o
=E h ψ(t), A0 (t, s)X(s)ds + Ai (t, s)X(s)dWi (s) i dt
0 0 i=1 0
Z T Z T
=E h A0 (t, s)T ψ(t), X(s) i dtds
0 s
Z T Z t d Z
X t
+E h Eψ(t) + ζ(t, s)dW (s), Ai (t, s)X(s)dWi (s) i dt
0 0 i=1 0
Z T Z T d
X Z T Z t
=E h A0 (s, t)T ψ(s), X(t) i dsdt + E h ζi (t, s), Ai (t, s)X(s) i dsdt
0 t i=1 0 0
Z T Z T h d
X i
=E h A0 (s, t)T ψ(s) + Ai (s, t)T ζi (s, t) ds, X(t) i dt
0 t i=1
Z T Z T Z T
=E h η(t) + ζ(t, s)dW (s), X(t) i dt = E h η(t), X(t) i dt.
0 t 0
Since X(·) ∈ L2F (0, T ; lRn ) is arbitrary, and η(·) ∈ L2F (0, T ; lRn ), we obtain (4.15).
Representation (4.16) can be proved similarly.
Let us now consider the following FSVIE:
Z t d Z
X t
X(t) = f (t) + A0 (t, s)X(s)ds + Ai (t, s)X(s)dWi (s), t ∈ [0, T ], (4.21)
0 i=1 0
with f (·) ∈ L2F (0, T ; lRn ). It is clear that our state equation (1.1) is a special case
of the above with
Z t X d Z t
f (t) = ϕ(t) + B0 (t, s)u(s)ds + Bi (t, s)u(s)dWi (s), t ∈ [0, T ]. (4.22)
0 i=1 0
The following result is called a duality principle.
Theorem 4.2. Let Ai (· , ·) (0 ≤ i ≤ d) satisfy (H3) and f (·) ∈ L2F (0, T ; lRn ),
g(·) ∈ L2 ((0, T ) × Ω; lRn ). Let X(·) ∈ L2F (0, T ; lRn ) be the solution of FSVIE (4.21),
and (Y (·), Z(· , ·)) ∈ L2F (0, T ; lRn )×L2 (0, T ; L2F (0, T ; lRn×d )) be the adapted solution
to the following BSVIE:

Z T h d
X i
Y (t) = g(t) + A0 (s, t)T Y (s) + Ai (s, t)T Zi (s, t) ds
t i=1 (4.23)
Z T
− Z(t, s)dW (s), t ∈ [0, T ],
t
Account: ns145102
60
satisfying
Z t
Y (t) = EY (t) + Z(t, s)dW (s), t ∈ [0, T ]. (4.24)
0
Then the following relation holds:
Z T Z T
E h X(t), g(t) i dt = E h f (t), Y (t) i dt. (4.25)
0 0
We call (4.23)–(4.24) the adjoint equation of FSVIE (4.21), and call (4.25) a
duality relation between (4.21) and (4.23)–(4.24). We point out that process g(·)
appeared in the above is not required to be lF-adapted. The above result modifies
that in Ref. 33.
Proof. By Theorem 4.1, we see that (4.23) with (4.24) can be written as
Y = g + A∗ Y.
Since (I − A)−1 exists and bounded, we have
Y = (I − A∗ )−1 g.
Hence,
h X, g i = h(I − A)−1 f, g i = h f, (I − A∗ )−1 g i = h f, Y i,
proving (4.25).
5. A Maximum Principle and FBSVIEs

In this section, we would like to seek some alternative conditions for the solvability
of Problem (LQ). We begin with a necessary condition for optimality. Suppose
(X̄(·), ū(·)) is an optimal pair of Problem (LQ). Then, by a standard variational
technique, we have the following: For any u(·) ∈ U,
nZ T h i o
E h Q(t)X̄(t) + S(t)T ū(t), X(t) i + h S(t)X̄(t) + R(t)ū(t), u(t) i dt = 0,
0
(5.1)
where
Z th i d Z th
X i
X(t) = A0 (t, s)X(s) + B0 (t, s)u(s) ds + Ai (t, s)X(s) + Bi (t, s)u(s) dWi (s)
0 i=1 0
Z t d Z t
X
≡ f (t) + A0 (t, s)X(s)ds + Ai (t, s)X(s)dWi (s), t ∈ [0, T ],

0 i=1 0
(5.2)
with
Z t d Z
X t
f (t) = B0 (t, s)u(s)ds + Bi (t, s)u(s)dWi (s), t ∈ [0, T ]. (5.3)
0 i=1 0
Account: ns145102
61
By the duality principle established in the previous section, we have the following
result which can be regarded as a Pontryagin’s maximum principle.
Theorem 5.1. Let (H2)-(H3) hold. Let (X̄(·), ū(·)) be an optimal pair of Prob-
lem (LQ). Then there exists adapted solutions (Y (·), Z(· , ·)), (λ(·), µ(· , ·)) to the
following BSVIEs:
Z T h d
X i
T T
Y (t) = Q(t)X̄(t) + S(t) ū(t) + A0 (s, t) Y (s) + Ai (s, t)T Zi (s, t) ds
t i=1
Z T
− Z(t, s)dW (s),
t
(5.4)
and
Z T h d
X i Z T
λ(t) = B0 (s, t)T Y (s) + Bi (s, t)T Zi (s, t) ds − µ(t, s)dW (s), (5.5)
t i=1 t
such that
S(t)X̄(t) + R(t)ū(t) + λ(t) = 0, a.e. t ∈ [0, T ], a.s. (5.6)
We see that (1.1) (with (X(·), u(·)) replaced by (X̄(·), ū(·))) together with (5.4)–
(5.5) is a system of coupled forward and backward stochastic Volterra integral equa-
tions. The coupling is given through (5.6). We call such a system an FBSVIE.
Proof. By Proposition 4.1, BSVIEs (5.4) and (5.5) admit unique adapted
solutions
(Y (·), Z(· , ·)) ∈ L2F (0, T ; lRn ) × L2 (0, T ; L2F (0, T ; lRn×d )),
(λ(·), µ(· , ·)) ∈ L2F (0, T ; lRm ) × L2 (0, T ; L2F (0, T ; lRm×d )),
respectively. By Theorem 4.1, with g(·) = Q(·)X̄(·) + S(·)T ū(·), and f (·) given by
(4.22), we have
Z T
E h X(t), Q(t)X̄(t) + S(t)T ū(t) i dt
0
Z T Z t d Z
X t
=E h B0 (t, s)u(s)ds + Bi (t, s)u(s)dWi (s), Y (t) i dt
0 0 i=1 0
Z T Z th d
X i
=E h B0 (t, s)T Y (t), u(s) i + h Bi (t, s)T Zi (t, s), u(s) i dsdt
0 0 i=1
Z T Z T h d
X i
=E h B0 (s, t)T Y (s) + Bi (s, t)T Zi (s, t) ds, u(t) i dt.
0 t i=1
Account: ns145102
62
Hence, (5.1) implies

Z T
0=E h S(t)X̄(t) + R(t)ū(t)
0
Z T h d
X i
+ B0 (s, t)T Y (s) + Bi (s, t)T Zi (s, t) ds, u(t) i dt
t i=1 (5.7)
Z T Z T
=E h λ(t) + µ(t, s)dW (s) + S(t)X̄(t) + R(t)ū(t), u(t) i dt
0 t
Z T
=E h λ(t) + S(t)X̄(t) + R(t)ū(t), u(t) i dt, ∀u(·) ∈ U[0, T ].
0
Consequently, (5.6) follows.

Let us look at the above in a little point view, using the results from Section 3.
To this end, let us denote X u (·) and X ϕ (·) to be the solution of (1.1) corresponding
to (u(·), 0) and (0, ϕ(·)), respectively. Then X(·) = X u (·) + X ϕ (·). Now, by the
definition of Φ2 , we have that for any u(·) ∈ U,

(Φ2 u)(t) = B ∗ (I − A∗ )−1 [QX u + S T u] + SX u + Ru (t)
= (B ∗ Y u )(t) + S(t)X u (t) + R(t)u(t) = λu (t) + S(t)X u (t) + R(t)u(t),
(5.8)
with


 Z Th Xd i

 u u T T u

 Y (t) = Q(t)X (t) + S(t) u(t) + A0 (s, t) Y (s) + Ai (s, t)T Ziu (s, t) ds

 t

 i=1
 Z T
u
 − Z (t, s)dW (s),

 t

 Z Th Z T

 X d i

 u T u T u

 λ (t) = B 0 (s, t) Y (s) + B i (s, t) µ i (s, t) ds − µu (t, s)dW (s),
t i=1 t
(5.9)
satisfying
 Z t



 Y u
(t) = EY u
(t) + Z u (t, s)dW (s),
Z 0t t ∈ [0, T ]. (5.10)


 u u
 λ (t) = Eλ (t) + µu (t, s)dW (s),
0
Thus, Φ2 ≥ 0 is equivalent to the following:

Z T
E h λu (t) + S(t)X u (t) + R(t)u(t), u(t) i dt ≥ 0, ∀u(·) ∈ U, (5.11)
0
On the other hand,

(Φ1 ϕ)(t) = B ∗ (I − A∗ )−1 QX ϕ (t) + S(t)T X ϕ (t). (5.12)
Account: ns145102
63
Hence,

(Φ2 u + Φ1 ϕ)(t) = B ∗ (I − A∗ )−1 [QX + S T u] + SX + Ru (t)
(5.13)
= (B ∗ Y )(t) + S(t)X(t) + R(t)u(t) = λ(t) + S(t)X(t) + R(t)u(t),
with


 Z Th Xd i

 T T T

 Y (t) = Q(t)X(t) + S(t) u(t) + A 0 (s, t) Y (s) + A i (s, t) Z i (s, t) ds

 t

 i=1
 Z T
 − Z(t, s)dW (s),

 t

 Z Th Z T

 Xd i

 T T
 λ(t) =
 B0 (s, t) Y (s) + Bi (s, t) µi (s, t) ds − µ(t, s)dW (s),
t i=1 t
(5.14)
satisfying
 Z t



 Y (t) = EY (t) + Z(t, s)dW (s),
Z 0 t ∈ [0, T ]. (5.15)

 t

 λ(t) = Eλ(t) + µ(t, s)dW (s),
0
Note that the equation for (λ(·), µ(· , ·)) in (5.14) is different from the equation (5.5).
But, as we explained in the previous section, we can redefine µ(· , ·) in (5.5) so that
it takes the same form as that in (5.14). From the above, we obtain the following.
Theorem 5.2. Problem (LQ) is solvable if and only if (5.11) holds and for the
given ϕ(·) ∈ L2F (0, T ; lRn ), there exists a u(·) ∈ U ≡ L2F (0, T ; lRm ) such that
λ(t) + S(t)X(t) + R(t)u(t) = 0, t ∈ [0, T ], (5.16)
with λ(·) given through (5.14)–(5.15). In this case, (X(·), u(·)) is an optimal pair.
The above gives alternative conditions for the solvability of Problem (LQ). In
some sense, conditions in Theorem 5.2 are a little less abstract than those presented
in Section 3.
Note that if R(t)−1 exists and uniformly bounded. Then (5.16) is equivalent to
h i
u(t) = −R(t)−1 S(t)X(t) + λ(t) , t ∈ [0, T ]. (5.17)
Account: ns145102
64
Substituting the above into (1.1) and (5.14), we obtain the following:
 Z tn
 o

 = + [A0 (t, s) − B0 (t, s)R(s)−1 S(s)]X(s) − B0 (t, s)R(s)−1 λ(s) ds

 X(t) ϕ(t)

 0

 d Z tn

 X o



 + [A j (t, s) − B j (t, s)R(s) −1
S(s)]X(s) − B j (t, s)R(s) −1
λ(s) dWj (s),


 j=1 0

Y (t) = [Q(t) − S(t)T R(t)−1 S(t)]X(t) − S(t)T R(t)−1 λ(t)

 Z Th Z T

 X d i

 T T

 + A0 (s, t) Y (s) + Ai (s, t) Zi (s, t) ds − Z(t, s)dW (s),

 t t

 i=1

 Z Th d i Z T

 X

 = (s, T
(s) + (s, T
(s, − µ(t, s)dW (s),
 λ(t) B 0 t) Y B i t) µ i t) ds
t i=1 t
(5.18)
with constraints:
 Z t



 Y (t) = EY (t) + Z(t, s)dW (s),
Z 0 (5.19)

 t

 λ(t) = Eλ(t) + µ(t, s)dW (s).
0
This is a coupled FBSVIE. The general solvability problem for the above equation is
still under our careful investigation. We hope to report some further results related
to this in the near future.
References
1. M. Ait Rami, J. B. Moore, and X. Y. Zhou, Indefinite stochastic linear quadratic
control and generalized differential Riccati equation, SIAM J. Control Optim. 40,
1296–1311 (2001/02).
2. B. D. O. Anderson and J. B. Moore, Linear Optimal Control, Prentice-Hall, Englewood
Cliffs, New Jersey, 1971.
3. S. A. Belbas, Iterative schemes for optimal control of Volterra integral equations,
Nonlinear Anal. 37, 57–79 (1999).
4. S. A. Belbas and W. H. Schmidt, Optimal control of Volterra equations with impulses,
Appl. Math. Computation 166, 696–723 (2005).
5. R. Bellman, I. Glicksberg, and O. Gross, Some Aspects of the Mathematical Theory
of Control Processes, Rand Corporation, Santa Monica, CA, 1958.
6. M. Berger and V. Mizel, Volterra equations with Itô integrals, I, II, J. Int. Eqs. 2,
187–245; and 319–337 (1980).
7. D. A. Carlson, Infinite-horizon optimal controls for problems governed by a Volterra

integral equation with state-and-control-dependent discount factor, J. Optim. Theory
Appl. 66, 311–336 (1990).
8. S. Chen, X. Li, and X. Y. Zhou, Stochastic linear quadratic regulators with indefinite
control weight costs, SIAM J. Control Optim. 36, 1685–1702 (1998).
9. S. Chen and J. Yong, Stochastic linear quadratic optimal control problems with ran-
dom coefficients, Chinese Ann. Math. Ser. B 21, 323–338 (2000).
Account: ns145102
65
10. S. Chen and J. Yong, Stochastic linear quadratic optimal control problems, Appl.
Math. Optim. 43, 21–45 (2001).
11. S. Chen and X. Y. Zhou, Stochastic linear quadratic regulators with indefinite control
weight costs, II, SIAM J. Control Optim. 39, 1065–1081 (2000).
12. N. Dunford and J. T. Schwartz, Linear Operators, Part II, Spectral Theory, Selfadjoint
Operators in Hilbert Space, John Wiley & Sons, Inc., New York, 1963.
13. D. Duffie and C.-F. Huang, Stochastic production-exchange equilibria, Research paper
No. 974, Graduate School of Business, Stanford Univ, 1986.
14. R. E. Kalman, Contributions to the theory of optimal control, Bol. SOc. Mat. Mexi-
cana 5, 102–119 (1960).
15. M. T. Kamien, and E. Muller, Optimal control with integral state equations, Rev.
Economic Studies 43, 469–473 (1976).
16. A. M. Letov, Analytic design of regulators, Avtomat. i Telemekh, (1960), 436–446,
561–571, 661–669 (in Russian); English transl. in Automat. Remote Control, 21 (1960).
17. I. Lasiecka and R. Triggiani, Control Theory for Partial Differential Equations:
Continuous and Approximation Theories, I. Abstract parabolic systems; II. Abstract
hyperbolic-like systems over a finite time horizon, Encyclopedia of Mathematics and
its Applications, 74, 75, Cambridge University Press, Cambridge, 2000.
18. X. Li and J. Yong, Optimal Control Theory for Infinite-dimensional Systems,
Birkhäuser, Boston, 1995.
19. J. Lin, Adapted solution of a backward stochastic nonlinear Volterra integral equation,
Stochastic Anal. Appl. 20, 165–183 (2002).
20. J. L. Lions, Optimal Control of Systems Governed by Partial Differential Equations,
Springer-Verlag, New York, 1971.
21. N. G. Medhin, Optimal processes governed by integral equations, J. Math. Anal. Appl.
120, 1–12 (1986).
22. L. W. Neustadt and J. Warga, Comments on the paper “Optimal control of processes
described by integral equations. I” by V. R. Vinokurov, SIAM J. Control 8, 572 (1970).
23. E. Pardoux and P. Protter, Stochastic Volterra equations with anticipating coefficients,
Ann. Probab. 18, 1635–1655 (1990).
24. A. J. Pritchard and Y. You, Causal feedback optimal control for Volterra integral
equations, SIAM J. Control Optim. 34, 1874–1890 (1996).
25. P. Protter, Volterra equations driven by semimartingales, Ann. Probab. 13, 519–530
(1985).
26. V. R. Vinokurov, Optimal control of processes described by integral equations, I, II,
III, Izv. Vysš. Učebn. Zaved. Matematika, 7 (62) (1967), 21–33; 8 (63) (1967), 16–23; 9
(64) (1967), 16–25; (in Russian) English transl. in SIAM J. Control 7 (1969), 324–336;
337–345; 346–355.
27. J. Willems, Least squares stational optimal control and the algebraic Riccati equation,
IEEE Trans. Auto. Control 16, 621–634 (1971).
28. W. M. Wonham, On a matrix Riccati equation of stochastic control, SIAM J. Control
6, 681–697 (1968).
29. W. M. Wonham, Linear Multivariable Control: A Geometric Approach, 2nd Edition,
Springer-Verlag, 1979.
30. H. Wu and X. Li, A linear quadratic problem with unbounded control in Hilbert
spaces, Diff. Int. Eqs. 13, 529–566 (2000).
31. D. Yao, S. Zhang, and X. Y. Zhou, Stochastic linear-quadratic control via primal-dual
semidefinite programming, SIAM Rev. 46, 87–111 (2004).
32. J. Yong, A leader-follower stochastic linear quadratic differential game, SIAM J. Con-
trol Optim. 41, 4, 1015–1041 (2002).
Account: ns145102
66
33. J. Yong, Backward stochastic Volterra integral equations and some related problems,
Stoch. Proc. Appl., to appear.
34. J. Yong and X. Y. Zhou, Stochastic Controls: Hamiltonian Systems and HJB Equa-
tions, Springer-Verlag, New York, 1999.
35. Y. You, Quadratic integral games and causal synthesis, Trans. Amer. Math. Soc. 352,
2737–2764 (2000).
Account: ns145102
August 1, 2007 10:46 Proceedings Trim Size: 9.75in x 6.5in Chap16-ChenZengjing
AN ADDITIVITY OF MAXIMUM EXPECTATIONS

AND ITS APPLICATIONS
ZENGJING CHEN∗
Department of Mathematics
Shandong University, Jinan
250100, P.R. China
E-mail: zjchen@sdu.edu.cn
MATT DAVISON, MARK REESOR

Department of Applied Mathematics
The University of Western Ontario
London, Canada
YING ZHANG
Department of Mathematics and Statistics
Acadia University, Canada
In this paper, we explore an additivity of a class of maximum (minimum) expectations

which come from the pricing of contingent claim in incomplete markets. We give exam-
ples to show this property can be used to calculate maximum (minimum) expectations
and Choquet integral. Furthermore, we also explore its applications in partial differential
equation (shortly PDE) and the pricing of the contingent claims in incomplete markets.
Keywords: Maximum expectation, partial differential equation, backward stochastic

differential equation; contingent claim, Malliavin calculus.
1. Introduction
Fixed time horizon T > 0, let {Wt }0≤t≤T be a d-dimensional standard Brown-
ian motion defined on a completed probability space (Ω, F, P ) and {Ft }0≤t≤T be
the natural filtration generated by Brownian motion {Wt }0≤t≤T , that is, Ft =
σ(Ws ; s ≤ t), we assume F = FT . To ease of exposition, in this paper, we assume
d = 1 and Ω is a Wiener space, that is Ω := C0 (0, T ), the set of all continuous

functions {ωt } defined on [0, T ] with w0 = 0 and for any ω ∈ Ω, Brownian motion
Wt (ω) = ωt . We present the following notations that will be used in the rest of this
∗ Financial support partly from NSF of China (10325106) and (10131030), FANEDD (2001059).
67
Account: ns145102
68
paper. Let p ∈ (1, ∞) and set

n RT o
Lp (0, T ) := V : all Ft −adapted processes {Vt } such that E 0 |Vs |p ds < ∞ ;
B 2 :=nL2 (0, T ) × L2 (0, T );
P := Q : all probability measures Q s.t. {E[ dQ
dP |Ft ]} is a square integrable
o
martingale .
Lp (Ω, F, P ) := {ξ : all F−measurable random variables such that E|ξ|p < ∞} ;
We define maximum (minimum) expectations by
E[ξ] := sup EQ [ξ], E[ξ] := inf EQ [ξ]; (1)

Q∈P Q∈P
and conditional expectations by
E[ξ|Ft ] := ess sup EQ [ξ|Ft ], E[ξ|Ft ] := ess inf EQ [ξ|Ft ]. (2)

Q∈P Q∈P
In order to make maximum expectation E[ξ] be well-defined for any ξ ∈

Lp (Ω, F, P ), we further need some assumptions on P. Indeed, for any Q ∈ P,
let Mt := E[ dQ
dP |Ft ], because {Mt }0≤t≤T is a square integrable martingale, by Mar-
tingale Representation Theorem (see for example, Theorem 4.3.3, p. 53, Ref. 13),
there exists a unique predictable process h ∈ L2 (0, T ) such that
Z t
Mt = 1 + hs dWs .
0
hs
RT 2
Let as := Ms , if E 0 |as | ds < ∞, then
Z t
Mt = 1 + as Ms dWs .
0
Solving the above linear stochastic differential equation (shortly SDE), we obtain
Mt is of the following form:
Rt Rt
1
|as |2 ds+ as dWs
Mt = e − 2 0 0 , 0 ≤ t ≤ T.
dQ
Which implies that for Q ∈ P, there exists a process {at } such that dP is of the
form:
dQ 1
RT 2
RT
= e− 2 0 as ds+ 0 as dWs . (3)
dP
We rewrite Q as Qa and call Qa the probability measure generated by {at }. Hence,
P actually is the set of probability measures generated by some processes {a t } via

(3). In this paper,
RT we further assume there exists a deterministic positive function
{k(t)} with 0 k 2 (s)ds < ∞ such that
a

a dQ − 12 0T as 2 ds+ 0T as dWs
R R
P= Q : =e , |at | ≤ k(t), t ∈ [0, T ] . (∗)
dP
Account: ns145102
69
Immediately, the maximum (minimum) expectations defined in (1) and (2) are well-
defined. Indeed, for any Qa ∈ P and ξ ∈ Lp (Ω, F, P ), by Schwartz inequality and
the fact that
RT RT
1
|qas |2 ds+ qas dWs
E(e− 2 0 0 ) = 1, ∀q ∈ (1, ∞),
we get
a
|EQa ξ| ≤ E|ξ|| dQ
dP |
1 a 1
≤ (E|ξ|p ) p (E| dQ q q
dP | )
1 2 −q 1
qas dWs + q
RT RT RT
1
|qas 2 |ds+ |as |2 ds
≤ (E|ξ|p ) p (Ee− 2 0 0 2 0 )q
1 q−1 RT
|k(s)|2 ds
= (E|ξ|p ) p e 2 0
q RT
1
|k(s)|2 ds
≤ (E|ξ|p ) p e p 0 < ∞,
1 1
where p + q = 1. Thus
q RT
p 1
|k2 (s)|ds
|E[ξ]| = | sup EQ ξ| ≤ (E|ξ| ) p e p 0 < ∞,
Q∈P
which implies E[ξ] and E[ξ|Ft ] are well-defined for any ξ ∈ Lp (Ω, F, P ), so are E[ξ]
and E[ξ|Ft ].
In this paper, we only consider maximum expectation E[·] as minimum expec-
tation E[·] can be treated by the fact that
E[ξ] = −E[−ξ], E[ξ|Ft ] = −E[−ξ|Ft ], ∀ξ ∈ Lp (Ω, F, P ). (∗∗)
To simplify notations, let us write E[·] and E[·|Ft ] simply as E[·] and E[·|Ft ]
respectively. Obviously, by the definition of E[·], E[·] is a nonlinear operator on
Lp (Ω, F, P ) and
E[ξ + η] ≤ E[ξ] + E[η], ∀ξ, η ∈ Lp (Ω, F, P ),
moreover, the equality holds if and only if P = {P }.
We are interesting in finding some conditions on ξ and η under which the fol-
lowing relations hold.
E[ξ + η] = E[ξ] + E[η],
or
E[ξ + η] < E[ξ] + E[η].
We will indicate the applications of the above results in Section 3.
The maximum (minimum) expectations defined in (1) and (2) actually come
from the pricing of contingent claim in incomplete markets. As we know, for given
contingent claim ξ, which usually depends on the price of stocks, if the market is a
complete market, then there exists a unique equivalent martingale measure Q such
that the pricing of this contingent claim at time t is given by conditional expectation
EQ [ξ|Ft ]. However, if the market is an incomplete market, such probability measure
Account: ns145102
70
Q is not unique. Thus it is impossible to price and hedge this claim. In that case,
an investor hopes to know the superpricing E[ξ] and subpricing E[ξ] (cf. El Karoui
and Quenez10 ; Karatzas8; Cvitanic6 ; Föllmer and Kramkov7). Hence, to explore
the pricing of contingent claim in incomplete markets, it is important to study the
properties of maximum expectations E[ξ] and minimum expectations E[ξ].
2. Main Result
In this paper, we will use backward stochastic differential equations (BSDE) and
Malliavin calculus to study the properties of maximum expectations. We refer
readers to Refs. 14 and 9 for BSDEs and Refs. 11 and 12 for Malliavin calculus. Let
us now give the following Lemmas:
Lemma 2.1. If ξ ∈ L2 (Ω, F, P ) and (yt , zt ) is the unique solution of BSDE,

Z T Z T
yt = ξ + k(s)|zs |ds − zs dWs , (4)
t t
where k(t) is defined in (∗), then
E[ξ|Ft ] = yt ; E[ξ] = y0 .
Proof. For any n > 0, let k (n) (t) = k(t) ∧ n and

v

v dQ − 12 0T |vs |2 ds+ 0T vs dWs
R R
(n) (n)
P := Q : =e , |v(t)| ≤ k (t) .
dP
Obviously, P (n) ⊂ P (n+1) ⊂ · · · ⊂ P and P (n) ↑ P.

Since |k (n) (t)| ≤ n, thus BSDE(5) below satisfies uniform Lipschitz condi-
tion. By Theorem 2.1 of Ref. 9, for each n > 0 , there exists a unique solution
(yt (n) , zt (n) ) ∈ B 2 such that
Z T Z T
(n) (n) (n)
yt =ξ+ k (s)|zs |ds − zs (n) dWs . (5)
0 0
Moreover, by Proposition 35.2 of Ref. 15, there exists a constant c > 0 such that
Z T
2
E |zs (n) | ds ≤ c E|ξ 2 |.
0
(n)
Let E [ξ|Ft ] := ess supQ∈P (n) EQ [ξ|Ft ], by Proposition 3.2.2 of Ref. 10, or
Lemma 1 in Ref. 4,
E (n) [ξ|Ft ] = yt (n) .
Note that E (n) [ξ|Ft ] ↑ E[ξ|Ft ], as n → ∞, it remains to prove that the limit of
{yt (n) } is the solution of BSDE (4).
Account: ns145102
71
In fact, for any n, m ≥ 1, by Proposition 35.2 of Ref. 15, there exists a constant
C > 0 such that
2 RT 2
E[sup0≤t≤T |yt (n) − yt (m) | + 0 |zs (n) − zs (m) | ds]
h R i2
T (n) (n)
≤ C E 0 (|k (n) (s)|zs | − k (m) (s)|zs |)ds
RT R T (n)
≤ C 0 |k (n) (s) − k (m) (s)|2 dsE 0 |zs |2 ds
R T
≤ CcE|ξ|2 0 |k (n) (s) − k (m) (s)|2 ds
→ 0, as n, m → ∞,
which implies that (y (n) , z (n) ) is a Cauchy sequence of B 2 , thus there exists (y, z) ∈
B 2 such that (y (n) , z (n) ) → (y, z), as n → ∞. Since
R 2
T
E 0 |k(s)|zs | − k (n) (s)|zs (n) ||ds
R RT 2
T
≤ E 0 k(s)|zs (n) − zs |ds + E 0 |k (n) (s) − k(s)||zs (n) |ds
RT RT 2 RT 2 RT .
≤ 0 k 2 (s)dsE 0 |zs (n) − zs | ds + 0 |k (n) (s) − k(s)| dsE 0 |zs (n) |2 ds
RT 2 R T (n) 2 R T 2
≤ 0 k (s)dsE 0 |zs − zs | ds + cEξ 2 0 |k (n) (s) − k(s)| ds
→ 0, as n → ∞.
Thus,
Z T Z T
(n) (n)
k (s)|zs |ds → k(s)|zs |ds, as n → ∞
0 0
in L2 (Ω, F, P ).
Set n → ∞ on both sides of BSDE (5), then (y, z) is the solution of BSDE(4).
Thus yt = E[ξ|Ft ]. In particular, let t = 0, we get y0 = E[ξ].
Using the same method and the fact (∗∗), we can get
Corollary 2.1. Assume ξ ∈ L2 (Ω, F, P ), then E[ξ|Ft ] is the unique solution of

BSDE:
Z T Z T
yt = ξ − k(s)|zs |ds − zs dWs .
0 t
RT
Lemma 2.2. Let ξ ∈ L2 (Ω, F, P ) be of the form ξ = f ( 0 σs dWs ), where f (x) is
a continuous differentiable function, {σt } is a deterministic function. Let (yt , zt ) be
the solution of the following BSDE
Z T Z T
yt = ξ + k(s)|zs |ds − zs dWs . (6)
t t
(i) If f is increasing (resp. decreasing), then σt zt ≥ 0, (resp. σt zt ≤ 0).

(ii) Assume f is strictly increasing (resp. decreasing), if σt 6= 0, then zt 6= 0, a.e.
t ∈ [0, T ].
Account: ns145102
72
Proof. The proof of (i): For any ε > 0,

(
|z|, if |z| ≥ ε;
gε (z) := 1 z2
2 (ε + ε ), if |z| < ε.
Obviously, gε (z) → |z|, as ε → 0. Let (yt (ε) , zt (ε) ) be the solution of BSDE
Z T Z T Z T
yt (ε) = f ( σs dWs ) + k(s)gε (zs )ds − zs (ε) dWs .
0 t t
(ε)
Thus, by Proposition 35.2 of Ref. 15, (yt , zt ) → (y, z) in B 2 as ε → 0, where (ε)
(y, z) is the solution of BSDE (6). Thanks to Proposition 5.3 of Ref. 9, we have for
ε > 0,
zt (ε) = Dt yt (ε) .
where Dt yt (ε) denotes the Malliavin derivative of yt (ε) . Moreover, Dt yt (ε) satisfies
the following linear BSDE
Z T Z T Z T
Dt yt (ε) = Dt f ( σs dWs ) + k(s)gε 0 (zs (ε) )Dt zs (ε) ds − Dt zs (ε) dWs ,
0 t t
where gε (z) denotes the derivative of gε (z) with respect to z.

0
Solving the above linear BSDE, we have

Z T
(ε)
Dt yt = EQ [Dt f ( σs dWs )|Ft ],
0
RT 2 ε 2 RT (ε)
dQ − 12 0
[k (s)gε (zs )] ds+ 0 k(s)gε (zs )dWs 0
where dP =e 0 .
Let f be the derivative of f with respect to x, by the definition of Malliavin
0
derivative, we have
Z T Z T
Dt f ( σs dWs ) = f 0 ( σs dWs )σt ,
0 0
thus
Z T Z T
zt (ε) = Dt yt (ε) = EQ [f 0 ( σs dWs )σt |Ft ] = σt EQ [f 0 ( σs dWs )|Ft ], (7)
0 0
which implies if f 0 (x) ≥ 0, ∀x ∈ R, then σt zt (ε) ≥ 0 and σt zt (ε) ≤ 0 for f 0 (x) ≤

0, ∀x ∈ R. It then follows by the fact z (ε) → z in L2 (0, T ) that we complete the
proof of (i).
The proof of (ii):R Without loss of generalization, we assume f 0 > 0, from
T
(ε)
(i), σt zt = σt2 EQ [f 0 ( 0 σs dWs )|Ft ], note that |gε0 | ≤ 1, thus Q ∈ P, by strict
(ε) RT
Comparison Theorem 35.5 of Ref. 15, σt zt ≥ σt2 E[f 0 ( 0 σs dWs )|Ft ] > 0, the
proof of (ii) is complete.
RT RT
p
Let ξ, η ∈ L (Ω, F, P ) be of the forms ξ := f ( 0 σs dWs ), η := h( 0 vs dWs ),
where {σt } and {vt } are two deterministic functions, f and h are two functions
Account: ns145102
73
with continuous differentiation, under the above assumptions, We have the following
main result.
Theorem 2.1. Suppose that ξ and η are the random variables defined above, f and
h are both increasing (decreasing) functions, (i) If σt vt ≥ 0, t ∈ [0, T ], then
E[ξ + η] = E[ξ] + E[η].
(ii) If f and h are strictly increasing (decreasing) functions and σt vt < 0, t ∈
[0, T ], then
E[ξ + η] < E[ξ] + E[η].
Proof. We turn our proof into three steps.

Step 1. We assume f and h are bounded by some constant N > 0, i.e. |f | ≤
N, |h| ≤ N.
Let yt = E[ξ|Ft ], y t = E[η|Ft ]. By Lemma 2.1, {yt } and {yt } are the solutions
of BSDEs respectively,
Z T Z T
yt = ξ + k(s)|zs |ds − zs dWs (8)
t t
and
Z T Z T
yt = η + k(s)|z s |ds − z s dWs . (9)
t t
By Lemma 2.2(i), we have

σt zt ≥ 0,
and
vt z t ≥ 0, a.e. t ∈ [0, T ].
It then follows by σt vt ≥ 0 that we have z t zt ≥ 0, which implies
|z t | + |zt | = |z t + zt |, a.e. t ∈ [0, T ]. (10)
Adding BSDE (8) with BSDE (9), we have
Z T Z T
yt + y t = ξ + η + k(s)|zs + z s |ds − (zs + z s )dWs ,
t t
which means (yt + y t , zt + z t ) is the solution of BSDE

Z T Z T
yt = ξ + η + k(s)|zs |ds − zs dWs ,

t t
by Existence and Uniqueness Theorem of BSDE (Theorem 2.1 of Ref. 9) and

Lemma 1,
yt + yt = E[ξ + η|Ft ].
Account: ns145102
74
That is,
E[ξ|Ft ] + E[η|Ft ] = E[ξ + η|Ft ]. (11)
Step 2. For sufficient large N > 0 and N = 1, 2, · · · , let us denote
(n) (N ) (n) (N )
(x) (x)
fN (x) := xe−nd , hN (x) := xe−nd , (12)
where d(N ) (x) is defined by

(N ) 0, |x| ≤ N ;
d (x) :=
(x − N )1+α , |x| > N, 0 < α < 1.
(n) (n)
It is easy to check that fN (x) → xI(|x|≤N ) and hN (x) → xI(|x|≤N ) as N → ∞,
(n) (n)
here and below IA is an indicator function. Obviously fN and hN are bounded
and continuous differentiable. By Step 1,
Z T Z T Z T
(n) (n) (n)
E[fN ( σs dWs ) + hN ( vs dWs )|Ft ] = E[fN ( σs dWs )|Ft ]
0 0 0
Z T
(n)
+ E[hN ( vs dWs )|Ft ].
0
(n) R T (n) R T
Note that fN ( 0 σs dWs ) → ξ and hN ( 0 vs dWs ) → η as n, N → ∞ in
Lp (Ω, F, P ) . We have
E[ξ + η|Ft ] = E[ξ|Ft ] + E[η|Ft ].
In particular, let t = 0,
E[ξ + η] = E[ξ] + E[η].
Step 3. If f and g are strictly increasing (decreasing), by Lemma 2.2(ii) and
the assumptions, we have σt vt < 0. Thus, z t zt < 0 , ∀t ∈ [0, T ]. Moreover, Equality
(10) becomes
|z t | + |zt | > |z t + zt |, t ∈ [0, T ], (13)
this with strict Comparison Theorem 35.5 of Ref. 15, we have
E[ξ|Ft ] + E[η|Ft ] > E[ξ + η|Ft ]; t ∈ [0, T ].
Let t = 0, the proof is complete.
Using the same method as the proof of Theorem 2.1, we can get the following
Corollary:
RT
Corollary 2.2. Suppose that ξ = f ( 0 σs dWs ) is defined in Theorem 2.1, let a1 ≤
a2 ≤ · · · < an be a sequence, then
n
X n
X
E[ I(ξ≥ai ) ] = E[Iξ≥ai ) ].
i=1 i=1
Account: ns145102
75
3. Applications
To illustrate the applications of the above main results, we give the following ex-
amples:
3.1. Compute maximum expectation

In general, it is not easy to calculate maximum expectations, but for special case,
we can calculate maximum expectations by using main results in this paper, let us
give the following examples.
RT RT
Example 3.1. Compute supQ∈P EQ e−2 0 k(s)dWs − e2 0 k(s)dWs .
In fact, let f (x) := e−2x and h(x) := −e2x , then f and h are monotone, by
Theorem 2.1,
RT RT RT RT
supQ∈P EQ e−2 0 k(s)dWs − e2 0 k(s)dWs = E[e−2 0 k(s)dWs ] + E[−e2 0 k(s)dWs ]
RT RT
k(s)dWs k(s)dWs
= E[e−2 0 ] − E[e2 0 ].
RT
k(s)dWs
By Lemma 2.1, E[e−2 0 ] is the value of (yt ), the solution of the following
BSDE at time t = 0,
RT
Z T Z T
k(s)dWs
yt = e−2 0 + k(s)|zs |ds − zs dWs .
t t
Solving the above BSDE, we get
Rt Rt
(yt , zt ) = e−2 0 k(s)dWs , 2k(t)e−2 0 k(s)dWs .
RT RT
Thus E[e−2 0 k(s)dWs ] = y0 = 1. Similarly, by Corollary 2.1, E[e2 0
k(s)dWs
] is the
value of {yt }, the solution of the following BSDE, at time t = 0,
Z T Z T
2 0T k(s)dWs
R
yt = e − k(s)|zs |ds − zs dWs .
t t
Solving the above BSDE, we obtain
Rt Rt
(yt , zt ) = e2 0 k(s)dWs , −2k(t)e2 0 k(s)dWs ,
RT RT RT
k(s)dWs k(s)dWs k(s)dWs
and E[e2 0 ] = 1. Thus E[e−2 0 − e2 0 ] = 1 − 1 = 0.
Remark
RT 2
3.1. From
RT
this example, we get the following interesting facts: (1) Since
Ee−2 0 k (s)ds−2 0 k(s)dWs = 1, thus
RT RT
k(s)dWs k2 (s)ds
Ee−2 0 = e2 0 ,
which depends on k, but by Example 3.1,

RT
k2 (s)dWs
E[e−2 0 ] = 1,
which does not depend on k.
−2 0T k(s)dWs
R RT
(2) Since E[e ] = E[e2 0 k(s)dWs ] = 1, thus for any Q ∈ P,
RT RT
k(s)dWs k(s)dWs
EQ [e−2 0 ] ≤ EQ [e2 0 ].
Account: ns145102
76
3.2. Compute Choquet integral

For any A ∈ F, let V (A) := supQ∈P Q(A). Obviously V (·) is a capacity and V (A) =
E[IA ]. Choquet5 introduced the following integral (shortly Choquet expectation)
which generalizes the usual mathematical expectations,
Z 0 Z ∞
C[ξ] := (V (ξ ≥ t) − 1)dt + V (ξ ≥ t)dt. (14)
−∞ 0
RT
k(s)dWs
Example 3.2. Compute C[e−2 0 ].
RT
k(s)dWs
Let ξ := e−2 0 , for sufficient large N and n = 1, 2, · · · , and
n
2
X
(n) iN
ξN := I( iNn ≤ξ∧N < (i+1)N ,
i=0
2n 2 n 2)
(n)
then ξN → ξ as N, n → ∞ in L2 (Ω, F, P ). Note that ξ ∧ N ≤ N and
P2n iN P2n iN
i=0 2 n I iN (i+1)N
( 2n ≤ξ∧N < 2n )
= i=0 2 n I iN
(ξ∧N ≥ 2n ) − I (i+1)N
(ξ∧N ≥ 2n )
P2n iN P2n +1 (i−1)N
= i=1 2n I(ξ∧N ≥ iNn ) − i=2 2n I(ξ∧N ≥ iN
2 2n )
P2n iN P2n (i−1)N
= i=1 2n I(ξ∧N ≥ iNn ) − i=2 2n I(ξ∧N ≥ iNn )
2 2
N P2
n
= 2n i=1 I(ξ∧N ≥ iNn ) .
2
Applying Corollary 2.2,

n n
2
X 2
iN N X
E[ I iN (i+1)N ] = E[I(ξ∧N ≥ iNn ) ]
i=0
2n ( 2n ≤ξ∧N < 2n ) 2n i=1 2
We have, from the definition (14) and applying the fact that
V (ξ ∧ N ≥ t) = E[I(ξ∧N ≥t) ],
RT R∞ RT
k(s)dWs
C[e−2 0 ]= 0
V e−2 0 k(s)dWs > t dt
RN
= limN →∞ 0 V (ξ ∧ N ≥ t)dt
P 2n
= limN →∞ limn→∞ i=0 2Nn V (ξ ∧ N ≥ iN 2n )
P 2n N
= limN →∞ limn→∞ i=0 2n E[I(ξ∧N ≥ iNn ) ]
2
P2n iN
= limN →∞ limn→∞ E[ i=0 2n I( iNn ≤ξ∧N ≤ (i+1)N )
]
n 2 2
RT
k(s)dWs
= limN →∞ E[ξ ∧ N ] = E[e−2 0 ].
RT RT
k(s)dWs k(s)dWs
According to Example 3.1, E[e−2 0 ] = 1, thus C[e−2 0 ] = 1.
Remark 3.2. This example also shows that under some assumptions on ξ, the
maximum expectation of ξ is equal to Choquet expectation.
Account: ns145102
77
3.3. Application to PDE

Suppose that σt , bt , K(t), Φ are continuous and bounded functions.
Let UΦ be the solution of the following PDE with respect to terminal value Φ,
( ∂U 1 2 ∂2U ∂U ∂U
∂t + 2 σt ∂x2 + bt ∂x = k(t)| ∂x σt x|;
(15)
U (T, x) = Φ(x), x ∈ R+ , 0 ≤ t ≤ T.
Obviously, In general, UΦ1 +Φ 6= UΦ1 + UΦ2 . However, we have
Example 3.3. Assume Φ1 and Φ2 are two continuous bounded functions, UΦ1 and
UΦ2 are the solution of PDE (15) corresponding to Φ = Φ1 and Φ = Φ2 . If Φ is an
increasing or decreasing function, then
UΦ1 +Φ2 = UΦ1 + UΦ2 ,
that is, the solution of PDE (15) is linear with respect to Φ.
In fact, by Theorem 12.3 of Ref. 1,
UΦ (t, x) = yt t,x ,
where {ys t,x }0≤s≤T is the solution of BSDE

Z T Z T
yt t,x = Φ(XT t,x ) + k(s)|zs t,x |ds − zs t,x dWs .
t t
And {Xs t,x } is the solution of SDE

dXs t,x = bt Xs ds + σs Xs dWs
Xt = x, t ≤ s ≤ T.
By Lemma 1 and Proposition 4.2 of Ref. 9,
yt t,x = E[Φ(XT t,x )|Ft ] = E[Φ(XT t,x )].
That is, UΦ (x, t) = E[Φ(XTt,x )].

Thus if Φ1 and Φ2 are increasing (decreasing) functions with continuous differ-
entiation, applying Theorem 2.1, we have
E[Φ1 (XTt,x ) + Φ2 (XTt,x )] = E[Φ1 (XTt,x )] + E[Φ2 (XTt,x )].

That is,
UΦ1 +Φ2 = UΦ1 + UΦ2 .
Account: ns145102
78
3.4. Application to European Option

As explained in introduction, maximum expectations defined in this paper come
from the pricing of contingent claim in incomplete market, for given contingent
claims ξ and η, the superpricing of contingent claims given the maximum expecta-
tion E[ξ] and E[η], since
E[ξ + η] ≤ E[ξ] + E[η],
thus, in an incomplete market, it is better for an investor to buy ξ and η together
than to buy them separately.
Let us now further consider the application of our result in incomplete markets.
If given two contingent claims XT and YT which satisfy the following SDEs

dXt = at Xt dt + σt Xt dWt ;
X0 = x > 0, 0 ≤ t ≤ T.
and

dYt = bt Yt + vt Yt dWt ;
Y0 = y > 0, 0 ≤ t ≤ T.
Where at , bt , σt , vt are continuous bounded functions.
R Solving
T
the above linear R T SDEs, we get XT and − YTR are of Rthe forms XT =
T 1 T 2
f ( t σs dWs ) and YT = h( 0 vs dWs ). where f (x) = e 0 at dt− 2 0 σt dt+x , h(x) =
RT 1
R T 2
e− 0 bt dt− 2 0 vt dt+x .
The maximal pricing for investors to buy XT and YT together is E[YT + XT ].
However, by Theorem 2.1, if σt vt ≥ 0, then
E[YT + XT ] = E[YT ] + E[XT ],
which implies that the maximal price that the investors buy XT and YT together is
equal to buy them separately. However, if σt vt < 0, then
E[YT + XT ] < E[YT ] + E[XT ].
In this case, the investors should buy the two contingent claims together to hedge
risk.
References
1. E. Barles and E. Lesigne, SDE, BSDE and PDE, Pitman Research Notes in Mathe-
matics Series No. 364, 47–80 (1997).
2. Z. Chen, A property of backward stochastic differential equations, C.R. Acad. Sci. Paris
No. 1, 483–488 (1998).
3. Z. Chen and L. Epstein, Ambiguity, risk and asset return in continuous time,
Econometrica 70, 1403–1443 (2002).
4. Z. Chen and S. Peng, A general downcrossing inequality for g-martingales, Statist.
Probab. Lett. 46, no. 2, 169–175 (2000).
5. G. Choquet, Theory of capacities, Ann. Inst. Fourier (Grenoble) 5, 131–295 (1955).
Account: ns145102
79
6. J. Cvitanic, Minimizing expected loss of hedging in incomplete and constrained

markets, Preprint, Columbia Univ. New York (1998).
7. H. Föllmer and D. Kramkov, Optimal decomposition under constraints, Probab.
Theory & Rel. Fields 109, 1–25 (1997).
8. I. Karatzas, Lectures in mathematical finance, Providence, American Mathematical
Society (1997).
9. N. El Karoui, S. Peng, and M. Quenez, Backward stochastic differential equations in
finance, Math. Finance No. 1, 1–71 (1997).
10. N. El Karoui, M. Quenez, Dynamic programming and pricing contingent claims in-
complete markets, SIAM J. Control Optim. 33, 29–66 (1995).
11. D. Nualart, The Malliavin Calculus and Related Topics, New York and Berlin,
Springer-Verlag (1995).
12. B. ∅ksendal, An introduction to Malliavin Calculus with application to economics,
Working paper No. 3/96, Norwegian School of Economics (1997).
13. B. ∅ksendal, Stochastic differential equations, 5th Edition, Springer (1991).
14. E. Pardoux and S. Peng, Adapted solution of a Backward stochastic differential
equation, Systems and Control Letters 14, 55–61 (1990).
15. S. Peng, BSDE and related g–expectation, Pitman Research Notes in Mathematics
Series No. 364, 141–159 (1997).
16. S. Peng, Monotonic limit theorem of BSDE and nonlinear decomposition theorem of
Doob-Meyer type, Probab. Theory & Rel. Fields 113, 473–499 (1999).
Account: ns145102
July 17, 2007 10:55 Proceedings Trim Size: 9.75in x 6.5in Chap17-HuYing
STOCHASTIC CONTROL AND BSDES

WITH QUADRATIC GROWTH
MARCO FUHRMAN
Dipartimento di Matematica, Politecnico di Milano
piazza Leonardo da Vinci 32, 20133 MILANO, Italy
E-mail: marco.fuhrman@polimi.it
YING HU
IRMAR, Université Rennes 1
Campus de Beaulieu, 35042 RENNES Cedex, France
E-mail: ying.hu@univ-rennes1.fr
GIANMARIO TESSITORE
Dipartimento di Matematica, Università di Parma
Parco Area delle Scienze, 53/A - 43100 Parma, Italy
E-mail: gianmario.tessitore@unipr.it
In this talk, we study a stochastic optimal control problem where the drift term of the
equation has a linear growth on the control variable, the cost functional has a quadratic
growth and the control process takes values in a closed set. This problem is related to
some BSDE with quadratic growth. We prove that the optimal feedback control exists
and the optimal cost is given by the initial value of the solution of the related backward
stochastic differential equation.
1. Introduction
This talk is based upon the paper 6 . The interested readers can see the full paper 6
for all the proofs of the results presented here.

In this talk, we consider a controlled equation of the form:

dXt = b(t, Xt ) dt + σ(t, Xt ) [dWt + r(t, Xt , ut ) dt], t ∈ [0, T ],
(1)
X0 = x.
In the equation, W is a Rd -valued Wiener process, defined on a complete probability

space (Ω, F, P) with respect to a filtration F := (Ft )t≥0 satisfying the usual condi-
tions; the unknown process X takes values in Rn ; x is a given element of Rn ; u is
the control process, which is assumed to be an Ft -adapted process taking values in
a given nonempty closed set K ⊂ Rm . The control problem consists in minimizing
80
Account: ns145102
81
a cost functional of the form

Z T
J =E g(t, Xt , ut ) dt + E φ(XT ). (2)
0
We suppose that r has a linear growth in u, g has quadratic growth in x and u,

and φ has quadratic growth in x.
The main novelty of the present work, in comparison with the existing literature
is that, on the one hand we neither assume K nor r to be bounded, on the other
hand we consider a degenerate control problem (since nothing is assumed on the
image of σ). Moreover we also allow φ to have linear growth.
Nonlinear Backward Stochastic Differential Equations (BSDEs) were first intro-
duced by Pardoux and Peng 9 . They have found applications in stochastic control,
see, e.g., Refs. 3 and 10. To have an overview of recent applications of backward
stochastic differential equations techniques to control problems being by several
aspects more general than the one considered here but involving “bounded con-
trol image” assumptions see for instance Ref. 4 or Ref. 7 and references therein.
The special “unbounded” case corresponding to the assumptions K = Rm and
g = 21 |u|2 + q(t, x) is treated in Ref. 5 by an ad-hoc exponential transform. We
notice that in Ref. 5 φ is allowed to take value +∞. Finally this same special case
(in which the Hamiltonian is exactly the square of the norm of the gradient) was
treated in Ref. 8 by analytic techniques, under non-degeneracy assumptions and in
an infinite dimensional framework.
The difficulty here is that the Hamiltonian corresponding to the control prob-
lem has quadratic growth in the gradient and the terminal cost is not bounded.
Thus the backward stochastic differential equation corresponding to the problem
has quadratic growth in the Z variable and unbounded terminal value. To treat
such equation we have to apply the localization procedure recently introduced by
Briand and Hu in Ref. 2. We notice that for such BSDEs no general uniqueness
results are known: we replace uniqueness by the selection of a maximal solution.
Moreover the usual application of Girsanov technique is not allowed (since the
Novikov condition is not guaranteed) and we have to develop specific arguments
both to prove the fundamental relation, see §4, and to obtain the existence of a
(weak) solution to the closed loop equation, see §5. Our main result is to prove
that the optimal feedback control exists and the optimal cost is given by the value
Y0 of the maximal solution (Y, Z) of the backward stochastic differential equation
(BSDE) with quadratic growth and unbounded terminal value mentioned above.
Moreover we show that we can construct an optimal feedback in terms of the pro-
cess Z. Finally we prove that if we fix a particular optimal feedback law then the
solution of the corresponding closed loop equation is unique, see Proposition 5.2.
The talk is organized as follows: in the next section, we describe the control
problem; in section 3, we study the related BSDE; in section 4, we establish the
fundamental relation between the optimal control problem and BSDE; and the last
section is devoted to the proof of the existence of optimal feedback control.
Account: ns145102
82
2. The Controlled Problem

We consider here the optimal control problem given by a state equation of the form:

dXt = b(t, Xt ) dt + σ(t, Xt ) [dWt + r(t, Xt , ut ) dt], t ∈ [0, T ],
(3)
X0 = x,
and by a cost functional of the form

Z T
J =E g(t, Xt , ut ) dt + E φ(XT ). (4)
0
We work under the following assumptions.
Hypothesis A.
(1) The process W is a Wiener process in Rd , defined on a complete proba-

bility space (Ω, F, P) with respect to a filtration (Ft ) satisfying the usual
conditions.
(2) The set K is a nonempty closed subset of Rm .
(3) The functions b : [0, T ] × Rn → Rn , σ : [0, T ] × Rn → Rn×d , r : [0, T ] × Rn ×
K → Rd , g : [0, T ] × Rn × K → R, φ : Rn → R, are Borel measurable.
(4) For all t ∈ [0, T ], x ∈ Rn , r(t, x, ·) and g(t, x, ·) are continuous functions
from K to Rd and from K to R respectively.
(5) There exists a constant C > 0 such that for every t ∈ [0, T ], x, x0 ∈ Rn ,
u ∈ K it holds:
|b(t, x) − b(t, x0 )| ≤ C|x − x0 |, |b(t, x)| ≤ C(1 + |x|), (5)
|σ(t, x) − σ(t, x0 )| ≤ C|x − x0 |, (6)
|σ(t, x)| ≤ C, (7)
|r(t, x, u) − r(t, x0 , u)| ≤ C(1 + |u|)|x − x0 |, (8)
|r(t, x, u)| ≤ C(1 + |u|), (9)
0 ≤ g(t, x, u) ≤ C(1 + |x|2 + |u|2 ), (10)

0 ≤ φ(x) ≤ C(1 + |x|2 ). (11)
(6) There exist R > 0 and c > 0 such that for every t ∈ [0, T ], x ∈ Rn , and
every u ∈ K satisfying |u| ≥ R
g(t, x, u) ≥ c|u|2 . (12)
Account: ns145102
83
We will say that an (Ft )-adapted stochastic process {ut , t ∈ [0, T ]} with values
in K is an admissible control if it satisfies:
Z T
E |ut |2 dt < ∞. (13)
0
This square summability requirement is justified by (12): a control process which

is not square summable would have infinite cost.
Next we show that for every admissible control the solution to (3) exists.
Proposition 2.1. Let u be an admissible control. Then there exists a unique con-
tinuous, (Ft )-adapted process X satisfying E supt∈[0,T ] |Xt |2 < ∞ and, P-a.s,
Z t Z t Z t
Xt = x+ b(s, Xs ) ds+ σ(s, Xs ) dWs + σ(s, Xs ) r(s, Xs , us ) ds, t ∈ [0, T ].
0 0 0
Proof. The proof of Proposition 2.1 relies on an approximation procedure.

The stochastic control problem associated with (3)-(4) consists in minimizing
the cost functional J(x, u) among all the admissible controls.
3. The Forward-Backward System

Let us consider again the functions b, σ, g, φ satisfying the assumptions in Hypoth-
esis A. We define the Hamiltonian function
ψ(t, x, z) = inf [g(t, x, u) + z · r(t, x, u)], t ∈ [0, T ], x ∈ Rn , z ∈ Rd , (14)
u∈K
where · denotes the usual scalar product in Rd . We collect some immediate prop-
erties of the function ψ.
Lemma 3.1. The map ψ is a Borel measurable function from [0, T ] × Rn × Rd to

R. There exists a constant C > 0 such that
−C(1 + |z|2 ) ≤ ψ(t, x, z) ≤ g(t, x, u) + C|z|(1 + |u|), ∀u ∈ K. (15)
Moreover the infimum in (14) is attained in ball of radius C(1 + |x| + |z|) that is
ψ(t, x, z) = min [g(t, x, u)+z·r(t, x, u)], t ∈ [0, T ], x ∈ Rn , z ∈ Rd ,
u∈K,|u|≤C(1+|x|+|z|)
(16)
and
ψ(t, x, z) < g(t, x, u) + z · r(t, x, u) if |u| > C(1 + |x| + |z|). (17)
Finally, for every t ∈ [0, T ] and x ∈ Rn , z → ψ(t, x, z) is continuous on Rd .
Next we take an arbitrary complete probability space (Ω, F, P◦ ) and a Wiener

process W ◦ in Rd with respect to P◦ . We denote by (Ft◦ ) the associated Brownian
filtration, i.e. the filtration generated by W ◦ and augmented by the P◦ -null sets of
F; (Ft◦ ) satisfies the usual conditions.
Account: ns145102
84
We introduce the forward equation:

dXt = b(t, Xt ) dt + σ(t, Xt ) dWt◦ , t ∈ [0, T ],
(18)
X0 = x,
whose solution is a continuous (Ft◦ ) adapted process, which exists and is unique by
classical results. Next we consider the associated backward equation

dYt = −ψ(t, Xt , Zt ) dt + Zt dWt◦ , t ∈ [0, T ],
(19)
YT = φ(XT ).
The solution of (19) exists in the sense specified by the following proposition.
Proposition 3.1. Assume that b, σ, g, φ satisfy Hypothesis A. Then there exist

Borel measurable functions
v : [0, T ] × Rn → R, ζ : [0, T ] × Rn → Rd
with the following property: for arbitrary complete probability space (Ω, F, P ◦ ) and
Wiener process W ◦ in Rd , denoting by X the solution of (18), the processes Y, Z
defined by
Yt = v(t, Xt ), Zt = ζ(t, Xt )
satisfy
Z T
2
E ◦
sup |Yt | < ∞, E ◦
|Zt |2 dt < ∞,
t∈[0,T ] 0
Y is continuous, nonnegative, and finally, P -a.s, ◦
Z T Z T
Yt + Zs dWs = φ(XT ) +
◦
ψ(s, Xs , Zs ) ds, t ∈ [0, T ].
t t
Proof. We apply the localization procedure together with a priori bounds,

which is introduced by Briand and Hu 2 .
4. The Fundamental Relation

In this section we revert to the notation introduced in the first section.
Proposition 4.1. For every admissible control u and for the corresponding trajec-
tory X starting at x we have
Z T
J(u) = v(0, x) + E [−ψ(t, Xt , ζ(t, Xt )) + ζ(t, Xt ) · r(t, Xt , ut ) + g(t, Xt , ut )] dt.
0
Proof. We apply an approximation procedure together with the Girsanov the-

orem.
Account: ns145102
85
Corollary 4.1. For every admissible control u and any initial datum x we have
J(u) ≥ v(0, x) and the equality holds if and only if the following feedback law holds
P-a.s. for a.e. t ∈ [0, T ]
ψ(t, Xt , ζ(t, Xt )) = ζ(t, Xt ) · r(t, Xt , ut ) + g(t, Xt , ut ),
where X is the trajectory starting at x and corresponding to control u.
5. Existence of Optimal Controls: The Closed Loop Equation

Let us consider again the functions b, σ, g, φ satisfying the assumptions in Hypothesis
A. We recall the definition of the Hamiltonian function
ψ(t, x, z) = inf [g(t, x, u) + z · r(t, x, u)], t ∈ [0, T ], x ∈ Rn , z ∈ Rd . (20)
u∈K
Lemma 5.1. There exists a Borel measurable function γ : [0, T ] × Rn × Rd → K

such that
ψ(t, x, z) = g(t, x, γ(t, x, z)) + z · r(t, x, γ(t, x, z)), t ∈ [0, T ], x ∈ Rn , z ∈ Rd .
(21)
Moreover there exists a constant C > 0 such that
|γ(t, x, z)| ≤ C(1 + |x| + |z|). (22)
Proof. It suffices to apply the Filippov theorem 1 .
Next we address the problem to find a weak solution to the so-called closed loop
equation. We define
u(t, x) = γ(t, x, ζ(t, x)), t ∈ [0, T ], x ∈ Rn ,
where ζ has been introduced in Proposition 3.1. The closed loop equation is

dXt = b(t, Xt ) dt + σ(t, Xt ) [dWt + r(t, Xt , u(t, Xt )) dt], t ∈ [0, T ],
(23)
X0 = x.
By a weak solution we mean a complete probability space (Ω, F, P) with a filtration
(Ft ) satisfying the usual conditions, a Wiener process W in Rd with respect to P
and (Ft ), and a continuous (Ft )-adapted process X with values in Rn satisfying,
P-a.s.,
Z T
|u(t, Xt )|2 dt < ∞, (24)
0
and
R T such that (23) holds. We note that by (9) it also follows that
0
|r(t, X t , u(t, X t ))| dt < ∞, P-a.s., so that (23) makes sense.
Proposition 5.1. Assume that b, σ, g, φ satisfy Hypothesis A. Then there exists a

weak solution of the closed loop equation, satisfying in addition
Z T
E |u(t, Xt )|2 dt < ∞. (25)
0
Account: ns145102
86
Corollary 5.1. By Corollary 4.1 it immediately follows that if X is the solution

to (23) and we set u]s = u(s, Xs ) then J(x, u] ) = v(0, x) and consequently X is an
optimal state, u]s is an optimal control and u is an optimal feedback.
Next we prove uniqueness in law for the closed loop equation. We remark that
the condition (24) is part of our definition of a weak solution.
Proposition 5.2. Assume that b, σ, g, φ satisfy Hypothesis A. Fix γ : [0, T ] × Rn ×

Rd → K satisfying (21) (and consequently (22)) and let u(t, x) = γ(t, x, ζ(t, x)).
Then the weak solution of the closed loop equation (23) is unique in law.
References
1. J.-P. Aubin, H. Frankowska, Set-valued analysis, Systems & Control: Foundations &
Applications, vol. 2. Birkhäuser Boston Inc., Boston, MA, 1990.
2. P. Briand, Y. Hu, BSDE with quadratic growth and unbounded terminal value.
Preprint no. 05-07, IRMAR, Université Rennes 1, 2005.
3. N. El Karoui, S. Peng, M. C. Quenez, Backward stochastic differential equations in
finance. Math. Finance 7, 1-71 (1997).
4. N. El Karoui, S. Hamadène, BSDEs and risk-sensitive control, zero-sum and nonzero-
sum game problems of stochastic functional differential equations. Stochastic Process.
Appl. 107, 145–169 (2003).
5. M. Fuhrman, A class of stochastic optimal control problems in Hilbert spaces: BSDEs
and optimal control laws, state constraints, conditioned processes. Stochastic Process.
Appl. 108, 263-298 (2003).
6. M. Fuhrman, Y. Hu, G. Tessitore, On a class of stochastic optimal control problems
related to BSDEs with quadratic growth. Preprint.
7. M. Fuhrman, G. Tessitore, Existence of optimal stochastic controls and global solu-
tions of forward-backward stochastic differential equations. SIAM J. Control Optim.
43, 813–830 (2004).
8. F. Gozzi, Global regular solutions of second order Hamilton-Jacobi equations in
Hilbert spaces with locally Lipschitz nonlinearities. J. Math. Anal. Appl. 198, 399-443
(1996).
9. E. Pardoux, S. Peng, Adapted solution of a backward stochastic differential equation.
Systems and Control Lett. 14, 55-61 (1990).
10. J. Yong, X. Y. Zhou, Stochastic Controls: Hamiltonian Systems and HJB Equations.
Springer, New York, 1999.
Account: ns145102

Presentación 2 Webconference

Diunggah oleh

Informasi Dokumen

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Presentación 2 Webconference

Diunggah oleh

Hak Cipta:

Format Tersedia

July 27, 2007 10:5 Proceedings Trim Size: 9.75in x 6.

Part Four: Dynamics and Optimal Control of Partial

Optimal Control of Quasilinear Elliptic Obstacle Problems . . . . . . . . . . . . . . . . . . . 263

Controllability of a Nonlinear Degenerate Parabolic System

Near-Optimal Controls to Infinite Dimensional Linear-Quadratic

Shape Optimization of Stationary Navier–Stokes Equation . . . . . . . . . . . . . . . . . . . 323

Solution Map of Strongly Nonlinear Impulsive Evolution Inclusions . . . . . . . . . . .338

Study on Repairable Series System with Two Components — A Semigroup

Finite Dimensional Reduction of Global Dynamics and Lattice Dynamics

Recent Progress on Nonlinear Wave Equations via KAM Theory . . . . . . . . . . . . . . 387

This page intentionally left blank

This page intentionally left blank

This page intentionally left blank

XUNJING LI’S ACADEMIC LIFE

A TRIBUTE IN MEMORY OF PROFESSOR XUNJING LI ON HIS

Professor Xunjing Li passed away in February, 2003, at age 68, in Shanghai,

ment of Mathematics, Fudan University. He was promoted to Lecturer in 1962, to

Professor of Fudan University in 1997. Professor Li retired in 2001.

1. 1959—1976. Almost as soon as Professor Li finished his graduate study

ciated with Shanghai Petroleum Refinery Factory, as well as Jin-Shan Petroleum

In addition to his mathematical research, Professor Li’s life-long pedagogical

research and in educating younger generations of graduate students. We are proud

1. Ordinary Differential Equations, Shanghai Science and Technology Press,

2. Control Theory Applied to Computers, Fudan University press, 1988 (with

3. Theory of Ordinary Differential Equations for Optimal Control Systems,

4. Optimal Control Theory for Infinite Dimensional Systems, Birkhäuser, 1995

5. Basics of Control Theory, Higher Education Press, 2002 (with Jiongmin

6. Collection of Mathematical Papers, Fudan University press, 2003 (forwarded

II. Edited Conference Proceedings

1. Control Theory of Distributed Parameter Systems and Applications, Lecture

2. Control of Distributed Parameter and Stochastic Systems, Kluwer Academic

III. Research Papers

1. Stability of second-order linear ordinary differential equations with periodical

2. Four-point common circles and circles of invariance under bilinear transfor-

3. An extension of Nikolisky’s theorem, Journal of Fudan University (Natural

4. Cesàro summability in Banach space, Acta Mathematica Sinica, 10 (1960),

5. Uniform approximation of a function by the Cesàro means of its Faber series,

6. On the absolute stability of the indirect control systems Journal of Fudan

7. On the absolute stability of systems with time lags, Acta Mathematica

8. Time optimal control of linear systems, Journal of Fudan University

9. Stability and time optimal control for automatic regulated systems,

10. Mathematical principle for electro-magnetic power-driven compressors (I) —

11. Period-analysis approach to time series, Selection of Long-Term Forecast

season), Eastern China Meteorological Observatory Center, (1973), 71–73

13. Analytical design of control equations, Journal of Fudan University (Natu-

15. An analysis of the mathematical model to the furnace of thermal dissocia-

17. Time optimal control of distributed parameter systems, Scientia Sinica, 24

23. Maximum principle of distributed parameter systems with time-lags,

25. N-person differential games governed by infinite dimensional systems, J. Op-

26. Maximum principle of optimal periodic control for functional differential

27. Maximum principle of optimal controls for functional differential systems,

28. On optimal control of functional differential systems, Springer Lecture Notes

30. Maximum principle for optimal control of nonlinear generalized systems —

31. Necessary conditions for optimal control of distributed parameter systems,

34. A realistic model of neural networks, J. of Electronics, vol. 9, no. 4 (1992),

35. Progress in control theory (distributed parameter systems), 1990 Yearbook

Shanghai, 1992 (in Chinese).

38. A linear quadratic optimal control problem with disturbances — an algebraic

40. Maximum principle for optimal control of distributed parameter stochas-

43. Stochastic verification theorems within the framework of viscosity solutions,

49. Synthesis of upper-triangular non-linear systems with marginally unstable

50. Research into unexpected events: a key project of mathematical finance,

IV. Conference Proceedings Papers

1. On optimal control for distributed parameter systems, Proc. of 8th IFAC

Et,T [ + X + Y ] = Et,T [ + X] + Y a.s. . (8)