Advances in Deterministic and Stochastic Analysis (2007)

Advances in
Deterministic
and Stochastic
Analysis
EDITORIAL BOARD
N g u y h Minh Chuang
Phillipe G. Ciarlet
Takeyuki Hida
Peter Lax
David Mumford
Duang H6ng Phong
Roger Temam
Nguyzn VBn Thu
Nguy6n Minh Tri
Vii Kim Tudn
Advances in
Deterministic
and Stochastic
Analysis
Editors
N M Chuong
Institute of Mathematics, Vietnamese Acad. of Sci. &Tech., Vietnam
P G Ciarlet
City University of Hong Kong, Hong Kong
P Lax
Courant Institute, USA
D Mumford
Brown University, USA
D H Phong
Columbia Universitv, USA
N E W JERSEY * LONDON *
K6 World Scientific
SINGAPORE * -
BElJlNG * SHANGHAI HONG KONG * TAIPEI * CHENNAI
Published by
World Scientific Publishing Co. pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA oflce: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601
UK oflcet 57 Shelton Street, Covent Garden, London WC2H 9HE
British Library Cataloguing-in-PublicationData

A catalogue record for this book is available from the British Library.
ADVANCES IN DETERMINISTIC AND STOCHASTIC ANALYSIS

Copyright 0 2007 by World Scientific Publishing Co. Re. Ltd.
All rights reserved. This book, or parts there% may not be reproduced in any form or by any means,
electronic or mechanical, includingphotocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In t h s case permission to
photocopy is not required from the publisher.
ISBN-13 978-981-270-550-1
ISBN-I0 981-270-550-3
Printed in Singapore by World Scientific Printers ( S ) Pte Ltd

V
PREFACE
In Junc 4-9, 2005, the Sccond International Conference 011 Abstract and
Applied Analysis was held at Quy Nhon, Vietnam. The conference brought
together Vietnamese mathematicians working on analysis, as well as many
distinguished foreign visitors. It covered all areas of analysis where Viet-
namese mathematicians have been playing an active role, including pure
and applicd analysis, and dctcrministic and stochastic approaches. The pro-
ceedings of the conference are gathered in the present volume. The spccific
topics discussed at the confcrcncc rcflcct thc most activc rcsca,rch directions
in Vietnam today, and as such, are quite varied. Seventeen papers appear
here. For convenience, they have been loosely organized into four different.
chapters.
Chapter I, Integral and Pseudodifferential Operators, includes the pa-
pers where these operators play a major role, particularly in the theory of
partial differential equations. The paper by Nguyen Minh Tri provides a
criterion for the hypoellipticity for a class of operators with sign-changed
characteristic form, generalizing earlier works of Beals and Fefferman, The
paper by Nguyen Minh Chuong and Dang Anh Tuan deals with general-
izations of the oblique derivative problem in several directions, including
non-lincar settings. The last two papers in the chapter deal more with the
properties of the operators themselves, with the boundedness of commu-
tator integrals of mixed homogeneity studied by Lubomina Softova, and a
classification of integral transforms of both convolut.ionand non-convolution
types provided by Vu Kim Tuan.
Chapter 11, Partial Differential Equations , deals of course also with
partial differential equations, but with essentially different methods than
the pseudodifferential and integral operator methods of Chapter I. The pa-
per by Martin Schechter provides a unified min-max approach to variational
problems, through the idea of linking sets separating a given funct.iona1.
The papcr by Mikio Tsiji and Peter Wagner studies generalizcd solutions
vi Preface
and the possible appearance of shock waves to a class of equations gen-

eralizing Burgers equation. Boundary value problems for the biharmonic
equation are studied in the paper by Dang Quang A and Le Tung Son,
by a reduction to a boundary value problem for a second order equation,
combined with an iterative scheme. A parabolic boundary value problem,
arising from an inverse problem in engineering for finding the heating regime
for a given desired final temperature, is studied in the paper by Dinh Nho
Hao, Nguyen Trung Thanh, and H. Sahli, by the DC (difference of two
convex functions) method. The paper by Peter Massopust describes ex-
plicit solutions, in various coordinate systems, for the Maxwell equations
describing magnetic fields generated by a hard ferromagnet. The paper by
Ha Tien Ngoan and Nguyen Thi Nga deals with the local solvability of the
non-characteristic Cauchy problem for weakly hyperbolic classical Monge-
Ampere equations. The asymptotic analysis of the solutions of the Navier-
Stokes problem is considered by Makram Hamouda and Roger Temam when
the viscosity goes to zero and the flow in a channel of R3 is also considered
by them in the case when the boundary is non-characteristic.
Chapter 111, Geometric Analysis, groups together the papers with a
more pronounced geometric flavor. The paper by Le Hong Van provides
a construction of obstructions to imbedding a given statistical manifold
into another one. A statistical manifold is a Riemannian manifold with an
equivalence class of symmetric 3-tensors, a geometric concept which arises
from information geometry. The paper by Do Ngoc Diep constructs a 22-
graded Cech cohomology, building in ideas of Rosenberg and Kontsevich
from noncommutative geometry. The paper by Nguyen Minh Chuong and
Le Duc Thinh studies Sobolev spaces with weights on Riemannian mani-
folds. Finally, as its title indicates, Chapter IV, Stochastic and Infinite-
Dimensional Analysis, groups together papers more related to either prob-
ability or analysis in infinite-dimensional spaces. They include the paper by
Situ Rong, on the existence of solutions to some reflexive stochastic differ-
ential equations with jumps and its applications to stochastic population
control; the paper by Shigeyoshi Ogawa, which provides a survey of devel-
opments related to a notion introduced by the author; the paper by Andrei
Khrennikov, which discusses the foundations of quantum mechanics from
the point of view of infinite-dimensional spaces, and the paper by Karl
Gustafson, which discusses the ideas of the author on Bells inequality and
contraction semigroups.
The Editors
vii
CONTENTS
Preface V
Chapter I Integral and Pseudodifferential Operators
$1.Pseudodifferential Operators of Second Order with Sign-Changed

Characteristic Form 3
Nguyen Minh Tri
$2. A Semilinear Nonclassical Pseudodifferential Boundary Value

Problem in Sobolev Spaces 1 < p < 00 15
Nguyen Minh Chuong and Dang Anh Tuan
$3. Singular Integral Operators in Functional Spaces of Morrey

Type 33
Lubomira Softova
$4. Classification of Integral Transforms 43

Vu Kim Tuan
Chapter I1 Partial Differential Equations
$5. Unified Minimax Methods 75

Martin Schechter
$6. Some Remarks on Single Conservation Laws 91

Mikio Tsuji and Peter Wagner
viii Contents
$7. Iterative Method for Solving a Mixed Boundary Value Problem

for Biharmonic Type Equation 103
Dung Quang A and Le Tung Son
$8. Numerical Solution to a Non-Linear Parabolic Boundary

Control Problem 115
Dinh Nho Hao, Nguyen Trung Thanh and H. Sahli
$9. A Class of Solutions to Maxwells Equations in Matter

and Associated Special Functions 131
Peter Massopust
510. On the Cauchy Problem for a Quasilinear Weakly Hyperbolic

System in Two Variables and Applications to that for Weakly
Hyperbolic Classical Monge-Ampkre Equations 177
Ha Tien Ngoan and Nguyen Thi Nga
511. Some Singular Perturbation Problems Related to the

Navier- Stokes Equations 197
Makram Hamouda and Roger Temam
Chapter I11 Geometric Analysis
$12. Monotone Invariants and Embeddings of Statistical

Manifolds 231
Le Hong Van
$13. Graded Cech Cohomology in Noncommutative Geometry 255

Do Ngoc Diep
$14. Sobolev Spaces with Weight on Riemannian Manifolds 269

Nguyen Manh Chuong and Le DUC Thinh
Chapter IV Stochastic and Infinite-Dimentional Analysis
$15. Stochastic Population Control and RSDE with Jumps 281

Situ Rong
Contents ix
$16. Noncausal Stochastic Calculus Revisited - Around the

So-called Ogawa Integral 297
Shigeyoshi Ogawa
5 17. Infinite-Dimensional Stochastic Analysis and Foundations of

Quantum Mechanics 321
Andrei Khrennikov
818. Noncommutative Trigonometry and Quantum Mechanics 341

Karl Gustafson
This page intentionally left blank
Chapter I
INTEGRAL AND
PSEUDODIFFERENTIAL OPERATORS
Advances in Deterministic and Stochastic Analysis 3
Eds. N. M. Chuong et al. (pp. 3-13)
@ 2007 World Scientific Publishing Co.
51. PSEUDODIFFERENTIAL OPERATORS OF SECOND

ORDER WITH SIGN-CHANGED CHARACTERISTIC FORM
NGUYEN MINH TRI

Institute of Mathematics
18 Hoang Quoc Viet, Cau Giay, 10307 Hanoi, Vietnam
E-mail: triminhOmath.ac.vn
In this paper we prove hypoellipticity for pseudodifferential operators of second

order with sign-changed characteristic form. The obtained results extend the
previous ones of Beals-Fefferman and Ganja for differential operators.
In this paper we consider operator of the form

k
in a bounded domain R c R$' = IRE x @, where XI,. . . ,xk are proper

pseudo differential operators (p. d. 0.) of first order with real principle
symbols Xj(x,t,J,~),j = and c ( z , t ) is a smooth function in R. In
Ref. 1 Hormander proved that for a differential operator of second order to
be hypoelliptic it is necessary that its characteristic form does not change
sign at a fked point in 0. However, Kannai's example in Ref. 2 showed that
characteristic form of a hypoelliptic operators can change sign from point to
point in 52. Necessary conditions for hypoellipticity of operators generalizing
the Kannai example were studied in papers of Helffer and Z ~ i l yand , ~ Beals
and Fefferman.4 In Ref. 4 it is proved that for the differential operator P in
the form (1) to be hypoelliptic it is necessary that XI,. . . ,xk satisfy the
following conditions
j = 1 , ..., k.
Xj(~,0,0,1)=0,
2000 Mathematics Subject Classification Primary 35H10; Secondary 35A08, 35B65,

35C15. 35D10.
4 N . M. Tri
In Ref. 4 Beals and Fefferman also established some sufficient conditions for
the differential operator P in the form (1) to be hypoelliptic. In Refs. 5,6
Ganja obtained further sufficient conditions for the differential operator P
in the form (1) to be hypoelliptic. In this note we extend the results of
Beah and Fefferman, and Ganja for the pseudodifferential operator P in
the form (1).Now we introduce the transformation:
%,t) + fi;d,t,)
(x,t ) -+ (21,t); 5 = 5 , t = 6t; 6 E (0,1].
For a function u E C r ( 0 ) we can define the function E C r ( 0 ) as

follows:
ug(z, t) = u ( 5 ,t6).
Note that when 6 = 1 we can identify the coordinates ( z , t )and ( d , t ) .
we can get easily the desired estimates.
Let Ps be the operator in the coordinates (d,

t) defined by the formula:
P.d.0 of Second Order with Sign-changed Characteristic Form 5
Lemma 2. Let the operator Pa have the form ( 2 ) . Then f o r arbitrary s E R

and for arbitrary K c R there exists a constant C(6)= C(s, K , 6 ) such that
c k
j=1
ll6txj,6911,2 6 21(6t%71g>sl + C(~)lIgl1,2, vg E C,-(K), (3)
where 11 . ]Is, ( , ) s are taken in the (x,t) coordinates.
Proof. Put Bj,6 = 6tXj,s - (dtXj,s)*. Obviously, ord Bj = 0 (since Xj,a

has real principle symbol). Note that
k
6tPs = C ( 6 t X j , , ) 2 +Qs,
j=1
where
a k
Qs = -t:
at
+ btcs(x,t) - S2tXj,s(t)Xj,6; ord Qs = 1.
j=1
Thus, Lemma 2 is proved. 0
Corollary 1. From the inequality (3) it implies that
c
j=l
k
lltxj,sgIl: 6 2lltPsgII,2 + C(~)11g11,2~
Lemma 3. Let X j be the proper homogeneous p. d. 0. with real principal
symbol Xj(x,t ,El r ) . Moreover assume that
axj
-h
37- t , E, t , E, 7)
= t c p j (5, (4)
6 N . M. Tri
for j = 1,.. . ,Ic; ord pj = 0. Then operator Pa can be written in the form
P&=
l d
6 at f S ck
j=1
+
Xj,&tXj,& 6t c
j=1
k
Pj,&Xj,S + C&(Z, t),
, J, $).
where cpj,g(z, t,J, 7) = p j ( ~ td,
Proof. We have
+
tX& = Xj,&tXj,& [t,Xj,&]Xj,&.
By a theorem on composition of p. d. 0. it is easy to see that
where o([t,X j , & ]denotes

) the symbol of the operator [t,Xj,s].From above
we get the desired equality.
Lemma 4. Assume that X j satisfies the condition of Lemma 3. For every

s E R, K G 52, there exists a constant C = C(s, K, X j ) such that
llxj,agll: 6 c11g11:+1 + c(6)llgll:-ll vg E GYK).
The constant C is independent of 6.
Proof. We have
t, E, 7) = xj (Z,tS, J,
Xj,&(Z, -)S .
7
ax. O , O , 1) = 0. Since
By the assumption of Lemma 3, we have +(z,
X j ( ~ 1* ., zn, ti
1 .. i Jnl7)
is a homogeneous function of degree one in ((1, . . ., Jnl 7) we have the fol-

lowing Euler formula
Putting in Eq. (5) (1 = . . . = Jn = 0 , =~1 we get

ax.
0 = --3-(2,0,0,1) = X j ( X , o,o, 1).
a7
Consequently
1
Xj,&(Z,t, J, 7) = -Xj(Z, tb, Jb, 7)
6
1 dX .
= 6 [Xj(Z,0, SJ, 7) S t at +
L ( z ,0 , SJ, 7) + O(S2)l.
Note that if X j satisfy the condition (4) in Lemma 3, then, equivalently,

X j can be written in the following form
xj (x,t , t,). = tq (x,t ,tl ). + zj (x,t , 6).
Lemma 5 . Let the operator X j have the form
x j ( x , t , t , T )= tu,(.,tJ) +Zj(.,t),
where Y,(x,t,O) = 0 , j = n.
Then f o r every s E 1w and a n arbitrary
compact K in R there exist constants C = C(s, K ) ,60 such that f o r 6 E
(0760)
where 11 . / I s , (., .)s are taken in the coordinates (d,

t).
Proof. Consider the expression
By Lemma 3, the operator P6 may be rewritten in the form
where ord (pj,a = 0. It follows that

8 N . M. Tri
where @a may be estimated by
Let us estimate the term -6

k
Cj=l ,, )o. We have
(t'Xj,ag, X~JA'~'-'&
It remains to estimate the second term in Eq. (10)
-6 C a
t'Xj,sg, -[Xj,a, (12)
j=1
k ( at'
Put Dj,3,6= [Xj,hlA'23-1]. Since ord Dj+,g = 2s - 1, the second term in

Eq. (12) will be estimated by
k
j=1
It remains to estimate the first term in Eq. (12)

We have
Consequently
By Lemma 4
where C is the constant appearing in Lemma 4. By choosing 60 =

min{ 1,&} and combining (7) - (14) we get (5). 0
Corollary 2. Under the hypothesis of Lemma 5 the following inequality

holds
for all g E C r ( K ) .
10 N . M. Tri
Corollary 3. Under the hypothesis of Lemma 5 the following estimate

holds
for some constant C = C(s, K ) .
Proof. First choose 6 = 2 , g = f6 in Eq. (15), then apply Lemma 1 we

immediately get the result.
Corollary 4. Let the hypothesis of Lemma 5 be fulfilled and the algebra

Lie generated by { XI,.g, +
. . ,X k } has rank n 1 in each point of 0. There
exists a number u > 0 such that following estimate holds
for some constant C = C(s, K).
Proof. We can easily deduce the desired estimate from Corollary 3 and
the apriori estimate achieved in Ref. 1 (see also Refs. 7-9). 17
Theorem 1. Let P have the f o r m (1). W e impose o n the operators X j the

following conditions
X ~ ( I C , ~ , E , ~ ) = ~ Y ~ ( I C , Jw, ~
h e) r+e Z
Y ,~( z(,~J , O
J ))=, O , j = p .
. . ,X,} has rank n + 1 in
The algebra Lie generated by {&,XI,.
each point of R.
Then P is a hypoelliptic operator.
Proof. We shall show that iff E D'(R), Pf E Hb,,(Q),then f E H,S:"(R),

X j f E H",' (a),j = for some u > 0. First we shall prove this assertion
in the case f E E'(R). Since f E E'(R), there exists y E R such that
+
f E HY(Rn+l). We can assume that y 6 s D . Set 77 = inf(s,y). Then
f,P f E H". For the reason of convenience we will work in the coordinates
( d t ' ) for 6 < 60. Instead of f we write g in the ( d , t ' ) coordinates. Let
xE*be an averaging kernel. We have

*
~ S ( X E 9) = --
j=1
where all TE,6,T:,6,0 < E 6 EO belong to a bounded set in the spase of p.

d. 0.of order zero, if EO is sufficiently small. By choosing an appropriate xE
from Lemma 5 we obtain
12 N . M. Tri
Therefore
uniformly in E. Passing to the limit as E -+ 0 we obtain g,Xj,ag E HV.

That means that g , X j f E H". From (16) it follows that
+IlXE * fll, G Ml
uniformly in E. Passing to the limit as E 4 0 we obtain f E Hq+'. By the
Schwarz inequality
[(XC * tPf,XE * f),+q I 6 C(llX8 *tw;

+ llxe * f112,+,).
Hence
uniformly in E . Passing to the limit as E -+ 0 we obtain 2,

Xj f E H,+%.
We can repeat our argument with inf(s, q+a) instead of q and get a similar
+
assertion, the replace inf(s,q n) by inf[s,inf(s,q n) 4,. + + . . . At last
we obtain the needed assertion for all f E E'(R). Now let fi be a relative
compact set in R, and f be the distribution on R, not necessarily with
compact support. By the general theory of distributions it follows that the
restriction o f f to 0 belongs to HLc(fi)for some y- E R. Consider a sequence
of Coo-function {gi} with compact support in R such that gi+l = 1 in a
neighbourhood of supp gi, (i = 0 , 1 , . . .,). We have
j=1 j=1
By a theorem on order of commutators we get
We assume as above that Pf E H,",,(s1). Set 71 = inf(s,y). Then the right

hand side in the ( i + l ) t h inequality belongs t o H q - l , i. e. P(gi+lf)
E Hq-l.
Hence by the previous step we have
Xj(Si+lf)E Hq-l+s.
Therefore the right hand in Eq. (i) belongs t o Hq-l+s whence P ( g i f ) E
H"-l+%. But by result proved above Xj(gif)E Hq-l+u,gif E Hq-l+%.
Repeating this argument again and again we get P(g0f)E Hq. It follows
that Xj(g0f)E Hq+q,gofE Hq+'. But we can choose function go E
C r ( 6 ) arbitrarily. Therefore
Using this assertion again and again we shall come to t = s and thus we
obtain the desired assertion. Our theorem is proved completely. 0
References
1. L. Hormander, Acta Math., 119,147 (1967).
2. Y. Kannai, Zsr. J , Math., 9,306 (1971).
3. B. Helffer, C. Zuily, C. R. Acad. Sci. Paris Sir. A - B , 277,1061 (1973).
4. R. Beals, C. Fefferman, Comm. Part. D i f f . Equat., 1, 73 (1976).
5. 4. E. ramca, Becmuulc M T Y , (1985), pp. 96-99.
6. kl. E. raaza, YMH, 41, 217 (1986).
7. J. J. Kohn, Proc. Symp. Pure Math., 23,61-69 (1973).
8. 0. A. OneCiiHm, E. B. Pameswr, Y p a e u e u u x Bmopozo n o p x d x a c Heompu-
yamenauto6 xapaxmepucmuueclco6 @ o p ~ 0 6klTorH , Haym. MaT. a ~ a n m .
-M.: BMHklTkl, 1971.
9. F. Treves, Pseudo-Differentia1 Operators and Fourier Zntegml Oprators,
(Plenum Press, 1982).
$2. A SEMILINEAR NONCLASSICAL

PSEUDODIFFERENTIAL BOUNDARY VALUE PROBLEM
IN SOBOLEV SPACES He+, 1 < p < 00
NGUYEN MINH CHUONGf and DANG ANH TUAN

18 Hoang Quoc Viet, Cau Giay, 10307 Hanoi, Vietnam
f E-mail: nmchuong@math.ac.vn
We study a semilinear nonclassical pseudodifferential value problem in Sobolev

spaces H e , p , 1 < p < a.
This paper is thecontinuationof Refs. 7,8,10,12,where
the problem was investigated in He,z.
1. Introduction
In Ref. 5 A. V. Bicatze firstly on the world dropped the classical boundary
condition imposed on the oblique derivative D,u = g,
namely, the con-
dition that the vector field D, must be not tangent to the boundary. He
investigated such problem, event for Laplace equations in three dimensional
region with the assumption that the vector field D, may be tangent to the
boundary. Then many mathematicians such as R. Borelli,' L. HOrmander,l4
Yu. V. Egorov and V. A. Kondratiev," Yu. V. Egorov and Nguyen Minh
C h ~ o n g Le
, ~ Quang Trung1lgNguyen Minh Chuong and Tran Tri Kiet18
Nguyen Minh Chuong and Dang Anh Tuan,' etc., continue to extend this
problem in several directions. For instance in Ref. 10 the problem was ex-
tended to an elliptic differential operator of second order for a bounded
domain R in Rn with a smooth boundary dR, in Ref. 9 the problem was
studied for a parabolic differential operator of second order, in Ref. 11 a
non-classical boundary value problem for a second order elliptic equation
in Sobolev space of variable order, in Ref. 12 a semilinear boundary value
problem for an elliptic singular integro - differential equation of order 2m in
He,z was investigated, in Ref. 8 a priori estimates were established for non-
classical derivative problems for elliptic and parabolic differential operators
of second order in Sobolev spaces He,p, 1 < p < 00.
The purpose of this paper is to investigate a semilinear non classical
16 N . M. Chuong and D. A . Tuan
pseudodifferential boundary value problem in Sobolev spaces He,prq,1 <

p < 00, q is a complex parameter, (YO 5 arg q 5 ,&
This problem was investigated in Ref. 12 by Yu. V. Egorov and by the
second author of this paper, with the singular integro - differential opetators
of M. S. Agranovich type3 and the Sobolev spaces He,2. In Ref. 3 M. S.
Agranovich used the singular integral operators corresponding to spherical
functions on a unit sphere Sn-l in R". He used the symbols expanded into
sequences of spherical functions converging in Lz(S"-~). Here by means of
a new class of singular integral operators it is able to solve the non classical
pseudodifferential boundary value problem in Ref. 12 in Sobolev spaces
1 < p < 03. The approach used in this paper is quite different from
Ref. 3, even from Ref. 12.
It is noticed that, by writing a non classical boundary value problem for
a Laplace equation in pseudodifferential operator form, in Ref. 13 by the
first author Yu. V. Egorov, an interesting estimate of subelliptic type was
obtained.
2. Sobolev Spaces
Let He,p(Rn),C E R, 1 5 p < 00, be the completion of the space C?(Rn)
of functions with compact support with respect to the norm
where Fu = C(E) is the Fourier transform of function u(x),IEI =

Jc; + . . . + t:.
Let p , C be in R such that 0 5 C, 1 5 p < 00. We denote by He,p(RI",)the
completion of the space C ~ ( I Wwith
~ ) respect to the norm:
Ilulle,p = Ilulle,p,w; = inf lIlulle,p,w,
when the infimum is taken over all extension I from "3 to R".
Let R be a compact domain in R" with (TI - 1)-dimensional smooth
N M
boundary d o , { ~ j } ~ a partition
, ~ of identity on R and {$j}j=l a parti-
tion of identity on dR such that for each support of +j , j = 1,. . . ,M , dR
can be locally transformed into a superplane.
Let p , e be in R such that 0 5 e, 1 5 p < 00. We denote by He,*(R) the
completion of the space C"(G) with respect to the norm
Semilinear Nonclassical P.d. BVP in Sobolev Spaces He,p 17
where C' denotes the sum over all j such that the neighbourhoods Uj
do not intersect the boundary dR, C" denotes the sum over all j such
that the neighbourhoods Uj intersect the boundary 80.
Let p , l be in R such that 0 5 l , 1 5 p < 03. Let He,p(as2)the completion
of the space C"(ds2) with respect to the norm
where Il$julle,p,Rn-lis the norm of the space He,,(R"-l) of $ju in the

locally coordinate system.
Let q be in C , p , l be in R such that 1 5 p < 03. In the spaces He,p,,(Rn),
He,p,q(R)or He,p,,(Rn+),He,p,q(8s2) (l 2 0 ) we define a respective norm
with parameter q:
We give now some properties of the spaces He,p.
Proposition 2.1. Let l , k , p be in R such that 1 5 p < 03, k < t. T h e

following bounded imbeddings hold true
Proposition 2.2. Let p , l be in R such that 1 5 p < 03, 0 5 l. T h e

restricted operator M f r o m R" to RI;" i s a bounded linear operator f r o m
He,p(R") t o There exists extended bounded linear operator L f r o m
He,p(Rn+)t o He,p(Rn).
Proposition 2.3. Let p , l be in R such that 1 5 p < 03, (1- k)< l. The
operator defined by u H uIzn---o i s a bounded linear operator f r o m He,p(E%n)
(or H e , p ( R y ) )t o He-(l-$),,( Rn-1 ). T h e operator defined by u H uI, i s a
bounded linear operator f r o m Ht,,(O) to He-(l-L),p(Xl).
P
where C is a positive constant not depending on u.

18 N . M . Chuong and D. A . Than
= ( 2 7 r - + & F u ( S ) d & for each u E C r ( R n )

Note that F(uIZn=o)(J,O)
and the integral
/(l+ It[+ IEnl)-*dEn = (1 + ltl)-*+l s,(1+ Itl)-*dt,

w
(t = (1+ lEl-l&J,
converges for (1 - k) < l , so by Holder inequality we have
hence,
(1 + lS/l)(t-(l-~))l~(ulzn=o)(~/, 0)l ICs,(l + lEJ)e*IFu(E)IPdtn,

by integrating both side of the above inequality we get the inequality
(1).
3. Pseudodifferential Operators in R
Definition 3.1. Let m be in Z+, 71, 7 2 be in R and

Q = (2 E CIyt I argz 5 7 2 ) .
The following operator is called a pseudodaflerential operator in R of
order m, with parameter q if it is defined by
Au(x,4 ) = 4 5 , D , 4 ) 4 x ,4 )
= (27-v [ c
lal+Plm
qp
R;
J K a d x , 5 - Y)qu(Y)dY], (2)
with
4E Q, K a p ( x ,z ) = H a p ( z ) + k ap( x : ,z ) ,
where
(i) H a p ( z ) , kap(x, 2 ) are homogeneous of order (-n) with respect to z ,
(ii) kap(x, z ) belongs to Cm(Rn) with respect to 2,
D,Yk,p(z, z ) --$ 0 when 1x1 + 00, for all y (multi-index)
Semilinear Nonclassical P.d. BVP in Sobolev Spaces He,* 19
(iii)
Izl=l n: lz1=1
where o ( z ) is an element area of the sphere { z E IWnllzl = 1).
Remark 3.1. Under conditions (i), (ii), (iii) and (iv), the pseudodifferen-
tial operator (2) is defined on C r ( R n , )and has another form
Au(x,Q) = (27rPn/' / c,
ei(z'E)aA(z, q ) W E ) & , (3)
R[g;
where
= g:p(x, E ) +gAa(l), (5)
gEp(x, = (2n)-"/2(-1)n / e-+?E)k,p(x, z)dz, (6)

R
:
gk p(c ) = ( 2 ~ ) - " / ' ( - 1 ) ~ 1 e-i(Z>E)Hap(z)dz. (7)

R:
Theorem 3.1. Let l , p be in IK, m be in Z+ such that 1 < p < 00. T h e
pseudodifferential operator A of order m of form ( 2 ) is a bounded linear
111. l\le-m,p). This OP-
operator from (C,"(R"), 1 1 1 . Ille,P) t o (He-m,p,q(Rn),
erator can be extended t o a bounded linear operator from He,p,g(Rn)t o
He-m,p,q(IKn) acting as a n operator of order m of form ( 3 ) .
Moreover, the estimate
I ll A ~ lle--m,p
l 5 CI lbllle,p, u E He,P,dIK"), (8)
20 N . M. Chuong and D. A . %an
holds true, where C is a constant not depending OIL u,q .
then
First, we prove (10). We have
Since gAp(r]) is homogeneous of order 0, we have
This implies that
Hence, we have (10).

Now, we prove (9). We have
Because of
and D,Y(x,z ) -+ 0, when 1x1 + co,for all multi-index y,so for all multi-
index y,
22 N . M. Chuong and D. A . %an
it implies that for all Ic in Z+, we have
where C5 not depending vl El q , so

We now study the pseudodifferential operator which has special form

called a homogeneous one with the symbol not depending on x
aA(<,4) =
bI+P=-m
qp<" /
yq
e-i(Zif)K,p(z)dz.
In this case it is not difficult we obtain the following result.
Theorem 3.2. Let l , p be in R, m be in Z+ such that 1 5 p < 00. Suppose

that the homogeneous pseudodifferential operator A of order m has symbol
+
O A ( [ , q ) # 0 , when I(l 141 # 0. Then, with q E Q \ {0}, A has the inverse
operator A-' which is a pseudodiflerential operator of order (-m) with
symbol a;'([, q ) . Moreover, the estimate
IIIA-lf II Il,P 5 CI Ilfllle-m,P, f E He-m,P,P(Rn)

holds true, where C is a constant not depending on f , q .
4. Pseudodifferential Operator in R y
Definition 4.1. Let m be in Z+, 71,7 2 be in R and
Q = { z E Q'lyl 5 argz 5 7 2 ) .
The operator defined by

A = MAL,
where
M is the restricted operator from R" to RT,
L is the extended operator from RT t o R" (in the Proposition 2.2),
A is the pseudodifferential operator of form (3) of order 'rrt in R" with
symbol
~ A ( J :I ,,41, J: E Rn,E E R",4 E Q I
is called the pseudodifferential operator of order m in RT with symbol
o A ( Z , E , q ) = oA(J:,I,4), J: Q.
E IE
Rw"+, Rnl 4E
The operator A is said to be homogeneous if the operator A is homoge-
neous.
The operator A is said to be admissible if its symbol has form
o A ( x l , 0, E , 4 ) = gap(x1i0i '?)<"qP
la I + B l m
m.
k=O
where OA, does not depend on It.

We now consider a boundary value problem in RT.
Let s, m l , . . . , ms be in Z+, 71,7 2 be in R and
Q =(2 E CIn Iargz 5 7 2 ) .
We now study the boundary value problem
A 4 J : ,4 ) = A b , D , Q)+I 4) = f(J:,4), J:" > 01 (13)
BjU(J:,4)JZn--0 = ~ j ( J : , D , 4 ) z 1 ( 5 , q ) I z , ==gj(J:',q),
0 j = l,...,S, (14)
where 4 E Q, A , Bj are homogeneous admissible pseudodifferential opera-
tors of order (respectively) 2s, m j with symbols (respectively) O A ( J : ,E , q ) ,
U B j (J:,1
I 4).
Put t o = m a x { 2 s , m l + I , . . . , m , I}, +
with respect to the norm

S
IIIV,g)IIIe,P = III(f,gi,...,gs)IIIe,p = I I I ~ I I I ~C- IIIgsIIIe-~j-(l-~),p.

~~,~+
j=1
It is easy to prove
Proposition 4.1. The operator U is a bounded linear operator f r o m

He,P,q(RT)to Hg,p,q(Ry,Rn-'). Moreover, the estimate
I IIUullle,P I CI ll4l e m u E Ht,P(R?)

holds true, where C is a constant not depending o n u, q .
Let us now consider the operators A , Bj possessing symbols CA(Z, <,q ) ,

<,q ) not depending on x,i.e.
U B (x,
~
gA(x, <,q ) = a A ( < , q ) , UBj (x,<,4 ) = U B j (6,q ) ,

we have the following useful results.
The problem (13) - (14) is said to satisfy the condition Shapiro-
Lopatinski if the problem on the halfline t 2 0
d
CA(<', iz,q ) v ( t ) = 0, t > 0 (15)
t 'd
CBj(< ,zz,q)v(t)It=o=hj, j = l , . . . , s ~ (16)
+
for [<'I 141 # 0, has a unique solution in the space M of all stable solutions
of (15) for arbitrary hj.
The problem (13) - (14) is said to be elliptic if
g A ( < , 4 ) # 0, for + Iql # 0,
0 the problem satisfies the condition Shapiro - Lopatinski.
Theorem 4.1. Let p , C be in R such that C, 5 C, 1 5 p < 00, and the

problem (13) - (14) is elliptic. Then, the following statements hold true.
( i ) If q E Q \ (0) then U has the inverse operator U-' is a bounded linear
operator f r o m H E , ~ , , ( RRn-')
~ + , t o H e , p , , ( R ~ )not
, depending o n p , C,
(aa) If q = 0 , there exists a bounded linear operator R f r o m H ~ , p ~ q ( R T ,
to He,P,q(RT),not depending o n p , C such that
UR=IfT,
where, I is the identity operator o n He,p(R"+RIW"-l) T is a bounded
linear operator f r o m He,p,q(RT,Rn-') to H ~ + I , ~ , , ( RRn-').
-~+,
26 N . M . Chuong and D. A . m a n
Proof.
5. Pseudodifferential Operator on a Compact Domain C2

Let R be a compact domain in R",with an ( n - 1)-dimensional, smooth
N
boundary 82,and { U j , qj}j=l be a partition of identity on R.
Definition 5.1. Let m be in Z+,

~ 1 E R~ and~ 2
Q = { z E C l ~I
i argz 5 7 2 ) .
Let A be the operator of form ( 3 ) . The operator A is said to be a
pseudodifferential operator on R of order m if
(i) for each p E Cm(R); P A - A9 is an operator of order ( m - 1) on

&p,,(R), e2 0, 1 < P < 00,
(ii) for each p, II,E Cm(R) whose supports in Uj
0 if U j n r = 0, in the local coordinates
PA[$.] = PA, M.1,

where A, are pseudodifferential operator of order m in R" with symbols
6 , 4 ) , 5 E R",E E R", 4 E Q 1
O A , (x,
0 if U, n r # 0, in the local coordinates
PANJ.1 = PA,M.1,
where A j are pseudodifferential operators of order m in R" with sym-
E E R n l 4 E Q.
bols b ~( z, , E ,q ) , 5 E Rn+,
Semilinear Nonclassical P.d. BVP in Sobolev Spaces He9p 27
The operator A is said to be admissible if for each 'p, ?1, E Cm(R) whose
supports in U j , Uj n r # 0, in the local coordinates
PA[$.]= 'pAj [?1,.1 ,

Aj are admissible.
We consider now the boundary value problem on R.

Let s , ml, . . . ,ms be in Z+, 71,7 2 be in R and
Q = { z E Clyi L argz I 7 2 } .
We now study the boundary value problem
It is obvious that
Proposition 5.1. Let l , p be in B such that l o 5 l , 1 < p < 00. Then U

is a bounded linear operator from He,p,q(R)to He,p,q(R,80).Moreover, the
estimate
holds true, where C is a constant not depending on u,q.

28 N . M. Chuong and D. A . m a n
Ellipticity C o n d i t i o n
Condition 5.1. Let xo E R with a neighbourhood U j . In the locall co-

ordinate system of 20, the operator A changes into the pseudodifferential
operator Aj of order 2s in R" having the principal part Aoj with the symbol
CAoj (2, < I q ) .
The problem (17) - (18) is said to satisfy the Condition 5.1 at xo if
CAoj ( 0 ,< I 4 ) # O1 for 1" -tIqI # O*

Condition 5.2. Let xo E dR with a neighbourhood Uj. In the locall co-
ordinate system of 20, the operator A changes into the admissible pseu-
dodifferential operator Aj of order 2s in RT having the principal part Aoj
with symbol C A (x, ~ El ~ q ) , the operator Bk changes into the admissible pseu-
dodifferential operator of order mk in RT having the main part Bokj with
symbol CB0k3 ( X I I ,4 ) .
The problem (17) - (18) is said t o satisfy Condition 5.2 a t zoif the problem
on half-line t 2 0
when 5'1 + Iq( # 0 , has a unique solution in the space M of all stable
solutions of (19) for arbitrary hk.
The problem (17) - (18) is said t o be elliptic if it satisfies Condition 1
for every xo E G and Condition 2 for every xo E dR.
The following proposition is proved without difficulties
Proposition 5.2. Let I , p be in JR such that l o 5 I , 1 5 p < 00, and the

problem (17) - (18) be elliptic. Then, the estimate
holds true, where C is constant not depending on u, q.

This implies that, when q having large enough 141, for every (f , g ) E
He,,(R, d o ) , if the problem (17)-(18) has a solution, then it is unique.
When the problem (17) - (18) is said to be elliptic, we also say the operator
U is elliptic. To obtain the next theorem we need the following easily proved
Lemma
Lemma 5.1. If U is elliptic, then U has a regularizer R satisfying

RU = I1 + Ti, U R = I2 -k Tz,
where I I , 1 2 are the identity operators (respectively) o n He,,,,(R),
He,p,q(R,dR);TI i s a bounded linear operator f r o m He,p,q(R)t o
He+l,p,q(R);T2 i s a bounded linear operator f r o m He,,,,(R,aR) t o
He+i, p , q (a,an).
Lemma 5.2. Let Bo, B , B' are Banach spaces with respective n o r m II.110,
11 . 11, 11 . 11'. A s s u m e that B L-) Bo i s a compact imbedding operator and
A ; B + B' i s a bounded linear operator. T h e n , the necessary and s u f i c i e n t
condition such that d i m ( K e r A ) < +CQ and ImA i s closed B', i s
llull 5 C(IIA4' + 11~110)l 21 E B, (21)
where C i s a constant not depending o n u.
By using Proposition 5.1, 5.2, Lemmas 5.1 and 5.2 we obtain
Theorem 5.1. Let e, p be in R such that lo5 e, 1 < p < 03. If U is elliptic
then for large enough IqI the operator U i s a n isomorphism f r o m He3p9q(fl)
t o He,p,q(Ql80).
6. A Linear Non-classical Pseudodifferential Boundary

Value Problem
We study now the following non-classical pseudodifferential boundary value

problem
Au = f(x) i n R (22)
B j u = Bj(D,u) = g j o n d R , j = 1 , 2 , . . . , s (23)
where A, Bj are of order 2s, mj respectively in the problem (17)-(18), D,
is a vector field that can be tangent to dR on a smooth manifold ro c 80
of dimension ( n - 2) and not tangent to ro. We use here the classification
of ro given in Ref. 10.
For the reader's convenience let us recall shortly this classification. Let
p be the inner normal unit vector of the boundary dfl. The real function
( p , v ) is defined on the boundary aR,. The submanifold is said to be of
the first class if on some neighborhood of ro, the function ( p , v ) is negative
on the negative part (with respect t o the direction v), and positive on the
positive one, the second class is positive on the negative part and negative
on the positive one, the third class does not change sign when passing ro.
If ro belongs t o the first class, then we add the following conditions

Dhu(x,q) q), k
= U~k(x, = 0,1,.. ., s - 1 on Ro. (24)
The above problem will be discussed in Sobolev spaces He,prq,1 < p < 00.
It is assumed that the problem defined by (A, B j ) is elliptic that is A is
an elliptic operator and the Shapiro - Lopatinski condition is satisfied.
Let r d be the intersection of a d-neighborhood of ro with R, h = hd(x) E
C" and
0, if x is outside of r d
hd(x) =
1, if x is in
Denote by IIe,p,q(R)the space with norm
where D, = g, N = Nd = {x E Gd13y E ro,3 is a normal vector to

6'0, d ( z , ro) I d } , if r o is of the first class and if r0 is not of the first class
the term hu in the norm is omitted.
Denote by G e , p , q ( d R )the spaces of functions defined on dR with norm
llullGe,p,q = Ill4lle,P,an + Illh~llle+l,P,an.
Denote ne,p,q(fl,
80) = 1 3 e - 2 s , p , q ( Q ) x fli=lGe-j-(l-~),p,q(dfl),
for C 2
Cl = max{2s, mj 2). +
Theorem 6.1. A s s u m e that the operator ( A ,B j ) i s elliptic, l 2 l , . T h e n
f o r large enough 141, f o r each (f,gj) E rIe,p,q(R,dR) ifro belongs t o the
third class, or additionally uo E He--(l--l/p),p,q(l?O), if ro belongs t o the
first class the problem (22)-(23) o r (22)-(23)-(24) has a unique solution
21 E l 3 e , P , m .
7. A Semilinear Non Classical Pseudodifferential Boundary

Value Problem
Let us now consider the semilinear non classical pseudodifferential boundary
value problem
D2m-1
Au(x7 4 ) = f(x,4, u , . ' . 7 u), in R, (25)
B j u ( x , q )= Bj(D,u(x,q)) = g j ( x , q , u ,. . ., Dm3-'u),on d R , j = 1,.. ., s
(26)
where A, Bj are pseudodifferential operators in (22), (23) in Sec. 6. If ro

belongs to the first class then we add the conditions
D:U(S, q ) = UOk(2, q ) , IC = 0,. . . , s - 1 on ro. (27)
For this problem we get
Theorem 7.1. Assume that ro is of the first class, !2 !I, ( A ,B j ) is el-

liptic, U O , A Hl-k-(l-L),p,q(ro),
P
f,g = (91,.. . ,g8) satisfy these conditions
(i) these maps
(2, q , G) H f(z, q , G),where G = ( ~ 1 ,... ,U N ) ,
(z, q , G) H gj(z, q , Gj), where Gj = ( ~ 1 ,... ,U N , ) , ~= 1 , 2 , . . . , s
satisfy Carathe'odory conditions, i.e. f,gj are respectively continuous
at G, Gj f o r almost (2, q ) , measurable at (z, q ) f o r every G, G j ,
(ii) the map u(z,4 ) (f(z, 4 , G(z, 4)),9j(Z, 4 , G j b , Q ) ) ) ,
i s f r o m rIl,p,q(R) t o rIg,,,,(R, aR) transforms each bounded set into a
relatively compact set, where
,(.'. 4 ) = ( u ( z ,Q ) , . . ., D2s-1u(5,Q)),
(iii)
Proof. According to Theorem 6.1 for 141 large enough, for each w E
(R) the linear problem
D2s-1
Q) = f ( z , 4 ,w,. . . , w) in 0, (28)
. ., Dm3-'w), on a R , j = 1 , . . . ,s,
Bju(z,q)= g j ( z , q , w,. (29)
q ), = u o k ( z ,q ) , IC = 0, I,.. . , s - 1 on ro
D;U(~C (30)
has a unique solution u E I I g , p , q ( Q ) .
So for IqI large enough, we get the map J : w H u from rIl,P,q(R) into
itself. By the assumptions ( i ) - (ii) the map J is compact. By the as-
sumption (iii) there exists a positive R such that the map J maps the
sphere SR = { w E r I ~ , p , ~ ( R ) ~ ~ = ~R
w }~into
~ ~ the ~ , ~BR
~ , ball ( ~ =) { w E
32 N . M . Chuong and D. A . %an
I R}. Therefore, t h e m a p J satisfies t h e Fixed Theo-

ne,p,~(S1)Illurlln,,,,,(n)
rem of Rothe so it possesses a fixed point which is a solution of t h e problem
(25)-(26)-(27).
References
1. R. A. Adams, Sobolev spaces, (Acad. Press, 1975).
2. S. Agmon, A. Dough, L. Nirenberg, Comm. Pure Appl. Math., 624 (1959).
3. M. S. Agranovich, Uspekhi Mat. N a u k 2 0 , 3 (1965).
4. M. S. Agranovich, M. I. Vishik, Uspekhi Mat. Nauk 19,53 (1964).
5. A. V. Bicatze, Dokl. Acad. Nauk. SSSR 148, 749 (1963).
6. R. Borelli, J . Math. Mech. 16,51 (1966).
7. Nguyen Minh Chuong, Dang Anh Tuan, A boundary value problem for sin-
gular integro-differential operators in He,p,l < p < 00, Preprint 2002/28,
(Institute of Mathematics, Hanoi, 2000).
8. Nguyen Minh Chuong, Tran Tri Kiet, On a priori estimates in L P ( R n ) , p 1 2
for nonclassical oblique derivative problems, Preprint 2000/2 6, (Institute of
Mathematics, Hanoi, 2000).
9. Yu. V. Egorov, Nguyen Minh Chuong, Diff. Urav. 24, 197 (1969).
10. Yu. V. Egorov, V. A. Kondratiev, Mat. Sbornik, 78,148 (1969).
11. Yu. V. Egorov, Nguyen Minh Chuong, Di8. Urav. 20, 2163 (1984).
12. Yu. V. Egorov, Nguyen Minh Chuong, Uspekhi Mat. Nauk 53,249 (1998).
13. Yu. V. Egorov, Mat. Sbornik 73,356 (1967).
14. L. Homander, Ann. of Math. 83, 129 (1966).
15. L. Homander, The analysis 0s linear partial differential operators, 111, Pseu-
dodifferential operator, (Springer - Ve,rlag, 1985).
16. J. Leray and J. Schauder, A n n . Sc. Ecole Norm. Sup. 51,45 (1934).
17. E. H. Lieb, M. Loss, Analysis, Vol. 14, (AMS, 1966).
18. G. 0. Okikiolu, Aspects of the theory of bounded integral operators in L p
spaces, (Academic Press, 1971).
19. Le Quang Trung, Uspekhi. Mat. Nauk 44, 169 (1989).
$3. SINGULAR INTEGRAL OPERATORS IN FUNCTIONAL

SPACES OF MORREY TYPE
LUBOMIRA SOFTOVA
Postal address: Politecnico d i Bari
Dipartimento d i Matematica
Via E. Orabona 4 ,
70125 Bari, Italy
E-mail: 1ubaQdm.uniba.it
Keywords: Generalized Morrey spaces, Singular integrals, Commutators,

Calder6n-Zygmund kernel, B M O , V M O .
MR(2000) Subject Classification: 42B20
1. Definitions and Preliminary Results

We consider the following integral operator
Kf(.) := P.V. Ln k ( z ;z - Y).f(Y)dY
and its commutator with essentially bounded functions
C b , kIf(.) := P.V.
1L k(.; .-y)[a(y)-a(.)lf(y)dy = K:(af)(.)-.(.)Kf(.).
(2)
The generating kernel k ( x ;5) : R" x {R" \ (0)) 4 R is the one studied by
Fabes and Rivibre' and is a generalization of the classical kernel of Calder6n
and Zygmund. Introducing a new metric p ( x ) as a solution of the equation
F(rc,p) = Cy=lz!p-2a' = 1, cri 2 1, they study (1) in LP(R"). The balls
with respect to p ( z ) , centered at the origin and of radius r are simply the
ellipsoids
with Lebesgue measure I&,.[ = C(n)ra.It is easy to see that E l ( 0 ) = S"-l

with respect to the Euclidean metric. This fact allows to impose on the
kernel k ( x , [) the Calder6n-Zygmund conditions on the unite sphere, in spite
34 L. Softova
ti,i = 1,.. . ,n.

of the lack of "symmetry" of k with respect to the variables
Let we note, that the standard parabolic metric = sup{lzl,&}, z E
R",t E ( 0 , ~does
) not permit t o define the mentioned above conditions on
kernels having homogeneity of parabolic type. Using the Fourier transform
in L2(a") and the Marcinkiewicz interpolation theorem, Fabes and Rivikre
obtained that the integral operator (1)is continuous in LP(Rn),p E (1,w).
Definition 1.1. The function k ( z ; C ) :R" x (R" \ (0)) + R is called a

variable kernel with mixed homogeneity if
i) for every fixed x the function k ( z ; .) is a constant kernel satisfying
ii) for every multiindex p : sup l D f k ( z ;

EEW-1
<)I 2 C(p) independently of z.
Let us note that in the special case ai = 1 and thus Q = n, Definition 29

gives rise t o the classical Calder6n-Zygmund kernels. One more example is
when a1 = . . . = ~ " - 1 = 1, an = 6 > 1. In this case we obtain the kernels
studied by B.F. Jones and discussed in Ref. 8.
Let w : R" x R+ + R+ and for any ellipsoid E we write w ( z , r ) =: w ( ) .
Definition 1.2. A function f E Lfbc(Rn),p E (1,m) belongs t o the gen-

eralized Morrey space LP+' (R") if the following norm is finite
The space LP+(fi) and the norm l l f \ \ p , U i are

~ defined by taking f E LP(fi)
and & n R instead of & in (3).
For w ( z , r ) = 1 we get the Lebesgue space Lkc(Rn)and for w ( z , r ) = r x ,

X E (0, a ) , Lp>"(Rn)coincides with the Morrey space LP2x(R"). However,
there exist weight functions, as w ( x , r ) = T ln(r + 2) for which LPl" (R")
does not coincides with any Morrey space.
Singular Integral Operators an Functional Spaces of Morrey Type 35
For a given measurable function f 6 Ltoc(Rn) define the Hardy-

Littlewood maximal operator M f and the sharp maximal operator f !I as
almost everywhere in Rn and the supremum is taken over all ellipsoids E

centered at x. Define also the operator M, f (x):= (MI f I"(x))'/" for s 2 1.
The next results are weighted variants of the well-known maximal and
sharp inequalities (see Refs. 3,5,15).
Lemma 1.1 (Maximal inequality"). Assume that there are constants

1 and
c c,
such that f o r any xo E Rn and f o r any r > 0
For 1 Is < p < 00, there is a constant Cp,, > 0 such that f o r f E Lpsw(Rn)
IlMJI l P , W I c P , S l l f IlP,W.
Lemma 1.2 (Sharp inequality). Let 0 < ~75 1 and E be an ellipsoid
centered at xo E Rn of radius r. Suppose that w ( x 0 , r ) satisfies (4)and
Then f o r p E ( 1 , co)and f E L?,,(Rn) there exists a constant C independent

o f f such that
Ilf IlP,W I cllf811p,w.
Proof. Let X& be the characteristic function of the ellipsoid and denote
by 2E an ellipsoid centered at xo and of radius 2r. It is easy to verify
) r " / ( p ( x - XO) - r)O 5 1 , for all z E Rn. Further, for any
that M X E ( X 5
x E 2"'& \ 2k&, k = 1 , 2 , . . . the maximal function of X E is comparable
with 2-ka for any 2 as above. From certain properties of the maximal
36 L . Softova
and sharp functions (see Ref. 9, p. 410) it follows
J := s, If(Y)IPdY= l" If(Y)IPX&(Y)dY
The assumptions imposed on the weight function give
Hence
One more background we need is that for spherical harmonics and their
properties (see for details Refs. 4,6,8,13). Recall that the restriction to the
unit sphere SnP1of the homogeneous polynomial P : R" -i R of degree m
is called n-dimensional spherical harmonic of degree m. The space Tm of
these harmonics is a finite-dimensional linear one with gm = dim Tm such
that go = 1, g1 = n and
(m+n-I)!- (m+n-3)!
Sm = < C(n)mnP2,m 2 2. (6)
( n - l)!m! ( n - l)!(m - 2)! -
The orthonormal base {Ysm(x)}zzl of Tm and the system { Y s m } ~ ~ lis~ = = o
a complete orthonormal system in L2(S"-') satisfying
If, for instance, q5 E COo(Sn-') then ~ , , , b s m ~ , ( ~ )is the Fourier series

expansion of 4(x) with respect to {Y,,(z)}. It converges uniformly to 4
Singular Integral Operators in Functional Spaces of M o n e y Type 37
(see Ref. 4) and for any integer 1
2. Singular Integral Estimates

Let k ( z ;[) be a kernel in the sense of Definition 29. In order to ensure the
existence of the operators (1) and (2) in LP(Rn) we restrict our consid-
erations to functions f E LP(Wn), 1 < p < 00 for which the norm (3) is
finite. For the sake of convenience we still denote these spaces by Lp9"(Rn).
Note that the last restriction is necessary only to obtain global estimates.
If we consider a bounded domain R E R" the operators (1) and (2) with a
density f E LP?"(R) will be well defined in LP(R) without any additional
assumptions.
.
Define the nonsingular operators K, and C,[a, k ] , a E B M O by
KEf(.) := /P("-Y)>E
k(z; - Y)f(Y)dY,
CE[% k]f(.) := L(af)(.) - a(z)K,f(z).
We are going to prove that ICE and C,[a,k] are bounded and continuous
from LP?"(R") into itself uniformly in E . This along with the properties of
the kernel will enable to let E + 0 obtaining as limits in the LP+'(Rn)-
topology the singular integrals (1) and (2). Moreover, we shall show that
the last ones are also continuous in LP@(Rn).
Let us note that K f exists in LP(R") as a limit of K,f when E + 0
in the LJ'-norm.8 Moreover, the operator K : LP(Rn) -+ LP(Rn) is con-
tinuous and this leads also to continuity in LP(Rn) of C [ a , k ] if a is an
essentially bounded function. As it concerns to the commutator we are go-
ing to derive a result similar to that of Coifman-Rochberg-Weiss (Ref. 7,
Theorem l),which asserts: if K is Calder6n-Zygmund operator in LP(Rn),
p E (1,m) and a E B M O than the commutator C[a,.] is a well defined lin-
ear continuous operator from LP(Rn)into itself. Later, this result has been
extended by Bramanti-Cerutti2 in the framework of homogeneous spaces.
Based on this background about Calder6n-Zygmund operators, we arc go-
ing to obtain continuity in L*+'(R") and boundedness in terms of (lull, for
the commutator (2) having a kernel of more general type.
Theorem 2.1. Let k ( z ; [ ) be a variable kernel of mixed homogeneity, f E

Lp+'(Rn),p E (I,..), w satisfies (4) and (5), and a E B M O . Then the
integrals K f , C [ a ,k]f E Lp?"(R") exist as limits o f K E f and CE[a,k ] f when
38 L. Softova
E -+0 with respect to the LP>"(R")-norm. The operators K and C [ a ,k ] are

bounded f r o m Lp+'(W") into itself and
IlKfllP,W ICllfllp,wr IIC[a, ~ l f l l P , W I Cllall*llfllP,w

where the constants depend on n, p , Q and k through the constant C(p).
Proof. Let 2 ,y E R" and = Y E S-', then

P(Y)
k ( z ;z - Y) = P(. - Y)-* Cbsm(x)Ysm(K)-

s,m
This way, the Definition 29 ii) and (8) imply
for any integer 1 > 1. Replacing the kernel with its expansion, we get
with xsm(z - y) standing for Ysm(F)y)p(a: - y)-" and being a constant

kernel in the sense of Definition 29 i ) . In order t o get series expansions of
K,f and C,[a, k ] f ,we let z E R" and y E R" t o be such that p(.: - y) > E .
Then (l),(7) and (9) yield
Singular Integral Operators in Functional Spaces of Morrey Type 39
Thus instead of K:f and C [ a ,k ]f we shall study the existence in Lp)"(Rn)

of
Csm[a,k]f(x) :=Ksrn(af)(x) - a(x)Xsmf(x).

For what concerns the c i, we dispose of Ref. 8(Theorem 11.1) and this
implies, through Ref. 2 (Theorem 2.5), boundedness in LP(Rn) of Csm[a,k ]
as well. The cited results however require the kernel to have some "inte-
gral continuity" , called the Hormander condition. It turns out that H
' ,(.)
satisfies even stronger condition as shows the following lemma.
Lemma 2.1. (Pointwise Hormander condition (see Ref. 13, Lemma 2.2))
Let & and 2& be ellipsoids centered at xo and of radius r and 2r, respec-
tively. Then
for each x E & and y 4 2E.

Remark 2.1. This result ensures that 'Hsmsatisfies also the integral con-
dition
s :dy)L4ds))
I'Hsm(Y - x) - 'Hsm(Y)l dY
with a constant independent of x (see Ref. 8, (1.1)).

5c
In view of the cited above results there exist K , , f , Csm[a,k]f E LP(Rn)

such that
lim
40
IIxsm,e.f - xmnf IILp(p) = Fz k]f
IICsrn,~[a, - C ~ n ~ [k]f
a rI I ~ p p n =
) 0.
Our goal is to show that this convergence is fulfilled also with respect to
the LP+'(R")-norm. The proof is broken up into several lemmas.
Lemma 2.2. (see Ref. 13, Lemma 2.4) The singular integrals K,,f and
Csm[a,klf satisfy
(K:smf)W I Cm"/2(M(lf lp)(Z))l/pl

(Csm[a,k]f)"x) 5 CIlall*{ (M(lK:smfIP)(x))11p+mn/2(M(lflp)(.)) l'p}
(11)
for all p E (1,co) and the constant depends on n, p , a but not on f
40 L. Softova
Lemma 2.3. The operators Ksrn and Csm[a,k ] are continuous, acting from
Lp,w(Rn),p E (1,oo),into itself and
IIKsrnfIIp,w L Cmn/211fIIp,wr I I ~ s m [ akIfIIp,w

r 5 Cmn/211aII*IIfIIp,w (12)
with constants depending on n, p , and a.
Proof. Let p E (1,m). Applying (11) for any q E (1, p ) and the maximal
inequality (Lemma 1.1) with s = 1 we get
S, 1 ( K s r n f ) n ( x )lpdx 1~mpn/2
S, l M ( l f l q ) ( x lp/ndx
)
-<Cmp"'24E)ll~(lfl~)ll;;;,w
ICmpn/2w()IIIf lql;;,w I Cmp"/2w(E)llf llg,w.
Dividing of w ( E ) and taking supElwe arrive a t
ll(~smf)nllp,w I Cm"/211f IlP,W

which implies the first inequality in (12) through Lemma 1.2. The Lp,w-
estimate for the commutator follows in the same manner. U
Lemma 2.4. The operators KSrn,and Cs m, E[a,

k] are continuous acting
from Lpiw(Rn)into itself and satisfy
IIKsrn,Efllp,w 5 c(n7P , a ) m n / 2 1 1 f I I p , w ,
II~sm,E[al klf Ilp,w 5 C(n1p1 Q)mn/211all*llfIlp,w. (13)
Proof. Let , and ,& I, be ellipsoids centered a t x E R" and of radius E

and ~ / 2 respectively.
, Writing f = fXE, fX(E,)= = f l +
f 2 we obtain +
:= 211(x1& / 2 ) + 12(2,542)
where 11 and 1 2 stand for the terms introduced a t the proof of Lemma 2.2,
and the same arguments as therein lead to
IKsrn,cf(x)I L ~ ( nP , a)mn/'
, (M(IfIn)(x))l'q
Singular Integral Operators in Functional Spaces of Morrey Type 41
for any q E ( 1 , ~ )It. remains to take the LP)w-normsof the both sides
above for 1 < q < p and to apply Lemma 1.1 in order to get (13). The
commutator estimate follows analogously. I7
Returning to the series expansions (lo), we are in a position now to

complete the proof of Theorem 2.1. First of all, note that
as it follows from (9), (1) and Lemma 16. Choosing 1 > (3n- 2)/4 the series
in (10) result totally convergent in Lp+'(Rn), uniformly in E > 0, whence
IIxEfllP,w i CllfllP,w, llCE[a, kIfll,,w i Cll~ll*llfllP,w.
Il~fllp,w I Cllfllp,w, IlC[a, kIfll,,w I Cllall*llfllP,w

through (9), (1) and Lemma 2.3. Finally, the total convergence in Lp?"(Rn)
of (lo), uniformly in E > 0, gives
lim CE[a,k]f(z) = C[a,k]f(x).

E+O
It is worth noting that singular integrals like (1) and (2) appear in the
representation formulas for the solutions of linear elliptic and parabolic
partial differential equations. To make the obtained here results applicable
to the study of regularizing properties of these operators we need of some
additional local results.
Corollary 2.1. Let R be a bounded domain in Rn and k : R x {Rn\{O}} +

R be a variable kernel of mixed homogeneity, a E BMO(R), p E (1,m) and
w satisfies (4) and (5). Then, for any f E LP+('R) and almost all x E 0,
42 L. Softova
the singular integrals
arc well defined i n Lp+(R) and
IlXfll*,w;n 5 Cllfllp,w;n, 11% kIfll,,u;n I Cllall*llfllP,w;n

with C = C ( n , p ,a , R, k).
To obtain t h e above assertion it is sufficient t o extend k ( z ; .) and f(.) as

zero outside R. One more necessary extension preserving t h e norm is t h a t
of a in BMO(Rn) and we have it according t o the results of Jonesl and
Acquistapacel (see Ref. 6 for details).
Another consequence of Theorem 2.1 is the good behavior of t h e
commutator for VMO functions a. Thus using t h e ideas in Ref. 6 (Theo-
rem 2.13) we can prove
Corollary 2.2. Suppose a E V M O with VMO-modulus y a . Then, for each
E > 0 there exists ro = TO(&, y a ) > 0 such that f o r any e E (0, ro) and any
ellipsoid &, of radius e holds
IICb, kIfllp,w;E, I CEllfllp,w;E,. (14)
References
1. P. Acquistapace,Ann. Mat. Pura Appl. 161,231 (1992). 286,A139 (1978).
2. M. Bramanti, M. C. Cerutti,Boll. Un. Mat. Ital. B 10,843 (1996).
3. W. Burger ,C. R. Acad. Sci. Paris 286, 139 (1978).
4. Amer. J . Math. 79, 901 (1996).
5. F. Chiarenza, M. Frasca, Rend. Mat. Appl. 7,273 (1987).
6. F. Chiarenza, M. Frasca, P. Longo, Ricerche Mat. 40, 149 (1991).
7. R. Coifman, R. Rochberg, G. Weiss,Ann. of Math. 103,611 (1976).
8. E. B. Fabes, N. RiviBre, Studia Math. 27, 19 (1966).
9. J. Garcia-Cuerva, J . L. Rubio De Francia, Weighted Norm Inequalities and
Related Topics: North-Holand Math. Studies, Vol. 116, North-Holand, Ams-
terdam, 1985.
10. F. John, L. Nirenberg,Comm. Pure Appl. Math. 14,415 (1961).
11. P.W. Jones, Indiana Uniu. Math. J . 29,41 (1980).
12. E. Nakai, Math. Nachr. 166,95 (1994).
13. D. Palagachev, L. Softova, Potent. Anal. 20, 237 (2004).
14. D. Sarason, Trans. Amer. Math. SOC. 207,391 (1975).
15. E. M. Stein, Singular Integrals and Differentiability Properties of Functions,
(Princeton University Press, Princeton, New Jersey, 1970).
16. L. Softova, C. R. Acad. Sci., Paris, Ser. I, Math. 333,635 (2001).
54. CLASSIFICATION O F INTEGRAL TRANSFORMS
VU KIM TUAN
Department of Mathematics,
University of West Georgia, Carrollton, GA 30118, USA
E-mail: vu@westga.edu
In this survey paper we give a classification of integral transforms based on

their composition structure.
1. Introduction
There is no consensus among mathematicians what should be called an inte-
gral transform, and what should not. Some assume that every linear integral
operator that appears frequently enough to bear some name can be consid-
ered as an integral transform. We accept here another concept. We believe
that the most distinguish character that enjoy integral transforms among
the vast of linear integral operators is their integral inverses. We call there-
fore integral transform every linear integral operator whose inverse is also
an integral operator of almost the same "complexity". To include Riemann-
Liouville fractional operators and Riesz potentials into integral transforms,
we should admit some operators of differentiation of finite order into the in-
verses. So, when we talk about integral transforms we have in mind pairs of
direct and inverse integral transforms. Because of the importance of appli-
cations and relations of integral transforms with differential equations it is
widely believed that new interesting integral transforms can be constructed
as eigenfunction expansions of some differential operators. Many classical
integral transforms are in fact arisen from self-adjoint Sturm-Liouville dif-
ferential operat01-s.~~We will show here another method to construct new
integral transforms, based on few well-known integral transforms. Many
nonconvolution (index) transforms, constructed by that way, seem impos-
sible to be obtained by eigenfunction expansions, since their kernels do not
satisfy any linear differential equations.
44 V. K. Tuan
The most classical transforms are:
The pair of Fourier transforms
The pair of Laplace transforms
rm
The Fourier and Mellin transforms are equivalent subject to some sub-
stitution of variables. But each of them has advantages in different areas
of applications. If the Fourier transform is considered in the complex do-
main, then the Laplace transform turns out to be the Fourier transform of
functions vanishing on the negative half-line. Otherwise they are distinct.
Other well-known transforms include the Hilbert, Hankel, Fourier-sine and
-cosine, Y- and H- transforms, Riemann-Liouville and Weyl fractional op-
erators, Kontorovich-Lebedev, Mehler-Fock and Olevskii nonconvolution
transforms. We have prepared a draft list of integral transforms coming
into our definition. It contained more than 300 integral transforms and was
included in final form in Ref. 27.
Of course it is impossible to have some classification of all known and

unknown integral transforms. In this work we restrict ourselves on classifi-
cation of just one-dimensional integral transforms listed in Ref. 27. Similar
treatment of multidimensional integral transforms is briefly given in Ref. 3 .
Our method of classification is based on composition structure of integral
transforms. Fortunately (or maybe unfortunately), almost all of these in-
tegral transforms have very simple structures. They are build up with the
help of very few operators: Fourier transform (equivalent Mellin transform
Classafieataon of Integral Transforms 45
and Laplace transform as a special case are also included), operator of mul-
tiplication and substitution of variables.
Throughout the paper, if we say that the integral transform
is mapping L,(R; w) into L,I (0;w), we understand that i f f E L,(R;w),

and 0, is any compact subset of R, over it k ( z , .) is continuous. then the
integral
converges to a function gn E Lp~(R;w),and moreover, if 01 c 0 2 c

n,
. . .Onc . . ., such that u,==, R, = then gn converges in L,, (0;w) to g.
All integral transforms can be divided into two classes:
-Convolution transforms
and
-NonconuoZution (index) transforms.
2. Convolution Transforms
A generalized convolution of functions f and k under three operators
K , K1, KZ, and with some weight function w is a function, denoted by the
symbol k * f , such that the following factorization property holds
If K = K1 = K2, we have the classical convolution.

Convolution transforms are called linear integral transforms of the fol-
lowing type
f +g=k* f.
Depending on the convolution, convolution transforms can be divided into
following types:
2.1. Fourier convolution transforms

46 V. K. Tuan
An example is the pair of Hilbert transforms
d z ) = (Hf)(z)=- fody,
Y-2
&dz.
Y-X
2.2. Laplace convolution transforms
An example is the pair of Riemann-Liouville fractional integral and differ-

ential operators
2.3. Mellin convolution transforms
dz)= k(zy)f(y)dy.
An example is the pair of Hankel transforms
f(Y) = (Hvg)(Y) =/ 0
fiJv(zy)s(z)dz. (15)
where J,,(x) is the Bessel function of the first kind.l
Similar to the relation between the Fourier and Mellin transforms, trans-
forms of Fourier and Mellin convolution types can be identified by a simple
substitution of variables. Transforms of Laplace convolution type are spe-
cial cases of transforms of Fourier convolution type (if k ( z ) = f(z)= 0 for
z < 0 in ( 7 ) ) .
Since most of convolution transforms are given in the form of Mellin
convolution type27 we identify the class of Fourier and Laplace convolution
transforms with the class of transforms of Mellin convolution type. The class
Classification of Integral Transforms 47
of Mellin convolution transforms is the most representative class among

known transforms. They have very simple composition structure
(Kf)(z)= (M-lk*(s)(Mf)(l- s))(z), (16)

where k * ( s ) = ( M k ) ( s )- the Mellin transforms of the kernel k ( z ) . Thus,
any Mellin convolution transform is a composition of a Mellin transform, a
substitution of variable s 4 1 - s, an operator of multiplication by k * ( s ) ,
and an inverse Mellin transform. Although the kernels k ( z ) in (13) are very
distinct in different transforms, their Mellin transforms k* ( s ) in general
behave very simple on the line %(s) = l/2. One can see that in almost
cases k*(s) is bounded and not equal to 0 on the line %(s) = 1/2. Based
on the asymptotic behavior of k * ( s ) on the line R(s) = 1 / 2 the class of
Mellin convolution transforms in the form (13) in its turn can be divided
into three subclasses:
2.3.1. Mellin convolution transforms of Watson type

These transforms are characterized by the condition4g
c > Ik*(s)l > c > 0, %(s) = 1/2. (17)
Since the Mellin transform (5) is an homeomorphism from Lz(0, co) onto
Lz(1/2--ico,1/2+ico) and the operator of multiplication by k * ( s ) under the
condition (17) is an automorphism on L~(1/2-i00,1/2+i00), from (17) it is
easy to see that convolution transforms of Watson type are automorphisms
on LZ(0,co).If, moreover,
lk*(s)I = 1, %(s) = 1/2, (18)

then convolution transforms of Watson type are unitary on Lz(0, co). In
the form of Fourier convolution (7) the condition (17) is replaced by
C > I(Fk)(z)l > c > 0, zE R, (19)
and transforms satisfying condition (19) are automorphisms on Lz( -03, co),
and unitary there if I ( F k ) ( z ) (= 1. Examples of unitary convolution trans-
forms of Watson type are the Hilbert and Hankel transforms, the Fourier-
sine transform
dz)=(Fsf)(z)= g p n z Y f ( Y ) d Y , (20)
(21)
48 V. K. Tuan
with the Bessel function of the second kind Yv(z)and the Struve function
H,,(x) in the kernels' are automorphisms, but not unitary in &(O, m).
Taking a composition of Fourier transforms in the form Feix3/3Ff we
obtain the following pair of convolution transforms of Watson type3'l4'
g ( z ) = (Aif)(z)
=
L 00
00
A+ +Y)f(Y)dY, (26)
f (Y) = (&)(Y) = /
--oo
A+ + Y)g(z)dzl (27)
with the Airy function Ai(z)in the kernel1 that are unitary on L2(-m, m).
Taking a composition of Fourier transforms in the form Fe-aaSinh Ff I
a > 0, we obtain another pair of convolution transforms of Watson type41
with the Macdonald function K,(a) in the kernels' that is unitary on

L2(-m, m).
Taking a composition of Fourier transforms in the form Feia'Osh Ff,a > 0,
we obtain one more pair of convolution transforms of Watson type3'
with the Hankel functions HL')(z) and HL2)(z)in the kernels' that are
unitary on L2(-m, m).
2.3.2. Mellin convolution transforms of Riemann-Liouville type

The Mellin transform k*(s) of the kernels of these transforms are charac-
+
terized by power decay on the line (1/2 - zoo, 1/2 zoo) as 3(s) + fw:
0 < c < Ik*(S)Sal < c < 03, a > 0,W(s) = 1/2. (32)
Many transforms of Laplace convolution type (10) such as Riemann-

Liouville fractional integral (11), and Weyl fractional operators
the Riesz potential
the Prabhakar transform27
with the confluent hypergeometric function 1F1 ( a ;c;x ) in the kernel, the
Saigo transform28
50 V. K. Tuan
with the hypergeometric function zFl(a, b; c; z) in the kernel, the Marichev

transform2'
Y X
F3(-a', n - a , -b', n - b; n - c; 1 - -, 1 - -)g(z)&, 0 < Rc < n,
X Y
with the Appell hypergeometric function of two variables

F3(a,a', b, b'; c; x , y) in the kernel are convolution transforms of Riemann-
Liouville type.
The following pairs of integral transforms27
X
g(x) = / x (x - y)"-l E l ( a , a', b; c; 1- -, A(z - y))f(y)dy, (43)
0 Y
and
-z z ( - a , -b; n -
X
c; 1- -, A(x - y))g(")(z)dx, 0 < RC < n,
Y
with the Humbert hypergeometric functions of two variables Z1 and Ez in
the kernels' are nonconvolution transforms, but their properties and com-
position structure containing only Riemann-Liouville fractional integrals
and derivatives are very similar to those considered in this subsection.
Some other special cases such as transforms with orthogonal polyno-
mials in the kernel are compositions of some Riemann-Liouville or Weyl
fractional operators with operators of multiplication. Feller and Riesz po-
tentials are, on the other hand, compositions of two fractional operators
of different type (Riemann-Liouville and Weyl). Their inverses contain, in
general, operators of differentiation of finite order.
Classaficataon of Integral Transforms 51
2.3.3. Mellin convolution transforms of Laplace type

The Mellin transform Ic*(s) of the kernels of these transforms is character-
+
ized by exponential-power decay on (1/2 - zoo,1 / 2 zoo) as 3 (s) + f-00:
0 < c < IIC*(s)eYlslsal < C < M, y > 0, a E R, %(s) = 1/2. (47)
Examples are the Laplace transform, the Stieltjes transform
with the modified Bessel function K,(z) and I,(z) in the kernels,' the
Meijer K , transform
ds) = 1 0
M
e-SyKv(sY) f(Y)dY, (52)
f(y) = A
47rva
/ y+im
y--ioo
sesy [(2v + l ) ~ - l ( s y )+ 4v1,(sy) (53)
+ (2v - l>~,+l(SY)lg(s)ds.
All of these transforms (Laplace, Stieltjes, Meijer, Meijer-Bessel,. . . ) map
integrable functions into functions analytic in some half plane. Their in-
verses, if considered strictly only on the real line, include some differential
operators of infinite order.
For example, the real variable inverse of the Laplace transform has the form
while the real variable inverse of the Stietjes transform has the form
Most of the Mellin convolution transforms are special cases of the fol-
lowing convolution G-transform
52 V. K. R u n
with the Meijer G-function in the kernel of Ref. 8. Kesarwani16-'* has found
the condition when this transform is of Watson type
In this case the inverse has the form
One can find conditions when this transform is of Riemann-Liouville type
and the inverse formula in this case has the form
or of Laplace type
c*=p+q-2(m+n)>0. (61)
with the real variable inverse of the form
(see Ref. 37 for more details).
As it was said, a Mellin convolution transform can be decomposed as

composition of direct and inverse Mellin transforms and an operator of
multiplication
(Kf)(x) = M-' @ * ( s ) M { f ( y )1;- s ) ; .I. (63)

As a special case of the Mellin convolution transform, the convolution G-
transform (56) enjoys more interesting composition structure. The Mellin
transform of the kernel, the Meijer G-function GE;" (x ):I,
is a quotient
of products of Gamma functions ( m+ n in the numerator and p +q - m - n
in the denominator). Each Gamma function in the numerator corresponds
to a modified Laplace transform, and each Gamma function in the denom-

inator corresponds to a modified inverse Laplace transform.2 Therefore]
the convolution G-transform (56) can be decomposed as composition of
+ +
m n direct Laplace transforms and p q - m - n inverse Laplace trans-
forms. If we do not want to include the inverse Laplace transform] that
contains integration in the complex plane, into the composition, then in
the composition we have to use some Riemann-Liouville, Weyl fractional
operators27 and Hankel transforms (the Mellin transforms of their kernels
are quotients of one Gamma function over one Gamma function). The con-
volution G-transform is of Riemann-Liouville or Watson type if and only
if it can be decomposed into composition of only Riemann-Liouville, Weyl
fractional operators and the Hankel transforms (without the Laplace and
inverse Laplace transforms).
Meijer G-function GEhn (x :;;) : 1
contains as special cases many clas-
sical special functions (orthogonal polynomials, Bessel functions] confluent
hypergeometric functions] hypergeometric functions] t o name just a few).
Therefore it has a very complicated asymptotic behavior near 0 and in-
finity,' and studying existence and mapping properties of convolution G-
transform (56) in classical L, spaces is a very challenging problem. However,
its Mellin transform k * ( s ) is a quotient of products of Gamma functions]
and therefore has a very simple asymptotic behavior a t infinity, power-
exponential decay. We introduce now a space of functions M;;(L,) based
on the power-exponential asymptotic behavior of k * ( s ) on %s = l / 2 . For
simplicity we assume p = 2. The general case p is considered in.37 This
space turns out to be very convenient in studying convolution G-transform
(56).
Definition 2.1. Let 2 signc+signy 2 0. By M,+(L2) we denote the set of
functions f on R+ = (Olm)lwhich can be represented as the inverse Mellin
transform
f(x) = / l/Z+im
1/2--im
f*(s)IC-Sds
+
of f*(s) from L2(1/2 - im, 1/2 im) with weight IsIYeTClsl.If the norm
o f f in M;:(L2) is defined through the norm of f*(s), then M;;(Lz) is a
Banach space.
If c = y= 0, then MCA(L2) = L2(R+).If c = 0, y > 0, then f E M;:(L2)

if and only if e"/2f(e")belongs t o the space of Bessel potentials L;(R). If
c > 0, then f E M;+(L2) if and only if f is infinitely differentiable a.e.,
54 V. K. Tuan
and moreover,
(2^c)2" ' ' ' N n 2
< 00.
dX
' L2(R+)
The importance of the space M.C^(LP) can be seen from the following
fundamental result
Theorem 2. 1.37 All convolution G-transforms, whether they are of

Watson, Riemann-Liouville, or Laplace types, are automorphisms from
M - i onto A^~
2.4. Fourier cosine and sine convolution transforms

There exist some different convolutions related to the Fourier cosine and
Fourier sine transforms.
The first convolution has the form29
1 f
(/**)(*) = -7= / [Ml x - y |) + k(x + y)]f(y)dy, (64)
V2.1t Jo
and it satisfies the convolution property
Fe(f*g)(x) = (Fcf)(x) (Fck)(x). (65)
Another convolution was first introduced by Churchill7
1 f
(f*k)(x) = -= \ (k(\ x-y\)-k(x + y)]f(y)dy, (66)
V27T Jo
and the respective convolution property for (66) has the form
F,(f*k)(x) = (F,f)(x)(Fck)(x). (67)
Recently, the third convolution for the Fourier cosine and sine trans-
forms has been discovered25
- = - x)k(\ y-x\) + k(y+ x)}f(y)dy. (68)

V 2?r
It was shown there that
)(x), xeR+ . (69)
Corresponding to these three convolutions we can construct three classes of
integral transforms of convolution type. For example, for the convolution
transform of the type (64) it has been proved43 that the condition
Jo * 4
is necessary and sufficient for the transform
g(x)=(l-^-]
ax
f (k1(x + y)+kl(\x-y\))f(y)dy, x e R+ ,
V ) Jo
(71)
to be unitary in L2(R+) with the symmetric reciprocal formula
/ d2 \ f t \
f ( x ) =[!- } I {ki(x + y)+ki(\x-y\)j g(y)dy, x R+ .
(72)
If moreover, k is twice differentiate, and k(x] = k$x] fcj (x) is locally
bounded on R+, then the integral transform
oo
(k(x + y) + k(\x-y$)f(y)dy, x & R+ , (73)
/
satisfies the Plancherel formula
\\9\\L,(R+) = ||/IU 2 (fl + ). (74)

and reciprocally,
f(x) = j (k(x + y)+k(\x-y\))g(y)dy, x R+ . (75)
As examples we can take
k(x] = -(Ai(x) + Ai(-x) + iGi(x) + iGi(-x)), (76)
and therefore,
k(x) = -(Ai(x) + Ai(-x) - iGi(x) - iGi(-x)). (77)

t
Since Ai(x) and Gi(x), the Airy functions, are bounded on the whole real
line,1 the kernel k(x) is bounded, therefore, transform (73) with the kernel
(18) is a bounded operator from Lp(R+), 1 < p < 2, into Lq(R+), p~l +
q~l = 1. Moreover, it is unitary on Z/2(-R+) and k ( x ) is the kernel of the
inverse operator.
Another example is
f,fx\ _ e~^rH^(a)
lx (78)
2 '
and
L/'.,.')
*'{><) _ -e^TfT-
9 i:c >(a}
\ '' \ <7Q}
'
56 V. K. Tuan
where H}, (a) and Hi, (a) are the Bessel functions of the third kind. v
Transform (73) with kernel (78) is a bounded operator from LP(R+), 1 <
p < 2, into Lq(R+), p"1 + q~l = 1, and moreover, it is unitary on L2(R+)
and (79) gives the kernel of the inverse operator.
Let
k(x) = -V2asinh7ra xia-1/2Kl/2_ia(x), x & R+ . (80)
Then
~k(x) = -\/2asinh7ra x'ia-1/2Ki/2+ia(x), x e R+ . (81)

7T
Transform (73) with kernel (80) is unitary on L2(R+) and its inverse has
(81) as the kernel.
For the convolution (66) it was shown that if fci e L2(R+), then the
transformation
g(x) = 1- (k^x - j/|) - ki(x + y)} f(y)dy, x ,R+,

\ O-X / JQ
is unitary on L2(R+), and the reciprocal formula

_ _
f(x)= (1--J-2a
/ [ki(\x-y\}-kl(x + y)}g(y)dy, x & R+,
\ X / JQ
holds almost everywhere, if, and only if, the kernel satisfies the condition
(18).
Suppose that the kernel k\ is twice differentiable, and k k\ k'[ is
locally bounded on R+, then the integral transform
oo
(k(\x-y\~k(x + y \ ) ) f ( y ) d y , x e R+ . (82)
//o.
satisfies the Plancherel formula
(83)
and reciprocally,
f(x) = I (k(\x-y\) - k(x+y)) g(y)dy, x R+ . (84)
k ( x ) = -r-r-r - [J(ia) - J_ ix (ia) + e^/_(o) - e

ii SllHi 7TX
where J p ( z ) is the Anger-Weber function.' For the convolution (68) it was

shown that if kl E Lz(R+),then the transformation
is unitary on La(R+), and the reciprocal formula
holds almost everywhere, if, and only if, the kernel satisfies the condition
Similar results are obtained when the kernel Icl is twice differentiable, and
Ic = Icl - Icy is locally bounded on R+.
1
+
k ( z ) = - ( G i ( z ) - Gi(-z) iAi(-z) - i A i ( ~ ) ) ,
2
and
3. Nonconvolution Transforms
3.1. W i m p nonconvolution G-transform.
The only class of nonconvolution transforms known until the eighties is the
Wimp nonconvolution G - t r ~ n s f o r r n ~ ' ~ ~ ~
Special cases of this class of transforms are the Kontorovich-Lebedev trans-

formlg
"f
the Mehler-Fock transform"
g(z) =l" Piz-I/2(Y)f(Y)dY,

58 V. K. Tuan
with the Legendre function of the first kind P,,(z)l in the kernel, the
Olevskii transform26
g(z) = J
; 2Fl (a + iz,a - iz;c; -Y)f(Y)dY, (90)
zB(a + iz,a - ix; c; -y)g(x)dz,

with the Gaussian hypergeometric function 2F1 in the kernel,' and the
Lebedev and Lebedev-Skalskaia transform^.'^ The conventional wisdom
is that this is the only class of nonconvolution transforms. You will see
soon how wrong it is. Analyzing the composition structure of the Wimp
nonconvolution G-transform, Y a k u b o ~ i c h has
~ ~ ,found
~ ~ that the Wimp
nonconvolution G-transform can be decomposed into the composition of
a Kontorovich-Lebedev transform and a convolution G-transform (56).
But the Kontorovich-Lebedev transform is a composition of two cosine
Fourier transforms and an operator of substitution, and the convolution G-
transform (56) is a composition of Mellin and inverse Mellin transforms and
an operator of multiplication, then the Wimp nonconvolution G-transform
is a composition of just Fourier transforms (Mellin and Fourier transforms
are the same, up to a change of variables), an operator of multiplication
and substitution of variables.
Using this composition structure we have proved a Plancherel theo-
rem for the Wimp nonconvolution G-transform as well as for its special
cases37$45
Theorem 3.1. The Wimp integral transform
where c* = y* = 0 , p # q , and the integral is under-

stood in L2(R+;xsinh2nx) is a homeomorphism from L2(R+) onto
L2(R+;x sinh2.rrx) and the inverse transform has the form
If, moreover, p = 2n, q = 2m, m # n; an+i = --ail ai E R, i =

1 , .. . ,n; b,+j = -bj, bj E R, j = 1,.. . , rn, then we have the Parseval
identity
rcsinh2.rrxJg(x)12dx= 1 00
If(y)12dy. (94)
In particular, put m = 1,n = p = 0 , q = 2, bl = a - 1/2, bz = 1/2 - a, we

obtain the Olevskii transform with c = a. Hence, we have
Theorem 3.2. Let c = a . T h e n the pair of Olevskii transforms

(go), (91), where the integrals are understood in L2(R+;z2(z +
and L2(R+;y1-2a) norms, respectively, is a homeomorphism between
L2(R+;Y-~)and Lz(R+;zz(z l)4a-3). + Moreover, the Parseval identity
holds
1 00
y1-2alf(y)12dy =
1 00
1 +
z s i n h 2 m p 2 ( a iz)g(z)12dz. (95)
In particular, taking a = 1 / 2 we get the Lz- theorem for the Mehler-Fock

transform established before by L e b e d e ~ . ~ ~
3.2. Transforms of Kontorovich-Lebedev type

Another generalization of the Kontorovich-Lebedev transform comes di-
rectly from its composition structure.37
Let cp(t)be a strictly increasing on R, cp(-00) = -00,cp(00) = 00, so
that the improper integral
k ( z , y) = -
27r Jm e-izt+iYV(t)dt (96)
converges uniformly on any closed interval of R\{O}. For example, if cp(t)is

odd and twice differentiable function, cp(00) = 00, cp(t) > c > 0 on [ N ,m)
and p(t)/(cp(t)) E L ( N ,cm),then the integral (96) converges uniformly
on any closed interval of R\{O}. Then
s(z)= 1, 00
k ( z ,Y ) f ( Y ) d Y , (97)
f ( Y ) =A / q-2,
m
-y)zg(z)dz, (98)
Y -m
is a pair of integral transforms.
- If cp(t)= a sinht, a > 0, we obtain the following pair of integral transforms
s(z)=;
1
s_,
lrz
e~-gnYKiz(lyl)f(y)dy, (99)
l m
J
lrI
f ( y ) =- e~sgnYKiz(lyl)zg(z)dz. (100)
TY -m
If f is an odd function, (99) and (100) become the pair of Kontorovich -
Lebedev transforms.
60 V. K. Tuan
- If cp(t) = sinht + at, a > -1, we obtain the following pair of integral
transforms
- If cp(t)= t 3 / 3 + a t , we obtain the following pair of integral transforms

(103)
f ( y) = - 1
Y
/
--m
00
y-1/3Ai(y-1/3(ay - x))zg(z)dx. (104)
3.3. The Buchholt nonconvolution G-transform

The following "odd" transform was introduced by Buchholz4
with the Whittaker function M p , v ( x )in the kernel.' It is a composition

of a Mellin transform, a multiplication operator, and a Hankel transform.
In its composition structure replace the Hankel transform by a more gen-
eral convolution G-transform (56) we could construct a nonconvolution G-
transform including the Buchholz transform37
as a special case.
Theorem 3.3. The G- transform (107) is a homeomorphismfrom Lz(R+)

onto La(R) and has the inverse
-_
f(Y) = (2.rr)rrrY
( em-i/2y I r 2 ,...,I,f - ( a g f l ) - i x ' 1
9
4- ( b ? + ' ) - i x ,
2
- (a,)-iz
-( b , ) - i x
yi"-
(108)
f g (5)dx.
Classification of Integral R a n s f o m s 61
If, moreover, p = 2 n , q = 2 n , a,+j = -aj E R,j= 1,..., n;bm+j = -bj E

R , j = 1,...,m, then we have the Parseval formula
The G-transform (107) is a composition of a Mellin transform, multiplica-

tion by e-iT-xl'r , and a convolution G-transform.
Put in formulas (107) and (108) m = 1,n = p = 0 , q = 2, bl = u, b2 =
-u, r = a = 1,we obtained the pair of Buchholz integral transforms. Hence,
we have
Theorem 3.4. The pair of Buchholz transforms (105) and (106) are home-
+
omorphisms between L2(R+;y) and L2 ( R ;(x2 l ) x ( v ) e T ( x - ~ z ~More-
)).
over, if u E R then the following Parseval equation holds
In particular, if we take the Fourier cosine (sine) transform as the inner

convolution G-transform then we obtain the pair of transforms
The Parseval equality
also holds.
Here is the composite structure
If we replace the Fourier cosine transform by the Hankel transform we

obtain an integral transform with the Legendre function in the
3.4. The cherry nonconvolution G-transform

In 1949 Cherry' considered an index transformation with & ( z ) as its kernel
and established that, if f E L 1 ( R )is of bounded variation and continuous
62 V. K. man
at x,then
In fact, Cherry's expansion (112) is still valid when the integrability condi-
tion on f is replaced by some condition on the asymptotic behavior o f f a t
infinity such as
where Q > $, c1 and c2 are constants, which may be different for x --t +m
and x .+ -CQ.
To derive composition structure and a Plancherel-type theorem for the
Cherry transform, we rewrite the Cherry expansion (112) in the following
form:
and
respectively. Substituting the integral representation (2.3.15.3) from Ref. 9

for the parabolic cylinder function Dv(z), that is,
and
into the formulas (7) and (8), and interchanging the order of integration
that is permissible if f ( y ) E S(R), the Schwartz space of all infinitely
differentiable functions which rapidly vanish at infinity together with all
derivatives, we arrive at the following composite representations for the
transforms D+, D-, 0; and DI1:
and
Here M and M- are the Mellin transform and its inverse, and F and F-
are the Fourier transform and its inverse. The composite representations
(13) and (14) of the transforms D+ and D- are key tools in studying the
Cherry expansion (6).
By L:(R;e*) (similarly, by L,(R;e*)) we denote a subspace of
&(R) consisting of all square-integrable functions f ( y ) such that the
Fourier transforms of f(y)e$ vanish on the negative (positive) part of
the real axis. These spaces are Hilbert spaces and their elements can be an-
alytically continued onto the whole upper (lower) complex half-plane. Any
function f E L2(R) can be uniquely factorized by
f(Y) = f+(Y) + f-h),

where
f+ E Lz(R;eg) and f- E L,(R;e*).
Applying the Parseval equalities for the Mellin and Fourier transforms
to the composite representations (13) and (14), we obtain the following
Parseval equalities for the transforms D+ and D-:
64 V. K. Tuan
and
From the composite representations (13) and (14), the Parseval equal-
ities, and the mapping properties of the Fourier and Mellin transforms,
by standard procedure we can expand these identities to the whole
spaces L t ( R ; e $ ) and L F ( R ; e * ) . Thus D+ and D- are isomor-
phisms from L$(R;e* ) and L, (R;e$), respectively, onto the space
L2(R;eY(cosh.rrz)-l), and the integrals (7) and (8) converge in the latter
Hilbert space.
Comparing the composite representations (13), (14), (15) and (123), it
is easy to see that (15) is the inverse of (13), and (123) is the inverse of
(14). Hence the operator 0;'is the inverse of D+, and DI1 is the inverse
operator for D-, and the integrals (9) and (117) converge in L2(R). Thus
DTID+f+ = f + and D I I D - f - = f-.
By the definitions of f + and f - , we have
F { f f ( y ) e $ ; -z} =0 and F{f-(y)e$;s} =0 (2 > 0).

Consequently, from (13) and (14) it is easy to see that
(D+f-)(x) =0 and (D- f+)(x) = 0.
Thus, for any f E L2(R), we have
D+lD+f + DI1D- f = D I l D +f + + D I I D - f - =f+ +f- = f.
We sum up our results in the following Plancherel-type theorem:
Theorem 3.5. T h e Cherry transforms (7) and ( 8 ) isomorphically m a p the

spaces L t ( R ;eP) and L,(R; e s ) onto the space L2(R;ef(cosh7rz)-'),
admit the inversion formulas (9) and (117), and satisfy the Parseval equal-
ities (124) and (125), respectively. Furthermore, the expansion formula (6)
holds true in L2(R).
We will show now that the Cherry transform is related to a singu-
lar Sturm-Liouville problem on the whole line. As a consequence, another
Plancherel-type theorem for the Cherry transform is established. For more
details see.3o
Let us consider the following singular Sturm-Liouville problem:
d2Y - q ( 2 ) y = - x y
Ly = -
dx2
(-oo < x < oo), (126)
lY(&:.)I < 00,

with q(z) being even on R such that the spectrum of the singular Sturm-
Liouville problem (1)is the whole R. Let $(z, A) and O(z,A) be the solutions
of the equation (1) such that
d(0, A) = 0, $'(O, A) = -1,

6(0,A) = 1, O'(0, A) = 0.
There exist two functions ml (A) and mz(A) analytic in the upper half-plane
such that, as a function of x,
$1 (X, A> = Q(x,A> + ml (A>4J(z,4

is in Lz(-m, 0), and
1cIz(z, A) = O(z, A) + mz(A)$(z,A>

is in LZ(0,m), for non-real A. Moreover, since q ( z ) is even, we have ml(A) =
-mz(X). In addition, there exist two non-decreasing functions E(A) and C(A)
such that
lim
6+0+
J*s
0
[ 2m2(21+26)
] du=J(A)
and
where the symbol '3 means the imaginary part of a complex-valued function.
It is a known fact31 that if f(z)E Lz(R),then
W
belongs to Lz(R,&), and

JYA) =
Lw 4 dz
f(x)4(z, (129)
belongs to Lz ( R ,@). The following inversion formula
holds true; and in addition, the Parseval equality:

66 V. K. Tuan
is valid.
Now we are ready to consider the following singular Sturm-Liouville
problem
d2y x2
d22 + 7 Y = -XY (--00 < x < a), (133)
Functions
are solutions of the equation (133). Since q ( x ) = -5 is even, q ( x ) 3 -03 as

x -+ f03,and the integral J ' , Iq(x)I-;dx is divergent, there is a continuous
spectrum over ( - 0 ; ) , 0 3 ) , ~ ~and the formulas (4), (5), (8) and (132) hold
true. It is easy to see that
and
satisfy the initial conditions (2) and (3). As x + 03, e y x has the argu-
ment 2,and - e y x has the argument -?.
Consequently, the following
asymptotic formulas for the parabolic cylinder functiong can be used:
(-? < arg(z)

7
< -- .
4
(137)
(x-+ cm).
If X is in the upper half-plane (A =T + id; b > 0), then
,ax-+ - i7--6--1
--z 2 E L2(1,00).
Consequently, we observe that

q-z, A) + m2(44(-z, E ~ ~ ( 0 ~ 0 0 )
for all X in the upper half-plane if and only if the term -z-ix-ie-$ van-
ishes, that is,
and
So, finally, the formulas (4),( 5 ) , (8) and (132) take the forms:
68 V. K. Tuan
respectively.
If we put
then the formulas (140), (141), (142) and (143) become

Classzficataon of Integral Transforms 69
and
respectively.
Hence we can sum up our results in the following Plancherel-type the-
orem.
Theorem 3.6. The integral transforms (144) and (145) are isomorphisms
f r o m L2(R) onto L,(R; e* (cosh AT)-'). Moreover, the inversion formula
has the form (146) and the Parseval equality (147) holds true.
One can show that the transforms (144), (145) and (146) are equivalent
to the Cherry expansion (112).
The composition structure of Cherry's transform gives rise t o a new
Cherry nonconvolution G - t r a n ~ f o r m . ~ ~ ~ ~ ~
Let p = 2n, q = 2m, m # n; an+i = -ai, ai E R, i = 1 , .. .,n; bm+j =
-bj, b j E R, j = 1 , .. . , m. Then Cherry G-transform
is an isomorphism from L2f(R) onto L2(R+)and the following inverse for-

mula holds
Similarly, Cherry G-transform
is an isomorphism from LT ( R )onto La (R,) and the following inverse for-

mula holds
The following expansion holds in L2 ( R )

70 V. K. Tuan
3.5. Other nonconvolution G-transforms

Let A, y @ (-a,
01. Then the integral transform3*
is a bounded operator from L2(0,m) onto L2(-00, m) and has the inverse
If, moreover, m = n , p = q, bj = - u j , j = 1,.. . , p , d j = - c j l j =

1,. . . ,m, X = -y = 26, b E R, then the Plancherel equality holds
l o o
-
2n 1, 19t4l2dz= .Im If(Y)12dY.
G-transforms (107) and (152) look completely different, but they have very
similar composition structure. In fact (152) can be obtained from (107) by
replacing the multiplicator e-zTZ1 byz.
Playing with only the Fourier transforms, operators of multiplications
and substitutions of variables we have found some more new classes of
nonconvolution G-transforms and prove L2 theorems for them. As special
cases about 60 new transforms with L2 theorems are added t o the list of
around 300 known transform^.^^
References
1. M. Abramowitz and I. A. Stegun Handbook of Mathematical Functions with
Formulas, Graphs, and Mathematical Tables. Reprint of the 1972 edition.
(Dover Publications, Inc., New York, 1992), pp. xiv+1046
2. Yu. A. Brychkov, H.-J. Glaeske and 0.1. Marichev Itogi Nauki i Tekhniki.
Mat. Anal., V I N I T I A N SSSR 21, 3 (1983).
3. Yu. A. Brychkov, H.-J. Glaeske, A.P. Prudnikov and Vu Kim Tuan, Mul-
tidimentional Integral Transformations, Gordon and Breach, (New York -
Philadelphia - London - Paris - Montreux - Tokyo - Melbourne - Singapore,
1992).
4. H. Buchholz, The Confluent Hypergeometric Function, (Springer Verlag,
Berlin - New York, 1969), pp. 240
5. I. W. Busbridge, J . London Math. SOC.9, 179 (1934).
6. T . M. Cherry, Proc. Edinburgh Math. SOC.(2) 8 , 50-65 (1948).

7. R. V. Churchill, Operational Mathematics. 2nd Ed. (McGraw-Hill Book Co.,
Inc., New York - Toronto - London, 1958), pp. ix+337
8. A. Erdklyi, W. Magnus, F. Oberhettinger, and F.G. Tricomi, Higher Tran-
scendental Functions, Vol. I. Based, in part, on notes left by Harry Bateman,
(McGraw-Hill Book Company, Inc., New York - Toronto - London, 1953),
pp. xxvi+302
9. A. Erdklyi, W. Magnus, F. Oberhettinger, and F.G. Tricomi, Higher Tran-
scendental Functions, Vol. 11. Based, in part, on notes left by Harry Bateman,
(McGraw-Hill Book Company, Inc., New York - Toronto - London, 1953), pp.
xxvi+396
10. A. Fock, Dokl. Akad. Nauk S S S R 39,279 (1943).
11. H.-J. Glaeske and Vu Kim Tuan, Serdica Bulgar. Mat. Publ. 16,143 (1990),
MR 91k: 44009.
12. H.-J. Glaeske and Vu Kim Tuan, Boundary Value and Initial Value Problems
in Complex Analysis: Studies in Complex Analysis and Its Applications to
Partial Diflerential Equations 1 , (Halle, 1988), pp. 209-220, Pitman Res.
Notes Math. Ser. N 256,1991, Longman Sci. Tech., Harlow., MR 93h: 42010.
13. H . 4 . Glaeske and Vu Kim Tuan, Math. Nachr. 152,179(1991).
14. G. H. Hardy and E. C. Titchmarsh, Proc. London Math. SOC.35,116 (1933).
15. Jr. Hirschman and D. V. Widder The Convolution Transform, (Princeton
Univ. Press, Princeton, 1955).
16. N. Kesarwani, Proc. Amer. Math. SOC.13,950 (1962).
19. M. I. Kontorovich and N. N. Lebedev, Zhurn. Eksperim. i Teoret. Fis. 8 ,
1192 (1938).
20. E. R. Love, Proc. Edinburgh Math. SOC.15,169 (1967).
21. 0. I. Marichev, Soviet Math. Dokl. 13, 703 (1972).
22. 0. I. Marichev and Vu Kim Tuan, Some Volterra equations with the Appell
function F1 in the kernel (Russian). Scientific Works of the Jubilee Seminar
o n Boundary Value Questions, (Collect. Artic., Minsk, 1985.) pp. 167-172.
23. 0. I. Marichev and Vu Kim Tuan, Reports of the Extended Sessions of a
Seminar of the I. N . Vekua Institute of Applied Mathematics (Tbilisi, 1985),
(Tbilisi Gos. Univ., Tbilisi, 1985), pp. 139-142.
24. 0. I. Marichev and Vu Kim Tuan, Complex Analysis and Applications '85
(Varna, 1985), pp. 418-433, ( Bulgar. Acad. Sci., Sofia, 1986).
25. Nguyen Xuan Thao, V. A. Kakichev, and Vu Kim Tuan, East- West J. Math.
1,85 (1998).
26. M. N. Olevskii, Dokl. Akad. Nauk S S S R 69, 11 (1949).
27. A. P. Prudnikov, Yu. A. Brychkov and 0. I. Marichev, Integrals and Series.
Vol. 5. Inverse Luplace Transforms, (Gordon and Breach Science Publishers,
New York, 1992), pp. xx+595.
28. M. Saigo, Math. Rep. Kyushu Univ. 11,135 (1977/78).
29. I. N. Sneddon Fourier Transforms, Reprint of the 1951 original, (Dover Pub-
lications, Inc., New York, 1995), pp. xii+542.
30. H. M. Srivastava, Vu Kim Tuan, and S. B. Yakubovich, Quart. J. Math.
Oxford Ser. 51,371 (2000).
72 V. K. Tuan
31. E. C. Titchmarsh, Eigenfunction Expansions Associated with Second-Order

Differential Equations. Part I, Second Edition, (Clarendon Press, Oxford
1962), pp. vi+203.
32. Vu Kim Tuan, Dokl. Akad. Nauk BSSR (Russian) 29, 584 (1985).
33. Vu Kim Tuan, Dokl. Akad. Nauk. Armyan S S R (Russian) 83, 7 (1986).
34. Vu Kim Tuan, Dokl. Akad. Nauk S S S R (Russian) 286,521 (1986), translation
in Soviet Math. Dolcl, 33, 103 (1986).
35. Vu Kim Tuan, Dokl. Akad. Nauk B S S R (Russian) 30, 689 (1986).
36. Vu Kim Tuan, Generalized integral transformations of convolution type in
some space of functions. Complex Analysis and Applications '85 (Varna,
1985), (Bulgar. Acad. Sci., Sofia, 1986) pp. 720-735.
37. Vu Kim Tuan, Integral Transforms and Their Composition Structure (Rus-
sian), Dr. Sci. Dissertation, (Belarusian State University, Minsk, 1987), p.
275.
38. Vu Kim Tuan, Dokl. Akad. Nauk S S S R (Russian) 299, (1988), translation in
Soviet Math. Dokl. 37, 317 (1988).
39. Vu Kim Tuan, Dokl. Akad. Nauk S S S R (Russian) 300,521 (1988), translation
in Soviet Math. Dokl. 37, 669 (1988).
40. Vu Kim Tuan, Some integral transformations with a Macdonald function
in the kernel (Russian) Current Analysis and Its Applications, (Naukova
Dumka, Kiev, 1989), pp. 16-22.
41. Vu Kim Tuan, Ukrain. Mat. Zh. (Russian) 4 2 , 9 9 0 (1990), translation in Ukr.
Math. J. 4 2 , 880 (1991).
42. Vu Kim Tuan, Airy integral transform and the Paley-Wiener theorem, in
Proceedings of the 2nd International Workshop on Transform Methods and
Special Functions, held in Varna, Bulgaria, August 24-29, 1996, Editors P.
Rusev, I. Dimovski and V. Kiryakova, (Institute of Mathematics, Sophia,
1998), pp. 523-531.
43. Vu Kim Tuan, J . Math. Anal. Appl. 229, 519 (1999).
44. Vu Kim Tuan, Dokl. Akad. Nauk S S S R (Russian) 313, 1299 (1990), transla-
tion in Soviet Math. Dokl. 42, 150 (1991), MR 9lk: 44006.
45. Vu Kim Tuan, 0. I. Marichev and S. B. Yakubovich, Dokl. Akad. Nauk SSSR
(Russian) 286,786 (1986), translation in Soviet Math. Dokl. 33, 166 (1986).
46. Vu Kim Tuan and Nguyen Thanh Hai, Dokl. Alcad. Nauk S S S R (Russian)
317, 797 (lYYl), translation in Soviet Math. Dokl. 43, 508 (1991).
47. Vu Kim Tuan and S. B. Yakubovich, Dokl. Akad. Nauk BSSR (Russian)29,
11 (1985), translation in Amer. Math. SOC.Transl. 137, 61 (1987).
48. Vu Kim Tuan and S. B. Yakubovich, Ukmin. Math. Zh. (Russian) 44, 697
(1992), translation in Ukrainian Math. J . 44, 630 (1993).
49. G. N. Watson, Proc. London Math. SOC. (2) 35, 156 (1932).
50. D. V. Widder, The Laplace Transform, (Princeton Univ. Press, Princeton,
1946).
51. J. Wimp, Proc. Edinburgh Math. SOC.(2) 14, 33 (1964).
52. S. B. Yakubovich, Izu. Vyssh. Uchevn. Zaved Mat. (6), 77-79 (1986).
53. S. B. Yakubovich, Indez Transforms, (World Scientific Publishing Co., Inc.,
River Edge, NJ, 1996), pp. xiv+248.
54. S. B. Yakubovich, Vu Kim Tuan., 0. I. Marichev , and S. L. Kalla, Rev. Tecn.
Fac. Ingr. Univ. Zulia, Edicion Especial 10, 105 (1987).
Chapter I1
PARTIAL DIFFERENTIAL EQUATIONS

Eds. N. M. Chuong et a2. (pp. 75-89)
55. UNIFIED MINIMAX METHODS
MARTIN SCHECHTER
Department of Mathematics University of California, Irvine
Irvine, C A 92697-3875
E-mail: mschechtOmath.uci.edu
The variational approach to solving nonlinear problems eventually leads t o

the search for critical points of related functionals. In case of semibounded
functionals, one can look for extrema. Otherwise, one is forced t o use minimax
methods. There are several approaches t o such methods. In this paper we unify
these approaches providing one theory that works for all of them. The usual
approach has used Palais Smale sequences. We show that all of them lead t o
Cerami sequences as well. Applications are given.
1. Introduction
Many problems arising in science and engineering call for the solving of the
Euler equations of functionals, i.e., equations of the form
G(u) = 0 ,
where G(u) is a C1 functional (usually representing the energy) arising from
the given data. The classical approach was to look for maxima or minima. If
one is looking for a minimum, it is not sufficient to know that the functional
G(u) is bounded from below. However, if it is bounded from below, one can
show that there is a sequence satisfying
G ( u ~-+) a , G ( u ~-+
) 0 (1)
for a = inf G. If the sequence has a convergent subsequence, this will pro-
duce a minimum.
A functional G is said to satisfy the PS condition if every sequence of
the form (3) has a convergent subsequence and consequently produces a
critical point.
AMS Subject Classification: Primary 35565,58305,49B27.

Key words and phrases: Critical point theory, variational methods, saddle point theory,
semilinear differential equations.
76 M.Schechter
Actually, one can do better. If the functional G(u) is bounded from

below, then there exists a sequence satisfying
G(N) -+ a, +
(1 l l u k I l ) G ' ( ~ -+
) 0 (2)
for a. = inf G. Such a sequence is called a Cerami sequence. It will have a
convergent subsequence even in cases when a PS sequence does not.
However, when the functional is not semi-bounded, there is no orga-
nized procedure for producing critical points. One approach is to search
for circumstances other than semibound- edness which will produce PS se-
quences (3). A successful method has been t o find subsets A and B having
the property that when they separate a functional, a PS sequence (3) re-
sults. Specifically, we want the subsets t o have the property that for every
G E C' (E l R) bounded on bounded sets and satisfying
a0 := sup G < bo := inf G, (3)
A I3
there is a sequence {uk} cE and a constant a such that
bo<a<m (4)
and (3) holds. When this is true, we say that A links B. Fortunately, such
pairs exist and help solve many problems.
There have been several approaches to linking;2~6~7~13~14 we describe
them below. Here we shall present an approach which includes all of them,
is more general and is easy t o apply.
The idea is as follows. For each A c E one wishes t o find a collection K
of sets K with the following properties.
1. If G E C'(E, R) satisfies
a0 := s u p G
A
< a := KEK
inf supG,
K
(5)
then there is a sequence {uk} c E such that (3) holds.

2. If
CLO 5 SUP GI K E Ic,
K\A
then there is a sequence {uk} c E such that (3) holds.

3. If A , B c E satisfy
AnB=4, BnK=+, K E K
and
a0 := sup G 5 bo := inf G,
B
A
Unified Minimax Methods 77
then there is a sequence { u k } c E such that (3) holds.

4. If
gK = { u E K\A: G ( v )2 ao} # 4,K E K, (9)
then there is a sequence { u k } c E such that (3) holds.
In the next two sections we determine collections K that have these
properties. Moreover, we shall show that all of these collections produce
not only PS sequences (3), but also Cerami sequences (4). In Sections 4-6
we describe known methods of linking and find the sets K in each case. In
Section 7 we give some applications.
2. Definitions and Theorems

In this section we begin with a Banach space E.
Definition 2.1. We shall say that a map 'p : E + E is of class A, if it is a
homeomorphism onto E, and both 'p, 'p-'are bounded on bounded sets.
Definition 2.2. For A c E, we define
R ( A ) = {'p E A : P(U) =U, u E A}.
Definition 2.3. For a nonempty set A c E , we define a nonempty collec-

tion K = K ( A ) of subsets K c E t o be a minimax system for A if it has
the following property:
p(K)E K, 'p E A ( A ) , K E K.
Every nonempty set has a minimax system. We have
Theorem 2.1. Let K be a minimax system for a nonempty subset A of E ,
and let G ( u ) be a C' functional on E . Define
a := inf sup G , (10)
KEK K
and assume that a is finite and satisfies
a > a0 := supG.
A
Let $(t) be a positive, locally Lipschitz continuous function on [0,co)
such that
Then there is a sequence { u k } c E such that

78 M. Schechter
Definition 2.4. We shall call a ( t ) E C([O,T]x ElE ) , T > 0, a flow if

a ( 0 ) = I and for each t E [O, TI , a ( t ) is a homeomorphism of E onto itself.
Theorem 2.2. Let K be a minimax system for a nonempty set A and let
G ( u ) be a C functional on E such that
gK = {w E K\A: G ( v ) L ao} # 4,K E K,

and
a = infsupG < 03.

I C K
Assume that f o r any b > a , K E K and flow g ( t ) satisfying
G(a(t)v)< b, w E K,t > 0,

and
G(a(t)u)< a , u E A , t > 0 ,
there is a K E K: such that there is a K E K such that
I? c U a ( t ) AU a(T)[EbU K ] ,
tE[O,tl
where
E, = {U E E : G(u) 5 0).
Then the conclusions of Theorem 2.1 hold.
3. Linking Subsets
We now show how linking can play a major role in finding critical points.
Definition 3.1. We shall say that a set A in E links a set B cE relative

to a minimax system K for A if
AnB=$ (20)
and
B n K # 4, K E K.
We shall say that A links B [mm] if there is a minimax system K: for A

such that A links B relative to K.
Theorem 3.1. Let K be a minimax system for a nonempty set A, and

assume that there is a subset B of E such that A links B relative to K .
Assume that G E C 1 ( E R)
, satisfies
a0 := sup G < bo := inf G , (22)
A B
and that the quantity a given b y (1) is finite. Then for each positive, lo-
cally Lipschitz continuous function Q ( t ) on [0,co) satisfying (12), there is
a sequence { u k } c E such that (13) holds.
Definition 3.2. A subset A of a Banach space E links a subset B of E if

for every G E @ ( E lR) bounded on bounded sets and satisfying (22) there
is a sequence { u k } c E and a constant a such that
bosa<co (23)
and (3) holds.
Definition 3.3. A subset A of a Banach space E links a subset B of E

strongly if for every G E C1(E,R) bounded on bounded sets and satisfying
(22) and each positive, locally Lipschitz continuous function +(t)on [0, co)
satisfying (12), there is a sequence { u k } C E such that (10) and (13) hold.
Theorem 3.2. If A links B [mml, then it links B strongly.

assume that A links a subset B of E relative to K . Let G be a C1 functional
satisfying (14)) (15) and (24)
a0 := sup G 5 bo := inf G. (24)
A B
Assume that f o r any b > a , K E K and flow a ( t ) satisfying (16), (17) and
such that
a ( t ) An B = 4, t E [0,TI, (25)
there is a K E K such that (18) holds. Then the conclusions of Theorem
2.1 hold.
Theorem 3.4. Let K: be a minimax system for a nonempty set A satsfying
K\A=4, KEK. (26)

Let G E C1 satisfy (14) and (15). Assume that for any K E K and flow
a ( t ) satisfying (16), (17), one has S ( K ) E K , where
S ( U )= a ( d ( u ,A))u,u E E . (27)
80 M.Schechter
Assume also that for each K E K there is a p > 0 such that G ( u )is Lipscluitz
continuous on
Then the conclusions of Theorem 2.1 hold.

assume that there is a subset B of E such that A links B relative to Ic.
Assume for any K E K andflow a ( t ) satisfying (16)) (17)) one has S ( K ) E
K , where S(u) is given by (27). Let G be a C1 functional satisfying (15)
and(24). Assume also that for each K E K there is a p > 0 such that
G ( u )is Lipschitz continuous on K p given by (28). Then the conclusions of
Theorem 2.1 hold.
Corollary 3.1. Let K be a minimax system f o r a nonempty set A , and let

G E C1 satisfy (14) and (15). Assume that f o r any b > a, K E K and flow
a ( t ) satisfying (16),(17)) and (25) for
B = { u E E \ A : G(u)2 ao}, (29)
there is a K E Ic satisfying K c a(T)Eb. Then the conclusions of Theorem

2.1 hold.
We present some linking methods which are special cases of linking with
respect to minimax systems.
4. A Method Using Homeomorphisms

One linking method can be described as follows. Let Ebe a Banach space
and let be the set of all continuous maps r = r(t)from E x [0,1] t o E
such that
1. r(0)= I , the identity map.

2. For each t E [0,1),I'(t)is a homeomorphism of Eonto Eand r-'(t)E C
( E x [0,1),E).
3. r(1)Eis a single point in Eand J?(t)Aconverges uniformly to r(l)Eas t
+ 1 for each bounded set A c E.
4. For each t o [0,1) and each bounded set A c E
We have
Theorem 4.1. Let G be a C1-functional o n E , and let A be a bounded

subset of E . Assume that
a0 := supG < a := inf sup G ( r ( s ) u )< 00.
rE+ O < S < l
uEA
Let $(t) be a positive, locally Lipschitz continuous function o n [0,m) sat-

isfying (12). Then there is a sequence {uk} c E such that (13) holds. I n
particular, there is a Cerami sequence satisfying (4).
Proof. In this case, we lct K be the collection of sets of the form

K = {r(tlu :t E [o, 1 1 , uE ~ , E ra}. (32)
To show that this is a minimax system, let 'p be a mapping in A(A), and
let r(t)be in a. Define the mapping
r,(t)u= 'p r(t) 'p-lu, E E.
Then rl(t)E a. Moreover,

rl(+ = 'p r(t)u,0 E A.
Consequently,
r l ( t ) A= ' p ( { r ( t ) u: u E A } ) ,
showing that the collection Kis a minimax system. Since (31) now becomes
( 3 ) , the result follows.
Definition 4.1. We say that A links B [ s t ] if A , B are subsets of E such

that A n B = 4 and, for each r(t) E i p , there is a t E (0,1] such that
r ( t ) An B # 4.
We have
Corollary 4.1. If A links B [ s t ] ,then it links B[mm].
We also have
Theorem 4.2. Let A, B be subsets of E such that A links B [ s t ] ,and let

G be a C1 -functional on E , satisfying
a0 := supG 5 bo := inf G. (33)
A B
82 M. Schechter
Assume that
a : = inf sup G ( r ( s ) u ) (34)
0<5-<1
uEA
i s finite. Let +(t) be a positive, locally Lipschitz continuous function o n

[ O , o o ) satisfying (12). T h e n there is a sequence {uk} c E such that (13)
holds. I n particular, there is a Cerami sequence satisfying (4).
Proof. The theorem follows from Theorem 3.3. If

~ = { r ( t ) ~[ O
: t, ~~ I , ~ E A } (35)
for some r E a, and a ( t )is any flow, let
1
F(s) = a(2Ts), 0 5 s I -2
1
= a ( q r ( 2 s- I), - < s I 1.
2
It is easily checked that F E @. Morever,
Kc u
tE[o,Tl
a(t)AUa(T)K.
Hence, (2.9) is satisfied.
5. A Method Using Metric Spaces

Another general procedure is described in Mawhin-Willem7 and Brezis-
Nirenberg2 as follows. One finds a compact metric space C and selects a
closed subset C* of C such that C* # 4, C* # C. One then picks a map
p* E C(C, E ) and defines
A = { p E C ( C , E ): p = p * on C*}
a = inf maxG(p(<)).
p d EEC
They assume
(A) For each p E A, maxEEc G(p(<)) is attained a t a point in C\C*.
They then prove
Theorem 5.1. Under the above hypotheses there i s a sequence satisfying

(1.1).
Unijied Minimax Methods 83
We shall prove
Theorem 5.2. Under the same hypotheses, let $(t) be a positive, locally
Lipschitz continuous function o n [0,m) satisfying (12). Then there is a
sequence {uk} c E such that (13) holds. I n particular, there is a Cerami
sequence satisfying (4).
Proof. If a > ao, this follows from Theorem 2.1. In fact, we can take
A = p * ( C * )and K as the collection of sets
K = {P(O :E E c, P E A}.
If cp E R ( A ) ,then
c p ( P ( 0 ) = cp(P*(E)) = P*(E), E E c*.
Thus, cp(p(<)) E K. Hence, K is a minimax system for A, and we can apply
Theorem 2.1 to come t o the desired conclusion.
Now assume a0 = a. Since each K E K is compact, the same is true of
K,, given by (28) for each p > 0. Moreover, if c(t)E C(R+ x E , E ) is such
) I , one has S ( K ) E K ,where
that ( ~ ( 0=
The result follows from Theorem 3.5. 0
6. A Method Using Homotopy Stable Families

Another approach is found in Ghoussoub.' The author condiders a closed
subset A of E , and a collection K of compact subsets of E having the
property that if K E K and $ E C ([0,1] x El E ) satisfies
$(O)U =U, uE E,
and
$ ( t ) =~ U , t E [0,1], u E A,
then $(l)K E K.He calls such a collection a homotopy-stable family with
extended boundary A. The author proves
Theorem 6.1. If K is a homotopy-stable family with extended closed

boundary A , B is a closed subset of E satisfying (9) and G is a C1 func-
tional which satisfies
sup G 5 a 5 inf G,
A B
84 M. Schechter
where the quantity a is given by ( l ) , then there is a PS sequence satisfying

(3).
We shall prove
Theorem 6.2. Under the same assumptions, f o r each f u n c t i o n +(t) satis-

fying the hypotheses of Theorem 2.1 there is a sequence satisfying (13). In
particular, there is a Cerami sequence satisfying (4).
Proof. It is easy to show that such a collection K is indeed a minimax

system. Indeed, if K E Ic, and cp E A(A), then
+(t)U = tcp(U) + (1 - t ) u
satisfies the stipulations above, and consequently cp(K) E K. Since each
of the members of K are compact, it follows that every C1 functional is
Lipschitz continuous on some set K p defined by (28), provided p > 0 is
sufficiently small. Moreover, if o(t)E C(R+ x El E ) is such that a(0) = I ,
one has S ( K ) E K ,where
S(u) = a ( d ( u ,A))u, u E E.
This follows from the fact that S(t)u = a ( t d ( u ,A ) ) u is in E C([O,11 x
E , E ) and satisfies
S(O)U= U , u E E,
and
S(t)u= U , t E [0,I ] , u E A.
Consequently, S ( K ) = S(1)K E Ic by hypothesis. It now follows that the

conclusion of Theorem 3.5 holds. 0
Note: It is not required to have a satisfy
a 5 inf G.
B
If it does, one obtains additional information as described in Ref. 6.
Definition 6.1. We say that A links B [ g h ]if A, B are subsets of E satis-

fying the hypotheses of Theorem 6.1.
7. Some Applications
We now apply the theorems of the preceding sections to various geometries
in Banach space. As before, we assume that G E C1(E,W) and that satisfies
the hypotheses of Theorem 2.1.
Theorem 7.1.
G(0) I(Y 5 G ( u ) , u E dBs (37)

and that there is a cpo E dB1 such that
R > 0.
G(Rpo) IT, (38)
Then for each function + ( t ) satisfying the hypotheses of Theorem 2.1, there
is a sequence {uk} c E such that
G(U) --+ c, cl Ic IT, G'(uk)/+(llukll) 0. (39)

Many elliptic semi-linear problems can be described in the following
way. Let R be a domain in Wn,and let A be a self-adjoint operator on
L2(R). We assume that A 2 A0 > 0 and that
Cr(R) c D := D(A1'2) c H"'2(R) (40)
for some m > 0, where Cm(R) denotes the set of test functions in R (i.e., in-
finitely differentiable functions with compact supports in R), and (0)
denotes the Sobolev space. If m is an integer, the norm in H"i2(R) is given
by
Here D@represents the generic derivative of order and the norm on the
right hand side of (7) is that of L2(R). We shall not assume that m is an
integer.
Let q be any number satisfying
and let f ( z , t ) be a CarathBodory function on R x R . This means that

f ( z ,t ) is continuous in t for a.e. z E R and measurable in z for every t E R.
Throught this section we make the following assumptions:
(A) The function f (2, t ) satisfies
86 M. Schechter
and
l l ~ l:=
l ~llA12~11 (46)
and q = q / ( q - 1). If R and Vo(z) are bounded, then (44) will hold au-
tomatically by the Sobolev inequality. However, there are functions Vo(x)
which are unbounded and such that (44) holds even on unbounded regions
R. With the norm (46),D becomes a Hilbert space. Define
F ( z ,t ) := Jo t
f(z,s)ds (47)
and
It follows that G is a continuously differentiable functional on the whole of

D (cf., e.g., Ref. 11).
For our first result, we assume further that
H ( z ,t ) = 2 F ( x , t ) - tf(x,t)2 -W1(x) E L1(R), x E R, t E R (49)

and
H ( z , t ) -+ a a.e. as It1 --f 03. (50)
, ( z )E L2(R) such
Moreover, we assume that there are functions V ( x ) W
that multiplication by V ( x )is a compact operator from D to L2(Q)and
+
F ( z , t ) I C(v(x)2(t12 V(z)W(z)ltl) (51)
We wish to obtain a solution of
, E D.
AU = ~ ( z , u ) u (52)
By a solution of ( 5 2 ) we shall mean a function u E D such that
(U,V)D = (f(.,u),v), E D. (53)

If f ( z , u ) is in L2(R), then a solution of (53) is in D ( A ) and solves (52) in

the classical sense. Otherwise we call it a weak (or semi-strong) solution.
We have
Theorem 7.2. Assume that Xo is an eigenvalue of A with eigenfunction

PO. Assume also
2 ~ ( z , t 5) X O t 2 , It1 5 6 f o r some 6 > o (54)

and
2 F ( z , t ) L Xot2 - Wl(Z), t > 0 , z E 0, (55)
where Wl E L1(R). Then (52) has a solution u = 0.
Proof. Under the hypotheses of the theorem, it was shown in Theorem

3.2.1 of Ref. 11 that the following alternative holds:
Either
(a) there is an infinite number of ~ ( xE)D(A)\{O} such that
AY = f(x,Y) = XOY (56)
or
(b) for each p > 0 suficiently small, there is an E > 0 such that
G(u) 2 &, llullD = p. (57)
We may assume that option ( b ) holds, for otherwise we are done. By (55)
we have
G(Rv0) 5 R2(11voll; - ~ o l l v o l l )+ 1R
Wl(Z)da:=
s, Wl(Z)da:= B1.
Thus (38) holds. By Theorem 7.1, there is a sequence satisfying (39). Taking
$(T) = l/(r +l ) ,we conclude that there is a sequence {uk} c D such that
G(%) -+ C, E 5 c 5 &, (1+ ((%((D)G~(WC)

---t 0. (58)
In particular, we have
and
88 M. Schechter
These imply
If ( ~ 1. Consequently there
D 00, let i i k = u k / p k . Then ( ( i i k ( =
P k = ( ( u ~ ( (+
is a renamed subsequence such that i i k 4 ii weakly in D ,strongly in L2(R)
a n d a.e. in 0. We have by (51)
Consequently
1 5 2 C L V(x)2iiEdx. (63)
This shows that ii $ 0 . Let Ro be the subset of R on which ii # 0. Then
I.k(.)I = PklG&)l 00, z E Ro. (64)
If R1 = R\Ro, then we have
This contradicts (62), and we see that P k = I l u k l ( ~is bounded. Once we

know that the P k are bounded we can apply Theorem 3.4.1 of Ref. 11 to
obtain the desired conclusion.
Remark 7.1. It should be noted that the crucial element in the proof of
Theorem 7.2 was (60). If we had been dealing with an ordinary PalaisSmale
sequence, we could only conclude that
- (f(,Uk),Uk) =
11~k112D
which would imply only
H ( z , urc)dz = O(Pk).
This would not contradict (65), and the argument would not go through.
References
1. P. Bartolo, V. Benci and D. Fortunato, Nonlinear Analysis T M A , 7,981
(1983).
2. H. Brezis and L. Nirenberg, C o m m . Pure A p p l . Math. 44, 939 (1991).
3. V. Benci and P. H. Rabinowitz, Invent. Math. 5 2 , 241 (1979).
4. K. C. Chang, Infinite dimensional Morse theory and multiple solution prob-
lems, (Birkhauser, Boston, 1993).
5. C. Corduneanu, Principles of Differential and Integral Equations, (Chelsea,

N.Y., 1977).
6. N. Ghoussoub, Duality and Perturbation Methods in Critical Point Theory,
(Combridge Tracts in Mathemtics, 107,1993).
7. J. Mawhin and M. Willem, Critical Point Theory and Hamiltonian Systems,
(Springer-Verlag, 1989).
8. L. Nirenberg, Bull. Amer. Math. Soc. 4 267 (1981).
9. P. H. Rabinowitz, Minimax methods in critical point theory with applications
to differential equations, Conf. Board of Math. Sci. Reg. Conf. Ser. in Math.
No. 65, Amer. Math. SOC.1986.
10. M. Schechter, Pacific J. Math. 171,529 (1995).
11. M. Schechter, Rend. Sem. Mat. Univ. Padova99, 255 (1998).
12. M. Schechter, Linking Methods in Critical Point Theory, (Birkhauser Boston,
1999).
13. E. A. de B. e. Silva, Nonlinear Analysis T M A 16, 455 (1991).
14. M. Schechter and K. Tintarev, Bull. SOC.Math. Belg. 44,249 (1992).[15] K.
Tintarev, Isotopic linking and critical points of functionals, in Proceedings
of the Second World Congress of Nonlinear Analysts, Part 7 (Athens, 1996),
Nonlinear Anal. 30,4145 (1997).
15. M. Willem, Minimax Theorems, (Birkhauser, 1996).
56. SOME REMARKS ON SINGLE CONSERVATION LAWS
MIKIO TSUJI
Department of Mathematics, Kyoto Sangyo University,
Kita-Ku, Kyoto 603-8555, Japan
mtsujiOcc.kyoto-su.ac.jp
PETER WAGNER
Institut fur Technische Mathematik,
Geometrie und Bauinformatik, Universitat Inns bruck,
Technikerstr. 13, A-6020 Innsbruck, Austria
First we consider the Cauchy problem for Burgers equation whose initial func-
tion is x 2 , and study the existence domain of a piecewise smooth weak solution
of the Cauchy problem. Then the existence domain of its real-valued solution is
different from that of the complex-valued solution. Next we consider the Cauchy
problem for a single conservation law which does not satisfy the convexity con-
dition, and study the regularity of weak solutions of the Cauchy problem. Our
question is whether the weak solutions become piecewise smooth, and whether
rarefaction waves and contact discontinuities would really appear in the
weak solutions. In this talk we solve an example as a tentative trial and con-
struct its piecewise smooth weak solution by the geometric method.
1. Introduction
We consider the Cauchy problem for an equation of conservation law in one
space dimension as follows:
au a
- + --f(u) =0 in {t > 0,x E R1},
at ax
u(0,x) = $(x) on {t = 0, x E R1} (2)

where f = f(u)is of class C. We assume that the initial functions
4 = $(z) are also sufficiently smooth. The global existence of weak so-
lutions of (1)-(2) has been well studied. As there exist too many papers on
this subject, we do not mention anything on references. In this talk we will
study the regularity of weak solutions of (1)-(2) and their domains of exis-
tence. We think that the most typical phenomenon in nonlinear hyperbolic
92 M . Tsujz and P. Wagner
equations is the appearance of shock wave. Moreover the shock is defined

only for the piecewise smooth weak solution. This is the reason why we
pay much attention on the existence of piecewise smooth weak solution.
Before considering the Cauchy problem (1)-(2), we state our program for
nonlinear hyperbolic equations and systems. Our first step is to lift the
equation into higher dimensional space so that the singularities of solution
would disappear. Next we construct a global solution of the lifted equation
in the higher dimensional space. We call this as a geometric solution of
the original problem. Finally we project the geometric solution to the base
space and construct a single valued weak solution which satisfies certain ad-
ditional condition called as the entropy condition. For single first order
partial differential equations, we have already proved in Refs. 9,lO and 11
that, if the equations are convex, our program is correct. But we do not yet
succeed in the realization of our program for non-convex first order partial
equations. For second order nonlinear hyperbolic equations and p-systems,
we could not construct weak solutions by our method.l4~l5By these inves-
tigations, we have now some questions on the mathematical formulation for
nonlinear hyperbolic equations and systems. Our final aim is to show that
our program is true for all kinds of hyperbolic problems.
In this talk we will first consider the Cauchy problem where an initial
function is analytic and not bounded, and study an existence domain of
solution. In analytic case, the complex method has been effective especially
for linear partial differential equations. Therefore we will study, in Sec. 2,
what would happen if we might accept a complex-valued solution. for linear
partial differential equations. Therefore we will study, in Sec. 2, what would
happen if we might accept a complex-valued solution.
Next we will construct piecewise smooth weak solutions of (1)-(2). In
the case where f = f(u)is convex, we think that we could understand
the mechanics of the propagation of singularities. See Refs. 9-11, and 12.
Therefore we will consider the case where f(u)changes the sign. This prob-
lem has been studied by J. Guckenheimer, G. J e n n i n g ~S.
, ~Izumiya and G.
T. Kossioris,2 etc. In these papers they have used a rarefaction wave and
a contact discontinuity, because they have reduced the starting problem
to the Cauchy problem whose initial functions have jump discontinuities.
Our question is on this reduction. If our program written in the above
might be accepted, we would not need to introduce the rarefaction waves
and contact discontinuities for the construction of weak solutions of (1)-(2),
because a family of characteristic curves covers the whole space. As we do
not yet arrive at the final conclusion on this problem, we have tried to solve
Some Remarks on Single Conservation Laws 93
a specific problem in an exact form as a tentative trial. This is the subject

of the second part of this talk.
2. Existence Domains of Solutions for Burgers Equation

In this section we consider the following Cauchy problem:
and study an existence domain of a piecewise smooth weak solution of (3).

What we would like to show in this section is the following: If we may
accept only a real-valued solution, then the solution does not exist in the
whole space R2 and its existence domain is surrounded by an envelope of
a family of characteristic curves. But, if we may accept a complex-valued
solution, then the Cauchy problem (3) admits a piecewise smooth weak
solution defined in the whole space R2. First we solve the Cauchy problem
(3) by the characteristic method. As a system of characteristic differential
equations is written by
x=u, u=o
whose initial condition is
40) =t I 4 0 ) = E2,
a family of characteristic curves is given by
x = I+
C2t, u = E2, E E R1. (4)
Here we denote D* = {(t1x);4tx+1 0) and C = {(t,z);4tx+ 1 = 0).
+
X 5 r 2 t } , and
Then we can easily see that D+ U C = U E ~ R I { ( ~ , X ) ; =
that the curve C is obtained as an envelope of the characteristic curves
+
{(t,2 ) ;x = E e2t} where 5 is a parameter moving in R1.Solving the first
equation of (4) with respect to t, we get the solution of (3) by
t , = 4x2/(1+ d
~ ( X) -)' (5)
in a neighbourhood of the initial line {t = 0). Though the right hand side
of (5) is a continuous function defined in the whole space R2, it is not dif-
ferentiable along the curve C and it takes complex values in the domain
D-. Therefore our problem is how to extend the solution (5) beyond the
curve C as a real-valued piecewise smooth weak solution of (3).
94 M . Tsuji and P. Wagner
Proposition 2.1. Assume that the solution (5), written b y u = u(t,x ) , can
be extended as a piecewise smooth real-valued weak solution of (3) beyond
the curve C . Then u ( t , x ) does not have a jump discontinuity along the
curve C .
Proof. Assume that u = u(t,x ) has a jump discontinuity along the curve
C. Here we write the curve C as 2 = y ( t ) = -1/4t. Then it holds [u]?=
+ +
[u2/2]where [u]= u(t,y ( t ) 0 ) - u ( t ,y ( t ) - 0 ) . As u(t,y ( t ) 0 ) = 1/4t2
+
and ? = 1/4t2, we get u(t,y ( t ) - 0 ) = 1/4t2, that is t o say, u ( t ,y ( t ) 0 ) =
u ( t ,y ( t ) - 0 ) . Hence u = u(t,x ) must be continuous along the curve C. CI
Next we consider the following Cauchy problem:
jg + a 1
&u2) = 0, ( t ,x) E D-
u ( t ,x ) = 4x2, on C
Proposition 2.2. The Cauchy problem (6) can not have a solution in
-
C'(D-) n C o ( 5 - ) where D- is the closure of D-.
Proof. Assume that (6) has a solution in C'(D-) n C o ( 6 - ) . Take any

point P on C and a point Q = (to,X O ) in D- satisfying u(t0,X O ) # 0. Then
a characteristic curve passing through the point Q is given by
x =20 + u(t0,m)(t - t o ) . (7)

As u(t0,xo)# 0, the characteristic line (7) meets with the curve C a t two
different points A and B. Obviously it holds U I A # ul~.But, as the solution
is constant along any characteristic curve, this is a contradiction. CI
We can localize the above proposition. Take any point P on the curve C
and any open neighbourhoon U of the point PI and consider the following
Cauchy problem:
au a 1 2
- + -(-u ) = 0 , ( t , x )E D- n U
at ax 2
L u(t,x ) = 4x2, on C n U
Proposition 2.3. The Cauchy problem (8) can not have a solution in
c ~ ( D -n U ) n ~~( 5-
n u).
Some Remarks o n Single Conservation Laws 95
The proof is almost the same as that of Proposition 2.2. Summing up

the above results, we get the following
Theorem 2.1. T h e Cauchy problem (3) can not admit a real-valued piece-
wise smooth weak solution defined in the whole space R2 and the boundary
of the existence domain of the solution is the envelope of the family of the
characteristic curves coresponding t o the Cauchy problem (3).
Remark 2.1. If we may accept a complex-valued solution, the solution (5)

is a piecewise smooth and continuous weak solution of (3) defined in the
whole space R2.
3. A Single Conservation Law without Convexity Condition

In the following sections we will study the Cauchy problem for a single
conservation law which does not satisfy the convexity condition. Assume
that the graphs o f f = f(u) and 4 = 4(x) are drawn as in Fig. 1 (i) and
Fig. 2 (i) respectively. The graphs of their derivatives are Fig. 1 (ii)-(iii)
and Fig. 2 (ii). In the above case, we assume that f ( a i ) = f ( a i ) = 0 (i =
1 , 2 ) , f ( b ; ) = 0 (i = 1 , 2 ) and #(c) = @(d) = 0. Moreover we assume
max4(x) > b2. For the Cauchy problem (1)-(2), the characteristic curves
X
are written by
x =X(t, Y ) = Y +t f ( W ) , 7J = v ( t ,Y) = $ ( Y ) . (9)

Therefore it follows that
8X
-(t, Y) = 1 + t f ( W ) 4 J ( y ) . (10)
8Y
We define h ( y ) ?Zf f ( 4 ( y ) ) # ( y ) , and assume that the graph of h = h ( y )
is drawn as in Fig. 3. Next, denote Ai = (yi, h ( y i ) ) ( i= 1 , 2 ) where h ( y i ) < 0
and h(yi) = 0. Then we have a t A1
4(Yl) > 0 and f ( 4 ( Y l ) ) < 0,

and a t A2
~(Yz)< 0 and f ( d ( y 2 ) ) > 0.

We now put ti = - l / h ( y i ) and X i = x(ti,y i ) (i = 1 , 2 ) . Then we can see
the appearance of a shock whose starting point is (ti,X i ) (i = 1 , 2 ) , and we
denote it by Si (i = 1 , 2 ) . We will briefly explain the construction of shocks
96 M . Tsujji and P. Wagner
Fig. 1.
Fig. 2.
Fig. 3.
at the beginning of Sec. 4. For the detailed explanation.lOill Our problem

is how to extend each shock Si (i = 1 , 2 ) for large t.
We give some comment on the above example. Let us consider the Rie-
mann problem for (l),that is to say, we assume that +(z) = c+ for z > 0
and +(x) = c- for z < 0 where c* are constant. If f ( c - ) < f(c+),the
region { ( t ,2); f(c-)t < z < f ( c + ) t } is not covered by the family of char-
acteristic curves (9). Therefore the rarefaction wave was introduced as a
solution in this region. On the other hand, as we assume that the initial
function (2) is in CF(R1), the family of characteristic curves covers the
whole space R2.This is the principal reason why we doubt the necessity to
use rarefaction waves in the process of construction of weak solutions.
4. Behavior of the Shock S1

In this section we will extend the shock S 1 for large t. To explain the
situation, we repeat briefly how we have constructed the shock 4 . The
graph of 2 = z ( t ,y) for t > ti is drawn as in Fig. 4. We explain the meanings
of the notations used in Fig. 4. Let y = &(t)(&(t) < < 2 ( t ) , t > t l ) be the
solutions of (&/dy)(t,y) = 0 with respect to y in a neighbourhood of
y = 91 and denote z i ( t ) = z ( t ,Ei(t)(i= 1 , 2 ) . Then we see that z l ( t ) >
z 2 ( t ) for t > t l . Solving the equation z = z(t,y) with respect to y for
z E ( ~ ( t z)l (, t ) ) we , get three solutions y = g i ( t , x) (i = 1 , 2 , 3 ) satisfying
g l ( t . 2 ) < g 2 ( t , 2) < g 3 ( t , z ) . As v(t,y) = b(y) for all (t,y) E R2,we define
98 M. Tsuji and P. Wagner
ui(t,z) = $(gi(t, z)) (i = 1 , 2 , 3 ) . Then, as qY(y1) > 0, we have
ul(ti x) < U2(tr x) < u3(ti x).

As we are looking for a single-valued weak solution, we jump from the first
branch {u = u l ( t , x ) } to the third one {u = u 3 ( t , x ) ) . A shock curve is
determined by the Rankine-Hugoniot jump condition as follows:
Though the right hand term is not Lipschitz continuous at the starting
point, we have proved in Ref. 12 that the Cauchy problem (4.1) has a unique
solution z = n ( t ) .This is the shock curve S1. To extend the shock & , we
consider the behavior of u i ( t , n ( t ) )= q5(gi(t,-yl(t>)) (i = 1,3). We put
u*(t) = u ( t , y l ( t )h 0 ) . Then u+(t)= U 3 ( t , Y l ( t ) ) and u-(t) = ul(t,-yl(t)).
Fig. 4.
Lemma 4.1. A s long as S1 satisfies the entropy condition, the function

g l ( t , -yl(t))is decreasing and g3(t, -yl(t))i s increasing.
Proof. By the definition of y = g i ( t , z), we have
z = gz(t, z) + tf($(9i(t,.))) (2 = 1,213).

Taking the derivatives with respect to t and x, we have
and
Hence it holds that
which leads us to
Then the entropy condition means
Hence the proof of the lemma is complete. 0
The initial function 4 = 4(x) has the properties as drawn in Fig. 2.

As long as the entropy condition is satisfied, we see by Lemma 4.1
that 4 ( g l ( t , n ( t ) ) ) = u-(t) is decreasing, and that 4(gs(t,n(t))) =
u+(t) is increasing since g3(t,x) < 0. Hence 4(gl(t,rl(t))) advances
to 0, and 4(g3(t,-yl(t))) goes to its maximum. We write here P+(t) =
) P-(t) = ( u - ( t ) ,f(u-(t))). The shock S1 satisfies the
( u + ( t )f,( u + ( t ) ) and
entropy condition for t > t l where t - tl is small. This means that the
graph of f = f(u)lies entirely on the upper side of the chord P+P- joining
the two points P+(t) and P-(t). If we extend the shock S1 further, then
the chord P+P- may become tangent to the curve {(u,f(u)) : u E R1}
in finite time. We assume that P+P- is tangent to the curve f = f(u)at
t = T . Since 0 < u - ( t ) < 4(yl) < bz and thus f ( u - ( t )< 0, we see that
100 M . Tsuji and P. Wagner
P+P- is tangent t o the curve f = f(u)a t the point P+(T),but not a t the
point P-(T). Therefore it holds that
d71
-(TI
dt = f(.+(T)). (13)
d271 d
Lemma 4.2. i) -(T) > 0, 22) -f(u+(t))lt=T = 0.
dt2 dt
Proof. i) Using the definition (4.1) of ( d y l / d t ) ( t ) , we have
d2Tl
-(t)
dt2
=
d dy1
-(-(t))
dt dt
= [ u+ - u- dt
Here we recall that the starting point of the shock S1 is the point ( t l ,X I ) .
Therefore it follows that ~ ( Y> I )0, u+(t)-u- ( t ) > 0 and ( a g l / a x ) ( t ,x) >
0. Hence we get ( d 2 y l / d t 2 ) ( T )> 0.
Part (ii) is easily obtained from (19). 0
By Lemma 4.2, we have
This implies that the entropy condition is satisfied for t > T . Summing up
the above results, we have the following
Theorem 4.1. T h e shock S1 starting f r o m the point ( t l ,X I ) always satis-

fies the entropy condition.
5 . Behavior of the Shock Sz

In this section we extend the shock 52 whose starting point is ( t 2 , X 2 ) . The
graph of z = z ( t ,y) for t > t 2 can be drawn as in Fig. 4. Though z = z(t,y)
in this section is different from z = z ( t ,y ) in the previous section, we use the
same notations introduced there. As in Sec. 3, we solve the Cauchy problem
for (16) whose initial condition is z(t2) = X 2 , and denote the solution by
5 = 7 2 ( t ) . We write here also u*(t) = u(t,7 2 ( t ) f0 ) ; that is to say, u+(t)=
q5(g3(trY 2 ( t ) ) ) and u-(t) = q5(91(tl7 2 ( t ) ) ) . Put P*(t) = (%(t),f (u*(t))).

The shock S2 satisfies the entropy condition for t > t 2 where t - t 2 is small.
In this case, as f(q5(y2)) > 0, the entropy condition says that the graph of
f = f(u)lies entirely on the lower side of the chord P + R .
Lemma 5.1. As long as the entropy condition f o r S2 i s satisfied,
g l ( t , 7 2 ( t ) ) i s decreasing and g 3 ( t l Y 2 ( t ) ) is increasing.
The proof of this lemma is almost the same as that of Lemma 4.1.
When t gets larger, q5(g3(tI72(t)))tends to 0 and q5(g1(t172(t)))tends to
the maximum of q5 = q5(z).Therefore we assume that the entropy condi-
tion is satisfied for t < T , and that P+P- becomes tangent to the curve
{(u, f (u));u E R1}at t = T . Then it follows that
dY2
-(T)
dt = f($(g3(t, 72(t))))lt=T= f ( U + ( T ) ) .
d
i i ) -f(u+(t))lt=T = 0.
dt
The proof is the same as that of Lemma 4.2. Using this lemma, we get
Therefore we get the following
Theorem 5.1. T h e shock S2 starting f r o m the point ( t 2 , X2) always satis-

fies the entropy condition.
It may happen that the shocks 51 and 52 collide in finite time. Then we
see that a new, uniquely defined shock appears. The method of the proof
is almost the same as that in Refs. 9 and 10.
Summing up these results, we can say that our program is true for the
above Cauchy problem.
102 M. Tsuji and P. Wagner
References
1. J. Guckenheimer, Solving a single conservation law, Lecture Notes in Math.,
Vol. 468, (Springer-Verlag, 1975), pp. 108-134.
2. S. Izumiya and G. T. Kossioris, Bull. Sci. math. 121,619 (1997).
3. G. Jennings, Adv. i n Math. 33,192 (1979).
4. S. N. Kruzhkov, Math. USSR Sb. 1,93 (1967).
5. S. N. Kruzhkov, Math. USSR Sb. 10, 217 (1970).
6. P. D. Lax, Comm. Pure Appl. Math. 10,537 (1957).
7. 0. A. Oleinik, Uspelchi Mat. Naulc 12,3 (1957) (in Russian). English transl.
in Amer. Math. SOC.Transl. 26,95 (1963).
8. D. G. Schaeffer, Adv. in Math. 11,358 (1973).
9. Tran D. V., M. Tsuji and Nguyen Duy T. S., The characteristic method
and its generalizations for first-order non-linear partial differential equations
(Chapman & Hall/CRC, USA, 1999).
10. M. Tsuji, C. R. Acad. Sci. Paris 289,397 (1979).
11. M. Tsuji, J . Math. Kyoto Univ. 26,299 (1986).
12. M. Tsuji, A n n . Inst. H. Poincare' - Analyse nonline'aire 7,505(1990).
13. M. Tsuji, Pitman Research Notes in Math. 381,164 (1998)(Longman).
14. M. Tsuji and T. S. Nguyen Duy, Acta Math. Vietnamica, 27,97 (2002).
15. M. Tsuji, Some remarks on nonlinear hyperbolic equations and systems, in
Abstract and Applied Analysisedited by N. M. Chuong, L. Nirenberg and W.
Tutschke (World Scientific, 2004), pp. 355-364.
37. ITERATIVE METHOD FOR SOLVING A MIXED

BOUNDARY VALUE PROBLEM FOR
BIHARMONIC TYPE EQUATION*
DANG QUANG A$ and LE TUNG SON

Institute of Information Technology
Vietnam Academy of Science and Technology,
18 Hoang Quoc Viet Road, Cau Giay Dist., 10307 Hanoi, Vietnam
E-mail: dangqaQioit.ncst.ac.vn
The solution of boundary value problems (BVP) for fourth order differential
equations by their reduction to BVP for second order equations with the aim
to use the achievements for the latter ones attracts attention from many re-
searchers. In this paper, using the technique developed by ourselves in recent
works, we construct iterative method for a mixed BVP for biharmonic type
equation. The convergence rate of the method is proved and some numerical
experiments are performed for testing it in dependence on the choice of an
iterative parameter.
Keywords: Iterative method; Mixed boundary value problem; Biharmonic equa-

tion
1. Introduction
The solution of fourth order differential equations by their reduction to
bouridary value problems (BVP) for the second order equations, with the
aim of using efficient algorithms for the latter ones, attracts attention from
many researchers. Namely, for the biharmonic equation with the Dirichlet
boundary condition, there is intensively developed the iterative method,
which leads the problem to two problems for the Poisson equation at each
iteration (see e.g. Refs. 2,8,9,11). Recently, Abramov and Ulijanova' pro-
posed an iterative method for the Dirichlet problem for the biharmonic type
equation, but the convergence of the method is not proved. In our previous
work^^-^ with the help of boundary or mixed boundary-domain operators
*This work is supported in part by the National Basic Research Program in Natural
Sciences, Vietnam
104 Q. A Dang and L. T.Son
appropriately introduced, we constructed iterative methods for biharmonic

and biharmonic type equations associated with the Dirichlet boundary con-
dition. It is proved there that the methods are convergent with the rate of
geometric progression. In this paper we develop our technique in Refs. 4-7
for a mixed BVP for the biharmonic type equation. Namely, we consider
the following problem
A2u + bu = f in R, (1)
u=g onr, (2)
aU
- = g1 on rl, (3)
all
AU = g2 on r2, (4)
where s1 is a bounded domain in Rn(n 2 2) with the Lipschits boundary
I? consisting of two smooth pieces rl and I?2, u is the outward normal
to I', b is a positive number and A is the Laplace operator. We propose
an iterative method for reducing the problem to a sequence BVP for the
Poisson equation, study its convergence and finally, present the results of
numerical experiments for the illustrating the effectiveness of the method.
2. Reduction of the Problem to Boundary - Domain

Operator Equation
As in Refs. 5 and 6 we set
AU = U,
~p = -bu
and denote
ulr, = vo. (7)

Then the problem (1)-(4) is reduced to the problems:
where vo and p are temperorily unknown functions.

The solution u found from the above problems should satisfy the boundary
condition (3) and the relation (6). To find U O , (p, we introduce an operator
Iterative Method for Solving a Mixed BVP f O T Biharmonic Type Equation 105
B as follows:
B : w + Bw,
e)
where
w= , Bw = (i."Ir1)
'p+bu
,
and u and v are the solutions t o the problems

Av = 'p, x E 0,
vlr, = 210, ulr, = 0 ,
The operator B primarily defined on couples of smooth functions is ex-

tended by continuity on whole space L 2 ( r 1 ) x L2(R) . Its properties will be
investigated later.
For the reduction of BVP (1)-(4) to an equation with the operator B
+ +
we set u = u1 u2, v = v1 u2 , where u1, v1 and u2, u2 are the solutions
to the following problems
Aul = 211, x E R,
ullr = 0.
From (13) and (14) we can determine 212 and u2, and from (15), (16), by
the definition of B,we have
(17)
106 Q. A Dang and L. T. Son
For u t o be the solution of (1)-(4), as mentioned above, the relations (3)

and (6) must be satisfied, i.e., there should be conditions
cp+bu=O inR. (19)

From these relations we derive
Setting
from (17) we get the equation
BW= F. (23)
Now, we study properties of B . First, we introduce the space H = L 2 ( r l )x

L2(R) with the scalar product
(w, V)H = (wo, G ) L z ( r l ) + (cp, c P ) ~ z ( n )

for the elements
Property 1. B is symmetric in H
Proof. We have
since Zi(1.1 = VO and V l r 2 = 0 in the problems (ll),(12) for the definition of

BV.
Iterative Method for Solving a Mixed BVP for Bihannonic Type Equation 107
Now, we transform the boundary intergral, taking into account of (11)

and (12). We have
Therefore,
Thus, the symmetry of B is proved.

Property 2. B i s positive in H .
Indeed, from (24) we have
( B w ,W ) = 1
R
(bv2 + p 2 ) d x 2 0.
The equality appears if and only if w = 0.
Property 3. B can be decomposed into the sum of a positive, symmetric
and completely continuous operator and a projection operator, namely,
where Bo and 12 are defined as follows
u being defined from (ll),(12).

The complete continuity of Bo easily follows from the theory of elliptic
problems and embedding theorems of Sobolev spaces (see Ref. 10).
Property 4. B i s bounded in H .
This fact is a direct corollary of Property 3.
Since B = B* > 0 but is not completely continuous in H the use
of two-layer iterative scheme13 to the equation (23) does not guarantee
its convergence. Hence, in the next section we will disturb this equation
and apply the parametric extrapolation technique (see Refs. 4 and 5) for
constructing approximate solution for problem (1)-(4).
3. Construction of Approximate Solution of the Original

Problem Via a Perturbed Problem
We associate with the original problem (1)-(4) the following perturbed prob-
lem
Aualr, = 92,
where S is a small positive parameter.
Theorem 3.1. Suppose that f E Hn-4(R), g E Hn-1/2(I'), g1 E
Hn-3/2(r1),9 2 E Hn-5/2(I'2);n2 4. T h e n f o r the solution of the prob-
lem (27)-(30) there holds the following asymptotic expanssion:
where yo = u i s the solution of ( l ) - ( d ) , yi (i = 1 ,2 , ...,N ) are functions

independent of 6, yi E H"-i(0), zg E H"-N(s2) and
llZ61IH~(S2, 6 c1, (32)
C1 being independent of 6 .
Proof. Under the assumption of the theorem there exists a unique solution
u E H " ( 0 ) of the problem (27)-(30).After substituting (31) into (27)-(30)
and balancing coefficients of like powers of 6, we see that yi and zg satisfy
the following problems:
A 2 yi + byi = 0, XER,
yilr = 0,
dyi (33)
b--Irl = -Ayz-llrl , i = 1,2, ..., N ,
dU
AyiIrz = 0.
A2z+
~ bzs = 0, x E 0,
zalr = 0,
(34)
Iterative Method for Solving a Mixed B V P for Biharmonic Type Equation 109
It is possible t o establish successively that (33) has a unique solution yi E

HnPi(fl) and (34) has a unique solution zg E H"-N(fl). Clearly, yi (i =
1,..., N ) do not depend on 6. It remains to estimate 26. For this purpose,
we reduce (34) to a boundary operator equation. Setting
Azs = us, ' ~ s= --b~s, uslrl = us0 (35)

we obtain
A u= ~ pa, x E 0,
4 r , = uso, U s l r , = 0.
azs= E a,
zslr = 0.
Now , denote
then by definition of B , we obtain
Using the second condition of (34), we obtain

(B + 6 1 1 ) ~=s h,
where
It follows that (see Lemma 1 in Ref. 3)

( B w s ,ws) 6 ( B w ,w), (37)
where w is the solution of the equation B w = h. This equation has a
solution because it is the equation to which the problem (34) with 6 = 0
may be reduced.
In Sec. 2, when investigating the properties of B , we have established that
In view of (35), we have

(Bws,ws) =
s,(bus2 + 'Ps2)dx. (38)
+
(blavsl2 blzs12)dx. (39)
Since the right-hand side of the above quality defines a norm in the class
of functions vanishing on the boundary, which is equivalent t o the norm
II.IIHz(n), we have II.zIIHz(n) 6 G , where C1 = GJ-, c2 being
independent of 6.
Thus the theorem is proved. 0
As usual, we construct an approximate solution U E of the original prob-

lem (1)-(4) by the formula
i=l
where
(_l)N+l-iiN+l
"li =
z. ! ( N + 1 - i)! ' (41)
ug/iis the solution of (27)-(30) with the parameter 6 / i (i = 1,...,N + 1).

Then it is easy t o obtain the following estimate
(IUE- UllHz(n) < C2dN+l, (42)

where u is the solution of the problem (1)-(4) and C2 is a constant inde-
pendent of 6.
4. Iterative Method for Solving the Perturbed Problem

First we notice that in the same way as for the original problem (1)-(4),
the problem (27)-(30) may be reduced to the operator equation
B 6 ~ g= F (43)
for wg = ( ' ~ g o , ' p g ) ~ ,where ' U ~ O = AugIrl, 'pg = -bug,Bg = B +
611, B and F are defined by (10) and (22), respectively. Clearly, Bg is
bounded and
Bg = B i 2 6 I , (44)
where I is the identity operator.
For solving (43) we can apply the general theory of two-layer iterative
scheme for equation with symmetric, positive definite operator. l 3 Namely,
we consider the iterative scheme
Iterative Method f o r Solving a Mixed B V P for Biharmonic Type Equation 111
where { i - ~ , k + l } is the Chebyshev collection of parameters according to

+
bounds 7:') = 6,$) = 6 llBll (see Ref. 13 for details). In the case of
simple iteration
we obtain
where
and as above H = L 2 ( r l )x L2(R) .

Using estimates for the solution of elliptic problems1' and taking into
account (47), we obtain the estimate
Ilup' - u611~5/2(n) < c ( P 6 ) k I I W p )- W ~ I I H , (48)

where C is a constant independent of 6.
The iterative scheme (45) can be realized by the following process:
(i) Given a couple (v~o(O),(ps(O)).

(ii) Knowing W I S O ( ~ and
) ( p 6 ( k ) , k = 0,1, ... solves successively two problems
112 Q. A Dang and L. T . Son
5 . Numerical Experiments
We perform some limited experiments in MATLAB for testing the con-
vergenve of the approximate solution given by (40) with fixed N = 2 and
6 will be chosen in agreement with the steps of the grid for discretizing
differential problems. We consider the computational domain R = (0,
with uniform grid and steps of the grid are hl = h2 = 1/M. The part r2
is the top side and rl is the remaining part of the boundary of the square
R . For the iterative process we choose by experimental way the parameter
r = 2 6 f 0 . 7 5 4 ' The Dirichlet problems (49), (50) are discretized by differ-
ence schemes of second order approximation and the normal derivative in
(51) is approximated by difference formula of the same order of accuracy.
The stopping criterion for the iterative process (49)-(52) is
max(IIvg+l) - vg)llool (k+l) -
IIYS )ol'& < & = hlh2.
and in view of the estimate (42) we choose 6 = E ~ / For ~ . solving systems of
grid equations the method of complete reduction] whose idea is the succes-
sive odd-even eliminations1l2is used.
We take some exact solutions and for b = 1 construct the right-hand
sides and the boundary data in respect with them. The results of com-
putation on PC Pentium 4 with CPU 1.80 GHZ for these examples are
presented in the following tables, where Error = llUE - uII, and in the
column "Number of iterations " we report the number of iterations for
finding the basic solution ugli (i = 1 , 2 , 3 ) , the time is in seconds.
Table 1. Case u = (z: - l)(z; - 1)

Grid Number of iterations Error Time
16 x 16 13 19 24 0.0023 0.88
32 x 32 23 37 48 0.0007 4.95
64 x 64 42 70 93 0.0003 42.51
Table 2. Case u = sin(m1) sin(xz2)

16 x 16 10 16 21 0.0050 0.78
32 x 32 17 26 33 0.0013 3.52
64 x 64 27 39 47 0.0003 23.59
It is interesting t o notice that for the above examples in the case if

6 = 0 our experiments show that the iterative process (49)-(52) does not
Iterative Method for Solving a Mixed B V P for B i h a m o n i c Type Equation 113
Table 3. Case u = 0.252: + 0.252; + x: + xz

16 x 16 16 25 32 0.0056 1.08
32 x 32 28 46 61 0.0022 6.17
64 x 64 50 84 115 0.0009 51.14
converge. It justifies the reason, why we have to consider the perturbed

problem (27)-(30) and extrapolate its solution by the parameter 6.
References
1. A. A. Abramov and V. I. Ulijanova, Journal of Comput. Math. and Math.
Physics, 32,567 (1992) (Russian).
2. Dang Quang A, Math. Physics and Nonlinear Mechanics, 44, 54 (1988) (Rus-
sian) .
3. Dang Quang A, Vietnam Journal of Math. 33,9 (2005).
4. Dang Quang A, Journal of Comput. and Appl. Math., 51, 193 (1994).
5. Dang Quang A, Vietnam Journal of Math. 26,243 (1998)
6. Dang Quang A, Journal of Comp. Sci. and Cyber. No. 4, 66 (1998).
7. Dang Quang A, Journal of Comp. and Appl. Math. 196,634 (2006).
8. A. Dorodnisyn, N. Meller, Journal of Comp. Math. and Math. Physics 8 , 393
(1968) (Russian).
9. R. Glowinski, J-L. Lions and R. Tremoliere, Analyse Numerique des Inequa-
tions Variationelles, (Dunod, Paris, 1976).
10. J-L. Lions and E. Magenes, Problemes aux Limites n o n Homogenes et A p -
plications, Vol. 1, (Dunod, Paris, 1968).
11. B. V. Palsev, Journal of Comput. Math. and Math. Physics 6, 43 (1966)
(Russian).
12. A. Samarskii and E. Nikolaev, Numerical Methods for Grid Equations, Vol. 1:
Direct Methods, (Birkhauser, Basel, 1989).
13. A. Samarskii and E. Nikolaev, Numerical Methods for Grad Equations, Vol. 2:
Iterative Methods, (Birkhauser, Basel, 1989).
8. NUMERICAL SOLUTION TO A NON-LINEAR

PARABOLIC BOUNDARY CONTROL PROBLEM
DINH NHO HAO*Z, NGUYEN TRUNG THANH and H. SAHLI

Vrije Universiteit B~ussel,
Department of Electronics and Informatics,
Pleinlaan 2, 1050 Brussel, Belgium
E-mail: haoOmath.ac.vn
The so-called difference of convex functions algorithm and the continuation

technique are applied to solving the nonlinear parabolic boundary control prob-
where a: > 0 is given and y(z, t ) = y(z, t ;u ) is the solution to the non-linear
parabolic problem
y t ( z , t ) = yzz(z,t), 0 < I < 1 , 0 < t < T,

y(z, 0 ) = Y O ( Z ) , 0 < I < 1,
yz(O,t)=O, ~z(l,t)=g(y(l,t))+z~(t)O , <t<T
with { u E Lw([O,T])Iumin(t) 5 u(t) 5 umax(t), for a.e. t E [O,T]}.Numerical
examples are given t o show the efficiency of the method.
1. Introduction
In Ref. 8 Kelley and Sachs proposed a trust region method for the problem
of minimizing
where Q > 0 is given and y(z, t ) = y(z, t;u)is the solution to the non-linear
parabolic problem
'Also at Institute of Mathematics, 18 Hoang Quoc Viet Road, 10307 Hanoi, Vietnam
116 D. N. Hao, N . T . Thanh and H. Sahli
In (1)-(2) u is constrained to be in the set
U = {U E LO([O,T ] ) I ~ m i n ( t )5 u ( t )5 u m a x ( t ) , for a.e. t E [0, TI} (3)

and the nonlinear function g is assumed to satisfy
g E C2(R), g, g E L(R). (4)

In order to use some known results by Sachs16 and Schmidt17 of the problem
(1)-(2) we will need some monotonicity conditions on g that we will describe
in section 2.
The problem (1)-(2) can be understood as a variational formulation
of the inverse problem where one wishes to find the heating regime u ( t )
at a part of the boundary such that at the final time the temperature of
the system reaches a given desired goal z. The particular case g(y) = -y4
corresponding to the Stefan-Boltzmann radiation condition seems to have
many applications in engineering.6>7J6
The method of Kelley and Sachs is a projected form of the Steihaug
trust-region-CG method with a smoothing step added at each iteration to
improve performance in the global phase and provide mesh-independent
sup-norm convergence in the terminal phase. However, whether their
method delivers a global solution to (1)-(2) is not clear.
In this paper we approach the control problem (1)-(2) from a quite
different point of view. Namely, from that of DC programming. We realize
that f is a DC one, i.e. it can be represented by the difference of two convex
functions. Therefore, we can consider (1)-(2) as a DC program and apply
the DC programming techniques to solve the problem. In fact we apply
the difference of convex functions algorithm (DCA).l2>I3 As DCA is a local
method, it cannot guarantee to deliver a global solution, in this work we
combine it with a so called continuation or cooling1 technique for the search
of a global solution. We note that DCA is applicable to finite-dimensional
problems only, but in order to solve (1)-(2) we need to disretize it somehow.
The convergence of the solutions of the discretized control problems of type
(1)-(2) to that of the continuous one has been proved for some cases in
Refs. 15 and 10. The problem (1)-(2) has been first considered by Sachs
and then by Schmidt,17 see also Refs. 3 and 14.
In the next section we prove that the functional f is twice Frkchet
differentiable and deliver the formulas for the gradients as well as prove that
f can be represented by the difference of two convex functionals. In section 3
the difference of convex functions algorithm (DCA) and the combination of
DCA with the continuation technique for solving non-convex optimization
Numerical Solution to a Non-Linear Parabolic Boundary Control Problem 117
problems will be described. Finally, in the last section numerical results

based on our method will be given.
2. The Control Problem and Its dc Representation

In this section we prove that (1)-(2) is a DC program. Further, to make
the algorithm more efficient we change the functional f by
F ( u )= f 1 1
Iy(z, T ;u)- z(z)12dz +- I' ( u ( t )- U * ( t ) ) Z d t . (5)
Here u* plays the role of a guess for the solutions or a selection criterion
for the solutions: as (1)-(2) could have many solutions to, we choose that
one which is nearest to u*, by minimizing (5).
Now we review some results on the solution of (2). Let I be a bounded
subinterval of the reals R, say, I = [a,b],--oo < a < b < 00. Let B be a
norm space. By writing y E BJ we mean that y E B with ranges in I . We
denote II . I l P ( 0 , l ) by II . II.
By H1(O,1) we denote the Sobolev space of all elements y E L2(0,1)
having weak derivative u, in L2( 0 , l ) . The space H'( Q ) is similarly defined.
Here, Q := ( 0 , l ) x (0, TI. Further, let
V''O(Q) := C([O,TI;L2(0,1)) n L 2 ( ( 0 T
, ) ;H1(O,1))
with the norm
IYI := vrai max II'u(., t)llL2(0,1) -k IIu~IIL2(Q).
OltlT
On the notion of Sobolev spaces we refer the reader to Ref. 9.
Let urnin,umax be given functions with ranges in I belonging to
L"([O, TI). To be consistent with the results of Schmidt17 we additionally
suppose that g is strictly decreasing in I .
Definition 2.1. Let yo E L;(O, 1). Then y E V;>O(Q)is said t o be a weak

solution of (2) if for all 4 E H1((O,1) x ( 0 , T ) )with = 0, then the
following integral identity holds
lT L 1 ( - Y ( 2 , t ) $ t ( x ,t ) + YZ(Z, t)$Z(Z, t ) ) d Z d t
= i' Yo(z)4(Z, O)dZ +
T
( d Y ( L t ) )+ u(t))4(1,W t .
It is proved in Ref. 17 that there is a unique weak solution t o (2). And it is
proved in Refs. 3 and 14 that y E L"(Q) n C ( G ) for every E > 0. Here
Q E , ~ = (0,1) x (6,T).
118 D. N. Hao, N. T. Thanh and H. Sahli
Now we prove that F is DC and propose a DC decomposition for it,

in doing so we note that F E C2(U). Denote by G : L(0,T) + V1io(Q)
the mapping associating the solution of (2) t o every function u.In what
follows, we denote the first- and second-order Frhchet derivatives of G a t a
point u in the direction h or ( h ,k ) by G(u)h and G(u)(h,k ) , respectively.
Following Refs. 16,17 and 4 it can be proved that G is twice continuously
Frkchet differentiable. If u,h E L(0, T ) ,cp := G(u)h,then cp = p(z,t; h )
is the solution of the problem
P~(x, t ;h) = (PZZ(Z, t ;h ) , 0 < x < 1,0 < t < T ,
cp(z, 0; h ) = 0, 0 < x < 1, (6)
cp,(O,t;h) = O , < t < T.
cp,(l,t;h) = g ( y ( l , t ; u ) ) c p ( l , t ; h ) + h , O
If h, k E L(0,T) and v =~ ( xt ;,h, k ) := G(u)(h,k ) , then w is the solution
of the problem
ut = v,,, 0 < x < 1 , O < t < T,
V ( X , 0; h, k ) = 0, 0 < z < 1,
~ ~ (t; 0h, k, ) = 0, 0 < t < T , (7)
U Z ( 1 , t ;h, k ) = g(y(1, t; u ) ) 4 1 , t; h, k )
+g(y(l, t ;u))cp(l, t;h)cp(l, t ;k ) , 0 < t < T.
From these results we obtain that F is twice continuously Frkchet differen-
tiable and, if h E L(O,T), then
rl rT
F(u)h = /o cp(x,T ;h)(y(z,2; u)- z ( z ) ) d x +a lo (u- u*)hdt, (8)
and
Jo.I
F(u)(h,h ) = Icp(x,T ;h)12dz
1 (9)
+ 4 x 7 T ;h, h ) ( y ( x , T ;u)- z ( z ) ) d J +
: aIIh112L2(O,T).
It can be also proved that the gradient of F is

F(u) = d(1, t ) + a ( u - u*), (10)
where d ( z ,t ) is the solution of the adjoint problem
-dt(x, t ) = d,,(x, t), 0 < x < 1,0 < t < T ,
d ( ~T, ) = y ( x , T ;U ) - z(x), 0 < x < 1, (11)
d,(O, t ) = 0, &(I, t ) = g(y(1, t ;u))d(l, t ) ,0 < t < T.
Now we estimate 11 F(u)11. We need the following result by Schmidt in
Ref. 17, Corollary A.3.
Lemma 2.1. Let w be the weak solution of the problem

W t = w,,, 0 < x < 1 , O < t 5 TI
-wZ(O,t)+al(t)w(O,t)=fi, ,,(l,t)+a2(t)w(l,t)=fi, 0 <t IT,
W ( . , O ) = wo, 0 < x < 1,
(12)
where 01~02are bounded, measurable, non-negative functions defined o n
[O, TI, wo E L 2 ( 0 ,1 ) fl L"(0, l ) ,f1, f 2 E L 2 ( 0 ,T ) n L"(0, T ) . Then there
exists a positive constant C such that
IIWIILm(Q) 5 C m a x { ~ ~ W o ~ ~ L - ( O , l~) ~ f l ~ ~ L - ( o~,~Tf)2~~ ~ L m ( 0 , T ) } ~

Remark 2.1. By inspecting the proof of Schmidt we get
5
C=2T+-.
4
Let
3. DCA and Continuation Techniques

3.1. DCA
3.1.1.
A function f defined in a convex set is called DC, if it can be written in the
form f(x) = g ( x ) - h ( x ) .Here, DC means Difference of Convex functions.
We suppose that f is defined in a Hilbert space X, although in the future
we mainly work with R". Note that the above representation of f is not
+ +
unique, since [g(z) allzl12]- [h(z) a11z112]for any positive a is also a
DC representation for it.
3.1.2.
We say that a convex function g is proper, if g > -ca and there exists x
such that g(z) # +ca. In this case, we write g E I'o(X).
3.1.3.
When we work with a convex function g defined in a convex set C we can
extend its domain into the whole X by setting its values to +ca outside C
so that the new function remains convex. Hence in the future we will work
with convex functions defined in the whole X .
3.1.4. Let g be a conwex function.

Its subdifferential is defined by
Here, < ., . > is the inner product in X .

It is known that g attains its minimum a t z* if and only if 0 E ag(x*).
3.1.5. Let g E I ' o ( X ) and lower semi-continuous (lsc).

The Fenchel transform of g is defined by
9*(Y) := SUP{< 2 , Y > -gb)).

2
It can be proved that g* belongs t o I'o(X) and is lower semi-continuous.

Furthermore, g** = g.
Note that y E ag(x) if and only if z E ag*(y), if and only if < x , y >=
g(z) + S*(Y).
3.1.6. Consider the DC program
(PI a := inf{f(z) = g(z) - h(z)}.

5
Numerical Solution to a Non-Linear Parabolic Boundary COntTOl Problem 121
Here g and h are proper convex functions and lower semi-continuous. With
the notation dom g := {xlg(x) < +00} we see that if la1 < 00, then
dom g c dom h and dom h* c dom g*.
The program (P) has its dual part
(D) a = inf{h*(y) - g * ( y ) } .
Y
3.1.7. DC algorithm:
Since (P) and (D) are not convex programs, we approximate them by convex
ones. Namely, the algorithm for solving (P) will use its dual part (D) and
their convex approximations. We construct two sequences { X k } and { Y k }
as follows.
Let xo E dom g.
(D)* ( D k ) mjn{h*(y) - [ g * ( Y k - l ) + < x k , Y - Yk-1 >I}.

We see that g * ( y k - 1 ) + < X k , y - y k - 1 > is an affine approximation to g * ( y )
at y k - 1 . Hence h*(y) - [ g * ( Y k - l ) + < x k , y - Y k - 1 >] is a convex function.
It follows that the minimum of the last function attains a t Y k which is
characterized by
0 E a h * ( y k ) - Xk
The last is equivalent t o
Xk E ah*(%)
which is the condition that
Yk E ah(xk).
so from x k and Y k - 1 we constructed Y k . Again to update X k , with the

same idea, we consider the k-th convex approximation t o (P)
(pk) mjn{g(x)- [ h ( Y k ) + < - xk7 Yk >I}.

The solution of the last is Xk+l which is characterized by the chain of
equivalencies
0E ag(Xk+l) - Yk @ Yk E ag(xk+l) zk+l E ag*(Yk).
We summarize the algorithm as follows

xo E dom g ,
122 D. N. Hao. N. T. Thanh and H. Sahla
zk+l = Argmin{g(z) - [h(yk)+< z - zk,Y k >I}.

The last equality is equivalent to
Yk E dh(Zk), Zk+1 E %*(Yk).
The convergence properties of DCA have been proved in Refs. 12 and

13.
3.2. The Continuation Technique

DCA provides however only critical points of f. In order to find a global
minimizer off we consider the regularized function f a ( z ):= f (z)+a/2)1z))2
with Q > 0. When Q is very large, in many cases it is obvious that fa is
convex. In these cases we can find a global minimizer of f a . After finding
this minimizer, which we denote by x,,, we reduce a to a1 < QO and
minimize fal by an iterative method starting from z ,, and so on. This is
the continuation method.' As DCA is globally convergent, if we apply DCA
to this procedure we can hope that we reach a global minimizer of f.
The combination of DCA and the continuation technique is summarized
in the following algorithm
Algorithm 3.1. minf(z)

X
(1) Initialization: xo E domf, p 2 )If"ll, and a finite decreasing positive

sequence { a,}$!O.
+
(2) Choose h ( z ) such that f(z) h ( z ) is convex (e.g., h ( z ) = g11x112)
(3) For i = 0 , 1 , 2 , .. ., M
+ +
3.1 gz(z) = f(z) 211z112 h ( z ) , Ic = -1
3.2 While stop criterion has not been satisfied
3.2.1 k = Ic +1
3.2.2 y k E ah(zk)
3.2.3 z k + l = argmin{gi(z)- < Y k , z >}
End (While)
+
3.3 zat = zk+l (minimizer of f(z) ~ ~ ~ z ~ ~ 2 )
3.4 20 = xat (change the initial guess in the continuation technique)
End (for).
The sub optimization problems in 3.2.3 of Algorithm 3.1 are convex. In this
paper, we used the Fletcher-Reeves nonlinear conjugate gradient method
with Amijo-type line search, which ensures global convergence of the algo-
rithm, for solving these sub-problems (see Ref. 5 for more details).
4. Numerical Results
To solve (2) we use an iterative technique to deal with the non-linear bound-
ary condition and the finite difference method for the linear boundary value
problem. Namely, let
yO(z,t ) := yo(z)
and yn(x, t ) ( n= 0 , 1 , 2 , . . .) be the n-th iteration. Solve
gyfl (z, t ) =
' : :y (z, t ) , 0 < z < 1 , O < t < T,
y"+l(z,O) = yo(z), 0 < z < 1, (17)
y,"+l(O,t) = O , yZf1(1,t) = g ( y n ( l , t ) ) + u ( t ) , 0 <t<T.
We stop the iteration procedure when some tolerance between yn and yn+l
is achieved. We note that for n > 1, the difference w = yn+l - yn satisfies
the simpler system
w ~ ( z t, ) = W Z ~ ( Z , t ) l 0 < z < 1 , O < t < T ,

w(z,O)=O, O < z < l , (18)
Wz(O,t)=O, w~(l,t)=g(y"(l,t))-g(y"-l(l,t)), 0 < t <T.
than y1 and yn, n > 1. For the convergence of this method we refer the
reader to Ref. 2.
To solve (17) we use the Crank-Nicholson finite difference scheme.
We devide [0,1] by N - 1 equal subintervals: z o = 0 , q = h, . . . ,zi =
ih, . . . ,X N = 1, where h = 1/N, and [0,TI by M - 1 equal subintervals:
t o = 0, tl = 7 , . . . , t j = j r , . . . ,t M = T . In order to approximate the bound-
ary conditions by central finite difference quotients, we extend the inter-
val [0,1] one more layer to each direction and denote the extended grid
points by 2-1 and Z N + ~ .For a function y defined on [0,1] x [0, TI, we set
$ := y ( z i , t j ) . For simplicity of notion, we denote the solution of (17) by
Y and the inhomogeneous boundary condition by V. We discretize (17) as
follows
y.j+l- yj -
-
yjfil- 2 q j + l + y!+l
2-1 + -2qj + qtl 7
7- 2h2 2h2
05i5 N,0 5 j 5 M - 1,
(19)
yi" = yo,i, 0 5 i 5 N,
It is well-known that this scheme is absolutely stable and has second order
approximation with respect to the both variables. l1 The Crank-Nicholson
scheme is also used to solve the adjoint problem (11).
124 D . N. Hao, N. T . Thanh and H. Sahli
4.1. Numerical Examples

In this subsection we illustrate DCA and the continuation technique for the
control problem (5) subject to (2) with some concrete examples.
Example 4.1. Let a be a number in (0,7r]. Then

y(z, t ) = e-a2t cos(az)
is the solution of the problem

< z < 1,0 < t < T,
) yz2(5, t ) , 0
y t ( ~ , t=
y(z,O) = cos(az), 0 < z < 1, (20)
yZ(o,t ) = 0, yZ(1,t ) = -ue-aztsina, o < t < T.
Let
u ( t ):= e-4aZt cos4 a - ae-aZt sin a.
We see that
y,(l,t) =-Y4(W +4t).

Thus, in this example we set g(y) = -y4. To test our method we set
z ( z ) = y(z, T ) = e-aZT cos(az).
In this example, T is set to be 1 and the space and time grid sizes are taken
to be 0.01 and 0.025, respectively, resulting in a nonlinear optimization
problem of 41 free variables. The control function u is then approximated
by linear interpolation from the discrete values for solving the main problem
(2) as well as the adjoint problem (11). We tested the algorithm with two
different values of the parameter a.
In the first case, we set a = 7r, then u(t) = e-4TZt and z(z) =
e-K2Tcos(7rz). The lower and upper bounds of the control function are
taken to be umin(t)= -1, and umaz(t) = 2, t E [0,1]. In implementing
the continuation technique, the regularization parameter a is reduced five
times at each iteration from 5 at the first iteration to 5e - 9 at the last one.
To analyze the effects of the initial guess and the regularizing function
u*(t)on the reconstruction of the control function, the algorithm was run
with two different initial guesses
uA(t)= 1 - t , ui(t)= 2, t E [O, 11,
and two different regularizing functions
u * J ( t )= 0, t E [O, 11,
2
~ * ? ~=(1t-) 8t, t E [ 0 , 0 . 1 ] , ~ * ' ~=
( t )-(1 - t ) , t E (0.1,1].
9
The results of the four cases are given in Table 1 and Figure 1.
(c) (4
Fig. 1. Reconstruction of the control function of example 1, with a = n
As we can see from the figures that the regularizing function u*(t)has
certain effects on the results. In the case that the regularizing function u*(t)
is taken to be zero, the estimated control function does not approximate
very well the exact solution at the points for which the exact solution is
far from zero (Figures 1-(a) and 1-(c)). In the cases that the regularizing
function u * ( t )is taken to be ~ * ~ ~which
( t )is, near to the exact solution,
the results are better (Figures 1-(b) and 1-(d)). The figures also show that
the algorithm possibly converges to the exact solution with different values
of initial guess, so we hope that it can converge to the global minimum
of the regularized objective function. However, we could not test the al-
gorithm with so many different cases of the initial guess because of its
126 D. N. Hao, N . T . Thanh and H. Sahla
time-consuming calculation.
Table 1. Result of example 1, with a = 7r

Test Objective Obj. func. Number of
function value gradient's norm iterations
a 3.6840e-008 8.3236e-007 2102
b 2.1032e-007 2.2014e-006 1129
C 3.4826e-008 8.3246e-007 2317
d 3.1505e-007 3.0112e-006 971
It is noted from Table 1 that the problem is severely ill-posed. The

values of the objective function is very small with different values of the
control function u ( t ), particularly, it is much smaller than the discretization
error in solving the main and the adjoint problems. It is well-known that
for severely ill-posed problems, as in this example, we cannot expect a
very good estimation for the control function. However, for the optimal
control problem, in which one only need to find a fiinction u(t) such that
the temperature of the system at the final state is close enough to a given
function, the algorithm works very well.
We also tested the algorithm on the previous example for a = 7r/2, then
u ( t ) = - ( ~ / 2 ) e - ~ ' ~ / ~z ,( z ) = e-"2T/4~os(7r2/2).The lower and upper
bounds of the control function are u,in(t) = -2, and u,,,(t) = 0, t E
[0,1].All the setup of the test is similar to that of the previous case. The
initial guesses and the regularizing functions in this case are taken t o be
~ ; ( t=) 0, u;(t) = -1.58 + 1.4t, t E [0,t ] ,

u*>'(t)
= 0, ~*?~
= (-1.58
t) + 1.4t, t E [0, t ] .
The results are given in Figure 2 and Table 2.
The figures show the estimated control function compared with the ex-
act one. In cases (a) and (b) ((b) and (d) resp.), the algorithm was started
at the same initial guess uA(t) (uA(t)resp.) while in cases (a) and (b) ((c)
and (d) resp.) it used the same regularizing function u*>'(t) ( ~ * > resp.).
~(t)
It is clear that the cases (c) and (d) gives better reconstructions than the
other two cases as the effect of the regularizing function u * ( t ) ,while the
initial guess does not effect so much the result. However, as we can see from
Table 2, the values of the objective function are very close to one another
and close to zero. This behavior again shows the severe ill-posedness of the
problem.
R__mdn.rnWMn
I " " " " ' I
(c) (4
Fig. 2. Reconstruction of the control function of example 1, with a = n/2
Table 2. Result of example 1, with a = n/2

T~~~caSe Objective Obj. func. Number of
function value gradient's norm iterations
a 4.7510e-008 9.5507e-007 16848
b 1.1387e-009 3.7500e-007 2866
C 2.0380e-009 1.2505e-007 40230
d 7.5335e-010 1.2505e-007 32760
Example 4.2. In order to compare our results with those of Kelley and
Sachs in Ref. 8, we coFsider here the same example as mentioned in Ref. 8,
10
in which g(y) = 1oo~++y4, yo(z) = 0, z(x) = 1, and T = 1. The constraint of
+
the control function is 0 5 u ( t )5 0.1 4t.
All the discretization parameters in this example are as in Example 1.

Unfortunately, the exact solution is not available in this example so we
cannot have any guess about the solution of the control problem. We also
tested the algorithm with two different initial guesses given by
uA(t) = 0, U i ( t ) = 4t, t E [O, 11,

and two different cases of the regularizing function u*(t)

u * J ( t )= 0 , = 4t,
u*12(t> t E [O, 11.
The results are given in Figure 3 and Table 3.
Fig. 3. Reconstruction of the control function of example 2
Table 3. Result of example 2

Test caSe Objective Obj. func. Number of
function value gradients norm iterations
a 9.9149e-004 4.3683e-004 2816
b 7.2189e-004 4.3675e-004 3108
C 9.9563e-004 4.5075e-004 2740
d 7.2604e-004 4.5646e-004 2710
It follows from Figure 3 that, for the initial guess uA(t)and regularizing
function u*>(t), (Figure 3-(a)), and a = 0.01, the result is the same as that
of Kelley and Sachs. The result is also very similar in the case of the second
initial guess u i ( t ) but with the same regularizing function (Figure 3-(b)).
However, the results are different for the regularizing function u*)(t)for
a = 0.01 (see the difference between (a) and (b) with (c) and (d)). Those
results show the effect of the regularizing function on the results. However,
when a is reduced, the four tests converge to almost the same solution. This
behavior ensures the efficiency of DCA and the continuation technique in
searching for the global solution of the problem.
In conclusion, the combination of DCA and the continuation technique
has been demonstrated to be possible for obtaining the global solution
of the control problem (1)-(2). We also showed that the obtained result
by this method is consistent with the result obtained by the trust region
method proposed by Kelly and Sachs in the case of the same regularization
parameter and function, and the same initial guess. However, our technique
is apparently proved to behave better because it reaches the same solution
with different initial guesses as well as different regularizing functions.
References
1. J. C. Alexander and J. A. Yorke, Trans. Amer. Math. SOC.242, 271 (1978).
2. W. F. Ames, Numerical Methods f o r Partial DiJSerential Equations, Third
edition, Computer Science and Scientific Computing, (Academic Press, Inc.,
Boston, MA, 1992).
3. E. Casas, S I A M J. Control Optim. 35,1297 (1997).
4. E. Casas and F. Troltzsch, Appl. Math. Optim. 39,211 (1999).
5. Yu-hong Dai, Journal of Conputational Mathematics, 9, 539 (2001).
6. Dinh Nho Him: Methods for Inverse Heat Conduction Problems, (Peter Lang,
Frankfurt am Main, 1998).
7. H. Jiang, T. H. Nguyen and M. Prudhomme, Control of the boundary heat
f l u x during the heating process of a solid material, Manuscript.
8. C . T. Kelley and E. W. Sachs, SIAM J. Optimization 9, 1064 (1999).
9. 0. A. Ladyzhenskaya, The Boundary Value Problems of Mathematical
Physics, (Springer-Verlag, New York, Berlin, Heidelberg, Tokyo, 1985).
10. P. Neittaanmki and D. Tiba: Optimal control of nonlinear parabolic systems.
Theory, algorithms, and applications, (Marcel Dekker, Inc., New York, 1994).
11. M. Necati Ozisik: Boundary value problems of heat conduction, (Dover Pub-
lications, Inc., New York, 1968).
12. T. Pham Dinh and L. T. H. An, S I A M J. Optim. 8 , 476 (1998).
13. T. Pham Dinh and El Bernoussi Souad, Dends in mathematical opti-
mization (Irsee, 1986), pp. 277-293, Internat. Schriftenreihe Numer. Math.,
(Birkhauser, Basel, 1988).
14. J. P. Raymond and H. Zidani, Appl. Math. Optim. 39, 143 (1999).
15. F. Troltzsch, Appl. Math. Optim. 29, 309 (1994).
16. E. Sachs, ZAMM 58, 443 (1978).
17. E. J. P. G. Schmidt, J . Dig. Eq. 78, 89 (1989).
9. A CLASS OF SOLUTIONS TO MAXWELLS EQUATIONS

IN MATTER AND ASSOCIATED SPECIAL FUNCTIONS
PETER MASSOPUST
GSF - Institute for
Biomathematics and Biometry
Neuherberg, Germany
and
Centre of Mathematics, M6
Technical University of Munich
Garching, Germany
massopust @ma.tum.de
The solutions for a class of Maxwells Equations in matter are presented. These
solutions describe the magnetic fields as generated by a hard ferromagnet of
finite length with missing mass and are important in the area of nondestructive
evaluation.
Keywords: Maxwells equations, ferromagnetism, Poisson equation, Bessel

functions, spherical harmonics, Mathieu functions.
1. Introduction and Preliminaries

In this section the basic model for a hard ferromagnet is presented. The
basic model is subject to modifications that describe the observed physical
phenomena more closely but these modifications result in sets of solutions
that are no longer expressible in simple closed analytical forms. For the
purposes of this paper, however, the basic model contains all the important
aspects of ferromagnetic theory.
1.1. Basic Magnetostatics

Let 52 c R3 denote a hard ferromagnet* and 852 its piecewise smooth
oriented surface. We assume that 52 is a compact subset of R3 and has
*A ferromagnet is called hard if its magnetization M is essentially independent of any

exterior applied field for moderate field strengths. One may treat such materials as having
a fixed specific magnetization M.
132 P. Massopust
nonempty simply-connected interior. Applying an exterior field H to R

0
induces a magnetization vector field M : R R3 in the interior R of R. It
is advantageous to assume that M vanishes on the boundary dR: Mian = 0.
Note that this introduces a discontinuity of M on the boundary] which is
not observed in the physical reality but provides a convenient mathematical
idealization. The vector fields H and M produce the magnetic field B in
R3 according to the constitutive relationshipt
B=H+47rM. (1)
In addition] there exists also a constitutive relation between the vector fields
H and B of the form
B = f (H),
where f is a multiple-valued function referred to as the hysteresis function.
Figure 1 shows an example of a hysteresis function. In the neighborhood of
Fig. 1. An hysteresis function.
any point Ho where f is single-valued] one may define the derivative

aB
p(H) := -,
aH
called the differential magnetic permeability. For an isotropic ferromagnet
and for small values of H and B, there exists a linear relationship between
the fields of the form
B=pH. (2)
Maxwells equations] which are based on theoretical and experimental
results, describe the behavior of electric and magnetic fields in matter. Their
differential and integral forms as follows.t(Here ( , ) denotes the canonical
tIn what follows, cgs units will be used.

$In the cgs system, the integration of the electric field over the surface of a sphere
produces the factor 47r (times the electric charge enclosed). This explains the ubiquitous
occurrence of 4 ~ .
Solutions t o Maxwells Equations 133
inner product on L2(R3).)
1~(EE) 4lr 1 me
VxH=-J+--
4~
C c at
f ( H, ds) = -c
I + - -,
c at
VxE=--,
dB
dt f ( E , ds) 1
1 awn
-- -
c at .
Here the vector function E : IR3 -+ IR3 is called the electric field and E
the electric permittivity. The constant c is the speed of light, Q : R3 -+
IR the electric charge density and Q the total electric charge inside the
closed surface S. The vector function J : IR3 + IR3 denotes the macroscopic
current density and I := Js J - d u the (macroscopic) current encircled by the
closed curve y bounding the surface S. Given any closed orientable piecewise
smooth curve y bounding a surface S , the quantities CPe, a : R3 -+ IR are
called the electric and magnetic flux, respectively:
CPe :=
s, (E, d u ) , CP :=
L
(B, do). (7)
Here the orientation on S is induced by that of y. Eqs. (7) allows the

interpretation of B as a magnetic flux density, i.e., as the number of
magnetic field lines enclosed by the curve y, whereas CPe can be interpreted
as the number of electric field lines piercing the surface S . For similar
interpretations of electrodynamic quantities and their modcrn differential-
geometric interpretations in terms of differential forms, the reader is referred
to Refs. 13,17.
Eqs. (3) and Eqs. (4) are referred to as Gausss law. In particular,
Eqs. (4) expresses the fact that there are no magnetic monopoles, i.e. that
B is an (axial) vector field or, more precisely, a 2-form: B = B,(z) dy A
+
d z By(z) +
dz A d z B,(z) d z A dy. The functions B,, B y ,B, : IR3 -+ IR are
called the 2, y, and z-component of B, respectively. Eqs. (5) is commonly
called Ampkres law with Maxwells extension and Eqs. (6) Faradays Law
of Induction.
For the purposes of this paper, it is assumed that there are no electric
fields present and that the magnetic field B is static, i.e., dB/at = 0. Hence
134 P. Massopust
the set of Maxwells equations reduces to
( V , B) = 0 ,
47T i ( H , d s ) = 4-7TI
(V,H)=-J,
C C
In the case when J vanishes, V x H = 0 and thus there exists locally a

continuously differentiable scalar function 9 : R3 4 R,called the magnetic
scalar potential, so that
H = -V9. (8)
Using Eq. (8) together with Eq. (11) transforms ( V , B) = 0 into the Pois-
son equation
A@ = 4~ ( V , M), (9)
where A denotes the Laplacian. The right-hand side of Eq. (12) can be
thought of as a magnetic charge density p : R3 -+ R,z H -( V, M)(z):
Choose a point p E 60 and apply the Divergence Theorcm to a cylinder
of radius E > 0 and height 2 E centered at p making use of the assumption
that M vanishes on 60 to obtain that the quantity ( M In) equals
and can therefore be thought of as a magnetic surface charge density. (Here

n denotes the unit (outward) normal to an.) A solution to (12) is then
given by (see for instance in Ref. 5)s
(Here Vz# refers to the gradient with respect to the variable X I E and, for
the remainder of this paper, all primed quantities are evaluated at X I E R.)
If we assume a linear relationship of the form (2), then one can compute
the magnetic field B from the above equation via
B = -V(p@) = -pV@,
Assuming that the interchange of V and is allowed, one obtains
Here one needs to assume that rk E C2(&)n C(0)

Solutions to Maxwells Equations 135
For the above-described idealized situation of magnetizing R, the magneti-

zation field M may, in first approximation, be considered to be uniform and
constant: M = MOe z , MO > 0, e, unit vector in the z-direction. Thus, (11)
simplifies to
< a < b < L, L >> 1, and A(r1,rz) := { ( I c , ~ ): 0 5 TI I

For 0
d m 5 7-2) assume now that R := A(r1~7-2)x [a,b].(For T I > 0 ,
such a hard ferromagnet is encountered, for instance, in nondestructive
evaluations of pipes.) The surface dR of R consists of the annuli A(rl , 7-2) x
{ a } , A(r1,r2) x { b } , and the two cylindrical surfaces dR1 := {(Ic,
y) : z2 +
y2 = r:} x [ u l b ]r1 , > 0, and dR2 := { ( q y ) : z2 + y 2 = r,}x [ a l b ] Now
.
suppose mass, represented by a volume element V , is removed from R in
such a way that part of dV belongs to either 601 or dRz, but not both.
Denote by C that part of the boundary of V which does not belong to dR
and assume C is piecewise smooth and orientable. Note that C is also part
of the boundary of R \ V .
Now if the dimensions of C are big enough and the applied field H is
sufficiently large, then the magnetic field lines of B will leave the interior
of R, penetrate its exterior, and reenter R. This process is called magnetic
~ T U I Cleakage. As removing mass from R is a local phenomenon in regards to
the overall geometry, focus is entirely given to the magnetic field B in the
vicinity of C. This field, BE,is called the magnetic flux leakage field and its
shape can be explained using AmpZlres Dipole Model. One assumes that BE
is generated by magnetic dipoles situated on opposite sides of the surface
C c R \ V . In other words, the distribution of magnetic dipoles on the
surface C, regarded as a Radon measure on C, generates BE. An explicit
expression for BE is given by Eq. (11) with only the slight modification of
replacing dR by C.
Here we also used the fact that the magnetic permeability p is approxi-
mately equal to one outside the ferromagnet R. The magnetic dipole dis-
tribution is given by the Radon measure p~ := (M, d a c ) whose support
is C. The distribution of the magnetic dipoles, i.e., p ~ is, determined by
the magnetization M in a neighborhood of C in R and the specific material
properties of the ferromagnet R. Precise knowledge of these two properties
together with exact information about the geometry of C would uniquely
136 P. Mussopust
-
determine the Radon measure pc. On the other hand, p~ does determine
the geometry of C once M and the ferromagnet's material properties are
known. In this case, one has the one-to-one correspondence C p x . Un-
fortunately, M and the material properties of R are only approximately, if
at all, known.
1.2. Coordinate Systems

Here a summary of the five types of coordinate systems employed in this
paper is given and the relevant quantities necessary for the evaluation of
the integral in Eq. (13) are presented. Since the integrals in Eq. 13 are
additive with respect to volumes and surfaces, the volume elements V asso-
ciated with these coordinate systems make up the building blocks for more
complex removed masses.
1.2.1. Cartesian Coordinates

The unit vectors in the Cartesian coordinate system (z, y, z ) are denoted by
{T,?,c}. Throughout the remainder of this paper, it is assumed that the
y-axis is the symmetry axis for R, that the magnetization M is along the
y-axis, and that is perpendicular to X l 2 .
Definition 1.1. If V has the form

Ku?"= [-w, w] x [--e, 4 x [-d, dl1
where 2w > 0, 2e > 0, and 2d > 0 denotes the width, length, and depth of
the missing mass element V, then it is called Cartesian.
1.2.2. Circular Cylindrical Coordinates

Definition 1.2. V is called cylindrical if it is of the form
V,,l = {(r,$, z ) : r = r g , 0 I 4 5 27r, 0 5 z I d } ,
where rg > 0 and d > 0 denote the radius and depth.
1.2.3. Spherical Coordinates

Definition 1.3. V is called (hemi)spherical if it is of the form
Vsph = { ( p , 4,O) : p = Po, 0 5 4 I 2.rr, T/2 I 0 I n},

with po > 0 denoting the radius/depth.
Solutions to Maxwell's Equations 137
1.2.4. Parabolic Coordinates

The parabolic coordinate system ((, q, 'p) consists of
(1) Confocal paraboloids about the positive z-axis: ( = constant, 0 5 ( <
00.
(2) Confocal paraboloids about the negative z-axis: q = constant, 0 5q<

00.
(3) Half planes through the z-axis: cp = constant, 0 5 'p 5 27r.
As usual, the azimuth angle 'p is measured counter-clockwise from the pos-
itive x-axis when viewed from the positive z-axis. The relationship between
Cartesian coordinate (x,y,z) and parabolic coordinates ((, v,cp) is given by
x = Jqcos 'p, E=\lJ.2.yz.,2-.

{ y
z
= Jqsincp,
= ;(q2 -52)
and q =
'p
Jm
= arctan ylx.
It is noted, that the parabolic coordinate system introduced above is a left-
handed system, i.e., i x i j = -+.
The unit vectors are denoted by i,6, and
+, and the nabla operator V and the Laplacian A are
and
respectively. The oriented surface element d u for a paraboloid ( = (0 =

const. is given by
d u = (07 qsincp r:") dqdcp,
where, due to the left-handedness of the parabolic coordinate system, n is

(14)
the inward normal
Definition 1.4. V is called circular parabolic of radius ro and depth d if

it is of the form
Vpav = { ( E , 71 c p ) : E = 60, 0 Irl I To, 0 Icp I 2 r } ,

138 P. Massopust
where <O = r o / a and 90 = a.

Remark 1.1.
(1) The values for and q in the above definition are obtained as follows.
When = x 2 + y2 = r; the depth d should equal the difference
<iqi
between z1 = z l , ~ + ~ 2 = ~=
; ( 1 / 2 ) ( r $ / < i- ti) and zo = ZI,=~=O =
--<$/2. This requirement then gives the values for and 90 in terms
of ro and d.
(2) A circular parabolic V is expressed then in cylindrical coordinates
(9,cp, by
Note that -ri/4d 5 z 5 d - r;/4d.
1.2.5. Elliptic Cylindrical Coordinates

The elliptic cylindrical coordinate system (u, u,z ) consists of
(1) Elliptic cylinders u = constant, 0 5 u < 00.
(2) Hyperbolic cylinders u =constant, 0 5 u 5 2n.
(3) Planes parallel to the zy-plane, z = constant, -00 < z < +00.
The transformation equations between Cartesian coordinates (x,y, z ) and
elliptic cylindrical coordinates (u,
v , z ) are given by
x = acoshucosv, u = Re [arccosh (w)]
{ y = a sinh usin 21,
z=z
and u = Jm [arccosh I)*(
Here %e and Jm denotes the real and imaginary part, respectively. The
(15)
azimuth angle u is again measured in a counterclockwise fashion from

the positive x-axis when viewed from the positive z-axis. The parame-
ter a is the focal length of the ellipse given by a2 = (major semi-axis)2 -
(minor semi-axis)2.
The nabla operator V and the Laplacian A in this coordinate system
are
and
2
A=
a2(cosh 2u - cos 2v)
respectively, with 0,6 , and 2 denoting the unit vectors.
Definition 1.5. V is called elliptic parabolic of width w , length e, and depth

d if it is of the form
Vecyl= { ( u ,v , z ) : 0 5 u 5 arctanh ( e / w ) , 0 5 v 5 27r, 0 5 z 5 d } ,

or, more precisely,
KCyl: +
z = a2(cosh2u cos2v sinh2u sin2 v ) , 0 5 u 5 uo, 0 5 v 5 2n,
where
a= d w ,w > e, and uo = arctanh ( t / w ) .
Remark 1.2. The values for a and uo are obtained by requiring that for
z = d the boundary of V in this plane is an ellipse with major semi-axis
equal to w and minor semi-axis equal to e.
The surface element d u for an elliptic parabolic V is given as
cos 3~ cash u - cos v cash 3~

sin 3v sinh u - sinv sinh 3u
cos 2~ - cash 2~
with the unit normal n being
-2a cosh u cos v

1
n= -2a sinh u sin v
dl + 2a2(cosh2u + cos 2v)
Remark 1.3. Setting T = (a/2)e", Eqs. (15) can be expressed in the form
If the focal length a-+ 0, v -+ (x, in such a way that ae" remains finite,
and the usual cylindrical coordinates are recovered.
2. Cartesian V
In this section, the magnetic field BEgenerated by a Cartesian V is explicitly
computed. For this purpose, consider a Cartesian V centered at the origin
with length 2 4 width 2w, and depth 2d. To compute the magnetic field BE
140 P. Massopust
generated by this type of V, (13) is employed. For a function f : R3 4 R

and a triple a := ( a l , a2, a3) E R3, denote by { f , a} the function
y , ,Z) := [f(x- ai, y
{ f,a } ( ~ + a2, z - a3) - f ( x - a l , Y + a27 z + a3)I
- [f(x+al,y+az,z-a3) -f(x+al,y+a2,zfas)l
- [ f ( x- al, Y - a21 - a3) - f ( x - Y - a2, z + a311
+ [ f ( x+ a i l y - a2, Z - a3) - f ( X + a l , Y - a2, z + a3)I
Theorem 2.1. Suppose that V,,, i s Cartesian of width 2 w , length 2C, and
depth 2d and the magnetization i s of the form M = MY. T h e n the magnetic
flux leakage field BE = (BzlBy, B,) generated by V,,, can be explicitly and
exactly calculated. T h e exact expressions f o r the components are given by
& ( x , y , 2) = M{P, (%C, d ) ) ( X ,Y l z ) , Y , ). = M{741(w, C, d ) ) ( x ,Y , z )
Proof. Under the assumption on MI (M, du) = M(daIyl=-e-daI,,=e)l

and thus
(M, du)
9 ( x )=
fB Ix-xl
= (Lr
f w
s_dd J ( X
d
- Ic)2
dzdx
+ ( y + !) + ( 2 -
dzdx
Z)2
-L w / d J(X + ( y - C)2 + ( 2 - Z)2

- X)2
).
Since the functions r - ( x , x) = l / d ( x - x)2 + ( y + C)2 + ( z - z ) ~and
r + ( x ,x) = l / J ( x - x ) ~+ ( y - C)2 + ( z - z)2 satisfy all the hypotheses
for the application of the theorem (Interchangeability of Integral and Partial
Derivative) in Ref. 15, p. 59, the x-, y , and z-components of BEare obtained
as
By(5)= --
a9
dY
= -M 1: s_dd (e 2) - dzdx,
and
respectively.
Performing the partial differentiation in the expression for B, yields
(X - 5 )dzdx
- 1: [(z - + (y - l ) 2 + ( z -
Setting =x - x and integrating over E between [ = x + w and E = x - w
gives
B, =M [l d
( x - .1)2
dz
+ (y + + e)2 ( 2 - z)2
dz
-/ d J(X + W)2 + (y + e)2 + (Z - Z)2
d
dz
-I d J(X - W)2 + (y e) +
- (2 - 2)2
+ l d J(Z + 211)
dz
+ (y - e) + (Z -
Substituting C = z - z in each of the four integrals reduces each one to an
1.
integral of the form
dX
= log(x + J.2+x2>+ constant
with a different value for u2. Defining a function cp : R3\ (0, 0,O) 4 Rf by
cp(z, y, z ) = log(-z + d x 2 + y2 + 22),
yields the stated expression for B,.

To obtain the y-component By of BE,one proceeds similarly.
By = -M [/+w
-W
ld J(z
+ e) dzdx
(y
- x)2 + (y + e ) 2 + ( 2 - z)2
3
+w d
(y - t )dzdx
-L w L d J ( x - x)2 + (y - e ) 2 +(z-
Setting E = x - z and integrating over E between E = x w and E = x - w +
and then replacing z by C = z - z and integrating over C from z + d to
142 P. Massopust
z - d yields
B, = -M
[(z - w)(y +t) Ld [(y + l ) 2 +P]J(z
d4-
-4 2 + (y + l ) 2 + (2
z-d
-(. - ul)(y -
l+d [(y - t ) 2 + C](x - w)2 + (y - t) +c2
Each of the four integrals is of the form

r dX
with different values for a2 and b2. Such integrals may be solved employing
the substitution X = a t a n u . This substitution yields
I a2
a sec2 u d u
+
sec2 u db2 a2 tan2 u
du
The latter integral evaluates to
ad- 1 arcsin (dv sinu) , b2 > a2.
Introducing the function $ : R3 \ (0, 0,O) .-+ R
$(z, y, z ) = sgn (y) arcsin
where sgn (x)denotes the signum function, gives the exact solution for By.
The z-component B, is given by
(Z - z) dzdx
- /: [(x - x ) +
~ (y - + ( z - z)2]3/2
Notice that B, is just B, with x replaced by z . Hence, the conclusion
follows. U
Figure 2 below shows an example of magnetic field components for Cartesian

V. The shapes of the depicted surfaces are characteristic of general magnetic
flux leakage fields: The field component B, is bipolar and the component
B, of quadrupole nature. The field component By is symmetric about a
plane perpendicular to the axis of R (for a symmetric defect.)
o Bx
-1
-2
Y
X
0 By
-1
-2
Y
x
-2
Fig. 2. The I , y, r-component of BE for a Cartesian V.

144 P. Massopust
3. Cylindrical V
Suppose that V is cylindrical with radius r o and depth d , i.e., I/cyl =
{ ( r , $ , z ) : ?- = T o , 0 5 $ 5 27r, 0 5 z 5 d } . Assume again that M
is again of the form M = M j , and thus ( M, da) = -Mro sin $d#dz.
Note that since the magnetic field leaks into the interior of V&, n is the
inward normal n = (- cos +,- sin$, 0). Thus,
In order to evaluate the latter integral, the expansion of the Greens function
15 - zI-~ in terms of Bessel or cylinder functions is employed. (See, for
instance, Ref. 8.)
Lemma 3.1. T h e Greens function in cylindrical coordinates f o r the equa-

tion A@ = 0 is given by
Proof. See Ref. 8
Here I, denotes the modified Bessel function of the first kind of order u and
K, the modified Bessel function of the third kind of order u or Macdonalds
function. Exact definitions and properties of these functions may be found
in Refs.1,3,5,9.
With this expansion of 1 ~ - - 2 1 - ~ , the magnetic scalar potential becomes
1
00
+ COS[Y($ - $)]I,(kr<)K,(kr>) d$dzdk.

u=l
Note that the above - and all subsequent - interchanges of the integrals
and the infinite sum is justified. The above integral can be further rewritten
Solutions to Manuell's Equations 145
as
Q ( r , 4, z) = --4 M r o
7-r
1" .Id1'" cos[Ic(z - z')]sin$'
( k r < ) K o(Icr,)
1 d4'dz'dk
As
and
L'" sin 4'd@ = 0
-2 sin[v7-r]sin[v(7-r- 4 ) ]
1'" cos[v(4- 4')]sin +'d+' =
'v - 1
only the v = 1 term in (17) contributes. For this value of v
= 0, for v # fl,
and therefore,
I'" cos(4 - 4') sin +'d@ = 7-r sin 4
@(r,4 , z ) = -4Mro sin 4 1" Ld cos[Ic(z - z')]Il(Icr<)Kl(kr>)dz'dk
= -4Mro sin4 Lm(I d

cos[k(z - z')]dz'
) Il(kr<)K1(Icr>)dIc.
Hence, the next result is established.

Theorem 3.1. Suppose that V i s cylindrical with radius ro and depth, d .
T h e n the magnetic scalar potential generated by V& i s given by
sin Icz - sin[lc(z - d ) ]
Q ( r , 4 , z) = -4Mro sin 4 I1(kr,)K1 (kr,)dk
Ic
(18)
I n the case that r< # ro # r > , this above representation m a y be expressed
in terms of a n infinite series of the f o r m
TMror, sin 4 z2n+1- (z - d)'"+l ( 2 n + l)!!
*I(., 4, z) = - r : n=O E(-V r (2n)!!
146 P. Massopust
Proof. The first part is obvious from the above calculations; the second
part is established as follows. The Taylor series for the function (sin kz -
sin[k(z - d ) ] ) / kis given by
sin kz - sin[k(z - d ) ]
k
= cm
n=O
(-1)n
-p+l - ( z - d)2"+1
(2n + l)! k2"
Substitution of this series into (18) and interchange of infinite sum and
integral yields
m
- 4 ~ r osin4 C (-l)n[z2n+l( 2 n -+ (.l)!- I 22n-1 < >
r-2n-2
n=O
xr(n + 1/2) r ( n + 3/2) 2F1(n + 1/2, n + 3/2; 2; ( r < / r > ) 2 > ,

Here ~ F I (b;
Uc;,x) denotes Guufl's hypergeometric function:
Note that
- &r(n +
1/2) - 7r ( 2 n - I)!!
~
2 2 n + q n + 1) 2 3 n + q n 1 ) +
- 7r ( 2 n - l)!!
-
22n+l (2n)!! '
where the duplication formula for Gamma functions was used (Ref. 1,
6.1.18). Thus, after simplification,
7rMror< sin 4 + l)!!

w r , 4, 2 ) = - r : F(-Un
n=O
2 n + l - (2
7-p
- d)2"+1 ( a n
(2n)!!
x 2J'1(n + 1 / 2 , n + 3/2; 2; ( r < / r > 1 2 > ,

which proves the statement. 0
In order to obtain formulae for the magnetic field BE, two cases need
to be considered.
CASE I: T > T O
In this case, rc = min{r, T O } = TO and r> = max{r, T O } = T . Hence,
a(')( T , 4, z ) = -4M'r-o sin 4

Irn sin kz
The radial component B, of BE is the given by

- sin k(z - d)
k
11 (kro)K1 ( k r)d k .
Using the fact that dKl/dr = (-1/2)[Ko+Kz] (Ref. 3, p. 401,11.4.6) gives
Jo
The azimuthal component B,#,of BE is obtained in a similar fashion. As
B+ = (-1/r)(N/a4), the result is
The z-component B, of Bc is -dG/dz yielding
CASE 11: T < T O
In this case, r< = min{r, T O } = r and r , = max{T, T O } = ro. Hence,
The radial component B, of BE is the given by
Using the fact that d1l/dr = (1/2)[10 + 121 (Ref. 3, p. 397, 11.116) gives
*m
As above in Case I, the azimuthal component B4 is

148 P. Massopust
and the z-component
BL)(r, 4 , z ) = 4Mr0 sin4

Lm {cos k z - cos[k(z - d ) ] }I l ( k r ) K l ( k r o ) d k
The Cartesian components B, and By of BE are obtained via
4. Spherical V
In this section, an exact solution for the magnetic flux leakage field due to a
(hemi)spherical V is derived. To this end, assume that the radius/depth of
said V is po. Then V,,, = { ( p , 4 , 0 ) : p = p o l 0 5 4 5 2 ~ ~ , / 52 6 5 T } .
The magnetization M is again assumed to be along the y-axis: M = f i .
The normal to C is given by ( M, do) = -Mpz sin 4 sin2 Od4de (inward
normal!). Thus,
WPl41e) = as # (MI,
l--l
da)
pi JT/2
T J 0
2T sin2 9 sin dd4de
15-x2l
To proceed the next lemma is needed.
Lemma 4.1. I n spherical coordinates, the Greens function Ix - x1-l for

the potential equation A* = 0 has a n expansion in terms of spherical har-
monics yem of the form.
with p< = min{p, PO} and p> = max(p,po} (Refs. 3,5,6,8,15). Here * de-
notes complex conjugation. The spherical harmonics Yem(B,4) are express-
ible as
where Pem(x) are the so-called associated Legendre functions of degree C

and order m:
Proof. See Ref. 5 or Ref. 8, for instance. For definitions and results see
Ref. 1 (Ch. 8) or Ref. 3 (Ch. 12). 0
The right-hand side of (19) may be expressed completely in terms of

real functions by splitting off the C = 0 and (C, m) = (C, 0) term, combining
those terms in the sum over m that correspond to positive and negative
+
indexes, and using the fact that ecim4 eim4 = 2 cos 4. This yields
The functions Pi(.) = Pe,o(Z) are the Legendre polynomials of degree e.

With the above expression for (z- ~ 1 - lthe
~ magnetic scalar potential
now reads
2x sin2 0 sin 4dqYdO
x li2 sin2 8 sin $cos[m(+ - 4)]Pem(cos8)d4d8 Pt,(cos 8) .

1
(21)
The first and second term inside the bracket vanishes since the integral over
4 is equal to zero. Moreover,
I sin#cos[m(+ - $)]dq5 =
ifm>l
.rrsin+ if m = 1.
Therefore, in the third term inside the bracket only the m = 1 summand
contributes to the sum over m. Thus,
150 P. Massopust
The integral over 8' can be exactly evaluated using the following approach.
Let u = cos 8'. Then
sin2 B ' P e , l ( ~ ~ ~ ~ '=

) d-B ' /-'
0
d D P t , l ( ~ ) d ~
= 1' JsPt,l(-u)du
= ( - ~ ) ~ + l /d' 3 P e , l ( ~ ) d ~ .
0
Here the fact that Pe,(-X) = (-l)'+"P~,(x) was employed (Ref. 3, p.

437, 12.93). Using the definition of the associated Legendre functions Pe,
(Ref. 3, p. 435, 12.84),
the function Pe,l can be written as Pg,l(u)= d m d P e / d u . Hence,
( - l ) e + l0/ > m P g , l ( u ) d u = (-l)e+l

11(1 - u2)-du
dPe
du
= (-1)'+l [(l- U2)P'(U)IA 2 + 1 1

uP'(u)du]
= (--I)'+' [-Pi(O) +2 /'0

uPr(a)du] .
Now,
(Ref. 3, p. 424, 12.34 and 12.35) and the moment integral is equal to
11uP*(u)du= { ~ ~ ~ + ~ ~
if[= 1
i f e+= 2A X)
2
if l = 2 X + 1,
(23)
(Ref. 1,8.14.15).In the equation above, r denotes the real Gamma function,
which is the extension of the factorial to real numbers: r(z)= (z+l)!.Using
the fact that (Ref. 1, 6.1.17 and 6.1.12)
combining (22) and (23), and simplifying yields

(-l)+X(2X+1)(2X-3)!! i f e = 2 X
S,, (X+1)(2X)!!
sin2 epe,l(cos e)de =
(I
Therefore one arrives a t the following result.
ife=1
if C = 2X + 1.
(24)
Theorem 4.1. T h e magnetic scalar potential due to a (hemi)spherical de-

fect of radius po i s explicitly given by
a(p,4 , 0 ) = - 2 r ~ p i [f sin 0 sin 4
5
+
X=l
(-1)X+(2X - 3)!! p;x
(2X+2)!!
2x+1p2X,1(cos0) sin 4
p,
1
where
(-1)x+1(2X - 3)!!
Nx =
(2X + 2)!! .
Proof. Only Pl,l(cos 0) = sin 6 remains to be shown. This, however, follows

immediately from (20). 0
152 P. Massopust
In the derivation of Be the following recursion relation of associated Legen-

dre functions was used to replace the derivative of P1,1,respectively, P2x,1
with respect to 0 (Ref. 3, p. 437, 12.90).
d ~ ~ ~ ( 1~ ~ ~ e )
ae = +
--[Pe,m+l(cos0 ) - (t m)(t- m
2
+ l)P+-, (cos e)]. (25)
The Cartesian components B,, By, and B, of BE are obtained via
B, = sin Bcos r$ B, + c o d cos4 Be - sin$&,

By = sinesin4 Bp + cos Bsin4 Be + C O S B,~
B, =coseB,-sinBBe.
4.1. Semi-spherical V
To complete the discussion of spherical V s , an exact expression for the
magnetic flux leakage field BE generated by a semi-spherical V is derived.
Definition 4.1. V is called semi-spherical if it is either of the form V, =

{ ( ~ , 4 , 0 ): P = PO, 7r 5 4 I 27r, .rr/2 I 8 5 x} or V , = { ( p , 4 , @ ) : P =
p0, 0 I 4 5 T , 7r/2 I e I TI.
In the derivation of the exact solution for magnetic flux leakage, the first
form is used and the expansion in (21) is employed with the appropriate
modifications. (The magnetic flux leakage field for V, is obtained from that
of V, via the reflection y H -y.)
Solutions t o Maxwell's Equations 153
The scalar magnetic potential 9 is comprised of three terms Ti =

T i h $,el, T2 = T2(p74, e), and T3 = T3(p,4,e). More precisely,
9 = -MPi K ( P 7 476) + Tz(P,4 , q + T3(p,4, Q)] .
with
Ti= IT /2x sin' 8' sin 4'd+'dQ'
- --
7r
x/2 T P> 2P> '
where
Ke = l;2 sin26'Pe(cos e')de',
and
(e- m)! p:
e=i m=l
x
x/2
/x
2x
sin2d'sin 4'cos[m(4 - 4')]Pem(cosO')d$'dO' Pem(cos0)
154 P. Massopust
where
Ke,zP = 1:2 sin2 8' P e , 2 , 8')dO'

(~~~
and the fact that

y
l= sin 4' cos m(q5 - q5')dq5' =
2cos COSrn
m2 - 1
($ - 4)
=
i;sin4
&
2cos2
0
4
form
form=2p
otherwise
=1
was used. Here [.]I : IR -t Z denotes the greatest integer function. In the
above derivations previous results, in particular (24), was used.
To obtain the magnetic field components, the partials of TI, T2, and T 3
need to be calculated. To this end, note that
where
and
The derivatives dPe,,/dcos 0 were replaced according to (25).

156 P. Massopust
Theorem 4.2. The magnetic flux leakage field BE = ( B p B+, , Be) gen-
erated b y a semi-spherical V of radius po is explicitly obtained from the
expressions below.
where the derivatives of Ti, i = 1 , 2 , 3 , are given above.
The Cartesian components of BE can be obtained via (26).
5. Parabolic V
In this section the magnetic flux leakage field generated by the two types
of parabolic Vs introduced in this paper is calculated.
5.1. Circular Parabolic V

Assume that V,,, is a circular parabolic with radius ro and depth d :
Vpar = { ( E , ~ I P)
, : E = Eo, 0 I r~ I 770, 0 I 9 L 2 ~ } ,
with 50 = -& and q o = d%. Using the form of the oriented surface element
as derived in (14) and M = M3, one finds for the magnetic scalar potential
Next an expression for the Greens function l / / x - xI in parabolic coor-

dinates is derived. To this end, recall that the Dirac delta distribution S in
parabolic coordinates is given by
and that a solution to the (distributional) differential equation

is sought. For (I p) # ( [ I , q, cp) the above equation can be solved by the

q,,
ansatz G = E(<)H(q)@(cp).Thus, upon substitution, simplification, and
division by SH@ one obtains
<2q2
E2+q2 [E
1
(@
a%
+--< aE
1
a[) +-
As the variables <, q, and p are independent of each other, there exists a
constant m2 so that
and
The former equation has solutions of the form

=
The latter equation can be rewritten as
Separation of variables yields
and
for a constant c.
(29) is a modified Bessel Equation whose solutions are Bessel functions
I , and K , of the third kind, and (30) is just Bessels equation with so-
lutions J , and y,. The Greens function for a magnetic scalar potential
concentrated on the surface [ = 6 = const. must involve only the function
J, as Y, is singular at the origin.
To solve (27), recall that
(this is just the Fourier series for the Dirac delta distribution) and
158 P. Massopust
which is the Hanlcel transform or Fourier Bessel Integral of the Dirac delta
distribution (Refs. 8,14). Therefore, a solution of the inhomogenous equa-
tion (27) in the form
is sought. Substituting this ansatz into (27) and simplifying gives
+L+ -Jm
E2 + v2
IC.
77
b2Jm - e*im(v-v)Jm(k771)
(Here the dots indicate differentiation with respect to the variable upon
which the function depends.) Using the fact that
.. .
k 2 Jm + -k Jm ,2
- -Jm = -k2 Jm
77 v2
and (31) and (32), one finally obtains
Hence gm(E, E) is a Greens function for the ordinary distributional differ-

ential equation
whose solution is known to be (cf. for instance,8)
with E< = min{E, E l } and E> = max{E, E}. Finally, the following result is
obtained.
where
1 ifm=O
em={ 2 ifm>O
i s the N e u m a n n factor, [< = min{[, c}, and E> = max{[, [}.
With this expression for the Greens function, the magnetic scalar potential
becomes
)Km(kJ>)dv k d k
x?72Jm(kv)Jm(k?7)Im(ICE<
where & = min{J, Jo} and [> = max{[, t o } . As
-/rsin+ i f m = 1
cos[m(cp - p)]sin cpdcp =
otherwise
only the term m = 1 contributes to the sum. Moreover,
and thus the next theorem holds.

Theorem 5.2. T h e magnetic scalar potential due t o a circular parabolic
V of radius ro and depth d i s given by
where [< = min{f,[o} and [> = max{[,[o}. In the case [< < [>, the
above improper integral m a y be expressed as a n infinite series of the form
160 P. Massopust
Proof. Only the second statement needs to be shown. To this end, employ
the following series representation of the product of two Bessel functions
(Ref. 16, p. 148, ( 2 ) ) :
(fa.)"( f b z ) .
J p( a z )J,, (bz) =
r(Y + 1)
M
(-1)"(~az)"2F1(-n1 - p - n;v + 1;b2/a2)
x n=Oc m!I'(p m -t1 )+
Setting 1-1 = 1, u = 2, a = r], and b = 770 in the above equation, using the
fact that
yields
(n + l)!(n + 2)!
n=O
+
x 2 ~ 1 ( 2 n, 3 + n;2; &J:) 2 ~ 1 ( - n , --n - 1; 3; v02/v2).
Simplification now gives the result. CI
Theorem 5.3. The components of the magnetic field BE = ( B EB,, , Blp)

generated b y a circular parabolic V of radius ro and depth d are then given
by
The Cartesian components (Bx,By, B,) of the magnetic field BE are ob-
tained from the parabolic components (Bc,B,, BV)via the transformation
-t rl 0
5.2. Elliptic Parabolic V

Now assume that V is elliptic parabolic:
a2
Kpar: Z = -(cash 2 ~ +
cos 2 ~ ) , 0 5 U <_ UO, 05 V <_ IT,
2
with
a= d m and uo = arctanh (elw).
Assuming that M = Moj, the magnetic scalar potential * in elliptic cylin-
drical coordinates is given by
-
-
a3Mo LUl
(sin 3v sinh u - sin vsinh 3u)dudv
2n
2 Iz - 21
At this point, the Greens function G ( z ,2)= l/lz - zI must be ex-
pressed in elliptic cylindrical coordinates. For this purpose, note that this
Greens function satisfies the (distributional) differential equation
a2G -
A G ( z , d )= - a2G +-=+
a2G
-4742 - z),
ax2 ay2 a22
where 6 ( z - z)denotes the Dirac delta distribution. In elliptic cylindrical

coordinates the above equation is written as
AG(u,u) =
2
+-
-8.T
- S(u - u)6(v - v)S(z - 2 ) . (35)
a2(cosh 2u - cos 2v)
For u # u a solution of the above equation may be found by the ansatz
G ( u ,u) = V ( u ) V ( v ) Z ( z )Substitution
. into ( 3 5 ) and division through
UVZ yields
162 P. Massopust
where
2
G=
a2(cosh2u- C O S ~ U )
Since u, u,and z are independent variables, it follows that there exists a
constant k2 so that
This equation has solutions of the form Z ( z ) = e f i k r .Again, from the fact
that G = lc2 and the independence of u and u there exists a constant p such
that
1 d2U a2k2
u aU2 2 cash 2u = p
and
1 d2V a2k2
cos 2u = p .
v au2 2
Simplified, these two equations reduce to a Muthieu equation
62V
-+ ( p - 2qcos2v)V = 0
dU2
and an associated Mathieu equation
d2U
-- ( p - 2qcosh 2u)U = 0, (37)
au2
with q = -a2k2/4 < 0. (Note that k = 0 does not yield bounded solutions.)
It is known that in (36), for each given value of q, there exists a set
of eigenvalues p n = p,(q) and an associated set of complete (even or odd)
periodic eigenfunctions of period 7r or 27r. (The second integral is non-
periodic and for physical reasons ruled out.) The set of even eigenfunctions
of period 27r is denoted by {~e2,+~(v): n = 1 , 2 , . . .} and the set of odd
eigenfunctions of period 27r by {se2,+l(u) : n = 1,2,. . .}.q Moreover,
cezn+1 and sezn+1 satisfy the orthogonality conditions
and
1'" se,(u)ce,(u)du = 0,
TThe functions ceZn+l and segn+l do also depend on q. To simplify notation, however,
this dependence is suppressed when not explicitly needed.
and may be expressed as infinite series of cosines and sines

m
seZn+l(v) = c(-l)n+VAe$)
m
V=O
+ < sin(2v l)v, q 0.
( 2 n + l ) = ACA1)(q)and B2+l
where the coefficients A2v+l = B;r+;l)(q)
depend on q and its sign. Note that
and
m
V=O
Solutions for the modified Mathieu equation (37) for the same values of p
and q are of the form cehn,+l(u) associated with ce2,+1(v) and sehz,+l(u)
associated with ~ e 2 ~ + 1 ( vNon-periodic
). second integrals of Eq. (37) are
given by fekhz,+l(u) and gekh2,+,(u) associated with cehz,+l(u) and
sehZn+l( u ) , respectively. It is noted that these modified Mathieu functions
also depend on q and its sign.
The following behavior of unmodified and modified Mathieu functions
applies when q 4 0-.
ce,(v) + cosnw, for n 2 1,
se,(v) -+ sinnv, for n 2 1,
ceh,(u) &l,(kr), c, constant,
-+
seh,(u) -+ c ~ l , ( k r ) ck
, constant,
fekh,(u) + c:K,(kr), c: constant,
gekh,(u) t c:K,(kr), c c constant.
For proofs see, and for the verification of the argument in the modidied
Bessel functions I, and K, recall Remark (1.3) and the fact that k2a2 =
-49.
More details and additional information about Mathieu functions can
be found in,1,4 and.1
Set
V ( v )= seZn+l(v, q ) , with q = -a2k2/4 < 0.
164 P. Massopust
Since the system {sezn+l} forms an orthogonal complete set of eigenfunc-

tions, the Dirac delta distribution 6(v - v) can be written in the form
b ( v - w) = - Se2n+l(v)Se2n+l(V/).
n=O
(Refs. 3,8,14) In order t o solve (35), the Dirac delta distribution S(Z - 2)
is represented by its Fourier transform as
6(z - z) = -
21r Jm
-m
cos[k(z - z)]dk.
The following ansatz for the Greens function is then made.
where the function gn is determined so that (35) is satisfied.

Substituting the above expression for G into (35) yields (As above, the
dots indicate differentiation with respect to the variable upon which the
function depends.)
+ cos[k(z - ~)]~e~~+l(~)~e2~+1(~)9,]
-k2 cos[k(z - z)]~e2~+l(w)se2~+1(21)g,

dk
)
Employing (36), and simplifying gives
9Lm(
7T2
n=O
2 cos[k(z - z)lsezn+l(v)se2,+l(v)
a2(cosh2u - cos2w)
x [gn - ( p - q cosh 2u)gn +4d(u
)
- dk = 0.
u)]
Hence, gn satisfies the (distributional) differential equation
jn - ( p - qcosh2u)gn = - 4 ~ 6 ( u- u). (38)

As sezn+l was chosen as a first integral of (36), the functions sehzn+1 and
gekh2,+, are the only acceptable solutions of (38). A solution gn(u,u)
satisfying the symmetry requirement g,(u, u)= gn(u,u)and the finiteness
condition a t u = 0 and u + 00 must therefore be of the form gn(u,u) =
csehZ,+l(u<)x gekh,,+,(u,), where u< = min{u,u}, u> = max{u,u},
Solutions t o Maxwell's Equations 165
and c is a constant to be determined so that the Wronskian W of sehzn+l

and gekhzncl satisfies
= -4T/C.
To determine c, results concerning the expansion of sehzn+1and gekh2,+, in

terms of modified Bessel functions and the asymptotic behavior of sehzn+1
and gekh,,+, and their derivatives are employed.
Lemma 5.1. The functions sehz,+l and gekh2,+, have the following series
representations in terms of modified Bessel functions (for q < 0). (- =
d/du)
w
and
u=o
The functions sehzn+1 and gekh2n+l and their derivatives satisfy the fol-
lowing asymptotic estimates as u 4 03.
and
Moreover,
and
166 P. Massopust
Proof. See Ref. 4 for the Bessel series expansions and the asymptotic esti-
mates for seh2,+1 and gekh,,+,. To obtain the estimates on the derivatives,
consider (39) and (40) together with the following asymptotic estimates for
the derivatives of the modified Bessel functions 1, and K , (Refs. 1,3,9).
and
Using the asymptotic estimates provided in the above lemma, the Wron-
skian of sehzn+1 and gekh,,+, is computed as
Hence,
and therefore
implying the next result.

Theorem 5.4. T h e Greens function of the equation A@ = 0 i s elliptic

cylindrical coordinates i s given by
k)dk
x ~e~n+l(v,k)seh2,+l(u<,k)gekh2n+l(u>,
with
(-l)nk2A~fl)B~2n+)
G ( k )=
=;,+I (7r/2) Se2,+1(0) dezn+1(7r/2)
and u< = min{u, u},u, = max{u, u}.
Next the magnetic scalar potential for an elliptic parabolic V is computed.
To this end, note that
x [sinhusin3v - sinh3usinvl se2,+1(v, k ) se~,+l(v,k )
x seh~,+l(u<, k ) gekhZn+,(u>,
k ) dudvdk
-~T x l
u ~ MO0 c,(k)
~ (lu12n 1
cos [z- 5U 2( ~ ~ ~ h 2 u - ~ ~ ~ 2 w )
n=O
x [sinhusin3v - sinh3usinwl se2,+1(2), k ) s e ~ ~ + l ( wk),
x seh~,+l(u, k ) gekh2n+l(u,k ) dudv
+ 12*
luo cos [Z - :(cash 2.11 - cos 2 4
1
x [sinhusin3v - sinh3usinvl se2,+1(vl k)se2,+1(v, k )
x seh2,+1(u, k ) gekh2n+l(u,k ) dudw

168 P. Massopust
To simplify notation, the following definitions are made.
Definition 5.1. Let
S2n+l(U, z , k ) = (-In
7 lU12 [ f
cos k z - (cosh 2u - cos 271)
1
x [sinhusin3v - sinh3usinvl ~ e h 2 ~ + 1 ( uk ),s e ~ ~ + l ( vk),dudv.
and
G2n+l(U,2, k ; uo) = -
?I- 2
(cosh 2u - cos 2v)
x [sinhu sin 3v - sinh3usinvI gekh2n+1(u,k ) ~ e 2 ~ + 1 ( vk),dudv

1
To further simplify Szn+l and G2n+1, note that
1
1 2 T c o s k [ z - ~ ( c o s h 2 u - c o s 2 v ) sin3v~e2~+1(v,k)du
= cos k
[ ;
z - - cosh 2 ~ ] 127Tk [
cos cos 2v] sin 3v seZn+l(v, IC)
du
1 1 sink [fcos 2w] sin3v ~ e p , + ~ ( w Ic), du
and, in particular,
I cos /c [ cos 2v] sin 3w seZn+l(w, IC) du = c03 (-1)m (T) 2m
(2m)!
m=O
and
6 [:
2mfl
00 (-1)rn ($)
sink cos 2v] sin 3v sez,+l(w, k ) du = C (2m + l)!
m=O
00 27T
C O S ~ ~ + ~ (sin3v
~ ~ )sin(2v + 1)wdv.
u =o
Using the trigonometric identity
sin3vsin(2v + 1)v = [cos2(v + 2)v - cos2(v - I)v]/2,

one obtains
= f C O S 2 y 2 V ) [COS2( + 2)v - cos 2( - l)v] dv

C O S z~ [COS(Y
~ + 2)z - cos(v - 1)z]dz
(43)
and
C O S ~ ~ + ~ sin 3v
( ~ ~ sin(2v
) + 1)vdv
( ~ ~2(v
C O S ~ ~ + ~ [cos ) + 2)v - cos 2 ( -~ l)w] dv (44)
= ;14= C0s2m+1 z [cos(v + 2). - COS(Y - l)z]dz
To proceed, the following lemma is needed.
Lemma 5.2. Let a and m be integers. Then, for m > 0 ,
47r
fora=2b andb=O, ..., m.
otherwise.
and, f o r m 2 0 ,
10 otherwise.
Thus, applying the above lemma to (43) and (44) and simplifying yields
170 P. Massopust
and
(-1)nT 2m+1 2m+1

-- ~ % ( ~ ~ l ) [ ( m - V - l ) + ( m + Z + v ) j
2m+1 2m+1
m-u
Hence,
1
l r c o s k [z - %(cosh2u1- cos2v) sin3vse2,+l(v, k)du
)
m- 1
+A%(m2)]
m 2m+la4m+2
* (-1) k
24+3(2m + l)!
-ACA1) [
?:Ti) + (m2m+ li-+ v )I)).
The integral
[ cos k [z - %(cosh2u - cos 271)

1 sinvsez,+l(v, k)du
] l k [cos cos 2v] sin 21 sean+1(v,IC)du
] 12Tk:[
sin k) du
cos 2v] sin v~e2~+1(v,
is computed in a similar fashion, employing the trigonometric identity
sinvsin(2v + 1)v = [cos2(u + 1)v - C O S ~ V V ] / ~ .

The result is
m-u
m
x ~ ( A ~ ' ) + A [~r m
u=o
~ +
) l) )
m--v
+ ( 2m+1
m+l+v
.)I
For the sake of notational simplicity, define
CL2)(k)= c
O0
m=O
(-l)mk2m+la4m+2
24m+3(2m + I)!
2 u=O
(A=') [( 2m+1
m-v-1
)+( 2m+1
m+2+v
)]
[(2:':)+ ( m2m
+l+v + )])
m
+ c
u=l
(2n+l)
(A4u-1 -k
Then, using these four functions, the expressions for Szn+1 and Gzn+1 can be
written in a more succinct way. The results are summarized in the following
lemma.
172 P. Massopust
Lemma 5.3.
S2n+l(U, z , k)
1
cosh 2u sinh u~eh2~+1(u,
k) du
- Ci2)(k) lu sin k [ z -
a2
1
cosh2u sinhuseh2,+l(u, k ) du
+ CL3)(k) 1 cos k [ z -
a2
1
cosh 2u sinh 3u ~ e h 2 ~ + l ( uk),du
- Ci4)(k) lu sink [z -
a2
1
cosh2u sinh3useh2,+1(u, k) du
and
G2n+l(U, z , k ; 210)
= Cil)(k) luo[ cos k z -

a2
1 k) du
cosh 2u sinh ugekh2,+, (u,
- Eiz)(k) / u o
U
sink [ z -
a2
1
cosh 2u ~ in h u g e k h ~ ,+ ~ (uk ,
) du
+ EF)(k) J, cosk [ z - a2
1
C O S ~ ~ U ~inh3ugekh,,+~(u,
k) du
- Ei4)(k) luo sink [z -

a2
1
cosh 221 sinh 3 ~ g e k h ~ , + ~ ( uk ),du
Proof. The formula for Gzn+l is essentially given by Szn+1 with sehzn+1
replaced by gekhzncl.
In summary,
Theorem 5.5. Suppose that V is elliptic parabolic of length 1, width w,

and depth d . T h e n the magnetic scalar potential generated by V i s given by
+ G2,+1(u, z, k;U O )sehz,+l(u)] sean+1(~)

dk.
where uo = arctanh (e/w), a = d w and

The components (BulB,, B,) of the magnetic flux leakage field BE gen-
erated by this defect are obtained as usual via differentiation.
Theorem 5.6. T h e magnetic flux leakage field BE generated by a n elliptic
parabolic V of length e, width w,and depth d i s given by BE = (BulB,, B,)
where
and
with
S2n+l(U, z , k)
1 Ic) du'
cosh 2u' sinh u'seh2,+1 (u',
+ EL2)(Ic) Lu [
cos k z -
a2
1
cosh 2u' sinh u' seh2,+1 (d,
Ic) du'
+ Cr)(k) 1"
sink [ z -
a2
1
cosh2u' sinh3u'seh2,+1(u', Ic) du'
+EL4)(Ic) Au a2
1 Ic) du'
cosk [z - - cosh 2u' sinh3u'~eh2~+l(u',
2
174 P. Massopust
and
=k (Cil)(k) loIc [ z sin -

a2
1
cosh 2u sinhu gekh2n+l(u, k) du
+ CA2)(k) luo cosk [ z -

a2
1
cosh2u sinhugekh2n+l(u,k) du
+Cr)(k) luo sink [z -

a2
1 k) du
cosh 2u sinh3~gekh~,+~(u,
+Ci4)(k) luo cosk [z -

a2
1 k) du
cosh2u ~inh3ugekh~,+~(u,
The Cartesian components (BZ,B,, B,) of Bc are obtained from the ellip-
tic parabolic components (Bu,B,, B,) via the transformation matrix
2 sinh u cos u -2 cosh u sin u 0
1
T= 2 cosh u sin u 2 sinh u cos Y 0
a(cosh 221 - cos 2v)
0 a(cosh 2u - cos 271)
as
(3) (3) =T 7
with the right-hand side completely expressed in Cartesian coordinates.
References
1. M. Abramowitz and I. Stegun, Handbook of Mathematical Functions, Dover
Publications Inc., (New York, 1972).
2. R. F. Arenstorf, Lectures o n Special Functions (Bessel Functions, Spherical
Harmonics, Elliptic Functions), Lecture Notes, (Vanderbilt University, 1973).
3. G. Arfken, Mathematical Methodsfor Physicists, Academic Press, (New York,
1966).
4. R. Campbell, Theorie Generale de Ldquation de Mathieu et de Quelques
Autres Equations Differentielles de la Mecanique, Masson et Cie Editeurs,
(Paris, 1955).
5. R. Courant and D. Hilbert, Methoden der Mathematischen Physik 111,
(Springer Verlag, Heidelberg, 1968).
6. E. W. Hobson, The Theory of Spherical and Ellipsoidal Harmonics, (Chelsea
Publishing Company, New York, 1955).
7. G. Hammerlin and K. H. Hoffmann, Numerical Mathematics, (Springer Ver-
lag, New York, 1991).
8. J. D. Jackson, Classical Electrodynamics, (John Wiley & Sons, 2nd ed., New
York, 1975).
9. N. N. Lebedev, Special Functions and their Applications, (Dover Publications

Inc., New York, 1972).
10. N. W. McLachlan, Theory and Applications of Mathieu Functions, (Oxford
Clarendon Press, London, 1947).
11. F. W. J. Olver, Asymptotics and Special Functions, (Academic Press, New
York, 1974).
12. J. Stoer and R. Bulirsch, Introduction to Numerical Analysis, 2nd, ed., Texts
in Applied Mathematics 12, (Springer Verlag, New York, 1993).
13. B. Schutz, Geometrical Methods of Mathematical Physics, (Cambridge Uni-
versity Press, UK, 1980).
14. A. Terras, Harmonic Analysis o n Symmetric Spaces and Applications I,
(Springer Verlag, New York 1985).
15. W. Walter, Potentialtheorie, BI Hochschulskripten, 765/765a, (Bibliographis-
ches Institut, Mannheim, 1971).
16. G. N. Watson, A Treatise o n the Theory of Bessel Functions, (Cambridge
University Press, London, 1922).
17. K. Warnick, R. Selfridge and D. Arnold, IEEE Trans. o n Education, 40, 53
(1997).
Eds. N. M. Chuong et aI. (pp. 177-196)
@ 2007 World Scicntific Publishing Co.
$10. ON THE CAUCHY PROBLEM FOR A QUASILINEAR

WEAKLY HYPERBOLIC SYSTEM IN TWO VARIABLES
AND APPLICATIONS TO THAT FOR WEAKLY
HYPERBOLIC CLASSICAL MONGE-AMPERE EQUATIONS
HA TIEN NGOAN* and NGUYEN THI NGA*

18 Hoang Quo. Viet Road, 10307 Hanoi, Vietnam
E-mail: htngoan0math.ac.t.n
The Cauchy problem for a normal quasilinear weakly hyperbolic system in two
variables is considered. Sufficient conditions for its diagonalization are given.
The local solvability of the noncharacteristic Cauchy problem for some weakly
hyperbolic classical Monge-Amphe equations is proved.
1. Introduction
The classical hyperbolic Monge-AmpBre equation with two variables is that
of the form
F ( 2 1 , 2 2 , z 1 p l , ~ 2 , r , s l=Ar + Bs + C t t ( ~-ts2 ) - E
t) = 0, (1)
where z is an unknown function defined for ( 2 1 , 2 2 ) E R2,pl =
= z(z1,22)
2x1 i P2 = zxlxl
zx2, r , s = zZlx2and t = zx2x2.The coefficients A , B , C
=
and E are real C2-functions of ( X I , 2 2 , z , p l , p2) and satisfy the condition
of hyperbolicity:
A := B2 - 4 ( A C + E ) > 0.
In this case the characteristic equation of the Eq. (1)
X2+BX+(AC+E)=0
has two different real roots
*Supported in part by the National Basic Research Program in Natural Science, Vietnam
178 H. T. Ngoan and N. T. Nga
In this paper we consider the case A > 0, i.e. the characteristic roots
Ai,A2 may coincide at some poits. In this case the Eq. (1) is said to be
weakly hyperbolic and it can be written in the following equivalent form
Zxixi + C zXlX2 + AX
= 0. (3)
un + ^2 2x2*2 + -A
We are interested in looking for local classical solution to the Eq. (1). So we
suppose that we are given C2-functions X(ai), X(ai), Z(ati), P^(QI),
P^(a\), ai e /, where / is an interval (0, 5), 6 is some sufficiently small
positive number.
The Cauchy problem for the Eq. (1) consists in looking for z(x) C2
which is a solution of (1) such that
where
From (4) we have the following necessary condition for the initial Cauchy
data
which is assumed to be fulfilled for all a\ I.

We assume that the Cauchy problem (l)-(4) is not charateristic (see
Refs. 6,8), i.e.
0,Vai /
where the coefficients A,B,C are calculated at
Equation (1) was investigated in Refs. 1,2 by G. Darboux and E. Goursat

under the assumptions that A > 0 and there exist two independent first
integrals for the Eq. (l). We recall that a function tp(xi,X2, z,pi,p 2 ) is said
to be a first integral ion the Eq. (1) if it satisfies the following system of
equations
y t,^ &,~ Q{n Q(n
'--\-i-^-- A-~ = 0 .
z ap\ ops
Cauchy Problem f o r a Quasilinear Weakly Hyperbolic 179
However, there exist Monge-Ampitre Eq. (1) which do not possess two in-
dependent first integrals. For example, as it has been proved in Ref. 7 that
the following equation
rt - s2 + c2(21, 2 2 , 2 , P l , P2) = 0
has two independent first integrals if and only if
In the case A > 0, but without the assumption on existence of two

independent first integrals, the Eq. (1) had been also considered in Refs. 3,
5-9 by reducing it to a hyperbolic quasilinear system of first-order partial
differential equations with two variables.
In this paper when considering the case A 2 0 we do not assume exis-
tence of two independent first integrals for the Eq. (1). To avoid this dif-
ficulty in Ref. 8 we have proposed another solving method for the Cauchy
problem (3)-(4) which reduces it t o that for the following normal quasilinear
first-order system of equations in two variables
ax,
L aa2
with the following initial conditions
Here ( X I ,X 2 , Z , P I ,P2) are unknown functions of two variables a1, a2

and the functions X y ( a 1 ) , X z ( a l ) ,Z o ( a l ) ,P f ( a l ) ,Pi(c.1) are given as the
same as above, which satisfy the conditions (5), (6).
180 H. T . Ngoan and N. T. Nga
ax, ax,
--
aal aa2
det # 0. (9)
-__
From the implicit function theorem we can locally solve the following system
of equations
X l ( Q 1 , Q2) =21
Xz(a1, a2) =2 2
t o obtain a1 = $ 1 ( ~ 1 , z 2 ) ,a 2 = &?(zl,2 2 ) .
In Ref. 10 we have proved the following theorem.
Theorem 1.1. Suppose that the Cauchy problem (7)-(8), satisfying the
conditions (5), (6), possesses a C2-solution (XI,X 2 , Z , PI,P 2 ) . Then the
condition (9) is satisfied and a local C2-solutionz(x) to the Cauchy problem
(1)-(4) is given by the following f o m u l a r
z(51,22) = Z($l(Zl, 2 2 ) , $2(21, 2 2 ) ) .
Moreever,
Z q ( 2 1 , 2 2 ) = Pl(1cll(Tl, 2 2 ) , $ J 2 ( 2 l 7 X 2 ) ) ,
222 ( 2 1 , 2 2 ) = P2(7f!J1(21,5 2 ) , $ 2 ( 2 1 , 2 2 ) ) .
So, t o study the Cauchy problem (1)-(4) we first concentrate t o investigate

the solvability of the Cauchy problem (7)-(8).To do this in Sec. 2 we recall
definition on hyperbolicity for systems (7) and some results on solvability
of Cauchy problem for them. In Sec. 3 we first reduce the system (7) to
an intermediate one (25). Provided some restrictions on coefficients aij and
initial Cauchy date, the last system will be in Sec. 4 reduced in Theorem
4.1 to a diagonal one (42), the Cauchy prblem for which is well solved
locally according to the Theorem 2.3. In Sec. 5 we describe in Theorem
5.1 a class of weakly hyperbolic classical Monge-AmpBre equations, Cauchy
problem for which is well solved locally. In Sec. 6 we give some examples of
weakly hyperbolic classical Monge-Ampkre equations that satisfy conditions
of Theorem 5.1.
2. Hyperbolicity
From (3) we denote C by a l l , A 1 by a12, A2 by a21 and A by a22. From now
we consider instead of (7) the following quasilinear system in two variables
-
- (a12 -
8x1
1)-
8%
+ a22- ax2
aa
+ ap2
-
aa
8x1 + 1)-ak2 a51
= -all-
aal
- (a21
aa, - -
aa1
8x1
= (a12P1 - U I l P 2 ) Z
+(a22P1 - a21P2)-
ak2 - -
a2
-
- (411022 +
Here (XI, X2,Z, PI,P2) are unknown functions of the variables a1 and (212;
aij are C2-functions of (XI,Xa, 2,
PI, P2).
We set
u = (Xl, x2,z,Pl, p 2 y ,
I
a12 - 1 0
a22 0
-all 0
-a21-l -1
d(U)= a12Pl-allP2 -a21P2 -1 -P2 Pi . (12)
0 -(711a22 +
a12a21 0 a12 -1 --a11
a11a22 - a12a21 0 0 a22 -a21 - 1
We write the system (11) in the matrix form
Now we recall some definitions and results on quasilinear hyperbolic sys-

tems. To do this we will consider the following more general normal system
in two variables
where U and G ( U ) are m-dimensional columm-vectors, H ( U ) is a matrix

of size m x m. If the matrix H ( U ) is diagonal, the system (14) is said to
be a diagonal one.
The Cauchy problem for system (14) consists in looking for U ( a 1 ,a2) E
C 1 such that
where U o ( a l )is a given m-dimensional vector function.
Definition 2.1. (Ref. 4) The system (14) is said to be hyperbolic if for

any U E R" the following conditions hold:
1) All the eigenvalues of the matrix H ( U ) are real;
2) There exists a basis in R" consisting of its corresponding smooth
with respect to U left eigenvectors.
Theorem 2.1. (Ref. 11) If a12 # a21 f o r any ( X I ,X 2 , Z , P I ,P2) E R5,

then the system (14) i s hyperbolic.
Theorem 2.2. (Ref. 4) Suppose that the system (14) i s hyperbolic. T h e n it

can be reduced t o a diagonal system of 2m quasilinear equations with respect
t o 2m unknowns.
Theorem 2.3. (Ref. 4) Suppose the system (14) i s diagonal. T h e n there

exists locally a unique smooth solution to the Cauchy problem (14)-(15).
For the system (11) we do not assume that a12 # a21. This means
that these functions may coincide at some points ( X I ,X 2 , Z , P I ,P2). In
this case only condition 1) in Definition 2.1 is valid and the system (11)
is said to be weakly hyperbolic. We show below in Theorem 5.1 that under
some restrictions on coefficients aij ( X I2,P ) , the weakly hyperbolic system
(11) can be still reduced to a diagonal one of 7 quasilinear equations with
respect to 7 unknowns. From the Theorem 2.3 it follows that there exists
locally a unique smooth solution for the Cauchy problem (7)-(8).
3. Reduced System
Set
Cauchy Problem for a Quasilinear Weakly Hyperbolic 183
where aij are the same functions of X I ,X z , Z , PI,P2 as in (11). Then it is
1;
easy to verify that
0 000
1 0 0 0
C-l(X1, X 2 , Z , PI,P2) = 0 1 0 01 . (17)
all a21 0 1 0
a12 a22 0 0 1
Proposition 3.1. Suppose that m a t r i x d is defined by (12) and
I.
-1 0 0 0 1
- 0 -1 0 -1 0
A= 0 0 -1 -P2 Pl (18)
0 0 0 a12 -a21 -1 0
0 0 0 0 a12 - a21 - 1
Then
C-l AC = A.
Proof. The proof is obvious by matrix multiplication.
Set
That means
x1 =x1,
[-
x 2 =X2,
z =z, (20)
PI = Pl + 2,P ) X 1 +
Ull(X, a21(X,2,P ) X 2 ,
p2 = P2 + a12(X, z,f)X1+ P)X2.
a 2 2 ( X , 2,
We introduce now the following condition for the Cauchy problem (11)-
(8).
(Cl): For the coefficients a i j ( X ,2,P ) and the initial functions
X y ( a 1 ) ,X , ( a l ) Z
, o ( a l ) ,PF(al), f,(a1)the following inequality holds f o r
184 H. T . Ngoan and N. T . Nga
where the derivatives of functions aij are calculated at (X,"(al),X~(a.l),

ZO(al),P?(Ql),
Pi(Q1)).
Proposition 3.2. Suppose that the condition ((71) holds. Then for any
(XI2,P ) E R5, which are suficiently closed to (X0((al),Zo(al),Po( al) ) ,
the following system of equations with respect to P I ,P2
+ Ull(X, z,P)X1 + a21(X,2,P)X2

I
Pl = Pl,
(22)
P
2 + a12(X,2,P ) X l + a22(X,2,PIX2 =p2
has a unique smooth solution
where
and
Proof. Set
= Pl+ W ( X , 2,P ) X l + a21(X,2,P)X2 - Pl,
{
fl
f2 = P2 + a12(X12,P)Xl + a22(X,2,P)X2 - P2.

We have
of =
det -
DP
Since a i j ( X , 2,P ) and X , 2,P are C2-functions, from (21), (24) and from
the implicit function theorem the assertion of the proposition follows. IJ
Theorem 3.1. The vector function U ( a 1 , a z ) satisfies the system (13) i f

and only if the vector function d(a1,a2) satisfies the following system
Here d is defined by (18) and

I -(a12 - a21 - +
L -(a12 - a21 -
aa12
1)-
aal t -(a12 - a21 - +
where in (26) the variables P I ,P2 are replaced respectively by the functions
f(X1, X2,Z, fj, &), g(X1, Xz,Z, PI,&,), which are defined in (23).
Proof.
1) Suppose U is a solution of (13). Since 0 = C-lU we have
u = c0.
Therefore
dU a0 ac -
-=
aa2
c-aa2 + -u,
aa2
au a0 a c -
aal c-aal + -u.
-=
aal
By (13), (27) and (28)
C-
a0 + -Ua c - = A(C-
aG
+ ac -
-U).
aa2 aa2 aa1 aa1
Hence,
a0 a0 ac ac -
-- -
aa2 C-lAC-
aa1
+ C-l(A-aal -)u.
aa2 -
So we have
-1 0 0 0 1
0 -1 0 -1 0
o o -1 -p2 Pl
0 0 Oa12-a21-l 0
0 0 0 0 a12 - a21 - 1
and
23 = C-l(A(C0)-ac - -)a c
aal aa2
- -
-&Lz
aal
-&?.22
8%
000
h
aalL h
aa1
000
- *P2 - *Pl EPZ -2 P l 000 .
-(a12 - a21 - +& - (a12 - a21 - 11% + 2o o o
--(a12 - a21 - I)% + -(a12 - a21 - I)% + o o 0-
4. Diagonalization
It is clear from (18) that the system (25) is still not diagonal. We give now
some sufficient conditions under which the system (25) can be reduced t o
a diagonal quasilinear one.
We introduce now another condition for the system (11).
(C2) : Each function a , j ( X , 2,P ) satisfies the conditions
= 0,
a 1 1 -~a12% = 0, (29)
aazl bat, da,,
a21 ap, a22apz = 0.
We set
Proposition 4.1.
Suppose ( X I ,X 2 , Z , P I ,P2) is a solution of (13). Under conditions (C1)
and (Cz) we have
1)
2) Suppose ( X , p ) are suficiently closed to ( X o , P o ) . Then there exist

smooth functions bij ( X ,p) and cij ( X ,p) such that
from (33), (34) and (11),we have
(a12 -a21 - 1)-

aaij
- -
aaij
aa1 a a 2
aaij 8x1 aaaj ax2 daij aP1 daij d P 2
= (a12 - a 2 1 - l)(-- + -~ f --+ --
8x1 aa1 a x 2 aa1 ap1 aa1 d P 2 aa, 1
aa,j ax1
-
ax1 - 1)-
aa 1 + (1.22-
a x 2
8% +GI
aP2
aaij 3x1 a x 2 aP1

- -[-all-
a x 2 aQ:1 - ((1.21 + 1)-aa1 - 1-aa1
aaij
- 3j7[(-U11(122 + (1.12a21)- 8 x 12 + ( a 1 2 - 1)-3%
ap1
-
a p 2
all-]
aa1
8x1 ap1 (35)
aff + + 1)-]aa1
-
dUij dP2
--[(a11a22 - (1.12021)- - (a21
ap2 1
a22-
8%
- daij aa.. aaaj 8x1
[-a21- + a 1 1 2 - (Ull(1.22 -a12(121)--]-
-
8x1 (9x2 dP2 aa,
aaaj aaaj a x 2
+ [a12-8 x 2 a22- ax1 + ( a 1 1 u 2 2 U12U21)-]-ap1
dUij
- -
aa1
aaij - a22-1-aaij ap1
+ [-a21- aaij + -aP1 8 x 2 aP2 aa1
aaaj ap2
+ [a12-
aaij
dP2
aaij
- -+all-]-.
ax1
aP1 aa1
It follows from (29) and (35) that
daij -
1)-aaij - - 8x1 8P1
(a12 -a21 - - 8 x 2
0.-+o.-+o.-+o.- dP2
=o.
act1 d a 2 aa, aa, aa1 aa1
2)
From (22) we have
ap1 -
- dull
- -_ ax1 - a21- 8 x 2 - -
aa21x2 I aPl
aa1 x1 -all-
_
aa1 aal aa1 aa1 aa1 (36)
ap2 -
-
aa,
- --
da12
aa,
x1 - a12- 8x1 - (1.22-8
aal aa1
x 2 - -dxa22 2
da1
+-
aP2
aal . (37)
Applying (29), (36), (37) to (33) we get
aaij - daij 8x1
daij
+--
aal ax, aal ax2aal
13x2
- u l l - - a ~ l8-x1
+ ~a a( .-. - x laa,, -
aP1 aa1 aal
8 x 2
aa,
aa21
-x2
aa1 + Fl)
+ aa..
~ ( - - xaa12
l - a l a - axl
- aaz2x 2
-- +
aP2 da1 aa1 aa,
a22- a x 2
aa1 F2)
+ ( - =aa11
XI- -x2
aa21 + P1)-daij
M + ( - =X1
aa12 -
aa1 ap1
= 0.-
ax, + 0.-ax2 + ( -
dull
%X1 - -x2
aa21
N
+ P1)daij
aal aa, aQ.1 ap1
Hence,
(38)
Letting i and j vary from 1 to 2 in (38), we obtain the system
(39)
This is a linear system with s,
aa.. . .
Z,J 1 , 2 as unknown functions.
=
Set
1
d=
Elemeritary calculations yield

190 H. T . Ngoan and N . T . Nga
Since (X, p ) are sufficiently closed to (X', Po),from (41) and (21) it follows
that d # 0. The assertion 2) follows now.
We introduce new dependent variables

"
w = (Xl I x2I 2,Pl ,P2 7 P 1 I p2 T .
Theorem 4.1. Assume that the conditions (Cl) and (Cz) hold. Then the
system (25) can be diagonalized, i.e. it may be reduced to the following
diagonal one:
here
-1 0 0 0 0 0
0 -1 0 0 0 0
N
N
0 0 -1 0 0 0
A= 0 0 0 a12 - a 2 1 - 1 0 0
0 0 0 0 a12 - a21 - 1 0
0 0 0 0 0 a12 - a21 - 1 0
0 0 0 0 0 0 a12 - a21 - 1
(*
and
with
I n the representations for F I ( W ) ,F z ( W ) the variables P I ,P2 must be re-

placed by f ( X I ,X 2 , Z , P I , P 2 ) and g ( X 1 , X z , Z , P I , p ~ respectively.
)
P2 -
and
The following result is a direct consquence of the last theorem, theorem

2.3 and proposition 3.2.
Theorem 4.2. Assume that the conditions (C1) and (C2) hold. T h e n there
exists locally a unique C2-solution to the Cuuchy problem (11)-(8).
Cauchy Problem f o r a Quadanear Weakly Hyperbolic 193
5. Application to the Classical Weakly Hyperbolic

Monge-AmpGre Equation
Applying Theorems 4.1 and 1.1 we obtain the following theorem on solv-
ability of the Cauchy problem (1)-(4) for weakly hyperbolic Monge-Amphre
equations.
Theorem 5.1. Suppose that besides the conditions ( 5 ) , ( 6 ) the following

conditions also hold:
1) The coeficients A, B , C, E are C2-functions and do not depend o n z ;
2) A 2 0 ;
5) For all cq E I it holds
where the derivatives of functions A,C,Xl,Xz are calculated at

( x : ( a l ) ,x;(w),
Z O ( a l )p, : ( 4 , m a l ) ) ;
4 ) Each function A ( x ,z , p ) , C ( x ,z , p ) ,Xl(x,z , p ) , X 2 ( 2 , z , p ) is a first in-
tegral for the Eq. (1); i.e., due to the condition 1 ) above, each of them
satisfies the following system
XIT =o,
A- & =O.
dP2
Then there exists locally a smooth solution z ( x ) to the Cauchy problem

(1)-(4).
6. Examples
Example 6.1. (Ref. 5) The coeffiients A , B , C, E are constants with A 2

0. It is easy to see that all the assumptions of Therem 5.1 are satisfied.
From (26) it follows that B = 0. From (25) and (22) we have the following
quasilinear system
194 H. T.Ngoan and N. T . Nga
AX2 - P2)-
aPl + ( 4 x 1-
aa1
with the following initial conditions
where po(al)= ( F : ( a l ) , F;(a1)), and

P,"(a1)= P,"(a1) + CX,"(a1)+ A2X,O(a1)
f$(ai) = J'i(ai) + + AX,O(ai).
AiX,"(ai)
From the two last independent linear equations of (49) with initial con-
ditions (50) it is easy to solve F l ( a 1 , a ~ &(a,, ), 0 2 ) . Then we substitute
them into the first two equations of (49) to obtain independent linear equa-
tions with respect to X l ( a 1 , a g ) and X ~ ( a l , a 2 )At . the last, substituting
& ( a l , a 2 ) , &(a1,a2),X l ( a 1 , a 2 ) , X 2 ( a 1 , a 2 ) , which have been found, into
the third equation of (49) we can solve Z ( a l , a 2 ) . Then by the Theorem
1.1the local solution ~ ( xto ) the Cauchy problem (1)-(4) can be obtained
from the functions X l ( a 1 ,ag),X2(a1,a2) and Z(a1,a2).
Example 6.2. Suppose that ~ ( yt,) is a solution of the Burger equation

Vt + 2121y = 0 ,
which is C2-function in a neighbohood of the point (clP:(O) -
c#;(O), czX,0(0) + c l X Z ( 0 ) ) and satisfies the condition -v,(clP:(al) -
c2P,"(a1),c2X?(a1) + +
Clx,O(al))(C2X,O(Ql) c l X ; ( a l ) ) +1 # 0, for all
a1 E I and where c1, c2 are some given constants, cf c$ > 0. +
Then the Cauchy problem for the Monge-AmpBre equation
rt - s2 + 212(C1ZZ, - C 2 Z Z 2 , c2x1 + ClZ2) =0
+
with A = B = C = 0, E = -v2(c1p1 - c 2 p 2 , c2z1 c 1 z 2 ) , A = 4 V 2 ( C 1 p 1 -
c z p 2 , c 2 z i + cizz), XI = -A2 = ~ ( c l p -
i c 2 p 2 , czz1+ ~ 1 x 2 satisfies
) all con-
ditions of Theorem 5.1.
Example 6.3. Suppose that ~ ( yt,) is a solution of the following equation

Vt + f(V)Vy = 0,
which is C2-function in a neighbourhood of the point
(-c1@(0) - c 2 @ ( 0 ) , c l x m +CZX,o(O))
and satisfies the condition
- f'(v(-clP:(al) - C2P,"(C.l), C l X ? ( C . l ) + CZX20("1))
x vy(-ClP:(al) - C2P,"(C.l),
+
clx;(Ql) C 2 X , o ( a l ) ) ( c l X : ( a l ) + CZX%l)) + 1 # 0,
for all a1 E I and where f ( v ) is a given C2-function of one variable, c1, c 2
are some given constants, c; +
c; > 0. Then the Cauchy problem for the
following equation
f(~(-clZ,I - C2ZZ2, c1z1 + .2.2))(. + t ) + (.t - s2)
+ f2(+C1Z,, - c2zz2, C l Z l + c2z2)) = 0
with A = f(v(-clz,, - c 2 z Z 2 , clzl +CZZZ)), C = f ( ~ ( - c l z , ~ -c2zz2, c ~ z ~ f

, X i = A2 = 0
2 ) ) 0,
~ 2 2 2 ) ) ,B = 0, E = - f 2 ( ~ ( - ~ l ~-,c, ~ z , ~~, 1 ~ 1 + ~ 2 ~ A
satisfies all conditions of the Theorem 5.1.
References
1. G. Darboux, LeGons sur la the'orie ge'ne'rale des surfaces, tome 3 (Gauthier-
Villars, Paris, 1894).
2. E. Goursat, LeGons sur l'inte'gration des e'quations aux de'rive'es partielles du
second ordre, tome 1 (Hermann, Paris, 1896).
3. J. Hadamard, Le problbme de Cauchy et les e'quations aux de'rive'es partielles
line'aires hyperboliques (Hermann, Paris, 1932).
4. B. L. Rodgestvenski, N. N. Yanenko, Quasilinear hyperbolic systems (Nauka,
Moscow, 1978).
5. M. Tsuji, Bull. Sci. Math., 433 (1995).
6 . M. Tsuji and H. T. Ngoan, Integration of hyperbolic Monge-Ampdre equa-
tions, in Proceedings of the Fifth Vietnamese Mathematical Conference
(Hanoi, 1997) pp. 205-212.
7. Ha Tien Ngoan, D. Kong and M. Tsuji, Ann. Scuola Norm. Sup. Pisa C1.
Sci., 27,309 (1998).
8. M. Tsuji and N. D. Thai Son, Acta Math. Viet., 2 7 , 97 (2002).
196 H. T . Ngoan and N . T.Nga
9. H. T. Ngoan and N. T. Nga, O n the Cauchy problem f o r hyperbolic Monge-

Ampdre equations, in Proceedings of the Conference on Partial Digerential
Equations and their Applications (Hanoi, December 27-29, 1999) pp. 77-91.
10. Ha Tien Ngoan and Nguyen Thi Nga, Actu Mathematica Vzetnamica, 29,
281 (2005).
11. Ha Tien Ngoan and Nguyen Thi Nga, Vietnam Journal of Mathematics, 34,
109 (2006).
$11. SOME SINGULAR PERTURBATION PROBLEMS

RELATED TO THE NAVIER-STOKES EQUATIONS
MAKRAM HAMOUDA
Laboratoire d 'Analyse Nurnh-ique, Universite' de Paris-Sud, Orsay, France
ROGER TEMAMx
T h e Institute for Scientific Computing and Applied Mathematics,
Indiana University, Bloomington, IN, USA
E-mail: temam@indiana.edu
In this article, we consider the asymptotic analysis of the solutions of the

Navier-Stokes problem, when the viscosity goes to zero; we consider the flow
in a channel of R3,in the case where the boundary is non-characteristic. More
precisely, a complete asymptotic expansion, a t all orders, is given in the lin-
ear case. For the full nonlinear Navier-Stokes solution, we give a convergence
theorem up to order 1, thus improving and simplifying the results of Ref. 11.
'-
at
- EAUE + (U".V) u~ + v p " = f, in R,,
div u" = 0, in R,,
' u" = (o,o, -u), on r,, (1)
uE is periodic in the x and y directions with periods L1, La,
\ uEl,,o = u o .
198 M . Hamouda and R. T e m a m
Throughout this article we assume that f and u g are given functions as

regular as necessary in the channel R,, and that U is a given > 0 constant;
a t the price of long technicalities, we can also consider the case where U is
nonconstant, U > 0 everywhere.
By setting uE = u"- (0, 0, -U), we obtain the following problem for u":
'
EAU"- UR3v" + (v'.V) ue + Vp" = f,
aVE
-_ in R,,
at
div u" = 0, in R,,
' u'=o, on roo, (2)
uE is periodic in the x and y directions with periods L1, Lz,
\ uEl*=O= uo.
Linearizing in u" we find:
I U" = 0, on r,,
w" is periodic in the z and y directions with periods L1, L z ,
0
ItZO = uo.
(3)
In the linear case, the limit problem ( E = 0) corresponding t o (3) is the Euler
problem those solutions may not verify the boundary conditions (3)s. More
precisely it follows from this work that the limit solution, denoted by u o , is
Singular Perturbation Problems Related to the Navier-Stokes Equations 199
solution (for all time) of the following system:
1 -U +
D ~ P 0p0 = j , in R,,
I~30div u0 = 0, in R,,
I 0
u =O,
= 0, on
on
ro,
rh,
(4)
vo is periodic in the x and y, directions with periods L1, Lz,
where l?o = R2 x (0) and rh = R2 x { h } .

Our aim in this article is to derive asymptotic expansions for u",when the
viscosity E goes to zero, for either the linear case (3) or the nonlinear case
(2).
It is obvious that we can not expect a convergence result of u" to uo in the
uniform or H 1 norm spaces since u" and uo do not have the same traces
on r0.This question was addressed in Ref. 11 in which the authors gave a
representation of the NS solution U" up to the boundary and proved con-
vergence results in several Sobolev spaces.
In this article, we continue this work and we first derive the complete asymp-
totic expansion of the linearized NS solution v" up to the boundary at all
orders. For this purpose, we will use the classical tools of the singular per-
turbation theory based on the use of correctors; see Refs. 4 and 3, or
Refs. 12 and 7 for a related point of view. A similar work is performed in
the nonlinear case but we then restrict ourselves to orders 0 and 1 (and to
limited time), although we believe that higher orders can be treated in the
same way.
This article is organized as follows. In the next section, we show how to
choose and construct the correctors of orders zero and one for the linearized
NS equations. The third section deals with the description of the correc-
tors of order N >_ 2 and the fourth section is devoted to the convergence
theorem which validates our (empirical) choice for the correctors (linear
case). Finally, in the last section (Sec. 5), we treat the nonlinear problem
at orders zero and one and give the adequate correctors by proving their
validity; this section improves and simplifies the results of Ref. 11.
2. Correctors of Order 0 and 1

Our method to construct the corrector of order zero is based on the ob-
servation made in Ref. 11 that, unlike the characteristic case, the Prandtl
equation for this problem is very simple, namely
-cD;O" - UD3P = 0, in R,,
{ div 8" = 0,
8 " = -U 0 ,
in SZ,,
on ro,
8" = 0 , on I'h.
However our method here is quite different than in Ref. 11;the correctors'
(5)
representation that we give hereafter is simpler than in Ref. 11 and this

simplification allows us to continue our analysis at higher orders. We intend
to give an asymptotic expansion of u",p" of the following form:
To determine the asymptotic expansion above we are guided by the general

techniques of boundary layer theory. Hence for the exterior expansion u"
c,"=, &j v j ,
-
p" C,"=,E? p l , we would have
N
avo
-- UD~U+
O Vpo = f,
at (7)
and, for j >_ 1:
The interior asymptotic expansion C,"=,& @+ is determined by the con-
-
dition that @ > = -d on roo,at all orders j and, for the equation we use
the ansatz 0 3 E - ~ which leads to (5) in Ref. 11. Hence, formally
-,cD;(j1," - uD3el," =
I ae03&
-&D$O0~"- UD307"= 0 , (9)
& at
and at orders j 2 2
However this general scheme has to be adapted to the present situation

because of the intricacies introduced by the incompressibility condition (and
the global nature of this equation); hence (7)-(10) will only serve as a general
guideline.
We begin at order 0 by proposing a corrector solving (approximatively)

this system
--ED;eo$c- u ~ ~ e ~on , (0,~h),

= 0,
div go>"= 0, in R,,
(11)
dotE= --you 0 , on r01
eosc= 0, on rh,
where -yo (resp. yh) is the trace on z = 0 (respectively on z = h).

Firstly, we derive the tangential component @+ of e0J,
by solving the sys-
tem composed of the tangential projection of (11)l and (11)3,41and then the
normal component (z-direction) O;E will be given by the incompressibility
condition (11)~.
Thus, an approximation denoted eT of the tangential component O:+
--O,E
up to an e.s.t. (exponentially small term) reads:
Here and in the sequel an e.s.t. is a function those norm in all classical
spaces H m , C' is exponentially small in E.
Notice that $" still satisfies equations (11)l and ( 1 1 ) ~

and
~ has an
exponentially small value at z = h:
i$!'E = O(e-uh/E), at z = h. (13)

Then, thanks to (11)2,we write
h
0': d i V T e E ( Z ly, c, t)d<,
and we deduce easily an approximation of QE up to an e.s.t.,
We note that this approximate corrector e"'" = (@, 82&)verifies the in-
--O&
compressibility condition (11)2 and the equation (11)l.However -yoen =
O ( E )and y h e E = O(e-Uh/E),these boundary values being different than
-o
the desired ones, namely en' = 0 at z = 0 , h.
-0,E
To correct the boundary conditions of 6 , we introduce an additional cor-
rector denoted by p0+ = &(pol which is solution of the following system:
I -Avo + V7ro = 0,
div po = 0, in R,,
(PO, = 0, on rou rh,
1 -0,E
p: = - - p O ,
p: = o , on rh.
1
in R,,
= -div,(yov:),
U on I'o,
Remark 2.1. The role of the corrector po>'is just to account for an O(E)-
error in the boundary value of 8" and this "error" in the boundary con-
dition will be actually corrected at the next order. There are also some
boundary value "errors" of order O(&-ae-uh/E); they will be corrected all
(15)
at once later on (at the final order N ) .
We now proceed to the first order. The function d appearing in (6), is

required to satisfy theses equations and boundary conditions:
UD3v1 + Vpl = --avo

at
+ UDypo + Avo, in R,,
div v 1 = 0, in R,,
{
v31 = 0, on ro,
v1 = O , on F h ,
,v1 is periodic in x and y.
on rourh.
As at order zero, we instead construct an approximation for 0'1' as follows:
@IE = +
e.s.t., with a tangential component
and, for the normal component,
As before $IE satisfies the desired boundary value exactly at z = 0 and up

-1 ,
to an e.s.t. at z = h, whereas On satisfies the desired boundary value up
to order O(e) at z = 0 and up to an e.s.t. at t = h. To correct the O(e)
error we introduce an additional corrector pl+ = up1 verifying
?)1 + # Y E
+
+
I -Aql
=
+
&
V d = 0, in a,
div cpl = 0, in a,
cp: = 0, on rourh,
1 -1,s
vn = 0, on I'h.
1
= -b;, on Fo,
Finally let us summarize what we obtained at order 1. We set G1+ =

pl>'. Thanks to (16), and (18)-(20), the functions G1+ and
pl' = p 1 7r1 satisfy the following set of equations and boundary conditions
(20)
3. The Corrector of Order N ( N 2 2)

In this section, we will focus on the N t h corrector for all integers N 2 2.
Starting with order N = 2, the equations for the vj and Oj?" are different
as observed in (8) and (10). More precisely for N 2 2, we propose to define
the tangential component 6;' of by O N i E
on roUrh,
(22)
where vN verifies for all N 2 1 the following form of the system (8):
' -dVN
- UD3vN +V p N
at
- -~
at
+ UD3pN-l + A v N - ' , in R,
div uN = 0, in 52, (23)
vf = o , on ro,
vN = o , on rh,
uN is periodic in x and y.
As for N = 0,1, the p N to be subsequently defined are introduced as cor-

rectors of the boundary conditions.
To solve system (22), we need a sequence of lemmas.
3.1. Preliminary results

The first two lemmas are elementary, they will be helpful for computing
the correctors.
Lemma 3.1. We have, for any j E N,
Lemma 3.2. A particular solution of the differential equation

is given by
Both lemmas are proved by induction on j. Based on these lemmas, the

next lemmas provide the explicit form of the corrector ON?&. We start by
describing the tangential component and the normal component is then
deduced from the incompressibility condition.
Lemma 3.3. A n approximation up to exponential order of the tangential

component OFEof the corrector solution of (22), is given as follows:
O N Y E ,
Proof. By induction on N . For N = 2, ( 2 7 ) is verified thanks to Lemma

3.2 and the explicit form of the first order corrector 8;" which was obtained
in (18).
The tangential component O?+llE satisfies the system ( 2 2 ) those first equa-
tion can be rewritten as follows:
Using Lemma 3.2, the resolution of this equation, taking into account the
corresponding boundary condition as given by (22)3, yields:
- N + ~ , E-
97- - +
--You,N+1 e - U ~ / ~ PS, (30)
where P S denotes a particular solution of (29); based on Lemma 3.2, its

expression reads:
at
d
p s = --(a;) cN
j=O
&J+1 ( j
.j+l
+ I)! U 2 N + 1 - j
---UZ/E-
By identification of (30)-(31) and by the general form of the corrector

-N+~,E .
8, , a.e.,
we deduce that
Finally, after some simple transformations of (31) and by identification we

obtain (28). This concludes the proof of Lemma 3.3. 0
Thanks t o the incompressibility condition and t o Lemma 3.1, we have
Lemma 3.4. A n approximation of the normal component of the corrector

reads:
elv+
where
Remark 3.1. For all N E N and 0 5 i 5 N , the quantities a?, b? are

functions o f t , x and y, but they are independent of z and E . Thanks to the
-N,E
incompressibility condition (22)2, the corrector 8 still satisfies the same
equation as given by (22)l. More precisely we have

-N-~,E
-1E { AS^-'^^},
I
- UD3gN'"= - on ( o , h ) ,
div 3"' = 0, in 0,
-N,E & N
-N,E
8,
N
= --You, 1 8, = ubo, on ro,
- o(&-Ne-Uh/E - O(&-N+le-Uh/E
-NiE
07- - ) I
-NIE
0, - >, on rh.
(33)
Finally, to recover the desired boundary conditions, we introduce an addi-
tional corrector, pN>e= &pN, rN+= & r N , defined by
-ApN +
V r N= 0 , in 0,
div pN = 0, in 0,
p,N =o, on rourh,
N 1 -N,E (34)
pn = - - ~ ~ e , = - b f , on ro,
&
p,N = o , on rh.
Notice that pN is independent of E.
We now define GN+ = uN

gN,E
+
p N + and p" = p N
+
+ 7rNiE which
satisfy the following equations and boundary conditions:
4. Convergence Result
In this section, we will validate our choice for the correctors by proving a
convergence result which concludes the study of the linear Navier-Stokes
problem defined in Sec. 1.
We start by introducing the following expression:
k=O
(36)
k=O
The correctors that we introduced contained exponentially small errors a t
the boundary z = h. To correct all these "errors", we now introduce an
additional corrector which will account for all the e.s.t. a t z = h. This
corrector, denoted by FN?' , satisfies the following system of equations:
-& A F N , " + v ~ N ~ -E 0 , in 0,
div FNYE = 0, in 0,
$1" = 0, on ro,
Note that by the classical results concerning the Stokes problem (see for
instance"), all the Sobolev norms of FN>"are exponentially small, namely
0 ( & - 2 N e-Uh/E
).
We then define new quantities, which will allow us t o prove the conver-
gence result and give a complete asymptotic expansion of the linearized NS
solution II", namely
,
, i j j N , ~= w N , ~+ & N @ N , E
(38)
PE = p~ + E N ~ N V E .
Thanks t o (3), (23), to our iterative construction for the correctors as de-
scribed by the system (22) (with N replaced by k), (34) and (37), the
equation verified by the pair (GNIE, F") reads:
Singular Perturbation Problems Related t o the Navier-Stokes Equations 209
Using equations (3)2, (22)2, (23)2, (34)2 and (37)2, we deduce the incom-
pressibility condition
div G N i E= 0. (40)
Moreover, our construction leads to the desired boundary conditions for
GN+,namely the homogenous ones: from (3)3, (22)3, (23)3,4, (34)3--5 and
(37)3-6, we conclude that
= 0, on rou rh. (41)

Also the periodicity condition is still satisfied for GN+:
GN3&is periodic in the x and y directions with periods L1, La. (42)
Finally, we state and prove the following theorem:
Theorem 4.1. For each N 1 1, the quantity GNvEdefined b y (38) tends

t o zero when E -+ 0. More precisely, we have the estimates:
c
IIGNIEIILm(O,T;LZ(n))I E N + 1 , (43)
<
-
ll~NEllL2(0,T;H(s2)) c EN+1/2, (44)
where L(C2) = (L2(R))3,H1(C2)= (H1(R))3, and C denotes a constant
which depends o n the data (and N ) but not o n E .
Since the (PN,& are exponentially small, the same results hold f o r wNJ.
Proof. We multiply (39) by G N } & and integrate over R. Thanks to (40),

(41) and (42) we see that the last two terms in the left-hand side of the
resulting equation vanish. Also all boundary terms vanish and we find:
I d
--ll#YE
2 dt
2
llLz(s2) +EIIVW -NE 2
llLz(s2) =E ~ + ~ +I E? N N
+ E
N N
, (45)
where
dPN
- + AvN + ATgN&
+ UD3pN]. GNiEdR,
at
Now, we proceed to estimate the terms I:, I:, I:. Using the Cauchy-
Schwarz inequality, we easily obtain that
I&N+lI?I + ;IIVGq;2(n).
I C&2N+2 (46)
210 M . Hamouda and R. Temam
Thanks to the explicit expression of our corrector e N I E , which is given by

Lemma 3.3 and Lemma 3.4, we easily see that
pTeN'EIILa(n)
5 cE1/2.
Using equation (37) for FN>' and the classical results on the Stokes problem,
we deduce that
I E ~ I 5~ CE-'
I e-Uh/E+ II~N%z(s2). (47)
We recall here that in order to obtain an estimate for the term aFNJ/&,
it suffices to differentiate the system (37) in time t , multiply by
and integrate over 0.
Finally, we deal with the most difficult term I?. For that purpose, we use
again the explicit expression of the corrector e"", for all N E N. Thus, we
have
This yields
5 (Thanks to the Hardy inequality)
+ 4((VG""lI;z(n).
5 C&2N+2
&
(49)
Combining (46), (47), and (49), we arrive at the following energy inequality
Applying the Gronwall inequality to (50), we obtain (43) and (44). This
concludes the proof of Theorem 4.1.
5. The Nonlinear Case

This section is aimed at deriving a similar asymptotic expansion of the
solutions of the full nonlinear Navier-Stokes equations; we believe that all
orders can be handled in the same way, but because of the complexity of the
calculations we restrict ourselves to order 1 thus recovering and improving
Singular Perturbation Problems Related to the Navier-Stokes Equations 21 1
the results of Ref. 11.
Proceeding as in the linear case, we consider v E = u" - (0,0, -U) which

is solution of the following system:
Ifi=o, EAV"- UD3v" + (v".V) IF + Vp" = f,
in ,a,
in R,,
IvE=O,
'JI
on ,,?I
is periodic in the x and y directions with periods

As in Ref. 11 our analysis shows that the leading terms remain the same
L1, L2.
as in the linear case and consequently the zeroth order corrector that we
propose here is the same as in Sec. 2. Of course we should treat the non-
linear term appropriately. For that reason, and besides its importance, the
derivation of a convergence rcsult at ordcr zero will bc helpful for the choice
of the corrector a t order one.
5.1. Corrector of order zero

Let
&"= v" - vo - p - Epo - p+ (52)
where, e"'",pa,p?" are respectively given by (12)-(14), (15), (37) with N =
0, and vo is now solution of the full nonlinear Euler equation
v30 = 0 , on r0, (53)
v0 = 0 , on Fh,
vo is periodic in the x and y, directions with periods L1, L2

Remark 5.1. The existence of a smooth solution for the system (53) is
proved in Ref. 8.
Thus, we deduce the following result:
Theorem 5.1. There exists a time T, > 0 such that we have

-0,'
1I.U" - V0 - f3 IILm(O,T*;Lyn)) 5 &,
where K is a constant independent of E.
Remark 5.2.
(1) The complete proof of Theorem 5.1 is similar to that of the correspond-
ing theorem at order one (that is Theorem 5.2), so we skip it. However,
it is helpful for the subsequent order to make explicit the form of the
nonlinear term. Using (52) we see that
(VE.0) V E = (uO.0) vo + ( u E . 0 ) wO," + (wO+.O) (21" - w y +

+ (P1'.O) 80'" + (v0.V) + ( 8 O " . 0 ) vo+
P"
+ & ( ( p O . 0 ) 80'" + & (eO'".V) (po + & (v0.0)PO+ (56)
+ & ( ( p O . 0 ) vo + E2 ( ( p O . 0 ) 'PO + (8O".V) p"+
+ (vO.0) p"+ & ((pO.0) p++ ( p @ . V ) p++
+ ( p ' . V ) uo + ( p E . 0 ) 80>' + & ( p . 0 )9 0 .
(2) In order to obtain (54)-(55), it is sufficient to obtain the same esti-
mates for wo+. Indeed, using the equations for (po and p?'which are
given respectively by (15) and (37) with N = 0, we have the following
estimates:
I
~ ~ ~ ( p O l l L z ( n C,
) I I V ~ ~ ~ I I L & Ic
~ -(2 ~N -) 1 / 2 e - U h '/ E (57)
5.2. Corrector of order one

In this section, we are interested in the derivation of an asymptotic repre-
sentation of v" up to order 1, that is t o find an approximation of 21" in R,
and up t o the boundary as follows:
vE vo + eo+ + & P O + &(d+ el+) + iz1>&,

p' N po +&PI.
First, we infer from (56) the equations satisfied by the mode w1 at order
one. These are described by the following system:
' -"l - UD3d + ( d . V ) + ( w O . 0 ) w1 + Vpl

at
- a(Po
~- +
UD3po - ( w O . 0 ) cpo - ( ( P O . ~ ) WO + AwO, in Cl,,
at
div w1 = 0, in R,,
< (58)
w31 = 0, on r0,
w1 = 0, on rh,
wl is periodic in the x and y, directions with periods L1, L2,
Remark 5.3. The existence,uniqueness and regularity of the solutions of

system (58) is proved in the Appendix.
5.2.1. Construction of the corrector 8'9'

We now present the equations that we propose for by considering the
next dominant terms in the nonlinear NS equation (51)i (of course without
the source term and the pressure already taken into account). Hence, we
obtain
1 4 E 4 E
-cp!D3e0,'E- -(8 ' .V) 8: , in R,, (59)
&
div 81iE= 0, in R,,
0I '&= - w 1 , on r O u r h .
Remark 5.4. Contrary to the linear case, it is important here t o empha-

size that equation (59)1 is valid only for the tangential component of the
corrector The equation satisfied by the normal component 82&is dif-
ferent and it will be derived later on. It will be slightly different than for
the tangential corrector equation since the nonlinear term is not divergence
free.
We rewrite equation (59)1 using the explicit form of go'&:
Thus, thanks to Lemma 3.2, we have an approximation, up to an e.s.t., of

the exact solution of (59)-(60) which is denoted, as in the linear case, by
-1,E
OT and given as follows:
Remark 5.5. 1. The approximation "3: of O:+ satisfies exactly the same
equation (60) as @:YE.
2. It is easy to observe, but important to mention, that the boundary values

of '3
: satisfy
where c > 0 is constant depending on the data but independent of E and z ;

(62) will be useful later on.
Using the incompressibility condition ( 5 9 ) ~and Lemma 3.1, we deduce

an explicit approximation g.',; of OkE.Thus, we have
-1,E
I roe, 15 c E , on rol
en'
-1e
= O(Ee-UhIE)I on rh.
(64)
Due to our choice of a simpler form for the corrector gl", we observe that,
in ( 6 2 ) ~and (64)2,3, we need to introduce an additional corrector in order
t o obtain the desired homogenous boundary conditions. For that purpose,
we define (Cpl+, Z1iE)as the solutions of the following system:
I -&
div
-
'p,
9
-
-";
'p;c
AG1,&+vZ1"=0, in R,
1,E
- 0,
-
=
== -&
in R,
'p;>" = 0, on ro,
- - -&-l
-1,s
-1
T h-o,&
e , - yhi$'
= O(E), on
-0,E
- rh~:E
=o ( &e-Uh/E),
r0,
-~
- O(e-uh/e),
By the properties of the Stokes operators, all the Sobolev norms of

of order E.
on
on
rh.
rh, (65)
FlyEare
5.2.2. Convergeme result: th,e main theorem

Finally, we state and prove the main result of this article which gives the
asymptotic expansion at order one of vE:
Theorem 5.2. For v E solution of the Navier-Stokes problem (51), there

exist a time T, > 0 and a constant K > 0 depending o n the data but not o n
E , such that:
Proof. Let
Before writing the equations satisfied by w ' ~ ~we, start by expanding the
nonlinear term:
(v'.V) WE = (vE.0) w1+ + (W'?V) (WE+ (v0.V) vo+

-w'q
4
+ E [(v".a) (Po + ((P0.V)vo + ( V O . 0 ) v' + ( v l . 0 )
+ (PE.0) 80" + & ((P0.V)8'"+ (80'&.V)210 + (vO.0) P'+
+ E (eo'".v) ((PO + v') + & (e"".v)(VO + & v')+
+ E [(vo + 8'' + ap0+ E v1 + ~8""+ E ( P ~ ) ~P1++ ).V] (68)
+ E [(vO + 80+ + &(PO + & + &P+ &cpl+).V]el,'+
211
+E + 21' + P ) . V ) e"?'+ & (8l'".O) (PO+

((@'I&
+ E (cpl>".V)(?I0 + & w1 + & (PO).

1
+ E2 [(@.V) (Po + ((P0.V) v' + ( V l . 0 ) (Po + (7Jl.V) v' +
Thanks to (51)1, (53)1, (9), (15)1,(58)~, (60)-(64)1 and (65)1, we deduce the
main equation verified by wl,&for which the boundary conditions are de-
rived from (51)3,4, (53)3,4, (15)3--5,(58)3,4, (62)-(64)2,3 and (65)3-6. Notice
that both the incompressibility condition and the periodicity property are
' --&Awl,E- UD3W1,E + V [ p " - (PO + T O ) - &(PI + ?+E)]+

at
+(v".V) wllE + (wl?".V) (v" - ~ ' 2 " )= -x$
j=O
2
Jj>",
{
div w'," = 0,
wl,&= 0 , on rowh,
, wljh is periodic in the 2 and y directions with periods L1, L2,
and
1
5 2 , E = -(
E
-+ UD3(pl,E) + Av' +
at
+ - ( ( p l ? E .V) (wO
&
+ E v1 +&YO)
+[(vO+8'E+&(pO+&w1 + &P
+ &(p'q.v] (p'+ + &2(p'+D381'"
+[ ((PO.0) Po + ((pO.0) v1
Remark 5.6. The underlined terms in J o ? &and

+ (vl.V) (Po + ( w l . 0 ) v' .
1
J1>&need to be split as
(72)
follows before estimating them later on:

21 + u3 D3v.
(2L.V) 11 = (7L7.VT)
Now, multiplying equation (69)' by ~ ' and integrating over 0, we
2 ~
obtain
218 M. Hamouda and R. T e m a m
where,
Thanks to the incompressibility, boundary and periodicity conditions

((69)2,3,4and (51)2,3,4),we deduce that the following terms vanish:
k3i' =- U D ~ W. ~wli"dR
~ ' = 0,
k'" .I = v [ p c - (po + TO) - ~(p'+ ? l " ) ] . w l i E d R= 0, (74)
b 5 7 & = L[(v'.V) w17"dR = 0.
Thanks to the Hardy inequality, we have
IJ6, ' l = I 1
(wli".V) (v' - wl,") . wl.EdRl
s,
n
0 ).~~~~]dRl+
51 [(w:yc.V,) (wE - w1iE).w13E W ? ~ D ~ (-W+"
w'," -
-0,"
w~ED3~'EW1>Edfi
+I&
I II v, (vE - w l , & )+ D3(vE- wile - 0 ' ) + D3eEIIL-(n) 11w1'Ell~2(n)+
Icllw
l a 2
llLZ(n) + 1I;u
2
2 -Uz/e
e IlL-(O,h)IIYO~~IIL~,IIV
wl,E 2
llLZ(n). (75)
Finally, for the term J 6 l E we impose the following condition:

U 0 4E 0 &
- 1Iz2 e--uz/"
&
IIL~(O,~)~~YO~,~IL~~ I ~ I l ~ n v , I I q ,I -.6 (76)
Since yo$ vanishes at t = 0, IIyov;ll~~,remains initially small and there-

fore there exists a time T, > 0 such that
Ue2
Ilro~%& 5 24' v t E [O,T*]. (77)
Consequently, we obtain
l a 2 E l a 2
lJ6,&l 5 CIIW ' IILyn) + ,IlVW IILz(n). (78)
We now estimate the right hand side of (73). We start by the easier terms
included in J2,&. First, by the properties of the solutions of the Stokes
Singular PeTtUTbatiOn Problems Related to the Navier-Stokes Equations 219
problem, the solutions of (65) satisfy:

,,-1,E
IIatIILz(n) Ic E , IIV,-lqILz(n) Ic E . (79)
The same estimates are also valid for the tangential derivatives of ( p l l E .
Hence, thanks to the Cauchy-Schwarz and Holder inequalities, we have:

1&2 J 2 q I /E4 + c IIW1q&q. (80)
For J 1 ) E , we use the Hardy inequality * combined with the Cauchy-Schwarz
and Holder inequalities and we conclude that:
Finally, since v: = v: = 0 on z = 0, we have:
(thanks t o the explicit expression of & given by (61) - (63))

5 c E52 p w l q L z ( n ) Ic &4 +6
&
-IIvW.E1l;z(n). (82)
Hence, inserting all the results (74), (78), (SO), (81) and (82) in (73), we
deduce the following new energy estimate:
~dtl l w E I I E z ( n ) + E I I VW l e IILZ(n)
2
5 I6.E4+ c IIW
l a 2
IILz(n), (83)
to which we apply the Gronwall inequality to conclude the proof of Theorem
5.2. 0
W
*Hardys inequality states that ll-llL~(n)
z
<211Vwll~z(n)provided w ( z = 0) = 0.
220 M . Hamouda and R. Temam
Appendix
In this paragraph, we prove the existence and regularity of the solutions

of the linearized Euler systems (4), (16) and (23). For the full nonlinear
Euler system (53), we refer the reader to Ref. 8.
We first want to apply the Hille-Yosida theorem to prove the existence
and uniqueness of the solution of (4). Thus, we start by introducing the
adequate function spaces
{
H = u E L2(s2); div u = 0, ullz=o = u11,=L,,
= W Z ~ , = L and
u~),=o
dU
~, w3 = 0 en z = 0, h
1,
{
D ( A ) = w E H ; z-d z E L2(n), u = 0 at z =h and
3 p E D'(R) such that - U -
Then for u E D ( A ) , we set Au = -UD3v

aU
dZ
+V pE H
1 .
+ V p , thus defining an unbounded
linear operator A : D ( A ) c H -+ H .
Remark 0.1. If s o p dR = 0 , then p is unique, otherwise it is unique up

to an additive constant.
Let us now recall the statement of the Hille-Yosida theorem (for more details
see for example Refs. 1,2,5 and 13):
Theorem 0.3. (Hille-Yosida Theorem). Let H be a Hilbert space and

let B : D ( B ) 4 H be a linear unbounded operator, with domain D ( B ) c H .
Assume the following:
( i ) D ( B ) is dense in H and B i s closed,
(ii) B i s a positive operator,
+
(iii) 3 po > 0 , such that B p o l i s onto.
T h e n ( - B ) is the infinitesimal generator of a semigroup of contractions
{ S ( t ) } t l o in H . More precisely, the solution u of the system:
satisfies the following properties:

(Ho): If uo E H and f E L1(O,T ;H ) then u E C ( [ 0 ,TI;H ) , 'd T > 0,
and,
( H I ) : If vo E D ( B ) and f' E L 1 ( O , T ; H ) then v E C ( [ O , T ] ; Hn
)
L"([O,T];D ( B ) ) ,and E L " ( [ O , T ] ; H ) ,'d T > 0.
In order t o apply the Hille-Yosida theorem t o the system (4) with B = A,

it is sufficient to prove the following proposition:
Proposition 0.1. The operator A verifies
+
(i) A XI is onto Q' ' X > 0 ,
+
(ii) ( A X I ) - ' E B ( H , D ( A ) ) , 'd X # {Ulkl, k E Z2},
(iii) A is closed.
Proof. The proof of the property (i) is equivalent t o finding a solution for
the following system:
+ +
-UD3u V p Xu = f , in R,,
1 div u = 0, in R,,
u3 = 0 , on ro,
u=O, on r h ,
u is periodic in the x and y, directions with periods L1, L z ,
where f is a given function of H .
Now, we decompose the solution of (84) in the Fourier basis as follows:
(84)
U = c
k=(ki ,kz)EZ2
't&(z)eklZ+k2Y,
and let
k=( k i , k z ) EZ2
Hence, the system (84) is equivalent to the following system of equations:

We multiply (85)1 by i k l , (85)2 by ik2 and we differentiate (85)3 with

respect to z , then we sum up the equations that we obtained and find:
-lkI2pk ;+; = divf = 0.
Since & ( h ) = 0 (from (85)3 and ( 8 5 ) ~ we

) have
where (Yk is a constant to be determined later on.
Remark 0.2. The existence of a k is guaranteed by (85)3.

Indeed, we multiply (85)3 by (-6e-6') and integrate the resulting equa-
tion on (0, h);we deduce
ak =
6e-+'f3kdz (87)
1 1 $ [e" k I- 6) z - e- (1 k 1-k 6) Ze21k 1h]dz '
for all Ikl # 0. If k = (0'0) then the pressure po reads
h
pO(z) = f30(z)dz.
From (87) we deduce the following estimate for a k :
(89)
where X is chosen such that X 6 {Ulkl, k E Z2}.
Lemma 0.1. We have p E L2(R).
Proof. By definition of p we have
There exits a constant C such that

Singular Perturbation Problems Related to the Nauier-Stokes Equations 223
Hence, we conclude that
This ends the proof of Lemma 0.1. 0
Lemma 0.2. The solution u of (84) satisfies:

u E L2(f2) and Z&U E L'(f2). (94)
Proof. We multiply (85)3 by (-$e-"/') and integrate on ( 0 , h ) ;we

deduce the following expression of U3k:
Thanks to (86) we have
I(thanks to (89))
I
I
Hence u3 E L2(R).
Now, since u1 and '112 satisfy similar equations ((85)i and ( 8 5 ) ~ it
) is suffi-
cient t o prove for example that u1 E L2(R). For that purpose, we multiply
(85)l by (-6 e-"/') and integrate on (0, h ) , and we obtain the expression
Of 'Ulk:
Thus, we infer from (86) that
(98)
where
224 M. Hamouda and R. Temam
Thanks to (89) and for llcl large, we deduce the existence of a positive
constant C > 0 such that:
Combining (98) and (99), we obtain
We conclude then that u E L2(!2).
Now, since we have the explicit expressions of u and p (given respectively

by (97)-(95) and (86)), we prove exactly as below that z D 3 ~ 1 , z D g ~Ez
L2(Q) and by the incompressibility condition (85)4, we deduce that
zD3u3 E L2(Q).
This ends the proof of Lemma 0.2.
Consequently, we have proved that for every f E H there exists u E

+
D ( A ) such that ( A XI)(u) = f. This proves ( i ) . The property (ii) is
obvious. To prove (iii),we consider a sequence (u,), c D ( A ) such that
urn+ u and Aum + cp in H . Thanks to (ii)we have
u, = +
( A X I ) - l ( A + XI)um -+m++m ( A+ XI)-l(cp + Xu). (101)
By the uniqueness of the limit, we conclude that
( A + XI)(u) = cp + Xu * u E D ( A ) and Au = cp. (102)
This proves that A is closed and consequently ends the proof of the Propo-
sition 0.1.
Theorem 0.4. Under the following hypotheses:
(H2) : v o ~ C ~ ( a , ) n D ( A ) and f ~ C ~ ( [ 0 , T ] x f i , ) n H ,
the solution ( v , p ) of the s y s t e m (4) satisfies:
w,p E C,([O,T] x L). (103)
Remark 0.3. The hypothesis (H2) on f can be replaced by f E
C - ( [ O , T ] x Gm). Indeed, we decompose f as follows:
f =PHf +PH'f,
where PH is the Leray-Hopf projector on H . Thus, only the pressure p is
changed if we replace f by P H f. We define a regular function 8 such that
PH1f = V8, and then p is just replaced by p - 8.
Let us state and prove the following Lemma in order to prove Theorem
0.4:
Lemma 0.3. For T E {z, y} and o E {T, t } , we have:

(2) I.@ belongs to the space Cm(2,)nH (respectiwelyCCY([0,T]x;2,)nH
or C ,(am) n D ( A ) ) then dQ,/da belongs to the same space,
(22) I . @ E Cm([O,T] x1 ,
) nL1(O,T;H ) then 8 @ / dtoo. ~
Proof. First, we notice that the periodicity is conserved after differentia-

tion with respect to x,y, z or t.
To prove (i) it suffices to verify that: if Q, E Cm(fiz,) n H (resp.
C"(2,) n D ( A ) ) then 33/80 E H (resp. D ( A ) ) .
Hence, let Q, E C"(2,) n H ; by using the definition of H , we have easily
8@/80E H since we differentiate the boundary conditions only in the tan-
gential direction on the boundaries z = 0 and z = h.
Now, let Q, E C "(G,)nD(A). Then, in particular Q, E H and consequently
aQ,/da E H . Moreover, since Q, = 0 on z = h, and we have = 0 on
z = h. From the condition @ E D ( A ) ,we deduce the existence of a regular
function p ( V p = ( I - P H ) ( U D ~ V such
) ) that -UD3v +
V p E H . This
+
implies that -UD3(8@/80) V(8p/80)E H , and thus a@/&
E D(A).
The property (ii) is a consequence of (2). Indeed, if Q, E C"([O, T] x 2,) n
L1(0, T ;H ) , then in particular Q, E H for almost all t E [0, TI. Thanks t o
(2) in Lemma 0.3, we have 8@/80E H for almost all t E [0,TI.
This concludes the proof of Lemma 0.3. 0
Now, we are able to prove Theorem 0.4:
Proof. of Theorem 0.4 We recall that under the hypothesis ( H z ) ,the Hille-
Yosida (Theorem 0.3) and more precisely ( H I )yields:
dv
v E C([O, TI;L2(0)) and E Lm(O, T ;L 2 ( 0 ) ) , V T > 0. (104)
Thanks t o Lemma 0.3, the linearity of the system (4) and after replacing
the source function f and the initial data vo by their tangential derivatives,
we infer that
amV
-E
a?-" C([O,T];L~(O)), v T E {z,y), v m E N.
On the other hand, since wo E D ( A ) and f is continuous a t t = 0, then we

deduce that (dv/dt)(t = 0) E H . Now, if we differentiate the system (4)
with respect to t, we obtain:
av
- E C([O, TI;LZ(0)).
at
From the incompressibility condition (4)2,we deduce that
D3213 = -01211 - 0 2 2 1 2 E C([O, TI;L2(s2)). (107)
Using (106)and (107),the third component of the equation (4)1 yields:
2
az
E C ( [ O , T ] ;P(s2)). (108)
Again, the invariance of the system (4)by differentiation with respect to T
gives:
LtZa7-m
E ~ ( [ oTI;
, L2(a)) \J 7 E {z, y), v m E N. (109)
In particular, we have
g,g E C ( [ O , T ] ;L2(s2)).
Hence, equation (4)1 projected on the tangential directions z and y implies

with (106)that:
03211, 03212 E L2(a))*

C([O,T]; (111)
Gathering (105),(107)and (1ll), we conclude the following regularity for
21:
IJ E C([O,T];H1(S1)). (114
Now, we reiterate the procedure; from (112) and the invariance of (4) by
differentiation in time t , we obtain
821
i.e.,
- E C([O,T];H1(S1)) w E C1([0,T];H1(0)). (113)
at
The invariance of (4)by differentiation with respect to 3: and y implies that:
03213 = -01211 - 0 2 2 1 2 E C1([0,TI;Hl(s2)). (114)
Then equation (4)l and (113) imply that
9
dz
E C ( [ O , T ] ;Hl(s2)). (115)
Thus,
and consequently,
w E C([O, TI;H2(C2)).
Finally, t h e invariance by differentiation in time gives
w E C1([0,T];H2(C2)).
This procedure allows us t o prove a higher regularity for w and more pre-
cisely (103).
This concludes the proof of Theorem 0.4. 0
Acknowledgements
This work was partially supported by the National Science Foundation
under t h e grant NSF-DMS-0305110, and by the Research Fund of Indiana
University.
References
1. H. BrBzis, Ope'rateurs maximaux monotones et semi-groupes de contractions
dans les espaces de Hilbert (North-Holland Publishing Co., Amesterdam,
1973).
2. N. Burq and P. GBrard, Contr6le optimal des e'quations aux de'rive'espartielles
(Ecole Polytechnique, Palaiseau, France, 2003).
3. K. 0. F'riedrichs, The mathematical strucure of the boundary layer problem
in Fluid Dynamics, eds., R. von Mises and K.O. F'riedrichs, Brown Univ.,
Providence, RI (reprinted by Springer-Verlag, New York, 1971). pp. 171-174.
4. J. L. Lions,Lectures Notes i n Math, 323 Springer-Verlag, New York, (1973).
5. J . L. Lions, Les Presses de l'universite' de Montre'al (Montreal, Que., 1965).
6. J. L. Lions, Selected work, Vol 1. (EDS Sciences, Paris, 2003).
7. R. E. O'Malley, Communications of the Mathematical Institute, Rijksuniver-
siteit Utrecht, 5. Rijksuniversiteit Utrecht (Mathematical Institute, Utrecht,
1977).
8. M. Petcu, Advances in Differential Equations (to appear).
9. L. Prandtl, Veber Fliissigkeiten bei sehr kleiner Reibung, in Verh. 111 Intern.
Math. Kongr. Heidelberg, (Tuber, Leipzig, 1905) pp. 484-491.
10. R. Temam, Navier-Stokes Equations (North-Holland Pub. Company, 1977),
Reedition in the AMS-Chelsea Series, AMS (Providence, 2001).
11. R. Temam and X. Wang,J. Differential Equations 179, 647 (2002).
12. M. I. Vishik and L. A. Lyusternik,Uspekki Mat. Nauk, 12,3 (1957).
13. K. Yosida, Functional analysis, (Springer-Verlag, Berlin, 6th edition, 1980).
Chapter I11
GEOMETRIC ANALYSIS
12. MONOTONE INVARIANTS AND EMBEDDINGS OF

STATISTICAL MANIFOLDS
LE HONG VAN *
Max-Planck-Institute f o r Mathein Sciences
InselstraJe 22-26
0-04103Leipzig
E-mail: hvle@mis.mpg.de
In this note we prove certain necessary and sufficient conditions for the ex-
istence of an embedding of statistical manifolds. In particular, we prove that
any smooth (Cl resp.) statistical manifold can be embedded into the space of
probability measures on a finite set. As a result, we get positive answers t o the
Lauritzen question on a realization of smooth (C' resp.) statistical manifolds
as statistical models.
Keywords: Fisher metric, Chentsov-Amari connections, statistical manifolds,

statistical models.
1. Introduction
A statistical model is a family A4 of probability measures on a measur-
able space 0. There are two natural geometrical structures on any statisti-
cal model equipped with a differentiable manifold structure. They are the
Fisher tensor and the Chentsov-Amari tensor.
The Fisher tensor was given by Fisher in 1925 as an information charac-
terization of a statistical model. Rao13 proposed to consider this tensor as a
Riemannian metric on the manifold of probability distributions. This Fisher
metric has been systematically studied in Refs. 1,4,9 and o t h e r ~ect. ,
in the field of geometric aspects of statistics and information theory.
Chentsov4 and Amari2 independently also discovered a natural structure
on statistical models, namely a 1-parameter family of invariant connections,
which includes the Levi-Civita connection of the Fisher metric. This family
of invariant connections is defined by a 3-symmetric tensor T together with
the Levi-Civita connection of the Fisher metric.
*Mathematical Institute of ASCR, Zitna 25, 11567 Praha 1, E-mail: hvleOmath.cas.cz

232 L. H. Van
Motivated by the question how much we can describe a statistical model

via their Fisher metric and Chentsov-Amari tensor T , in 1987 Lauritzen
proposed to call a Riemannian manifold ( M ,g) with a 3-symmetric tensor
T a statistical manifold. Since two 3-symmetric tensors T and k . T , k # 0,
define the same family of Chentsov-Amari connections, we shall say that
two statistical manifolds ( M ,g, T ) and ( M ,g, IcT) are conformal equivalent.
A natural and important question in the mathematical statistics is to
understand, if a given family M of probability distributions can be consid-
ered as a subfamily of another given one N . In the language of statistical
manifolds, this question can be formulated as a problem of isostatistical em-
bedding of a statistical manifold ( M ,g , T ) into another one ( N ,g', TI).Here
we say that an immersion f : ( M ,g, T ) -+ (z, ij, T ) is called isostatisticul,
if f*(?j)= g and f*(T) =T.
We shall see in Sec. 2 that the problem of the existence of an isostatisti-
cal embedding includes also the Lauritzen question in 1987, if any statisti-
cal manifold is a statistical model. It also concerns the following important
problem posed by Amari in 1997, if any finite dimensional statistical model
can be embedded into the space CapN of probability distributions of the
sample space oNof N elementary events for some finite N .
We shall construct a class of C" (and C') monotone invariants of sta-
tistical manifolds, which present obstructions to embedding of a given C k
statistical manifold M into another one N". Here a C kstatistical manifold
( M ,g, T ) is a smooth differentiable manifold with C k sections g E S2T*M
and T E S3T*M. These invariants measure certain relations between the
metric tensor g and the 3-symmetric tensor T . In particular, using these in-
variants we show that no statistical manifold which is conformal equivalent
to the space CapN can be embedded into the product of m copies of the
normal Gaussian manifolds for any N > 3 and any finite m. In the Main
Theorem (Sec. 5) we prove that any smooth (C' resp.) statistical manifold
M" can be isostatistically embedded to a the space CapN for some N big
enough.
As a consequence we also get a new proof of Matumoto theorem on the
existence of the contrast function for a statistical manifold (see subsection
2.6).
2. Statistical Models and Statistical Manifolds

In this section we recall the definitions of the Fisher metric and the
Chentsov-Amari connections on statistical models. We introduce the no-
tion of a weak Fisher metric and a weak potential function. At the end
Monotone Invariants and Embeddings of Statistical Manifolds 233
of the section we discuss the problem, if a given statistical manifold is a

statistical model. Most of the facts in this section can be found in Ref. 1.
Suppose that M is a statistical model - a family of probability measures
on a space R. We assume throughout this note that M and R are differ-
entiable manifolds, and R is equipped with a fixed Borel measure dw. We
also write
where p ( z , w ) in LHS of 1 is a Borel measure in M and p ( z , w ) in the RHS

of (1) is a non negative (density) function on M x 52 which satisfies
l p ( z , w ) d w = l'dz E M .
The Fisher metric g F ( x ) is defined on M as follows. For any V,W E TxM

we put
g F ( V ,W ) X = /R
(8vlnp(z, W ) ) ( a wlnp(z,W ) ) P ( Z ,w ) . (3)
The function under integral in Eq. (3) is well defined, if
Denote by Cap(R) the space of all probability measures on R. Clearly we

can consider the density function p ( z , w ) as a mapping M -+ Cup(R). Thus
we shall call a function p ( z , w ) a probability potential of the metric g F , if
p ( z , w ) satisfies (2), (3), (4). It is known1y4 that for a given Riemannian
metric gF on a smooth manifold M there exist many probability potentials
f(z, w ) for g F , even if we fix the space (a,dw).
Some time it is useful to consider functions p ( z , w ) which satisfy (3) and
(4) but not necessary (2). In this case, the Riemannian metric g F will be
called weak Fisher metric, and the function p ( z , w ) will be called a weak
probability potential of gF.
2.1. Example of a W e a k Fisher M e t r i c

The standard Euclidean metric go on the positive quadrant RT(zi > 0).
It is straightforward to check that go admits a weak probability potential
{pi(z)= iz:,i = m.} Here R = RN - the sample space of N elementary
events.
234 L. H. Van
2.2. The Fisher Metric on the Space ( C a p N ) + of all

positive probability distributions on ilN (see also
Refs. 1,496)
By definition we have
Cap? :={(PI,.-. ,pnr)lpi > 0 for i = I , & C p i = 1).
We define the embedding map
f :Capy -+ s N - l ( 2 ) ,
( p l , " ' , p N ) H ( 4 1 = 2 ~ , . ' ., q N = 2 m .

It is easy to see that the Fisher metric in the new coordinates ( q i ) is the
standard metric of constant positive curvature on the sphere SN-' ( 2 ) .
2.3. Divergence Potential (see 1 , l d )

A function p on M x M with the following property
p(x, y) 2 0 with equality iff x = y (5)
is called a divergence function. A divergence function p is called a divergence
potential for a metric g on M , if
g ( X ,Y ) z = Hf=(P)(il ( X ) ,21 ( Y ) ) . (6)

where
T(,,,)(M, M ) = (T,M, 0) CB (0,T,M) = ( i l ( T , M ) ) CB ( i 2 ( T z M ) ) .
An example of a divergence potential for a Fisher metric is the Jensen func-
tion J$"(x, y) of the entropy function H ( x ) on M , or a Kullback relative
entropy function K ( z ,y) on M x M .
2.4. Chentsov-Amari connections

Let p(x,w ) be a probability potential for a Riemannian metric g . We define
a symmetric 3-tensor T on M as follows
T ( X , Y ,2)=
I (axlnp(~,w))(dyInp(~,w))(~zln~(~,w))p(~,w).
(7)
We denote by VF the Levi-Civita connection of the (weak) Fisher metric

g F . We define
< V k Y ,2 >:=< VSY,2 > +t . T ( X ,Y,2). (8)
The connections V t are called the Chentsov-Amari connections.
Remark 2.1. (Refs. 1,8) Any divergence function p ( z , y ) on M x M defines

a tensor T on M via the following formula
T ( X ,y, Z)Z = - & * ( z ) H e s s ( P ) ( i l ( X )i ,l ( Y ) ) ( Z , Z )

+ & l ( z ) H e s s ( p ) ( i z(XI,iZ(Y))(Z,Z).
If g and T are defined by the same divergence function p ( x , y ) , we shall
call p(z, y) a divergence potential f o r the statistical manifold ( M ,g , 7'). It
is a known fact that the Kullback relative entropy function is a divergence
potential for the associated statistical model.
2.5. Statistical Submanif olds

A submanifold N in a statistical manifold ( M ,g , T ) with the induced Rie-
mannian metric g1N and induced tensor TIN is called statistical subman-
ifold of ( M ,g , T ) . Clearly, if f(z,w ) is a (weak) probability potential for
( M ,g, T ) ,then its restriction to any submanifold N c M is a (weak) prob-
ability potential of the induced statistical structure.
2.6. Statistical Models and Statistical Manifolds

Since any probability function p ( x , w ) defines a map M + Cap(R), we
shall say that a statistical manifold ( M ,g , T ) is a statistical model, if there
probability potential p ( x ,w ) for g and T . By the remark in subsection 2.5,
we get that a statistical submanifold of a statistical model is also a statistical
model. Furthermore, if a statistical manifold (M, g , T ) is a statistical model,
then it must admit a divergence potential. Hence the following theorem of
Matsumoto is a consequence of our Main Theorem in Sec. 5 .
Theorem 2.1 (Ref. 8). For any statistical manifold ( M ,g , T ) there exists
a divergence potential p for g and for T .
3. Embeddings of Linear Statistical Spaces

An Euclidean space ( R n , g o ) equipped with a 3 -symmetric tensor T will
be called a linear statistical spaces. We observe that the equivalence class
of linear statistical spaces coincides with the orbit space of 3-symmetric
tensors T under the action of the orthogonal group O ( n ) .In this section we
discuss certain invariants of these orbits and we show several necessary and
sufficient conditions for the existence of embedding of one linear statistical
236 L. H. Van
space into another linear statistical space by studying these invariants. A

class of our necessary conditions consists of monotone invariants A, i.e. we
assign to any linear statistical space (R", go, T ) a number A(Rn, go,T ) such
that, if (R",go, T ) is a statistical submanifold of (R",go, T ' ) ,then we have
A@", go, T ) 5 X(Rrn,go, T').

Since a tangent space of a statistical manifold is a linear statistical man-
ifold, these invariants play important role in the problem of isostatistical
immersion.
3.1. R a c e Type of a Symmetric 3-tensor

Let us denote by R" the subspace in S3(Rn) consisting of the following
3-symmetric tensors
T V ( z , y , z )=< v , z >< y , z > + < v , y >< x , z > + < v , z >< x , y >,

where v E R".Using the standard representation theory (see e.g. Ref. 12)
we have the decomposition
S3(R") = R(37r1) @ R". (9)
To compute the orthogonal projection of a 3-symmetric tensor T on the
space R" in the decomposition (9) we can use the following Lemma. We
denote by 7r2 the orthogonal projection form S3(Rn) to R".
Lemma 3.1. W e have
Here we identify the 1-form T r ( S ) with a vector in R" by using the Eu-
clideun metric go.
We omit the proof of Lemma 3.1 which is straightforward. In view of
Lemma 3.1 we shall call any tensor T E R" of trace type.
We note that
dims3@") = C i + 2 C : + n =
n(n + 1)(n
6
+ 2)
Thus the dimension of the quotient S 3 ( R n ) / S O ( n )is at least C i +
C: +n. A direct computation shows that the dimension of the orbit
~ ] )= dimSO(n), if nui # 0. Here {vi} is an or-
S O ( n ) ( [ C ~ = = l uisi vC:
thonormal basis in R". Hence the dimension of S3(Rn)/O(n)= C:+C:+n.
This dimension is exactly the number of all complete invariants of pairs
consisting of a positive definite bilinear form g and a 3-symmetric tensor

T.
Since the dimension of G k ( R n ) = k ( n - k), it follows that generically
it is impossible to embed a linear statistical space (Rk, go, T) into a given
+ +
statistical linear space (R",go, T), unless k ( n - k) 2 Cz Cz k. Clearly
the dimension condition is not sufficient as the following proposition shows.
Proposition 3.1. A linear statistical space (Rk, go, T ) can be embedded

into a linear statistical space (RN,gO,T"),if and only i f N 2 k and T i s
also a trace type: T = T" with 1wI 5 1211.
Proof. The necessary condition follows from the fact that the restriction
of T" to Rk equals Tc, where V is the orthogonal projection of v to Rk.
Conversely, if IwI 5 IvI we can find an orthogonal transformation, such
that w equals the orthogonal projection of v on Rk.
3.2. Commasses as Monotone Invariants

Since the metric g extends canonically on the space S3(Rn),we can define
the absolute norm
Now we define comasses of a 3-symmetric tensor T as follows
M3(T):= max T(z,y , z ) ,

Izl=l,Jyl=l,lzl=l
M2(T):= max T(z, y , y ) ,

lxl=1,lYl=1
M1(T) := maxT(z, z, z).

(xl=1
Clearly we have
0 I M1(T)I M 2 ( T )I M 3 ( T )I IITII.
Proposition 3.2. The comasses are positive functions which vanish at T,

i f and and only i f T equals zero. They are monotone invariants of T, since
if T is a restriction of 3-symmetric tensor T o n R N , then
IITII 5 IITII, M i ( T )5 M i ( T ) ,V i = 1,2,3. (11)

238 L. H. Van
Proof. To prove the first statement it suffices to show that M 1 vanishes

at only T = 0. To see this we use the identity
+ + + + + +
- 1 2 T ( x ,Y,2 ) = T ( x Y 2 , ~Y z , 2 Y 2)
+ + + + +z,z- Y +
T ( x Y - Z,Z + Y - Z , T y - Z) T ( x - Y Z,% -y + 2)
+ T(-x + y + Z , -X + y + Z , -Z + y + Z )
- 2(T(s,5, + T(Y,Y,Y) + T ( z ,z , .I).
The second statement follows immediately from the definition.
Now for a space (R", go,T ) and for 1 5 k 5 n we put
We can easily check that if T is a restriction of T to a subspace I%" c R",

then
Xk(T) 2 Xk(T) 2 0 for all k 5 m.

Thus X k ( T )is a monotone invariant of linear statistical manifolds. These
invariants are related by the following inequalities
M 1 ( T )= X,(T) 2 X,-l(T) ... 2 Xz(T) 2 X1(T) = 0.
The last equality follows from the fact, that the function T ( z , x , z )is
anti-symmetric on S"-l(lzl = 1) c R" and S"-' is connected. We observe
that if T is of trace type, then Xn-l(T) = . . . = X1(T) = 0.
We are going to give a lower bound of the monotone invariant X,-1 of
a linear statistical space of certain type. The equality Xn-l(Rn,go, T ) 2 A
means that no hyperplane with the norm M 1 strictly less than A can be
embedded in (R"-', go, T ) .
Lemma 3.2.
(a) Let T = CZ1(N- ~ i ) ( x a )be
~ a 3-syrnrnetric tensor o n R" with n 2 4,
N 2 4 and 5 1/4. Then we have
(b) Let T = N Cy=,( x i ) 3 ,and H be a hyperplane in R" which is orthogonal

to ( k n , 1,1,.. . , I), and let n 2 5 , k 2 3. Then we have
N
X"-2(?H) 2 7- 1.
(c) Let x = ((1 - E ) , &,

... , &) E S"(1) c ItWn+', where n 2 4,k >_ n.
W e denote by H the tangential plane TxSn, and by T o the following
3-symmetric tensor on IW"++~:
Then we have
Remark 3.1. The tensor T o in (??) defines on (R", go) a statistical struc-
ture with a weak probability potential { ix:,
i = 1,n}.
Proof of Lemma 3.2. The reader shall see that a proof of Lemma 3.2
can be done in the same scheme of the proof of Sublemma 5.4. Therefore
we do not repeat this argument here.
Remark 3.2. Lemma 3.2.a holds also for n = 3 but not for n = 2, Lemma
3.2.b holds also for n = 4, but not for n = 3, and Lemma 3.2.c holds also
for n = 3 but not for n = 2.
There are also several obvious monotone invariants of T .
A ~ ( T:=
) max T ( x ,Y, 2)
Ix)=ly~=lzl=1,<z,y>=<y,z>=<z,x>=0
is well-defined for n >_ 3.
A ~ ( T:=) max T ( x ,Y,Y),

1x1=lyl=l,<x,y>=O
is well-defined for n 2 2. We can check that

ker A' = 72".
On the other hand we have
kerA2 c R(37r1).
Thus A' and A2 are different invariants. 0
Lemma 3.3. Let 7r1 be the first component o f T in decomposition (9). Then
llTll1 := Ilnl(T)II is a monotone invariant of T .
Proof. Let Rk be a subspace of R". We denote by .rr;CnT the restriction of

T to R k . Clearly
TE(T)= T , " ( T ~ +
T )~ i E ( 7 ~ 2 7 ' ) .
240 L. H. Van
We have noticed in Proposition 3.1 that the restriction of the trace form
7r2T to any subspace is also a trace form. Thus 7rc(7r2) is an element in
Rkc S3(Rk).Hence we have
~1( 4 Y ) = ~1 (r;El(TIT)). (13)
Since all the projections 7r1, 7 r c decrease the norm 11. I I , we get
IlrcTlll = ll~l(4Y)ll
= llm(7r:(~lT))ll 5 II7rl(T)II = IITII1.
Proposition 3.3. A statistical line (R,g",T) can be embedded into

( R N , g O , T 'if) , and only zf M1(T) 5 M1(T').
Proof. It suffices to show that we can embed ( R , g O , Tinto

) (RN,gO,T'),
if we have M1(T) 5 M1(T'). We note that T ' ( w , w , w ) defines an anti-
symmetric function on the sphere SN-l(IwI = 1) c I R N . Thus there is a
point w E SN-l such that T'(w, v , w) = M1(T). Clearly the line w @ IR
defines the required embedding.
Let us consider the embedding problem for 2-dimensional linear statis-

tical spaces. It is easy to see that
S"(IW2) = R2 @ R2.
Thus the quotient S3(R2)/SO(2)equals (R2@R2)/S1.Geometrically there
are several ways t o see this. In the first way we denote components of
T E S3(R2)via Tiii1Tii2,Ti22,T222.
Lemma 3.4. There exists an oriented orthonormal basic in R2 such that
TI11 = M 1 ( T ) > O,T112 = 0 for all non-vanishing T . These numbers
(T111,T122, T222) are called canonical coordinates of T. Two tensors T and
T' are equivalent, i f and only i f they have the same canonical coordinates.
Proof. We choose an oriented orthonormal basis (w1 , 212) by taking as w1 a

point on S1((x(= l), where the function T(x, x,x) reaches the maximum.
The first variation formula shows that in this case T112 = 0. This shows the
existence of the canonical coordinates. Clearly, if two tensors have the same
canonical coordinates, then they are equivalent. Next, if two tensors T and
T' are equivalent, then their norms M1 are the same. We need to take care
the case, when there are several points x a t which T ( x , x , x ) reaches the
maximum. In any case, they have the same first coordinates. Next we note
that
< T r ( T ) Tr(T)
, >= (T111+ 7'122)~ + 7-2222,
Thus if two tensors are equivalent and have the same first coordinates,
they must have the same third coordinate T122, and this third coordinate is
uniquely defined up to sign. The condition on the orientation tells us that
the sign must be +.This proves the second statement. 0
Proposition 3.4. W e can always embed the 2-dimensional statistical space

(R2,go, 0) into any linear statistical space (R",
go, T ) , i f n 2 7.
Proof. It suffices to prove for n = 7. We denote by O ( T ) the set of of all

unit vectors v E S6 such that T ( v ,v,v) = 0. Clearly O ( T )is a set of dimen-
sion 5 in S6.Since T is anti-symmetric, there exists a connected component
O o ( T )of O ( T )which is invariant under the anti-symmetry involution. Now
we consider the following function f on O O ( T )For . each v E O o ( T )we de-
note by A" the bilinear symmetric 2-form on the space TzOO(T) considered
as a subspace in R":
A"(Y, 2 ) = T ( v ,Y, 2 ) .
Then we define f(v) equal to det(A"). Since O ( T )has dimension 5, the func-
tion f(v) is anti-symmetric on O o ( T ) .Hence the set O;(T) of all E O o ( T )
with f (v) = 0 has dimension 4 and it contains a connected component which
is also invariant under the anti-symmetric involution. For the simplicity we
denote this connected component also by Og(T).Now we consider the fol-
lowing two possible cases.
Case 1. We assume that there is a point E Og(T)such that the nullity
of A" is at least 2. Then there are two linear independent vectors y , z E T,
such that the restriction of A" on the plane R 2 ( y ,2) vanishes. Since the set
O o ( T )is connected and anti-symmetric and of co dimension 1 in S"-', the
plane R ( y , z ) has a non-empty intersection with O o ( T )at a point w. Then
the restriction of T on the plane R2(v,w) is vanished, because
T ( v ,21, v) = T(w,
w,w)= 0
T ( v ,w,w) = 0 (since A"(w,w) = O),
, w)= o (since w E T,o'(T)).

~ ( vv,
Case 2. We assume that the nullity of A" on @ ( T ) is constantly 1.
Using the anti-symmetric property of A" we conclude that the restriction
of A" to the plane R4(v) which is orthogonal to the kernel of A" has index
242 L. H. Van
constantly 2. Thus there exists a vector z which is orthogonal t o the kernel

y of A" such that A " ( z , z ) = 0. Clearly the restriction of A" t o the plane
R2(y,z ) vanishes. Now we can repeat the argument in the case 1 t o get a
vector w such that the restriction of T to R2(v, w) vanishes.
Theorem 3.1.
a) A n y statistical space (Rn,gO,T) can be embedded in the statistical
space (Rn("+l),go,T' = 211Tll xzY)x:),
where xi are the canonical Eu-
clidean coordinates o n IW"("+l).
b) The trivial space (Rn,go,O) can be embedded into
(R2", go, C;zl(dxi)3) f o r all n.
Proof.
a) We prove by induction. The statement for n = 1 follows from Propo-
sition 3.3. Suppose that the statement is valid for all n 5 k.
Lemma 3.5. Suppose that T E S3(Rk+'). Then there are orthonormal

coordinates 21, ' ' . , xk such that
k+l
i=l l<i,j,k
Proof of Lemma 3.5. We choose v1 as the unit vector in Sk c Rk+', on

which the function T ( v ,w,w) reaches the maximum on the unit sphere S k .
The first variation formula shows that T(v1,wl,w ) = 0 for all w which is
orthogonal t o w 1 . We denote by Rk the orthogonal complement t o R .w1.
Now we consider a bilinear symmetric form A on Wk defined as follows
A ( X , Y) = S ( V l , X , Y).
There is an orthonormal basis on Rk, where we can write A ( z , y ) =
six!. Clearly in this orthonormal basis we can write T in the form in
(14).
Continuation of the proof of Theorem 3.1.a We shall show explicitly that

that any statistical space (IW2,g0,T = U ~ I C ~ ( X can
~ ) ~be
) embedded in
, ~ f = ~ ( y i )if~o )I
( R ~go, , la21 I1/2. We put
1 1 1 1
L(v1) := *(-, -, --, --) (15)
2 2 2 2
L(v2):= -/Tl
(/F, /?-, -/?). (16)
Here we take the sign+ in (15), if a2 > 0, and we take the sign-, if a2 < 0.
Clearly, L defines the required embedding R2 -+ R4.
This together with Proposition 3.3 and the induction assumption com-
plete the proof of Theorem 3.1.a.
Proof of Theorem 3.1.b. We decompose the embedding f : (R",go,O)

to (Pn,go, z;E1(xi)3) as follows
f ( x l , . . . ,xn) = ( f l ( x ~ ) , *, f*n*( x n ) )
where f i embeds the line
(R, ( d ~ z ) ~into ) (dx2i-1)
, ~ (R2, + ( d x 2 i ) 2 ,(dx2i-1) + ( d ~ ' ~ )Clearly,
~). f
is the required embedding. 0
4. Monotone Invariants and Obstructions to Embeddings of

Statistical Manifolds
Let K ( M ,e ) denote the category of statistical manifolds A4 with morphisms
being embeddings. Functors of this category are called monotone invariants
of statistical manifolds. Clearly any monotone invariant is an invariant of
statistical manifolds.
4.1. Examples
There are many monotone invariants which arise from our analysis in Sec. 3.
a) Trace type of a statistical manifold. A statistical manifold ( M ,g, T )
will be called of trace type, if for all x E M the form T ( z )is of trace type
(see 3.1) It follows from Proposition 3.1 that any statistical submanifold
of a statistical manifold of trace type is also of trace type. Thus the trace
type is a monotone invariant. In particular we cannot embed the statistical
space C a p N and the normal Gaussian space into any statistical space of
trace type. On the other hand, unlike the linear case, we cannot embed
a statistical manifold of trace type into another one of trace type, even if
the norm condition is satisfied. For example, if the trace form is closed (or
exact), then the trace form of its submanifolds is also closed (resp. exact).
Hence within a class of statistical manifolds of trace type we get a new
monotone invariants which can be expressed via the closedness and the
cohomology class of the corresponding trace form.
b) Decomposability of a statistical manifold. We note that the class of
3-symmetric tensors of trace form is a subclass of all decomposable tensors
T 3 which are a symmetric product of 1-forms and symmetric 2-forms. Any
244 L. H. Van
statistical submanifold of a statistical manifold with a decomposable tensor

T has also the (induced) decomposable tensor. Thus the decomposability
is also a monotone invariant. The Gaussian normal 2-dimensional manifold
is an example of decomposable type but not of trace type.
c) Rank and comass. We define for any statistical manifold ( M , g , T )
the following number
runk(T) = suprunk(T(z))
M 1 ( T ) o= SUP M 1 ( T ( z ) ) .
XEM
Clearly these four numbers are monotone invariants of statistical man-

ifolds.
We recall that the normal Gaussian statistical manifold is the two di-
mensional statistical model which is upper half of the plane R2(p,a ) with
the potential
here z E R.
Proposition 4.1. Any statistical manifold which i s conformal equivalent

t o the space C u p N cannot be embedded into the direct product of m copies
of the normal Gaussian statistical manifold for any N 2 3 and finite m.
Proof. It is easy to check that M 1 ( C u p N )= 00. Thus any statistical man-

ifold which is conformal equivalent to C a p N has also the infinite invariant
M1. On the other hand, we compute easily that the norm M1 of the Gaus-
sian normal manifold, as well as the norm M1 of a direct product of its
finite copies, is finite. Namely the norm M 1 ( p ,a ) is f i for all ( p ,a). [7
4.2. Diameters of Statistical Manifolds

For a positive number p > 0 and a statistical manifold ( M ,9, T ) we set
d,(M, g, T ) := sup{l E Rf U 00 I 3 an embedding of
([O, 11, d z 2 ,P ( d 4 3 ) to ( M ,9, T ) . )
We shall call d,(M, g , T ) the diameter with weight p of ( M ,g , T ) .Clearly

d, are monotone invariants for all p.
To estimate the diameter with weight p of a given statistical manifold
( M ,g, T ) we can proceed as follows. For each point x E M we denote by
D p ( x )the set of all unit tangential vector u E TxM such that T ( v ,v,u ) = p.
We denote by D % ( x )the connected components of D L ( x ) . We say that a
unite vector u in T x M is p-characteristic with weight c ( x ) , if there exists i
such that we have
c(x) = min < u , w >> 0.
WD;(X)
We shall say that a point x E M is p-regular, if there is an open neigh-

borhood U,(x)c M such that D,(U,) = U, x D p ( x ) . It is easy to see that
the set of all pregular points is open and dense in M for any given p.
Proposition 4.2. The diameter d , of (M", g , T ) is infinite, zf m 2 3 and

there exists a number E > 0 such that one of the following 2 conditions
holds:
+
a)There exists a ( p &)-regular point x E M such that the convex hull
Cou(D$+, ( x ) )of one of connected components DS+,(x) contains the origin
point 0 E TxM" as it interior point.
b) ( M " , g , T ) has a complete Riemannian submanifold ( N , g ) such that
there exists a smooth section x H (D,+,(x) n T N ) ouer N .
Proof. The statement under the first condition a) is based on the fun-
damental Lemma of the convex integration technique of Gromov. Namely
Gromov proved that (2.4.1.A, Ref. 5), if the convex hull of some path con-
nected subset A0 c Iwq contains a small neighborhood of the origin, then
there exists a map f : S1+ Iwq whose derivative sends S1 into Ao. 0
Lemma 4.1. Under the condition in Proposition 4.2.a there exists a small
neighborhood U s ( x ) in M and an embedded oriented curue S1 c U ~ ( xsuch )
+
that for all point s ( t ) E S1 we have M 1 ( T s ( t ) S 12) p ( ~ / 2 ) .
Proof of Lemma 4.1. We denote by E x p the exponential map TxM" +

M" and by D E x p the differential of this exponential map restricted to
S m - 1 x TxM" c T ( T x M " ) . Here Sm-l is the unit sphere in TxM". The
space TxM" is a linear statistical space, so we denote by M : the induced
norm-function on S"-l x TxM" as follows:
246 L. H. Van
Since DExp is a continuous function, whose restriction to S"-' x 0 is the

identity, there exists a ball B(O,6)with center in 0 E T,M such that
Ml(DExp(1)) <~ / 4
- M:(1)) (17)
for all 1 E S"-' x B(6) c T(T,M"). We can assume that b is so small
such that DExp is a homeomorphism on S"-l x B(0,S).
Now we apply the above mentioned Gromov Lemma (2.4.1.A, Ref. 5)
t o get a oriented curve S1(t)in the linear space T,M such that
for all t. Next we observe that for all a > 0 the curve a . S1(t)has the same
norm as S ' ( t ) , i.e.
M:(q(a.sll(t))= M ; ( q ( s l ) ( t ) =
) p+~.
Thus we can assume that our curve S1(t), which satisfies (18), lies in the
ball B ( 0 , b ) . By our choice of S (see (17)), we get from (18)
5
p +3
ZE Ip + ZE,
I M1(Exp(S1(t))) (19)
for all t. This curve E x p ( S 1 ( t ) )is an immersed curve. To get an embedded

curve we perturb the immersed curve such that the condition of Lemma 4.1
is satisfied. This is possible, since rn 2 3.
Now let us t o continue the proof of Proposition 4.2.a. We denote by S1( t )
the embedded curve in Lemma 4.1. Next by choosing a tubular neighbor-
hood of S1(t) we can get a (small, thin) oriented embedded solid torus
T 3 ( t ,s, r ) = S1( t )x S1( s ) x [0, R] in M" such that our embedded curve is
exactly the mean curve S ' ( t ) x (0) x (0) on the solid torus. We can choose
this torus T 3 so thin, such that for all s, t , T we have
&
M1(T;(t,s)) 2 p + 4. (20)
Using (20) we choose a smooth unit vector field V ( t , s ) on the torus
T 3 ( t ,s, r ) which is tangential to each torus T:(t, s ) such that T(V,V,V )=
p. The integral curve of this vector field is either a circle or an curve of infi-
nite length. If there exists an integral curve of infinite length, then this curve
is our desired curve for the Proposition 4.2.a. Assume now that all the inte-
gral curves are circles. Then there exist an embedding S1(t) x [0, p] x [0, p]
such that for all ( s ,T ) E [0, p] x [0,p] the circle S1(t)x {s} x { T } is an inte-
gral curve of V . Now we perturb V in a neighborhood [0, a] x [0, p] x [0, a]
with a very small a such that the perturbed unit vector field V' satisfies
T(V',V', V ' ) = p and the integral curve of vector field V' is not any more
periodic. This completes the proof of the first part in Proposition 4.2.
Using the same argument we can prove the second part b) of Proposition
4.2. First we get the existence of an embedded curve S ' ( t ) of arbitrary
+
length on M such that M ' ( T s l ( t ) )2 p ( 1 / 4 ) ~Now
. we consider a torus
tubular neighborhood of this curve in M and apply the same argument in
the first part, namely we get on each torus T2(t,s) an integral curve whose
unit tangential vector V = (a/at)S' (t;s, r ) satisfies the condition:
T(V,V,V ) = p.
If there exists an infinite integral curve, then we are done. If not, that means
all integral curve are circles, then we apply the perturbation method in the
proof of the first part and get our desired curve.
5. Existence of Isostatistical Embeddings into CupN

Main Theorem-Any smooth (C' resp.) statistical manifold ( M " , g , T ) can
be immersed into the statistical manifold (Cup?, g F , T A V Cf)o r some finite
number N . Hence any statistical manifold i s a statistical model.
We first deduce our Main Theorem for compact statistical manifolds
( M " , g , T ) from Theorem 5.1 and Theorem 5.4.
Theorem 5.1. Let (M", g , T ) be a compact smooth (C' resp.) statistical

manifold. T h e n there exist numbers N E N+ and A 2 0 as well as a smooth
(C' resp.) embedding f : (M", g, T ) -+ (RN,go, A.To) such that f* (go) = g
and f * ( A .TO)= T .
Our proof of Theorem 5.1 uses the Nash embedding theorem, the Gro-
mov embedding theorem and an algebraic trick. The existence of monotone
invariants prevents us extend Theorem 5.1 for non-compact case (in con-
trast to the Riemannian case.)
Theorem 5.2 (The Nash embedding theoremloill). A n y smooth

(C' resp.) -Riemannian manifold ( M " , g ) can be isometrically embedded
into (RN,go) f o r some N depending o n M n .
We denote by TOthe "standard" 3-tensor on Rn:

n
248 L. H. Van
Theorem 5.3 (The Gromov immersion theorem5). Suppose that

M" i s given with a smooth (C' resp.) symmetric 3-form T . T h e n there
exists a n embedding f : M" --f WN1(") with N l ( m ) = 3(n + (;+')
(;+'))+
such that TO) = T .
Proof of Theorem 5.1. First we shall take an immersion fl :

(M ", g , T)+ (RN1(m),go, To)such that
f?(To)= T.
The existence of f 1 follows from the Gromov immersion theorem.
Then we choose a positive number A-' such that
9 - A-l(f?(go))= 91
is a Riemannian metric on M , i.e. g1 is a positive symmetric bi-linear form.
Such a number A exists, since M is compact.
Now we shall choose an isometric immersion fi : (M", 91) --t (I@"go).
The existence of f2 follows from the Nash isometric immersion theorem.
Lemma 5.1. There i s a linear isometric embedding Lm+l : Rm+l 4
such that Ln+1(To)= 0.
Proof. We put
L m + l ( x l , . . . ,xm+1) = XI),... f"+'(xrn+l))
7
where fi embeds the line (R, (dxi)', 0) into

(It2,(dz2$-l) + ( d x 2 i ) 2 ,(dx2i-1 ) + ( d ~ ~ ~ ) ~ ) :
Clearly, Lm+l is the required embedding.
Completion of the proof of Theorem 5.1. Finally we take an embedding

f3 : M" --t W("+1)("+2)+" @ W2"+2
as follows.
f3(2) = A - l . f i ( x )@ 0f2).
Since f 2 is an embedding, f3 is the required embedding map for Theorem

5.1.
Theorem 5.4. Suppose that C i s a compact subset in Cap?. T h e n any
bounded domain D in a linear statistical manifold (Rn,go, A . TO)can be
realized as a n embedded statistical submanifold of Cap? \ C .
We denote by Cap:(X) the statistical manifold which is obtained by

the restriction of the statistical structure from (R", go, Ef=l
(xi)-l(d~i)~)
to the positive quadrant S:(X) of the sphere of radius X with center a t the
origin of R4.
Proof of Theorem 5.4. We put

-
A := m a x { 4 6 , 4 f i . A}.
Let U be an open neighborhood of (A,(2z)-',(2z)-',(2z)-')E
S 3 ( 2 / f i ) . Here X is the positive number such that
+
nX2 3 n ( 2 2 ) - 2 = 4. (21)
We denote by U+ the intersection U n C a p : ( 2 / f i ) . We now choose
U so small such that the product U x,times U C S4n-' lies in
the complement S4n-1\ C. Since V+ is a statistical submanifold of
(R:, go, Ci=l
4
~ i ' ( d x i ) ~the ) , direct product U+ x n t i m e s U+ is a statis-
tical submanifold of (RP, go,
4n
x ; l ( d ~ i ) ~ Hence
). U+ X , t i m e s U+ is a
statistical submanifold of (Cap4", g F , T A P C ) .
We denote by U ( A , r ) the ball of radius r at the point
(A, (22)-', (2z)-l, ( 2 z ) - l ) E S 3 ( 2 / J 5 i ) . First we prove
Lemma 5.2. For given positive numbers R > 0 and A 2 0 there exists
a positive number r such that the bounded domain [0,R] X , times [0,R] c
(Rn,go, A * To) can be realized as a n immersed statistical submanifold of
u + ( A , r ) Xntimes u + ( A , r ) C (Capyn,gF,TA-C).
Proof of Lemma 5.2. It suffices to show that there is a statistical im-

mersion f : ([O, R ] , d x 2 , A .d x 3 ) -+ U + ( A , r ) . On U + ( A , r ) we consider the
distribution D ( p ) which is defined by
D,(p) := {w E T,U+(A,r) : IvI = l,T(w,w,v) = p }
for any given p > 0. Clearly the existence of an immersion f : ([0,R], d x 2 , A-

d x 3 ) + U+ ( A ,r ) is equivalent to the existence of an integral curve with the
length R of the distribution D ( A ) . The existence of the desired curve is a
consequence of the following Lemma
Lemma 5.3. There exists an embedded torus T 2 in U + ( A ,r ) which i s pro-

vided with a unit vector field V o n T 2 such that T(V,V,V ) = A.
Proof. Let us denote xo := (A, ( 2 2 ) - ' , ( 2 z ) - ' , ( 2 2 ) - ' ) E S 3 ( 2 / f i ) with

X satisfying ( 2 1 ) . We shall need the following
250 L. H. Van
Lemma 5.4 (Sublemma). Let H be any %dimensional subspace in

T,,U+(A, r ) . Then there exists a unit vector w E H such that T ( w ,w , w ) 2
2A.
Proof of Sublemma 5.4. The subspace H can be defined by two linear

equations:
< w , z o >= 0 ,
+
<w,h>=O.
Here w is a vector in H c R4 and iis a unit vector in R4 which is not

co-linear with 30 and which is orthogonal t o H . Without loss of generality
we can assume that
+
h = (0 = h l , h2, h3, h4) and h: = 1.
i
Case 1. Suppose that not all the coordinates hi of iare of the same
sign, so we assume that hl = 0, h2 5 0 , h3 > 0. We put
w := ( W l , w2 = (1 - &2)k3,w3 = (1 - &2)k2,0= w4). (24)

The equation (23) for w is obviously satisfied. Now we choose w l , ~2 from
the following equations which are equivalent to (22) and the normalization
of w:
+
. w1 (1 - E 2 ) . (2A)-1 * (Ic2 + k3) = 0 ,
w:: = (2&2- E ; ) .
0 From (25) we get
w1=
(1- E2)(k2 + k3)
A . 2.4
Substituting this into (26),we get
+
(k2 k3I2
+ 1)E; - (2 + 2(k2 + k3)2 + (-----)
)&2
k2 + k3 2
= 0.
( (A. 2 4 2 ( A . 2A)2 A . 2A
We shall take one of ( 2 possible) solutions ~2 of (28) which is
0 From (29) we get

5
82 I-,Y
16
+
since 0 < k2 Ic3 I 2 (this follows from (24)), and X 2 2n-lI2 (this follows
from (21)), so A . 2 2 2 16. Since
T,,(w,
w,w)= X-lw; + (2Z)(w,3 + w,")
we get
5 1
T ( w , w ,W ) 2 8 f i . A ( 1 - -) . -> 2A.
16 2Jz-
So in this case 1 Sublenima 5.4 holds.
Case 2. Now we shall assume that h l = 0 and h2 2 h3 2 h4 > 0. We
set
a > 0 such that a2(a2 2) = 1. + (31)

The Eq. (30) ensures that < w, h >= 0. In order that w is a unite vector
and w E H the numbers w1,8 2 must satisfy the following equations
- Awl + (1 - &2)(a- 2)"

2x
= 0, (32)
w: = (282 - &;). (33)
0 From (32) we get
(1 - &z)(a- 2)"
w1= (34)
A271
Now substituting (34) into (33) we get a solution EZ
~2 = (1 + 2B2) - J(1 + 2B2)' - B2(1 + B 2 ) ,

where
(a-2)" 1
B= <-
A271 X2ii'
Since 0 < ( a - 2)" < 1 we have E Z < 1/(X221) and ~1 < l/(X2Z). Hence
T(w,
w ,w)= X-'W? + (2Z)u3(1 - E Z ) ~+(a 3~ ) > 2A. 0
252 L. H. Van
Completion of the proof of Lemma 5.9. First we choose a small embedded

torus T 2 in U+( A ,r ) such that for all x E T 2 we have
3
max { T ( v ,
w,w)} 2 TA. (35)
vE S (T,TZ)
This is possible thank t o Sublemma 5.4. Since T(w,w,w) = -T(v, w,w) and
T 2 = R2/Z2 is parallelizable, (35) implies that we can find a smooth vector
field V on T 2 sastifying the condition in Lemma 5.3.
Completion of the proof of Theorem 5.4. For a given A in Theorem 5.5 we
let A := A + & for some small positive E and we apply Lemma 5.2 t o ( R ,A)
which is in fact t o apply Lemma 5.3. We can show that the existence of an
isostatistical immersion f : ([0, R],d x 2 , A . dx3) --f U+(A,r ) implies the
existence of an isostatistical embedding f : ([0,R ] ,d x 2 , A . dx3)toU+(A,r )
by using the same argument in the proof of Proposition 4.4.a.
Proof of Main Theorem f o r the non compact case. We can deal with this
case by using the compact decomposition of M as Nash did for the iso-
metric embedding for the smooth case [lo]. Namely we cover M by disk
+
neighborhoods N,?,j = 1,m 1 in the following way. For each j let
Cj := UiNf. (36)
Then we require that the union in (36) is a disjoint union, i.e. N: nN l = 8,
if i # k. We also require that each N j overlaps only a finite number of
other Nf.Now we compactify N j via an surjective smooth mapping
: qbi : N! -+ S:, where S: is a sphere of the same dimension m. The map 4:
can be extended to the whole M, since it maps the boundary of N i into
the north point of the sphere. On the other hand, this map 4: is injective in
a large (enough) sub domain c Nf. We can furthermore use the unity
partition function to define a C1 statistical structure on each 5: such that
the (sum of) pull back via & is the given statistical structure on M . In
other words we can consider the C1 statistical structure on M as induced
from (infinitely many) spheres 5; via the smooth mapping 4:.
+
Now let for each j , 1 5 j 5 m 1, we put
sj := uas;.
Using Theorem 5.1 we can find an isostatistical embedding
inductively, since each 5: is compact.

Now let us consider the map
as a composition of the m a p 4: : M -+ 5': and the m a p $j.

Finally the product mapping
10 = I1 x * * * x L + l
is the desired isostatistical embedding, since the statistical structure on

SN(")-l is induced from the (non-linear) statistical structure on EXN(").
We can see this easily by noticing t h a t both symmetric forms go and TS
are decomposable w.r.t. the embedding i : Rn + Rn+' for any n, 1 > 0, i.e.
go(Rn) = i*(go)(Rn+'), Ts(RY)= i*(Ts)(R";t').
Acknowledgement
I am thankful t o Jiirgen Jost and Nihat Ay for their introduction t o t h e
field of information geometry and helpful discussions.
References
1. S. Amari and H. Nagaoka, Methods of Information Geometry, Trans. of
Math. Monograph (20QO).
2. S. Amari, Contemporary Math. 203, 81 (1997).
3. N. Ay, Ann. of Prob. 30, 416 (2002).
4. N. N. Chentsov, Statistical decision rules and optimal inference, (AMS.
1982), (originally published in Russian, Nauka, Moscow, 1972).
5. M. Gromov, Partial Differential Relations, (Springer-Verlag, 1986).
6. J. Jost, Information geometry, (lecture at MPI MIS 2005).
7. S. Lauritzen, Statistical manifolds, in Differential Geometry in Statistical
Inference, IMS Lecture Notes, Monograph Serie lo., (Inst. of Math. Stat.
Hayward, California, 1987).
8. T. Matsumoto, Hiroshima Math. J. 23, 327 (1993).
9. E. A. Morozova and N. N. Chentsov, Markou invariant geometry on manifolds
of states (in Russian), (Itogi Nauki i Techniki 36, 1990) pp. 69-102.
10. J. Nash, Ann. Math. 60, 383 (1954).
11. J . Nash, Ann. of Math., 63, 20 (1956).
12. E. B. Vinberg, A. L. Onishchik, Seminars on Lie Groups and Algebraic
Groups, (in Russian) (Nauka, Moscow, 1988).
13. C. R. Rao, Bull. of the Calcutta Math. SOC. 37, 89 (1945).
14. C. R Rao, Differential metrics in probability spaces, in Differential Geometry
in Statistical Inference, IMS Lecture Notes, Mongraph Serie lo., (Inst. of
Math. Stat. Hayward, California, 1987).
$13. GRADED CECH COHOMOLOGY IN

NONCOMMUTATIVE GEOMETRY *t
DO NGOC DIEP
Institute of Mathematics,
Vietnam Academy of Science and Technology,
18 Hoang QUOCViet Road, Cau Giay Dist., 10307 Hanoi, Vietnam
E-mail: dndiep@math.ac.vn
The Zz-graded Cech cohomology theory is considered in the framework of

noncommutative geometry over complex number field and in particular the
homotopy invariance and Morita invariance are proven. In some special case we
deduce an isomorphism between this noncommutative theory and the classical
Zz-graded Cech cohomology theory.
Keywords: Cech cohomology of a sheaf, cyclic theory

Mathematics Subject Classification 2000 14F20, 19K35, 46L80, 46M20
1. Introduction
Let us fix in this paper the field of complex numbers as the ground field for
algebras, modules, etc.
The general Cech cohomology presheaf 'l?q for an arbitrary presheaf X
on a category C, k Q ( X G)
, = 'l?q(G)(X)was introduced in Ref. 7.
The idea of Cech cohomology for noncommutative geometry was ap-
peared in Refs.2 and 6. In this paper we use this idea to define the corre-
sponding (periodic) Z2-graded Cech theory.
We prove the homotopy invariance and Morita invariance of Cech coho-
mology in the framework of noncommutative geometry. Our main result is
based on a detailed analysis of the structure of C*-algebras. A crucial ob-
servation is the fact that for C*-algebras the category of *-representations
*The text is presented a t the Second International Conference on Abstract and Ap-
plied Analysis, Quinhon, Vietnam, 4-9 June 2005. The author thanks the organizers and
especially Professor DSc. Nguyen Minh Chuong for the invitation t o give this lecture.
+The work was supported in part by Vietnam National Project of Research in Funda-
mental Sciences and The Abdus Salam ICTP, UNESCO.
256 D. N. Diep
defines exactly the C*-algebra itself, by the well-known Gelfand-Naimark-

Segal Theorem. From this we can deduce the Morita invariance and homo-
topy invariance of the Cech cohomology, the same properties of periodic
cyclic homology of the C*-algebra. Since our result is valid not only for
C*-algebras, we work in the general context of a noncommutative algebra
over complex numbers.
The paper is organized as follows. Taking the Cech cohomology in place
of the de Rham theory in the periodic cyclic theory of A. Connes, in Sec. 3
we define the Zz-graded Cech cohomology theory. Then in Sec. 4 we prove
two important properties of the theory as the homotopy invariance and
Morita invariance. In the last Sec. 5 we deduce also some kind of Connes-
Hochschild-Kostant-Rosenberg theorem. This lets us see a clear relation
with the classical case of commutative algebras and ordinary Cech coho-
mology theory.
Notations: We follow the notations from Refs. 7 and 4
2. Preparation: Differential Systems and the Cyclic Theory

2.1. Differential Systems, Following Kashiwara
From M. Kashiwara Ph. D. dissertation:l Linear polynomial differential
systems
N R
j=1 p=l
correspond to the V-modules

M =VN/VN'P JV= VR/VR'Q
In general if
2 3
is a system of partial differential operators from the vector bundle to 3
and set
= Homo(, 0),
then
V 8 & E { differential operators from to 0}
and we get D-module homomorphism
Graded Cech Cohomology in Noncommutative Geometry 257
Let M be the cokernel, then

HomD(27 @ &,0 ) = Homo(&", 0 ) = &
and the ?>-module M represents E P , 3, e. i. the following short 5e-
- --
quence is exact
P
0 HomD(M,O) E F
and HomD(&,0 ) appears as the sheaf of solution of P : & .+ F.
--
In the same way for a given
P Q
& 3 Q , such that Q o P =0
- - -
and
P Q
0 c- M 27@ 278.P 27@G
is exact, then the cohomology of the resolution
E-F-Q
is isomorphic to ExtD(M, 0 ) .Therefore, E x t & ( M ,0 ) represents the ob-
struction for an element f E F satisfying the compatibility condition
Qf = 0 t o be written as f = Pg for some g E E . Thus the computation of
E x t b ( M , 0 ) become an essential problem.
Let us recall the first Spencer sequence. Let M be a 27-module with a
filtration M = U k M k and we define the D-homomomorphic derivative
6 : 27 80 A P B @ o M + ?>@o A p-16' 80 M
+ ([vi,vj]A v 1
(-l)i+jp@ . A Gi A ... A c j . A vp) 18
u.
I<i<j<p
Theorem 2.1 (D. Quillens). Assume that the diflerential system M is

endowed with a good fifiltration M = U k M k . Then locally for suficiently
large lc the first Spencer sequence associated to this filtration is exact.
258 D. N . Diep
One of the famous applications of these ideas is the Cauchy-

Kovalevskaya Theorem: Let us consider an analytic space X , Y an an-
alytic subspace with defining Ox-ideal Z.For any Ox-modules F
. set
H b l ( 7 ) = l ~ E x t ~ x ( O x / l v +7l), .
v
There is a canonical homomorphism H b l ( F )-+ H $ ( F ) . In the case where

X and Y are complex manifolds, and f : Y 4 X an analytic map,
define the sheaf of holomorphic functions onf Y x X , we have
V f = Vy+x = H"[ y l (Q(0'")
y x x ) as the sheaf of differential operators of finite
order from Y to X. In local coordinate xi of X and yj of Y , the sections
of Vy,x may be written in form
P(Y,&)= c
bllm
aa(Y)a,",
where we use multiindex notation. For any p(x) E OX, we can define
( P P ) ( Y )= c0
a4Y>[~,aP(X)lz=rp(d.
This define the product action

vy,x x f-bx -+ Oy.
It is not hard to prove that
vy,x E! Oy Xf-10, f-lDx.
Define
f * M = Vy-x @f-lvxf - l M = Oy @f-10, f-lM,

then
f * O x = Oy, f*Vx = v y + x .
Theorem 2.2 (Cauchy problem). Let f : Y + X be a smooth holo-

mophic map. For any coherent Vx-modules M and N, the map
v : f-'RHomV,(M,N) 4 RHomv,(f*M, f * N )
is an isomorphism. Moreover, for any k,
f-l Extg,(M,N) -+ Ext&,(f*M,N)
is also an isomorphism. I n the particular case where Y is an submanifold

of the complex manifold X , let T * X be the cotangent vector bundle of X
and P * X = ( T * X \ X ) / C x the cotangent projective bundle of X , TGX the

kernel of the canonical projection
T*X xx Y -+ T * Y , P$X = (T;X \ Y ) / CC P*X.

The V = , V x - m o d u l e is non-characteristic if c h a r ( M ) n PGX = 8. If M
i s a coherent V)x-module and the submanafold Y i s non-characteristic, then
M y = Oy @ox M = V ~ + @vx X M is a coherent V y - m o d u l e and
HomD,(M, O X ) ~ +
Y H o m ~ , ( M y OY)
,
i s an isomorphism.
In the next we construct almost the same construction of homological al-

gebra of C*-algebras and projective finitely generated C*-modules.
2.2. Fields of C*-algebras and Sheafification

One of the best results of J. M. G. Fell we need is the following proposition.
Proposition 2.1. Every C*-algebra of type I can be presented as the C*-

algebra r ( X ,E ) of cross-sections o n the continuous field of elementary C*-
algebras &, o n its dual X = a
It is easy to see that using the Fourier-Gelfand transforms
a H &; & ( T )= ~ (a),
we can define for each open subset U c a a closed two-sided ideal

I ( U ) := 0 kerr.
TA\U
Proposition 2.2. Let A be a C*-algebra of type I, A its dual object. There

i s a bijection between the set of open subsets in A and the set of closed
%sided ideals in A .
Proposition 2.3. There are naturally localized algebras
A ( U ) = l%A/In(U)
n
with respect to the I ( U ) - a d i c topology.

260 D. N . Diep
3. Z2-graded c e c h Cohomology
3.1. Grothendieclc Topos
The main purpose of this section is to formulate and define the functor
of (periodic cyclic) Zz-graded cech cohomology. The well-known periodic
cyclic homology is based on the cyclic homology theory of A. Connes, which
is an algebraic framework of the Zz-graded de Rham cohomology theories.
In the algebraic context, it was defined by J. Cuntz and D. Quillen in
terms of X-complexes and it has become a new chapter of noncommutative
algebraic geometry. The most general Gronthendieck algebraic geometry is
purely based in terms of categories. In the generic case this turns out to the
algebraic version of the Cech cohomology in place of de Rham cohomology
theories. We follows the work of Orlov4 in particular to formulate the theory.
The main references are Refs. 7 and 4.
Many of our results could be obtained in the fields of other characteris-
tics, but we restrict ourselves to the complex case.
Let us denote by C a fixed category and Set the category of sets. Any
contravariant functor X from C to Set is called a presheaf of sets and
the category of all presheaves of sets on C is denoted by c^. The category C
can be considered as a subcategory of c^, consisting of representable functors
h = hR : C 4 Set. If R is an object of C, then there is a natural isomorphism
Honie(hR,X) = X ( R ) . For any object X E C the category over X is the
category of pairs ( R ,a), where R is an object of C and E X ( R ) ,and is
denoted by C I X .
Recall that a sieve in the category C is a full subcategory V C such
that any object of C for which there exists a morphism from it to some
object in V is contained in Obj(V). A sieve on R is nothing more than a
subpresheaf of R in the category c^
A Grothendieclc topology 7 on a category C is defined by giving for each
object R in C a set J ( R ) of the so called covering sieves satisfying the
following axioms:
(Tl) For any object R the maximal sieve C/R is in J ( R ) .

(T2) If T E J ( R ) and f : S 4 R a morphism in C, then the induced sieve
f*(T):= (U 01Slfa E T }
is in J ( S ) .
(T3) If T E J ( R ) is a covering sieve and U is a sieve on R such that f*(U) E
J ( S ) for all f : S -+ R in T , then U E J ( R ) .
A Grothendieck site @ = ( C , I ) is a category C and equipped with a

Grothendieck topology 7 .
It is reasonable to remind the Jacobson topology on the set of all rep-
resentations of a group or Zariski topology on algebraic varieties.
For the categories with fiber product, a Grothendieck topology can be
given by a Grothendieck pretopology which is defined by giving for each
object R in C a family Cov(R)of morphism to R such that
(Pl) For any family { R , -+ R } a Ein~ Cov(R)and S -+ R a morphism of C,

the fiber product family R, X R S -+ S is also in Cov(R).
(P2) If { R , -+ R } a Eis~in Cov(R) and {Roo -+ R,}P,~J,is in Cov(R,)
for each a E I , then the total family { R , 4 R } Y E ~ , E
J , I is in Cow(R).
(P3) The trivial family { i d R : R -+ R } is in Cow(R).
Any Grothendieck pretopology P on C generates a Grothendieck topology

7 such that a sieve is covering in 7 if and only if it contains some covering
family in P .
The topos on the category of functors from a category C to another one
D is defined by the usual rule.
3.2. The Standard Cosimplicial Complex of a Continuous

Functor
We define in this subsection the &-graded Cech cohomology theory. Let
us recall the definition of the Cech cohomology with coefficients in a sheaf
M . Let U = (Ui -+ X ) i E i be a covering sieve of X in the category C.
Suppose that the cover has the property that all fiber product and pushout
diagrams exist, (see Ref. 7, Exp. 4). Denote A. the associated standard
simplicia1 complex:
presheaves on C, E : C -
Let us consider again a ringed cite (category) ( C , A ) ,C the topos of
c^ the canonical functor associating to each object
X E C the functor h x , presented by X . Define H Q ( h x M , ) = HQ(X,M),
(see Ref. 7, Exp IV, 2.3.1),as derived functor of the projective limit functor:
H q ( S ,M ) := Rq 1Lm M J s
c/s
262 D. N . Diep
For any A-module M , denote C'(U, M ) := HomA(d., M ) :
One defines Hq(U, M ) := Hq(C'(U, M ) ) . If R is a covering sieve gen-

eralized by the family U then
H q ( U ,M ) E H q ( R ,M )
and the functor W ( U ,.) commutes with restriction of scalars (see Ref. 7,
Exp. IV, Proposition 2.3.4).
One defines, (see Ref. 7, Exp. IV, 2.4)
'FIO(M)(X):= HO(X, M ) = M ( X ) ,
W ( M ) ( X ):= H q ( X , M ) .
For an arbitrary presheaf G of A-module G , the groups H q ( R , G ) are
called the Cech cohomology with respect to the covering sieve R, with coef-
ficients in G.
For a sheaf M of A-modules over C , the group
H Q ( UM
, ) := W ( U ,'FIO(M))
is defined as the Cech cohomology group of the sheaf M with respect to he
the cover U .
One has also
'Fio(G)(X)= l@ G ( R )
R-X
and therefore
H q ( G ) ( X )= l@ H q ( R , G )
R-X
which is called the presheaf of Cech cohomology

Define
P ( X , G ) := 'Rq((G)(X),
one has also
P ( X ,G ) = l@Hq(U, G).
U
For a sheaf M of A-modules, one has

kqx,M ) = kyx,'FI"M)), 3iP(M) = 3iQ(3i0(M)).
The groups H q ( X , M ) are called the Cech cohomology groups of the sheaf
M.
3.3. The P e r i o d i c Cyclic Bicomplex

Lemma 3.1 (The action of & + I ) . There i s a natural action of the
cyclic group Zk+l on the Cech cohomology cochain complex C'(U, M ) as-
sociated with a covering U .
Proof. The action of the cyclic group Zk+l is defined a cyclic permutation
of indices of U's, i.e.
(XM)(Ui, x * * a x Ui,) := M(Ui, x Ui, x . . . x U i k p 1 ) .
It is not hard also to see that for a covering sieve U = {U, -i X} there
is a natural isomorphism
M(Uio x . . . x Ui,) M ( U . , ) @ . . . @ M ( U i ,.)
Therefore the Cech cohomology complex becomes the cyclic complex for
+ M (U).
lim
U
Corollary 3.1 (Hochschild differentials and Cyclic operations).

The well-known Hochschild difjerentials b' and b and Connes cyclic op-
+ + +
erators A, N = 1 X . . . A k , s are well-defined on Zz-graded Cech
cocycles
Definition 3.1 (Periodic bicomplex). Let ( C , A ) be a ringed U-cite, C
the topos of sheaves, M a sheaf of A-modules. Then the bicomplex
is well-defined and is called the (periodic) Cech bicomplex.

Definition 3.2 (The total complex and Zs-graded Cech cohomology).
The associated total complex of which is defined as
Tot C(U,M ) * := Ci?j
i+j=f(mod 2 )
264 D. N . Diep
where Ciij :=
(il
n
,..., i k ) E l k
M(Ui, x . . . x U i k ) and f = ev (even) or od (odd).
The cohomology of this total complex is called the Z2-graded Cech

cohomology of M and denoted by Z z H * ( U , M ) and Z 2 H ( X , M ) :=
lim-; Z2H(U, M ) . It can be also realized as the cohomology of the total
complex related with the process of passing through direct limits, i.e. the
direct limit bi-complex
For a C*-algebra A, we define it Zz-graded Cech cohomology Z a H ( A ) =

Z z H ( d ) as the Z2-graded Cech cohomology of the category of *-
representations of A.
Remark 3.1. In the first periodic bi-complex without direct limits, all the
horizontal lines are acylic, but it is in general not the case for the second
periodic bi-complex with direct limits.
4. Homotopy Invariance and Morita Invariance

We prove in this section two main properties of the (periodic cyclic) Z2-
graded Cech cohomology theory: homotopy invariance and Morita invari-
ance, which make the theory easier to compute and being a generalized
homology theory.
Definition 4.1 (Chain Homotopy of functors). Let us consider two

functors F,G : C -+ 2). Denote the corresponding chain functors between
0 -1 -1 -1 -
complexes by {F,}, {G,}, where F,, G, : C, -+ D , for complexes
co c1 c
2 ...
0 __t- -F~ , G ~
Do
F~ , G ~
D1
F~, G ~
__t
We say that F and G are chain homotopic i f there exist augmentation

D2 ...
functors sn : Cn -+ Dn-l such that for all n

Fn - Gn = S, 0 &-I+ an o Sn+1,
Lemma 4.1. Two functors F,G : A -+ B are homotopic if and only if

for any covering sieve U = (Ui + X ) i E ~there
, exists a chain homotopy of
chain complexes
c
iEI i,jEIxI c
and
c
Remark 4.1. In the case of smooth manifolds the chain complex homotopy
is realized by integration of the so called Cartan homotopy formula for the
Lie derivative
Lc = z ( E ) 0 d +d 0 z(()
between de Rham complexes.
Lemma 4.2. Let A and B be C*-algebras, and let A (resp. B) be the
category of *-modules . Then the categories A and B are homotopic one-
to-another i f and only i f the two algebras A and B are homotopic.
Proof. Because of the Gelfand-Naimark-Segal theorem, the C*-algebras

are exactly defined by the category of *-representations, the category
C[O, 11 is isomorphic to the category of *-
of *-representations of B 18
representations of B . 0
Lemma 4.3. Let A be a (?*-algebra and A the category of *-representations

(i.e. A-modules) of A, then
Z & ( d ) 2 HP*(A).
Proof. Let us consider affine covering sieve Ui -+ X =A = SpecA, the

dual object of A. 0
Theorem 4.1 (Homotopy Invariance). Let pt : A -+ B,t E I = [0,1]

be a homotopy of algebras, then
Z z H * ( A )2 Z z H * ( B ) .
266 D. N. Diep
Proof. Let A (resp. B) be the category of A-modules (resp., B-modules).

Step 1. Change the homotopy by a piecewise-linear homotopy in the
space of functors from the category A to the category B.
Step 2. A piecewise-linear homotopy gives rise to a chain complex ho-
motopy.
Step 3. Two chain complex homotopical functors induces the same iso-
morphism of Cech cohomology groups. It is an easy consequence from the
results of homological algebra: If F, and G, are chain complex homotopic
than the induced morphisms satisfies
n - G:
F* = o S: + s:+~ o a;.
The second summand is a zero morphism on cohomology and the first
summand is a boundary. The sum on the right is therefore a zero
morphism.
Lemma 4.4. Let us denote by Mat,(@) the algebra of all square n x n-

matrices with complex entries. Then we have a natural isomorphism
ZZH*(Mat,(@)) E ZzH*(@).
Proof. Every complex matrix can be homotopic to a unitary one. Then,

every unitary matrix can be by conjugation reduced to a diagonal matrix
of complex numbers of module 1. Every elementary block (in this case,
diagonal element) [eie]is homotopic to identity [l]by the classical homotopy
[ei8t IO<t<l, i.e.
[sine case]
cos 0 - sin 0
-I=[:;].
Lemma 4.5 (Adjoint functors). There is a natural equivalence of func-

tors Hom and @:
Hom(R 8 M,(@), M ) 2 Hom(R, Hom(M,(@), M ) ) .
Proof. This isomorphism of functors is a particular case of the general

adjointness between Hom and @ in homological algebra.
Lemma 4.6. There is a natural isomorphism of derived functors

Rq Hom(M,(@), M ) Rq Horn(@,M ) .
Proof. It is an easy exercise from homological algebra.

Graded Cech Cohomology an Noncommutative Geometry 267
Theorem 4.2 (Morita Invariance).
Z 2 H * ( A@I Mat,(C)) % Z2H*(A).
Proof. Let us remark that A A 8 Matn(C) is a fiber bundle. Now

.--)
apply the Grothendieck's Leray-Serre spectral sequence for this fibration.

Following the previous lemmas 4.5 and 4.6, there is a natural isomorphism
of functors
RpHom(R 8 Mn(@),R q ( M ) RPHom(R, Rq Hom(Mn(@),M ) ) ,
which are the E2 term of a Leray-Serre spectral sequence converging to the

Cech cohomology. 0
Corollary 4.1. The Zz-graded Cech cohomology theory is a generalized

cohomology theory.
5. Comparison with the Classical Cech Cohomology Theory

In this section we show that a generalization of the Connes-Hochschild-
Kostant-Rosenberg Theorem can be obtained easily.
Theorem 5.1. Let A be a stable continuous C*-algebra with spectrum a

smooth compact manifold X , in fact A = C ( X , K : ( P ) ) is the algebra of
continuous sections of a smooth, locally trivial bundle K ( P ) := P x p u K:
on X with fibre the algebra K: of compact operators o n a separable Halbert
space associated to a principal P U bundle P on X via the adjoint action
of P U o n K:. Let b ( P ) E H 3 ( X ;Z ) be the Dixmier-Douady invariant, that
classifies such algebras A and c ( P ) some closed %form o n X , that presents
the class 27rid(P) in the real cohomology. Let A be the category of all *-
representations of the C*-algebra A. Then the Z2-graded Cech cohomology
Z,H*(A) is isomorphic to the de Rham cohomology H * ( X ;c ( P ) ) which is
isomorphic to the classical Z2 -graded Cech cohomology Z2G* ( X ;c( P ) ) .
Proof. In this situation, the Z2-graded Cech cohomology of the category A

is isomorphic t o the Connes periodic cyclic homology H P , ( C m ( X , L ' ( P ) ) ) ,
where C m ( X , L ' ( P ) ) is consisting of all smooth section of the sub-bundle
L 1 ( P )= PxpuL' of K ( P ) with fibre the algebra L1of trace class operators
on the Hilbert space with the same structure groups P U , see Ref. 3 for a
more detailed proof in the language of periodic cyclic homology. 0
268 D. N. Diep
Acknowledgments
T h e main part of this work was done while the author was visiting T h e
Abdus Salam ICTP in Trieste, Italy. T h e author is grateful t o ICTP and
in particular would like to express his sincere thanks t o Professor Dr. Le
Dung Trang for invitation and support.
T h e text of this paper is presented at the Second International Confer-
ence on Abstract and Applied Analysis, Quinhon, Vietnam, 4-9 June 2005.
T h e author thanks the organizers and especially Professor Dr. Sc. Nguyen
Minh Chuong for the invitation t o give this lecture.
References
1 . M. Kashiwara, Algebraic study of systems of partial differential equations,
MBmoire No. 63, Suplkment Bull. SMF, Tome 123, No. 4, 1995.
2. A. Rosenberg and M. Kontsevich, Noncommutative smooth spaces, The
Gelfand mathematical Seminar, 1996-1999, pp. 85-107.
3. V. Mathai and D. Stevenson, O n a generalized Connes-Hochschild-Kostant-
Rosenberg theorem, arXiv: math.KT/0404329 v l , 19 April 2004.
4. D. Orlov, Quasicoherent sheaves in commutative and noncommutative geom-
etry, MPI-1999-31 Preprint, 1999.
5. D. Quillen, Formal properties of overdetermined systems of linear differential
equations, Thesis (1964).
6 . A. L. Rosenberg,Compositio Math. 112,93 (1998).
7. M. Artin, A. Grothendieck, and J. L. Verdier, The'orie des topos et cohomolo-
gie des sche'mas, SGA 4, tome 2, Springer Lecture notes in Malhematics, No.
270, 1972.
514. SOBOLEV SPACES WITH WEIGHT ON RIEMANNIAN

MANIFOLDS
NGUYEN MINH CHUONGt and LE DUC THINH

18 Hoang Quoc Vaet, Cau Giay, 10307 Hanoi, Vietnam
tE-mail: nmchuong@math.ac.vn
Sobolev spaces with weight are defined on Riemannian manifolds. The class of
such spaces, even in Rn, is very wide and very interesting (see Ref. 10, p. 353).
It includes the usual Sobolev spaces of nonnegative integer orders as special
cases. Properties of such spaces are studied.
1. Introduction
Many famous geometers, mathematicians have contributed to solve the well
known Yamabes p r ~ b l e r n . To
~ > solve
~ this geometrical problem one must
use various tools, among which mathematical analysis and Sobolev spaces
are very powerful. From this geometrical problem arises a very celebrated
elliptic problem with critical Sobolev exponent^.^?^
So far, the Sobolev spaces used on Riemannian manifolds are of noneg-
ative interger orders.
The aim of this paper is to introduce Sobolev spaces with weighted
norms on Riemannian manifolds including usual Sobolev spaces mentioned
above as a special case.
2. Sobolev Spaces with Weighted Norm on Riemannian

Manifold
Let (A&, g) be an n-dimensional smooth Riemannian manifold with the
metric tensor g. In local coordinates, g has the form:
n
iJ=l
Let D be the Levi-Civita connection on Adn. D is extended uniquely to co-

variant derivative of tensor fields, denoted also by D. For f E C(M,)(lc 2
270 N . M. Chuong and L. D. Thinh
0, an integer), let us denote by Akf the covariant derivative of f of order

k:
AOf = f, A1f = of,
A'flf = D(A'f ) , (0 I T 5 k - 1).
Thus Akf is a tensor field of kth covariance. In local chart (U,'p) of Mn,
we denote by Aal...akf the components of Akf :
n
A kf = C
al,...,ak=l
Aal...ak
f dxa' @ ... @ dxak
The metric on the tangent bundle of the Mn generates the metric on all
tensor bundles over M,. So, we can define the norm of Akf , denoted by
P"I :
p k f l z = C g a l B 1 . . p D 1 (Aal...ak,)(np,.../3kf),
where (gij) is the inverse matrix of the matrix (gij). For k = 0, ]Aof I = If I.
Definition 2.1. For a general Riemannian manifold M , , we call a weight
function a positive continuous function on Mn. In the case the Riemannian
manifold is R", a positive function w is called a weight function if there
exist positive constants C and N such that:
w(E + 77) 5 (1 + cl[l)Nw(77); vc, 77 E R"
(These weight functions are introduced by L. Hormander in Ref. 8).
We denote by IC(Mn) the set of all weight functions on Mn.
It is well known that w,([) = (1 + 1<12)$, s E R, belongs to x(Rn).
If w E K ( R n ) then the functions:
belong to K(Rn) (see Ref. 8).

Definition 2.2. Let w E IC(Mn), p is a real number, p 2 1 and k is a
nonnegative integer. Let .jFp,k,w(Mn)be the vector space of all functions
f E Coo(Mn)such that w l V e f l E L p ( M n ) ,for all integers C with 0 5 C 5
k . The Sobolev space with the weight function w , denoted by Hp,k>w(Mn),
is the completion of 3p,k,w(Mn)with respect t o the norm:
k
e=o
Sobolev Spaces with Weight on Riemannian Manifolds 271
Let us denote by C?(M,) the space of smooth function of M , with

) closure of C,"(M,) with respect t o the
compact support, H i , k , w ( M n the
above mentioned norm.
For k = 0, H,,O,~(M,)= LP,,(M,) is a Banach space equipped with the
norm:
Ilfllp,w = IbJfllP.
As in Ref. 2 we can get Theorems 2.1 and 2.2 below:
Theorem 2.1. cp(Rn)is dense in Hp,k+,(Wn), that is Hi,k,w(Rn)=
HP,k,W(R").
Theorem 2.2. Let M, be a complete Riemannian manifold. Then

C r ( M , ) is dense in HP,1+,(M,).
Theorem 2.3. I f w 1 , w2 E K(M,) and there exists a constant C > 0 such
that w 2 ( z ) I c w l ( x ) , V z E M,, then Hp,k,wl(Mn)C Hp,k,wz(Mn),and the
inclusion operator is continuous.
Theorem 2.4. If w1,wg E K(Rn) and Hp,k,wl(Rn)c Hq,e,wz(Rn)f o r some
p , q, k,C, then there exists a constant C > 0 such that:
w2(z) I C w 1 ( z ) , V z E R".
Proof. Consider the inclusion operator:
I : Hp,k,wl(Rn)L,H q , e , w z ( R n ) .
The operator I has closed graph. Indeed, if u, -+ u in H p , k , w l ( R n ) and
Iu, 2, in Hq,e,wz(Rn)
---f then u, --+ u in Lp,wl(Rn)and Iu, --+ Y in
Lq,wz(Rn). Consequently there exists a subsequence denoted also by {u,}
such that u, -+ u almost everywhere, and u, = lu,
where. So u ( z )=).( almost everywhere in R", and Iu = By the closed ..
Y almost every-
graph Theorem, the operator I is continuous. Therefore there exists a con-

stant A > 0 such that:
Ilfllq,t,wz 5 AllfI l p , k , w l , vf E Hp,k,wl(R") (1)

Fix a function f E CF(Rn), f $ 0. For each q E R", consider a function
f q , defined by:
f,(O = f(< - 7 ) .
We have:
D"f,(E) = (D"f>(E- q),V'I..

272 N . M . Chuong and L. D. Thinh
so:
Since f E CF(R"), we get f, E CF(R"), thus f, E Hp,k,wl(Rn).

On the other hand:
Hence:
Theorem 2.5. Assume that w l , w2 E IC(Rn) and p > n. If -

wz(') --t 0 as
W 1 (El
< 4 then the inclusion operator C r ( R n ) n Hp,2,w,(Rn)L-) Lp,wz(RWn)
00 is
completely continuous.
Proof. Let {fi}i?l c C r ( R n ) n Hp,2,wl(Rn)be an arbitrary sequence

such that 1 lfil lp,2,wl I1,V i . We shall prove that there exists a subsequence
converging uniformly on every compact subset of W".
Let K be an arbitrary compact set. Then there exists an open ball S such
that K C S. The closure S of S is a n-dimensional compact Riemannian
manifold with smooth boundary in W". Since p > n , applying Sobolev
imbedding Theorems to compact manifold with boundary (see Ref. 2), we
see that H z ( S ) is imbedded into C1(S). Consequently there exists a positive
constant C such that:
Note that:
e=o
By the continuity of w1, we have:
Hence the sequence {fi} and { IAfil} are uniformly bounded on K , so the
sequence {fi} is uniformly bounded and is equicontinuous on K . Since
Rn is separable and countable a t infinity, the sequence {fi} possesses a
subsequence converging uniformly on every compact subset of Rn. This
subsequence, may be denoted by {fi}. Given any E > 0, by wz([)/w1(<) 40
as [ -+ 00,there exists a large enough closed ball S such that:
By Minkowski incquality, we obtain:
where C = supwz
- > 0. Since llfi - fjllp,2,wl I2 and the sequence {fi}
S
uniformly converges on S we get:
So the sequence {fi} is Cauchy in Lp,wz(Rn)and because LPIW2(Rn)

is
(EXn).
complete, the sequence {fi} converges in Lp,wz 0
Remark 2.1.
(1) With the assumptions of the Theorem 2.5, it is obvious that the in-
clusion operator CF(Rn) n Hp,k,wl(Rn)-+ Lp,wz (Rn)is completely
continuous, for all integers k 2 2.
(2) When Mn is a Riemannian manifold, it is also a metric space that is
separable and countable at infinity. But a ball S in M,, is not necessary
to be a Riemannian submanifold with boundary, so we cannot apply
the Sobolev embedding theorem as above to S. However, the above the-
orem holds true for the class of Cartan-Hadamard manifold, that is, a
simply connected complete Riemannian manifold with non-positive sec-
tional curvature. By the exponential map at any point such a manifold
is diffeomorphic to an Euclidean space of the same dimension. Conse-
quently a closed ball in a Cartan-Hadamard manifold is a Riemannian
submanifold. So we get:
Theorem 2.6. Assume that M , is a Cartan-Hadamard manifold and

~ 1 E IC(Mn).
~ 2 >
~ If p and -W 2 ( Q ) --+ 0 as d(P,Q ) 4 c o ( ~
TZ i s a fixed
wi ( Q )
n
point of M,), then the inclusion operator f r o m C r ( M n ) Hp,2,wl(Mn)to
Lp,w2(M,) is compact.
Theorem 2.7. Assume that w l , w 2 E K(Rn)and the embedding operator
is completely continuous for reals p , q 2 1 and
from C?(Rn) to L4,wz(Rn)
for some non-negative Ic. Then -
w2(') -+ 0 as E --+ oo.
Wl(E)
Proof. It is sufficient to prove that if{,$,}u21 is an arbitrary sequence such

w2 ( t u 1
that Eu -+ 00 as u -+ 03 then --+ 0 as v -+ 00.
Wl(EU)
Let us fix a function f E C r ( R n ) ,f $ 0 . Consider sequence {fu} defined
by:
From the proof of Theorem 4 we have:
From the first inequality it follows that the sequence { f u } is bounded in

H P , k , w i (R")nC?(Rn). The assumptions of the theorem give us the relative
compactness of the sequence { f,,}in Lq,wz(R,). So there exists a compact

set G c R" such that:
where E is any positive number (see Ref. 1, pages 31-33). By the Minkowski
inequality, we obtain:
Wn\G ' G
Since G and supp f are compact sets, for u large enough, we have x - <,, q!
supp f , V x E G. So:
-(/w$(x)lf(z-&,)lqdx)'
1 +Oasu-+cc
wl(ru)
G
that implies 11 fVllq+,, -+ 0 as u + 00, and by the second above inequality
we obtain:
Similarly to the proof of Proposition 2.11 from [2] (pp. 36-37), we get:
1
Proposition 2.1. If Hqo,l,u(Mn)is imbedded into L,,,w(M,) for - -
Vn
In [2], Lemma 2.22, T. Aubin proved that:

SUP IVI I C(~)IIVIIq,l,~V
E C,"(Mn),
here q > n and C(q) is a constant, M , is a complete Riemannian manifold
having positive injectivity radius and upper bounded sectional curvature
( K 5 b').
If M , is a compact Riemannian, then all these above conditions are
satisfied, so we get the desired inequality.
However, in the case of a compact manifold M,, another proof will be
shown.
Theorem 2.8. Let Mn be a compact Riemannian manifold. Then there
exists a positive constant C such that sup I f 1 5 CII f l l q , l , V f E C,oO(Mn).
Proof. Since Mn is compact, there exists p > 0 such that V P E M,, the
polar coordinates exist on the closed ball:
BP(P)= { Q E MnI4C Q ) IP}

(see Ref. 9, p. 22).
Because of the compactness of Mn, there exists also the points Pi E
(-)}+,
Mn(i = 1,..., N ) such that {Bpi P N is an open covery of M , (Bp, P (-)
4 4
P
is an open ball of radius - with center Pi).
4
Let {ai}Ll be the unity partition corresponding to the above open
covery. It will be shown that for each i, there exists Ci > 0 such that:
suplazfl 5 ~ z l l ~ ( ~ z f ) l l qE, CF(Mn).

V~
Set f , = a , f , K , = suppa,. Obviously there exists Q, E K , such that

suplazfl = Ifi(Qz)l.
Consider a closed ball B p of radius p with center 0 in Rn. The expo-
nential map ezpp, a t P, is a diffeomorphism from Bp onto Bp,(p). The
function h ( z ) = f,(ezpp,(z)) is defined on B p . Set zo = ezp;%'(Q,). Let
( r ,8 ) be polar coordinate in Rn with center a t 2 0 . For every z E R",sat-
2P P
isfying d(z0, z) = -,
we have d ( 0 , x) > - and d ( 0 , x) < p, so z E Bp and
3 4
h ( z )= 0. Consequently:
h(zo)= - j h h ( r ,8)dr.
0
Hence:
Ifi(Qi)l = Ih(z0)l 5
f
0
I&h(r, 8)ldr.
Integrating of the two sites of this inequality over S"-l with respect to 8
yields:
1
un-llfi(Qi)l I I&h(r, 8)ldrd8,
B
where B is a closed ball of radius -,23P with center at z o in R" and u,-1 is
1 1
where dE is the canonical volume element in Rn and -
4
+ - = 1. Since Ki
4'
is compact there exists constants A, p with 0 < X 5 p < co,such that:
X1tI2 IC S i j ( P ) t i t j i plt12,Vt = (tl,...,tTL)
E R",V'P E Ki.
i,j
Consequently:
d~ I X - ? ~ V ,l ~ h I
l p f l ~ f i ol ezppi
on K i , where dV is the Riemannian volume element on Mn. Hence:
We can choose:
Therefore, Vf E C p ( M n )we get:

N
i=l i
N
References
1. R. A. Adams, Sobolev spaces, Academic Press, (New York, 1975).
2. T. Aubin, Nonlinear Analysis o n manifolds, Monge-Ampere Equations,
(Springer-Verlag,New York, 1982).
3. T. Aubin, A Course in Differential Geometry, AMS, Providence, Rhode Is-
land, (2000).
4. T. Aubin, Bull. Sci. Math. 2e SCrie, 100, 149 (1976).
5. H. Brezis and L. Nirenberg, C o m m . Pure A p p l . Math. 36,437 (1983).
278 N . M. Chuong and L. D. Thinh
6. N. M. Chuong and T. D. Ke, Existence and nonexistence of solutions f o r

a semelinear degenerate elliptic system, J. Abstract and Applied Analysis
(submitted).
7. S. Gallot, D. H u h , and J. LaFontaine, Riemannian geometry, (Spinger-
Verlag, 1993).
8. L. Hormander, Linear partial diflerential operators, (Springer-Verlag, Berlin,
1976).
9. Jurgen Jost, Riemannian goemetry and geometric analysis, (Springer, 1998).
10. H. Triebel, Theory of functions spaces, (in Russian) Akdemische Ver-
lagsgesellschaft Geest & Portig K. G, (Leipzig Birkhauser Verlag, Basel-
Boston-Stuttgart, 1986).
Chapter IV
STOCHASTIC AND
INFINITE-DIMENTIONAL ANALYSIS
$15. STOCHASTIC POPULATION CONTROL

AND RSDE WITH JUMPS
SITU RONG
Department of Mathematics, Zhongshan University,
Guangzhou 510275, China
E-mail: mcsstr@zsu.edu. cn
We obtain and improve results on the existence, convergence, stability of so-

lutions to reflecting stochastic differential equation (RSDE) with jumps and
apply them to discuss the properties of stochastic population solutions and
finally solve the optimal stochastic population control problem.
1. A Deterministic Population Dynamical System

Consider a population dynamics system. Let us denote by S: the size of the
+
population with age between [i,i l), and T, - the possible largest age
of the people. Suppose that the year variable is discrete as t = 1 , 2 , . . ..
Then as time t evolves, the size xi,i = 1 , 2 , . . . , r m also changes. Then
intuitively, the increment of Ax: caused by the time increment At = 1 can
be expressed as
nx;= [-(1 + v;)z; + c:L1-,
b:x;pt] nt ,
{ Ax: = [-(1 + +)xf + S ; - ' ] A t ,

i = 2 , . . . lrm;
where ,Ot is the specific fertility rate of females, b: = (1 - p!)kjhi # 0,

(1)
i = rl, . . ., 1-2; 1 < r1 < T 2 < r,; p: is the death rate of babies, 7: is the
forward death rate by ages, kj is the sex rate, hi - the fertility model,
and [rl,rZ] is the age interval that women can give babies. The absolute
+
value (1 vt)xi of the first term in the right hand side of the first equation
in (1) is the amount of persons who will leave the age interval [1,2)in
+
the next year t 1, and the value c:& b:x$t of the second term is the
amount of persons who will arrive a t age interval [l,2) in the next year tt 1.
Similar physical meaning can be made for Ax:, 2 5 i 5 r,. If one considers
that the time variable changes continuously, and writes the corresponding
282 S. Rong
ordinary differentialequation in a matrix form, then he will get the following

deterministic population control system:
dxt = (Atxt + Btxt,Bt)dt (2)
50 =x, t E [O,T],
where xt = (xi,.. ., xEm),and
Bt= [ 0 .. 0 bL1
0 .... ..
.. b y 0 .. 0
1
.. .. .... 0 . (4)
Notice that the size of the population should be non-negative, so one

should require that
Under appropriate conditons (5) is valid.
2. RSDE for Stochastic Population Dynamical System

In case that the size of the population is disturbed by some stochastic per-
turbation the situation is quite different. For example, in the simplest case
even it is disturbed by a standard Brownian Motion process (BM), caused
by the strange properties of the BM (5) still cannot be guaranteed. So, we
need to consider the reflecting stochastic differential equations (RSDEs) as
the stochastic population dynamic system model. Let us first introduce a
more general d-dimensional RSDE with jumps as follows: (Here we write
"RCLL" for "right continuous with left limit" .)
r dxt = b ( t ,Z t , w)dt + u(t,~ tw)dwt
, + Jz ~ ( ~t ,t ,Z- ,w ) ~ (Id,t ,d ~+) d+t ,
" 0 = z, xt E $,for all t 2 0 ,
where R$ = { z = (zl,...,zd) ' >0,l 5 i 5 d ,
E Rd : Z
+t is a Rd - valued & - adapted RCLL process with finite variation
1
(6)
I+l t on each finite interval [0,t] such that 40 = 0, and
I+lt Jot I a R p s P 1 4 s
= 1
+t = 1
;4 s ) d I& >
,n(t)E Nz,,
as ~t E aR$,
Stochastic Population Control and RSDE with Jumps 283
where w t is a d-dimensional BM, and f i k ( d t , d ~ is) a Poisson martingale

measure generated by a zt-Poisson point process k(.) with a compensator
7r(dz)dt, T ( . ) is a a-finite measure on a measurable space (2,R(2)) such
thatI
N k ( d t , d z ) = N k ( d t , d z ) - n(dz)dt,
and
N, = u,.+oN,,,.,Nx,r= { n E Rd : In1 = 1,B ( x - nr,r ) n 0 = S} ,
B(xo,E)= {y E Rd : 19 - 201 < E } .
Condition l$lt = I a R t ( x 3 ) d in (6) obviously means that l$ls in-
s,'
creases only when x, c aRd+,and the condition +t = J,"n(s)d141s means
that the reflection dq5t happens along the direction of the inner normal
vector n ( t ) ,so we may call such a reflection a normal reflection.
The geometrical meaning of N, is that it is the set of all inner normal
vectors a t x,when x E OR$. Obviously, when x E Rd,,by definition N, = CP
(the empty set). Actually, in the case that xt E Rd, no refelction is needed,
so we do not need an inner normal vector.
More precisely, for ( 6 ) , by definition as x E Rd+,
N,= CP (empty set).
Notice that
d
aRd, = '$=I Uil<a2<...<ik,ilra2,... ,ik=l Ak,
where
Ak = { ( 2 1 ~ x 2 ,... ,xd) Xi, = 0, Xj > 0, v j # i,, 1 5 p 5 k }
So, if x E Ak, then n = ( 7 2 1 , 1 2 2 , ' . . , nd) E N,, iff, for k 2 3,
nil = sin8coscpl coscp2.. .. . . . . . (20s(Pk-21
ni, = sin 8 cos (PI cos cp2 . . .cos (Pk-3 sin ( P k - 2 ,
nik-, = sin 8 cos cp1,

nikPl= sin 0 sin cp1,
ni, = cos 8, 8 E [o, ; ] 1 (P11$?2 ' ' ' I (Pk E [o, $1,
nj = 0,as j # i l l i z , . . . ,i k ;
and for k = 2,
nil = cos 8, ni, = sin 8, 8 E [O, $1,
n k = 0 , V k # i1,iz;
and for k = 1,
nil = 1, nj = 0,Vj # il.
From this one sees that the inner normal vector a t a boundary point
of the domain Rd, is not necessary unique. However, from (6) one knows
that there exists a unique n(t)such that as xt E dRd,, n(t) E N,, and the
284 S. Rong
reflection d d t happens along the dircction of the inner normal vector n(t).
Since all n, 2 0, i = 1 , 2 , . . . ,d. So we have the following Claims.
Claim 2.1. all &, i = 1 , 2 , . . . ,d, in (6) are increasing, where

$t . ,4 3 .
= (4L 4L..
Claim 2.2. If (xt,d t ) satisfies (6), then t 2 0 ,

t
so
zp4; = 0, i = 1 , 2 , . . . , d.
Claim 2 . 2 is easily obtained by (6) and by the property of N,.
From now on we always make the following Hypothesis on the jump
term:
( H ~ ) : z + c ( t , z , z , w ) ~f zo r, a l l t 2 0 , z E Z , w E R a n d x E R d , .
(Hz) ~ ( d z =) d z / Izld+', ( 2 = Rd - (0)).
The meaning of (HI) is that if the state xt- before the jump lives in Rd,,
+
then after the jump, the state ~t = xt- c ( t , Z,z , w ) still stays in nd,. The
Hypothesis (H2) gives a more concrete measure for the compensator of the
jump Poisson point process. Notice that under assumption (HI) if ( z t ,4t)
satisfies (6), then & should be continuousli Lemma 329. So under assumption
(HI) in (6) we may write that
q5t is a continuous function. (7)
Now for the solution of RSDE (6) with (7) we introduce the following
definition.
Definition 2.1.
-
1) We say that ( x t , 4 t ) is a strong solution of (6) with ( 7 ) ,if xt is RCLL
and $t is continuous and they are 3y1Nk -adapted, where
3W'fi'" -
t U ) ;s I t , u E B(Z)),
= d W s , N k ( ( 0 ,s),
and (xt,&) satisfies (6) with (7).
2) We say that the pathwise uniqueness for solutions of (6) with (7) holds,
if for any two solutions (xi,&), i = 1 , 2 , of (6) with ( 7 ) ,which are defined
on the same probability space (fl,3,(St),P ) with the same BM wt and
Poisson martingale measure ( d t ,d z ) ,
P(SUPtL01x2 - x;l = 0) = 1, P(sup,20 14; - qq
= 0) = 1.
3. Existence of Solutions to RSDE with Jumps

We can have the following general existence theorems of solutions for RSDE
with jumps which are some generalization of Theorem 335 and 336 for
0 = R$ in Ref. 1.
Theorem 3.1. Assume that for t 2 0 , x E @f, z E Z

1" b(t, x),a ( t ,x),and c ( t ,x , z ) are non-random, jointly measurable; more-
over, b(t, z), a ( t ,x) are jointly continuous, and as 1% - yl -+0 , It - sI 4 0
Jz Ic(t, 5 ,). - 4%Y , .)I2 .rr(dz) 0, -+
furthermore, b(t, x),o(t,x) and Jz Ic(t, x,. ) I 2 7r(dz) are locally bounded for
x; that is, for each r > 0 , as 1x1 5 r ,
+
I b ( t , . ) I 2 + Ila(t, %)112 JZ Ic(t, x,z)I2 .rr(dz)I ICT,
where kT 2 0 is a constant depending only on r ;
2" there exists a real function V ( t , x ) E C112([0,03) x R d ) ,i.e. it has a
first continuous derivative with respect to t and up to second continuous
derivatives with respect to x, such that one of the following conditions is
satisfied:
+ J Z [ V ( t2, + c ( t ,5 , z ) )- V ( t ),. - EL Ci(t, x,z)&V(t, 2)17r(dz)

ICl(t),
where c l ( t ) 2 0 , s,'cl(t)dt < 03,'v''T < 03;
(ii) LV(t,x) 5 c l ( t ) V ( t ,x),
where c l ( t ) has the same property as in the above (2);
3 there exists a a l ( t ) with al(lt1) 2 0 such that a l ( t ) 00, as t T
' 03; and
as 1x1 is large enough,
V ( t ),. 2 a1(lzl),
where V ( t x) , is the same function occured in 2'; and as xi 2 0 ,
L i V ( t ,Z) I k o ~ i i, = 1 , 2 , . . . , d;
-
4" 2(. -Y) . ( b ( 4).

- b ( t ,Y ) > +
Ila(t,). - a(t,dI12
+ Jz I C ( ~2, , z ) 2
c ( t ,Y , z)12~ ( d zI) ~ N , T ( ~ ) P N , T-(YI I~1,
-
as 1x1 IYI I N , t E [O,TI,

1
where for each T < 03, 0 I k N , T ( t ) , l T k N , T ( t ) d t < 03 and P N , T ( U ) i s

concave and strictly increasing in u 2 0 such that P N , T ( O ) = 0 , and
So+ d u / PN,T ( u )= 03, for each N = 1 , 2 , ....
Then (6) with (7) has a pathwise unique strong solution.
Theorem 3.1 has the following obvious applications.
Corollary 3.1. Under assumptions 1" and 4" in Theorem 3.1 if, further-
more, there exists a constant ko 2 0 such that
2 (x,b ( t , .)) +
Ila(t,.)1 2 +
J, Ic(t, 2 , .)I2 T ( d 2 ) 5 ko(1 + 1xI2),
then (6) with (7) has a pathwise unique strong solution.
286 5'. Rong
Corollary 3.1 obviously includes Theorem 335 for 0 = Rd, in Ref. 1. To

show this Corollary by Theorem 3.1 one only needs to take V ( t , z )= 1 +
1212 a l ( t ) = t 2 ,c l ( t ) = Ico. Then one easily sees that under the assumption
of Corollary 3.1
C V ( t ,). I Cl(t)V(t,x),
and Theorem 3.1 applies.
Corollary 3.2. Under assumptions 1' and 4" in Theorem 3.1 if, in addi-
tion,
nlZ"=,
( X I b ( t l .)) I C l ( t ) ( l + 1xI2 gk(x))l
Il4tl 4 1125 C l ( W + Ixl2nE1 gk(x))l
, I C ( t l 5 1 .)I2 4 d z ) I C l ( t > l
J
where
g k ( x )= 1 + l n ( l + l n ( l + . . . l n ( l + 1 x 1 ~ " ~ ) ) ) ~
k-times
(m and no are some natural numbers), and c l ( t ) 2 0 is non-random
s'
such that for each T < 00, c l ( t ) d t < 00;
then (6) with (7) has a pathwise unique strong solution.
Corollary 3.2 obviously includes the statement of the second part of
Theorem 336 for 0 = Rd, in Ref. 1. To show this Corollary by Theorem 3.1
one only needs to take V ( t , x )= gm+1(2), a l ( t ) = gm+l(t), c l ( t ) = c l ( t ) .
Then one easily sees that under the assumption of Corollary 3.2
q t ,). I k O C l ( t ) I
where Ico 2 0 is a constant. So Theorem 3.1 applies.
Now let us prove Theorem 3.1.
Proof. Let
where f = b,a;&nd
Then by Theore; 334 in Ref. 1 there exists a pathwise unique strong so-
lution (zpl
#) satisfying RSDE (6) with (7) but with coefficients b N , oN
and c N . Let
T ~= inf{t
, ~ 2 o : Ixrl > k} .
Then by Ito's formula one easily sees that under the assumption (ii) of 2"
in Theorem 3.1 for any given T < 00 as 0 I t I T
EV(t A xKrN,N) = E V ( 0 , X O ) E J o
rNIN, + tArN*N
L V ( S ,x,N)ds.
< kh
- + Ji
C l ( S ) E V ( S A T N ' N ,2 f A r N , ~ ) d S .
Hence as 0 5 t 5 T
EV(t A x E r N , N5) k,eI J 0T C l ( t ) d t
rNiN,
1
and
a ( N ) P ( r N J "< T ) 5 AT, z~N,N,,,T)I~N~N<T
~ ~ ( 7 ~ 1 ~
I EV(rNyNA T ,x,"N,NAT)I k,eI Jc:1 ( t ) d t *

Therefore, as N ---f 00,
P(rNiN< T ) + 0.
For the case that assumption (i) of 2" in Theorem 3.1 holds, the same
result is similarly derived and the derivation is even simpler. From this
result one easily sees that P - a s s .limN,, rNvN= co.Now set
xt = xtN , dt = @, as o _< t < T ~ , ~ .
It is easily seen that ( x t ,$ t ) is the pathwise unique strong solution of RSDE
(6) with (7). 0
4. Stochastic Population Solutions and Their Properties

Now let
where At and Bt are defined by (3) and (4), and let d = r,, then ( 6 )
with (7) and (8) becomes the stochastic population dynamic system. By
Corollary 3.1 we immediately obtain the following result.
Theorem 4.1. Assume that coeficients (T and c satisfy conditions 1" and
4" in Theorem 3.1, and assume that At and Bt are bounded and ,L$(x) is
bounded and jointly continuous such that
IPt(x) - Pt(y)I I ~ N , T ( ~ ) P N , T ( -
I XY I ' ) ,
as 1x1, lyl 5 N , t E [O,T],W'<00;
where kN,T(t) and P N , T ( U ) have the same properties as those in the con-
dition 4" of Theorem 3.1. Then the population SDE, i.e. (6) with (7) and
with the coeficient (8), has a pathwise unique strong solution.
A simple case for u and c t o satisfy the above conditions in Theorem

4.1 is given as follows:
288 S. Rong
Corollary 4.1. Let
i u(t,x ) = co + ( c p x , .. . , Cid)x)
+
c ( t , x , z ) = (CO G z ) f ( z ) ,
(9)
where Co, Cii),i = 1,* . . ,d and 7 7 1 are constant d x d matrices, and 770
is a constant Rd-vector, (i.e. all elements in the matrices or vectors are
constants), and , in addition, all elements in 770 and 771 are non-negative,
i.e. ?$,i$ (.A,..
2 O , i , j = 1 , 2 , . . . , d ; where 770 = -d T , 771 - (- icj ~d ) ~ ,
,co)
and f ( z ) 2 0 is such that J R d - { o ) S d z < 00. Then one easily checks
that assumption ( H I ) and all conditions for u and c in the above theorem
hold. So Theorem 4.1 applies.
Let us discuss the properties of the stochastic population solutions
1)Convergence Property. First we can consider the more general RSDEs
with jumps. If (z:, @) satisfies (6) with (7) and with coefficients bn(t, x ) ,
un(t,x ) , and cn(t, x , z ) and with the initial condition x; as well, then we
have the following convergence theorem.
Theorem 4.2. Assume that

1' Ibn(t, .)I2 + + +
/Inn@,x)1I2 J, I C n ( t , x , z ) I 2 T ( d z )Iko(1 1xI2),
n = 0 , 1 , 2 , .., where ko 2 0 is a constant;
2" there exists a real function 0 5 V ( t ,x ) E C1i2([0,00) x R d ) such that
and there exists a a l ( t ) with al(lt1) 2 0 such that a l ( t ) ,t 2 0 is strictly

increasing, and
V ( t ,x ) L a1(lxl),
moreover,
where 1g(y)l 5 k o , and c l ( t ) 2 0 , SoT cl(t)dt < 00,'v'T < 00;

3"
l i % + c o ~ ~ [ S U p xIb"(s,z)
~~d
2
- b o ( s , z ) ) + S U P X E R ~ ( ~ ~-~ ~ ( S I ~ ) ao(s~z)II 2
+ J
2
, Icn(s,z, z ) - co(s,z, z)I .rr(dz))]ds = 0,f o r all t > 0;
4" zt z:, as n -, 00,
where the initail values xz,n = 0 , 1 , 2 , * - . are all non-random constants.
T h e n for all t 2 0
sup,,, V ( t ,zy - z:) = 0.
I n particular, for any E > 0 and t 2 0
I n addition, assume that

5" limlQ-xl+OSUPs<t lgO(s,). - g0(s,Y)I = 0,
where g = b, a; and
lim
IY--zJ+O
sup
s_<t s, Ico(s,z, z ) - co(s,y, .)I 2
~ ( d z=) 0.
Proof. By condition 1" and using Ito's formula one easily obtains that1
EsuplzTl 5 kT, h = 0 , 1 , 2 , * * * ,
t<T
where kT 2 0 is a constant depending on T < 00 only. Now applying Ito's

formula to V ( t ,zy - zf), one has that as t 5 T
where
+ s, p ( S , z, z ) - c y s , 5 , .)I 2
T(dZ)]dS}.
290 S. Rong
So, by Gronwall's inequality'! Lemma 11*
5 V ( 0 ,J: - J::) + f ( t )+ I' eJt c l ( s ) d s c l ( sf) (s)
Hence the 1st and 2nd conclusions are easily derived.

[ + V ( 0 ,x; - z:)]ds.
Notice that
d ( x y - $) = (bn(t,J::, W ) - bo(t,J::, w))dt
+ (a"@, J::, W) - ao(t,J::, w))dwt
+s, (c"(t,J:?-,
+ 44: -43.
2 , ~-
) Co(t,J:y-, z , w ) ) f i k ( d t ,d z )
From this the last conclusion is also easily obtained.
Corollary 4.2. Under assumptions 1",3" and 4" in Theorem 4.2 if

6" 2(x - Y ) * (bO(t,J:) - bO(t,Y)) IIflO(t, +
J:) - aO(t,Y)1I2
2
+J~Ico(t,J:,z)--o(t,Y,Z)12T(dZ) I c 1 ( t ) l J : - y l 1
where c l ( t ) 2 0, c l ( t ) d t < 00, V T < 00; then f o r all t 2 0

2
limn+m sup,^^ (J:: - J:!() = 0.
I n addition, i f 5" in Theorem 4 . 2 is also satisfied, then for all t,E 2 0,
sup,^^ 14; - > E ) 3 0, as n --f 00.
To see that the first conclusion in Corollary 4.4 is true, one only needs
to take V ( t , z ) = (zI2, a l ( t ) = t2 and applies Theorem 4.4.The second
conclusion can be proved directly.
Corollary 4.3. Under assumptions l o ,3" and 4" in Theorem 4.2 if
(x - Y ) * (bO(t,J:) - bO(t,2/11 L C l ( t > 12 - Y I 2 r I L g k ( Z - Y)1
2
t , - " O ( ~ , Y ) / \ ~ + /z \c0(t,2 , ~-) co(tly ,
( ( a o ( J:) .)I
T(dz)
5 C l ( t ) IJ: - Y12 n:=.=, 9k(J: - Y ) ,
where c l ( t ) has the same property as that in Corollary 4.4, and gk(x) is
defined in Corollary 3.2; then for all t 1 0
Esupslt gm(z: - J::) = 0,
and for all t , E 2 0 ,
supslt P ( ~ z - : x:( > a) 3 0, as n -+ 00.
I n addition, if 5" in Theorem 4.2 is also satisfied, then f o r all t ,E 2 0 ,
~(14: -&I > c ) 4 0, as n +. 00.
To show Corollary 4.4, one only needs to take V ( t ,x ) = g m ( x ) , a l ( t ) =

g m ( t ) and applies Theorem 4.4.
Applying Corollary 4.2 we immediately have the following result for the
convergence of solutions of population SDEs.
1" there exists a constant ko 2 0 , for all n = 0 , 1 , 2 , ...
lA"(t)l + lB"(t)l + IP,"I I ko,
where A"(t), B n ( t ) ,and pz" are non-random such that they do not depend
o n x , and P ( t ,x ) and c n ( t ,x , z ) satisfy the condition 1" in Theorem 4.2;
2' oo(t,x ) and co(t,x , z ) satisfy the condition 6" in Corollary 4.2;
+
3" limn-+mJ:[IA"(t) - A o ( t ) ( ) B n ( t )- B o ( t ) f ) (pp- @l]dt = 0,
and the condition 3" in Theorem 4.2 for d ' ( t , x ) , c O ( t , x ) and
c"(t, z, z ) ,co(t,x , z ) holds;
4" the condition 4" in Theorem 4.2 holds.
Then for all t 2 0
l i m + m E(supslt Iz: - IC, 1
0 2 ) = 0.
I n addition, i f 5" in Theorem 4.2 is also satisfied, then f o r all t,E 2 0 , as

n-+m
P(SUP,<t 14; - 4:1 > E ) 0. +
This indicates that if coefficients in the population SDE can be ap-

proximatly calculated, then solutions of the population SDE also can be
approximatly (in probability) obtained. For more convergence results on
the solutions of population SDE one can refer to Ref. 1.
2) Stability Property.
For the stability of solutions to RSDE (6) with (7) we have the following
theorem.
Theorem 4.4. Assume that conditions l a ,3', 4" in Theorem 3.1 hold,
and assume that condition 2" in Theorem 3.1 is changed to be
2"' : there exists a real function 0 5 V ( t ,x) E C'v2([0,m) x R d ) such
that
C V ( t ,x ) I -.1(t)V(t, x),
where q ( t ) 2 0 , c l ( t ) d t = 03, and as xi 2 0 ,
V ( t , X) 5 k o ~ i i, = 1 , 2 , . . . ,d.
"ax,
Then (6) with (7) has a pathwise unique strong solution, which satisfies
that
E V ( t ,x t ) 5 e- Ji~ ~ ( ~ ) ' s E 2v 0()o. ,
Furthermore, if b ( t , 0 ) = 0 , a ( t ,0 ) = 0 , c ( t , 0 , z ) = 0 , and there exzkts a
real function a l ( t ) 2 0 , t 2 0 , which is increasing and a l ( t ) > 0 , as t > 0 ,
292 S. Rong
such that
V ( t ),. L a1(Ixl),
then (0,O) is the pathwise unique strong solution of (6) with (7) and with
the initial condition zo = 0; and f o r any given E > 0 ,
limt+oo P ( l x t ( > E ) = 0,
i.e. the str0n.q solution of (6) with (7) is stable in probability.
Proof. First, by Theorem 3.1 (6) with (7) has a pathwise unique strong
solution for any initial condition zo. Now applying I t 0 3 formula t o V ( t ,xt),
one has that as t 5 T
E V ( t ,Z t A T N ) 5 V ( 0 , x o )- J
; C l ( S ) E V ( S ,G A T N ) d S ,
where TN = infjt 2 0 : lztl > N}. By using Gronwall's inequality and
Fatou's lemma the first conclusion is obtained. Secondly, under additional
condition one has that for any given E > 0
al(&)p(lxtl> &) 5 E V ( t ,xt) I e- JiQ ( ~ ) ' s E v ( o5 ,0 ) .
So the second conclusion is also obtained. 0
An immediate corollary is as follows:
Corollary 4.4. Assume that l o ,4" in Theorem 3.1 hold, and

+
2 (x,b ( t , x)) + Ila(t,4112 Jz Ic(t, 2, z)I27r(d4 L -c1(t) 1bI27
Jr
where c l ( t ) 2 0 , c l ( t ) d t = 00.
Then (6) with (7) has a pathwise unique strong solution, which satisfies
that
E 1xtI2. < I x01 .
- e-s," cl(s)dsE (10)
Corollary 4.4 obviously follows from Theorem 4.4, if we take V ( t ,x) =

1x12,and a l ( t ) = t2.
Now let us apply the above Corollaries to give the stability results on
the stochastic population dynamics (6) with (7) and with b ( t , x ) = Atx +
BtxPt(z).Suppose that a ( t ,0) = 0 and c ( t ,0, z ) = 0 then under conditions
of the existence Theorem 4.1 (6) with (7) and with the coefficient (8),
has a pathwise unique strong solution (0,O) with 20 = 0. Furthermore, if
2
2 ( A t z , x ) + 2 (B,P,(2)2,2)+lla(t,2)/I2+JZ I c ( ~ , ~ , z ) I~ ( d z F) - c I ( ~ )
12121
where c l ( t ) 2 0, cl(t)dt = 00; then (10) holds, i.e. the solutions of

population SDE is exponentially stable in the mean square. More precisely,
if a ( t ,x) = (C:'),. . . ,C:'..") z,c ( t ,z,z ) = C 1 z I u ( z ) , where T ( U ) = 1,
and Ciz),i = 1,.. . , d ; are defined in Corollary 4.1, and 0 I: a ( x ) 5 PO.
Let
Stochastic Population Control and RSDE with J u m p s 293
60 = mintlo(; + rll(t),7 7 2 ( t ) , . . . ,r l T m - l ( t ) , ;+ 77Trn(t)),
b M = maxt>O(brl( t ) , 7 h-2 (t)).

Then as
2
61 = 260 - c:z1IIc,(i)II - 1C11 2
- PObM(r2 - r1+ 1) > 0,
E 1ztI2I E 1z0l2e-61t, for all t 2 0.
That is, the population dynamics of (6) with (7) and with the coefficient
(8) is exponentially stable in the mean square under the above assumptions.
In particular, for a given constant a > 0 we easily derive from the above
inequality that
t> +f
- ~61 l n ==+ E 1ztI2I a.
This indicates that if the stochastic perturbation is not too large (that
is, xf=llICii)II 2
+IF1 1 is very small), and the forward death rate is greater
than a positive constant (that is, 60 > 0), then the stochastic population
dynamics (6) with (7) and with the coefficient (8) can be exponentially
stable in the mean square if the fertility rate of females is small enough.
Furthermore, given a > 0, we can find out when the population size vector
in the mean square can be less than this target a.
3) Comparison property.

lo lla(t, .)1 2 +
J, Ic(4 z, . ) I 2 n ( d z ) I ko(1 + 1zI2);
where ko 2 0 is a constan,t;
2 a ( t ,z) = (aik(t,x))$=, satisfies the condition that
2 2
5 k N , T ( t ) P N , T ( l z i - yil 1,
( a i k ( t ,z) - cilc(t,y)(
as IzI,1y1 I N , t E [O,T],vi, Ic = 1 , 2 , . . . , d ;
where 0 5 k ~ , ~ ( s,kN,T(t)dt
t), < 00, for each T < 00,
and p ~ , ~ ( >u O,as
) u > 0; P N , T ( O ) = 0;and ~ N , T ( u )is strictly increasing
in u and such that
so+ d u / p ~ , ~ (= ) f o r N = 1,2, .... and T < 00;
u 00,
, z ) d) ~ =satisfies
3 c ( t ,z, z ) = ( c z ( t X, ~ conditions that
z + c ( t ,z, 2) E z:,vt 2 0 , Ic E Ti;, z E 2;
zi 2 yi ===+ zi f C i ( t , z,.) 2 yi C i ( t , y, z ) , +
vtrO,z,yERd,,ztZ,
where x = ( 5 1 , . . . ,z d ) , y = ( y l , . . . ,y d ) ;
4 LZ>yi (Pt(.) - P t ( Y ) ) I k v , T ( t ) P N , T ( C : l , , , ( z i - y i ) + ) , v z , y E z:,
as 1x1, IyI I N , t E [O,T],
where kN,T(t) and ~ N , T ( u )have the same properties as those in 2;
5 0 I v ~ , b ~ , P t (I ) Vz = 1 , 2 , . . . , d ; j = r l , . . . ,r2.
z ko,
294 S. Rong
If (xt,&) and (&,) , are solutions of the stochastic population dynam-

ics (6) with (7) and with the coeficient (8) and with the initial value X O ,
the fertility rate of females pt(z) and T o , P t ( z ) , respectively; then
x i >?i$,Qi = l , . . ., d ; and
A(.) 2 Bt(z),Qt2 0 , Q X E Ed+,
implies that P - a.s.
X : 2 8,Qt 2 0 , QJi = 1,.. . , d .
This indicates that larger initial size and lager fertility rate of female
always result larger population size.
Theorem 4.5 can be proved by Tanaka's formula as in Ref. 1. (See the
proof of Theorem. 346 in Ref. 1. However, now here ,Bt(z)depends on z).
5. The Optimal Stochastic Population Control

Now let us discuss the optimal stochastic population control problem for
the stochastic population dynamics (6) with (7) and with the coefficient
(8). Suppose that the admissible control set is
6={,B : ,O = ,&(x)
- is jointly continuous,
Po I P I Po1
where 0 I Do and Pa are constants. Let us minimize the following functional
J(P> = .I/' F ( t ,z f ) d t + G ( & ) ] ,
0
B , dt
P ) is the pathwise unique
among all ,B E 6, where 0 I T < 00, and (zt
strong solution of (6) with (7) and with the coefficient (8) corresponding to
the given j3 E 6;
where F ( t ,z) and G ( x )are Borel measurable.
Theorem 5.1. Assume that a ( t ,z) and c(tlx , z ) satisfy all conditions in
Theorem 4.1 and Theorem 4.5; F ( t , x ) and G ( x ) are jointly Borel mea-
surable functions defined on t E [O,T]and x E ad,
such that as xi I y i ,
1 <_ i 5 d ; x , y E $,Qt E [O,T],where we denote d = r,,
G ( z ) I G(Y),F(t1.) I F ( t ,Y).
Then the smallest constant Po is an admissible optimal control f o r the min-
imization
-
of the functional J(j3); that is,
Po E 6, and J ( & ) = infPGgJ ( P ) .
Furthermore, assume that
azrc(t,z) = C Z ; ~ X iC, i ( t , x , z ) = $ X J L r ( Z ) '
and xfO is the so-called optimal trajectory corresponding to the optimal

control Po. Moreover, for a given constant a > 0 we have that
E ~
t>&h a +X E ~ ~
I ~ ? I/ a.
~
~
This theorem actually tells us the following facts:

1) If a target functional monotonically depends on the population size,
for example, the energy, the consumption, and so on, spent by the popula-
tion, then to minimize this target functional we should control the fertility
rate of females to take a value as small as possible.
2) The optimal population size can be exponentially stable in mean
square, if we take the fertility rate of females small enough and when the
stochastic perturbation is not too large. Furthermore, given a > 0, we can
find out when the optimal population size vector in the mean square can
be less than this target a.
Theorem 5.1 can be derived by applying Theorem 4.1, Theorem 4.5 and
Corollary 4.4.
To summarize the discussion we have the following conclusion.
6. Conclusion
1) RSDE (6) with (7) and with the coefficient (8) is an appropriate
model for the stochastic population dynamic system.
2) The stochastic population size from this model continuously depends
on all parameters (i.e. the coefficients in the drift term). If one can ap-
proximately evaluate the parameters, then the population solution size
can also be approximately obtained.
3) The stochastic population size monotonely depends on the initail size
and the fertility rate of females.
4) The stochastic population size can be exponentially stable, if we
can take the fertility rate of females small enough, and the stochastic per-
turbation is not too large. Furthermore, we even can find out when the size
vector can be smaller than a given level in the mean square.
296 S. Rong
5) If some target functional, e.g. energy consumption etc., monotonically

depends on the population size, then t o minimize this functional one only
needs t o take the fertility rate of females as small as possible.
References
1. Situ Rong (Rong Situ), Theory of Stochastic Differential Equations with
Jumps and Applications. (Springer, 2005).
2. Situ Rong, Reflecting Differential Equations with Jumps and Applications,
Research Notes in Mathematics 408, Chapman & Hall/CRC, (2000).
3. Situ Rong, Reflecting differential equations with jumps and stochastic popu-
lation control, in Control Theory, Stochastic Analysis and Applications, eds.
S . Chen and J. Yong, etc., (World Scientific, Singapore, 1991), pp. 193-202.
4. Situ Rong & W. L. Chen, Stochastic Analysis and Applications 10,45 (1992).
5 . J. Y. Yu, B. Z. Guo, and G. T. Zhu, J . Syst. & Math. Scien. 7,214 (1987)
(In Chinese).
$16. NONCAUSAL STOCHASTIC CALCULUS REVISITED -

AROUND THE SO-CALLED OGAWA INTEGRAL
SHIGEYOSHI OGAWA
Dept. of Mathematical Sciences, Ritsumeikan University,
Kusatsu, Shiga, 525-8577 Japan
E-mail: ogawa-s@se.ritsumei.ac.jp
1. Introduction
The causal theory of stochastic calculus originated by K.It8 in 1942 is
founded on the Hypothesis of Causality;
In order that the stochastic integral / J ( t , w)d& is well defined the
function f ( t , u ) , (t 2 0 , w E a) should be adapted t o t h e increasing f a m i l y
of a-fields, generated by the underlying basic process Z t ( w ) , which is a
square integrable semi-martingale like the Brownian motion.
The hypothesis seems well fit t o the principle of causality in physics,
where the variable tappears as time parameter. Moreover it endows the
theory a remarkable feature of being in natural concordance with the notion
of martingale which plays indeed an essential role in It6s Calculus.
However it is also true that the Hypothesis of Causality has imposed a
significant restriction on the applicability of the causal theory of stochastic
calculus. Let us take for example the case of such SDEs driven by general
stochastic processes, like the fractional Brownian motion, that do not have
the martingale property,s or the case of usual SDEs but under noncausal
situations, that is the SDE with noncausal coefficients and/or noncausal ini-
tial values. Let us take moreover the case where the parameter t stands
for the space parameter, or the case where tis multi-dimensional (ie., a
stochastic calculus for the random field14). The notion of Causality looses
its sound meaning in such cases because of the lack of natural sense of t i m e
direction. Even in the case of physical problems where t appears as t i m e
parameter, we can find various situations of noncausal nature, such as the
Cauchy problem in the theory of Brownian particle equations, l 3 noncausal
298 S. Ogawa
version of the SDEs, the White noise analysis23 etc.

Guided by these motivations the author had presented the noncausal the-
ory of stochastic calculus in 1979,23by introducing the noncausal stochastic
integral which is now refereed by authors name. In this note we will give
a unified sketch of that noncausal stochastic calculus. We will show re-
cent development of the theory (Refs. 10 and 13) and we will also refer to
some typical applications of the theory to mathematical sciences including
mat hematical finance.
2. Noncausal Problems in Stochastic Analysis

To motivate our study we like to show in this paragraph some typical ex-
amples of noncausal nature in stochastic analysis. Hereafter we fix once for
all a probability space (a,F ,P ) on which all random quantities are defined.
Among these we understand by W t ( w ) ,t 2 0 the standard Brownian mo-
tion and denote by {Ft}the natural filtration generated by the w.,
namely
the family of o fields Ft = u { W s ; 0 5 s L t } c F.
By the random functions we understand those real valued functions
f ( t , w ) which are measurable in ( t , w ) with respect to the field B ~ , TxIF
and satisfy the condition,
P{lT f 2 ( t , z l u ) d t< CO} = 1
We will denote by H the set of all such random functions. A random func-
tion f ( t ,w ) E H is called causal (or non-anticipative) if, for each t E [0,T ]
it is adapted to the o field Ft, We will denote by M the totality of causal
random functions. For the random function f of this class, we have the
It6 integral which we will denote by
s f dWt throughout the paper, while
for the B-differentiable function f E M the integral of symmetric type (or
noncausal type) is denoted by the symbol,
Example 2.1 (Stochastic conservation law).

s f d,Wt (see, Paragraph 3).
Suppose given two stochastic processes, X t , Yt (t 2 0 ) which obey a

conservation law as follows;
E ( X t , Y t , t , a , p , y , ...)= C o n s t ,
where a , p, y etc are causal constants, that is the deterministic or at least
the random variables which are independent of the W. Now suppose that
Noncausal Stochastic Calculus Revisited 299
the process X t is unknown but another one is known to be generated by the

following mechanism,
d Y , = a(Y,)dt + b(Y,)dWt, Yo= yo E R1.
Given these we are to derive the equation of dynamics for the process
X t . W e are already familiar with such problem especially in mathematical
physics, but also in the mathematical finance we face to this. For example
some authors like M.Schweizer, E.Platen this idea f o r the modelling of the
price process in some specific situation, and we, S. Ogawa and M . Man-
cino," followed the same way to derive the SDE model for price process
in such market that admits a feedback of the future information about the
delta hedging strategy.
For the derivation of the desired SDE, we are only to apply the It6
formula to the conservation equation. But an essential problem can arise
when the parameters a , p, y etc. are allowed to arbitrary random variables
dependent of the W., for in such case we can no longer suppose that the
process X t is causal and we can not apply the It6 theory.
Let us take an example again from the mathematical finance.
Example 2.2 (Black-Sholes model in noncausal situation).
The SDE in Black-Sholes model is as follows;
dSt = rStdt +a2StdWt, (1)
which can be interpreted in the following f o r m of the SDE with the symmet-
ric stochastic integrals,
ffz
dSt = ( r - -)Stdt
2
+ uStd*Wt,
where the term
s d,Wt represents the symmetric integral,l8 that is; for the
B-differentiable causal function f (tlw),
-f
a is the B-derivative of the f . The symmetric integral is a special
awt
case of the noncausal integral that we are to study in the next paragraph.
To this simple model there correspond the following two variations of
noncausal nature;
(1) Case 1: The case where the constants r, u in the B-S equation are re-
placed by arbitrary random variables.
300 S. Ogawa
(2) Case 2: The case where the Brownian motion W in the equation is re-
placed b y such random process which does not have a martingale prop-
erty, for example the Fractional Brownian motion.
The first modification can arise when we need to study the B-S model in
such situation of admitting the insider trading, and the latter case has
been already discussed as the fractional B-S model b y many authors, like
A . Shiryaev,6 P. Cheridito for example, but without giving a justification
of such noncausal type SDEs in terms of the noncausal calculus.
For the study of these modified models, first of all we need the intro-
duction of the noncausal stochastic calculus and then we are to give precise
meaning to such noncausal SDEs as follows;
+
d X t = a ( t , X t , v ( w ) ) d t b ( t , X t , <(w))dWt,
(3)
Xo(w) = t(w).
I n view of the application to the SDEs of the cases 1 and 2, we are concerned
with the study of the noncausal SDEs, with a special interest o n the validity
of a noncausal It6 formula. W e will study these subjects in the paragraphs
3,4 and 5 in a much more general situation.
Example 2.3 (SIE of Fkedholm type).

Let us consider the boundary value problem for the second order SDE as
follows;
where ( 0 and (1 are arbitrary random variables.

where Zt, t E [0,1] is an arbitrary stochastic process defined o n the
(Q, F,P ) , with square integrable sample paths.
Then as we do for the boundary value problem of ordinary differential equa-
tions, by using the Greens function K ( t ,s, w ) corresponding to the above
situation,we may g e t in a very formal way the following integral equation
of Fredholm type;
where L ( t , s , w ) = h ( s ) K ( t ,s , w ) and Jd,Z(s) represents the stochastic

integral of noncausal type with respect to an orthonormal basis {cp,} in
Nonenusal Stochastic Calculus Revisited 301
LZ(0,l).
This is a very typical subject in the noncausal stochastic calculus, since

in such situation we can no more suppose that the solution X t is still causal
(it i.e. adapted) to the natural filtration generated by the underlying fun-
damental process Zi,hence the stochastic integral t e r m
meaning in the framework of the It6s theory.
s XtdZt loses its
I n the article of 198616 the author studied this subject in connection with the
boundary value problems of stochastic differential equations and he showed
the existence and the uniqueness of solutions (see Theorem 1 in Ref. 16),
under some reasonable assumptions on the choice of the fundamental pair
(2, { c p n } ) and on the sample regularity of the kernels K and L.
A s a natural extension of such SIE (stochastic integral equation) to the

case where the Zt, X , are the random fields, namely the stochastic processes
with multi-dimensional parameters t E J = [ O , l l d , we can think of the SIE
for the random fields. Imagine for example the case where the driving force
process Z,is the Brownian sheet.
X(t,u) = f(t,u) + L(t,s,w)X(s)ds + K ( t ,s , u ) X ( s ) d , Z ( s ) , (6)
For the review of this subject we would refer to the coming a r t i ~ l e which
,~
will appear in another monograph of lecture notes of the same conference.
So f a r the examples of noncausal problems are given where the noncausality
has entered into the situation through the noncausal quantities, such as the
noncausal initial values, the noncausal coeficients in SDE, or the driving
stochastic process that does not have the martingale o r semi- martingale
property. But the noncausal problem can arise even in the ordinary situation
where all the random quantities are supposed to be causal, that is adapted
to the natural filtration { 3 t } . Here is another example which is typical in
this sense.
Example 2.4 (The SPDE called BPE).

Given the real Brownian motion Wt and smooth coeficients
we consider the Cauchy problem of a stochastic partial differential equation

302 S. Ogawa
as follows;
a + { a ( t , + b(t,x)-}-u
--u X )
dWt a
= A(t,x)u(t,X , u)-
dWt
+ Bu + C ( t ,x )
at dt ax dt
4 0 ,x ,w ) = u o ( x )
(7)
B y the solution of this problem, we understand the random function
u(t,x , w ) measurable in ( t ,x , w ) with respect t o the field B p , ~ xl BR1 x 3,
especially causal in ( t , w ) for each fixed x E R1,and satisfies the follow-
ing weak solution equality with probability one for any test function q ( t ,x )
s
(cf.' 4 ) ;
LNVt + (acplx + A P M t , 2 , w ) + CPldtdX

+Ll d x L T ( b P ) 4 t ,x , w)d*Wt +
L 1
uo(x)Cp(O,x ) d x = 0,
(8)
a
where G = [O,T]x R1, pt = -p,
dX at
a
and the t e r m d,Wt stands
pX = -p
for the symmetric stochastic integral(see the Paragraph 3 below) as it is
s
remarked in the introduction.
The equation of this type, called the "Brownian particle equation", was
introduced by the author ('4) in the study of wave propagation in random
media as a mathematical model for such stochastic propagation phenomenon
carried by the Brownian or diffusive particles, namely the propagation along
the stochastic trajectory X t governed by the SDE,
dXt = a ( t ,X t ) d t + b(t, Xt)d,Wt.
Now the noncausal situation arises in this problem when we try to con-
struct the solution u(t,x , w ) by means of the method of stochastic charac-
teristics which was first studied by the author; B y the formal application of
the well known method of characteristics in the theory of the partial differ-
ential equation of the first order, or of hyperbolic type, we will get in this
case the following system of (symmetric) stochastic integral equations.
X ( t i " ) ( s )= x +
6"a ( r ,X ( t ' " ) ( r ) ) d r+
6"
b(r, X(t3")(r))d,W,,
+I t
A~(s,X(~~")(s),w)d,W,. (9)
Here the first equation gives the characteristic curve X ( t i x ) ( s )( s 5 t )

passing through the fixed point ( t ,x), along which the phenomenon propa-
gates, and the second equation is derived by integrating the PDE along that
stochastic calculus. Now remembre that the solution u(t,x,w ) is adopted to
the a-field Ft, while the characteristic curve X ( t i x ) ( s ) ,0 6 s 5 t 5 T is
measurable with respect to the a-field 3; := a { W ( t ) - W ( r ) s; 5 r 5 t } .
The composed function u ( s , X(tix)(s),w) appeared in the second equation
above is no more causal with respect to the Brownian motion and hence the
term of stochastic integral loses its meaning.
Such are the examples of noncausal problems, for the treatment of which
we need another theory of stochastic calculus that is free from the restriction
of the Causality. The calculus introduced by the author in 1979 is one of
such theories and has many advantageous properties compared t o other
calculus from the viewpoint of the tool for the modelling and analysis of
noncausal random problems. We are going t o give in what follows a resume
of that noncausal theory of stochastic calculus, brief but sufficient for us to
see that these problems can be treated in very natural way.
3. Review of the Noncausal Stochastic Calculus

For the rigorous study of those noncausal problems listed in the previous
paragraph, we need to introduce a noncausal stochastic calculus that is free
from the restriction of causality and one of such calculus was introduced
by the author in 1979 (,23,18,2220 and, etc.). We are going to give in
this paragraph a rapid review of some fundamental results in the theory
of noncausal stochastic calculus, mainly following the recent article. l 3 For
its special importance of the Brownian motion in the stochastic theory and
also for the concreteness of the discussion, we will show mainly the case of
the noncausal calculus with respect to the Brownian motion Wt, but as we
will see (cf. Remark 3.2) easily the formalism can work even for the ase of
more general stochastic processes instead of the Wt.
3.1. Causal functions and the B-diflerentiability

Following the,25 we will say that an H-class random function g ( t , w ) is
differentiable with respect t o the Brownian motion Wt (or B-differentiable)
provided that there exists an M-class random function say & g ( t , w ) such
that, for small enough h > 0,
304 S. Ogawa
where the integral

s dW stands for the It8s stochastic integral. The
function &g is called the B-derivative of the g . It is not difficult to see that
if the function g ( t , w ) is B-differentiable then its B-derivative is uniquely
determined (see25). The B-differentiability of the random function with
respect to the multi-dimensional Brownian motion is defined in a similar
way.
Remark 3.1. Let g ( t , w ) be a functional of the multi-dimensional Brow-

nian motion, Wt = (W,, W:, . . . , Wp) where the W i , (1 < i < n)
are independent copies of the l-dim. Brownian motion Wt. Then the B-
derivative of such function, say V,g, can be defined in the following way:
the V,g = ( a t g , w ag , . . . , m ag ) t is a causal random vector such that,
Here we notice that the It8 integral is defined for the causal random
functions f ( t , w ) E M and roughly speaking the symmetric integrals (i.e.
Z1/2 of O g a ~ and a ~ Stratonovich-Fisk
~ integral) are defined for the causal
and B-differentiable functions. That is, the symmetric integral Z 1 / 2 ( f ) of
a B-differentiable function was introduced as the limit (in probability)
limlal,o Ta(f) of the sequence {Ta(f)} of Riemannian sums,
<
where, A = (0 tl < . . . < t, 5 l} is a partition of the interval [O, 11 and
lAl = maxi(ti+l - t i ) .
The following result was established by the author in 1970,
Theorem 3.1 (Ref. 25). T h e l i m i t (in p r o b u b i l i t y ) T l p ( f ) = lim z~(f)

lAl-0
exists and is represented in the following f o r m :
3.2. Noncausal stochastic integral

Given a random function f ( t ,w ) E H and an arbitrary complete orthonor-
ma1 system (9,)in L 2 ( [ 0l]),
, we consider the formal random series
The stochastic integral of noncausal type introduced by the author in 1979

(23), is given in the following way,
Definition 3.1. A random function f ( t , u )E H is said to be integrable

with respect to the basis {pn} (or 9-integrable) when the random series
above converges in probability and the sum, denoted by
I' f ( t , w)d,Wt,
is called the stochastic integral of noncausal type with respect to the basis
{ p n ).
Remark 3.2. The validity of the above definition is not limited to the
case of Brownian motion or to other square integrable semi-martingales.
Indeed i t can apply even to the case of general square integrable processes
say Zt that do not posses the property of semi-martingale, as long as the
quantities
are well defined. The simplest example is when we employ the system of
Haar functions (Ho,o(t),H,,i(t),O 5 i 5 Z n - l - 1,n E N} as orthonormal
basis. Let us remembre that the Hn,i(t) are as follows;
Ho,o(t) = 1,
In this case the quantity above is defined in the natural way as follows;
2i 1 +22 22 2
Jo1Hn,2(t)dZt= 2n'2[{z(-) - Z(-)} - {Z(-)
+ 2i 1
Z(-)}I.
-
+
2" 2n 2" 2"
The noncausal integral with respect to the fractional Brownian motion can
be introduced by this way.8
306 5'. Ogawa
In general case, the way of convergence of the random series being con-
ditional, the integrability and the sum should depend on the basis, even
on the order of the same complete system of orthonormal functions. On
the relation between the noncausal integrals with respect to different bases,
very few is known except the following,
Theorem 3.2 (1984, Ref. 21). If the random function f ( t , u ) E H is
integrable in the L'- sense (ie. convergent in L1(R, P ) sense) with respect
to the system of trigonometric functions,
(1, JZcos2nnx, JZsin2nnx; n E N}.
Then the f is integrable with respect to the system of Haar functions and
the value of two integrals coincide.
If the function is integrable with respect to any basis {cp,} and the sum
does not depend on the choice of the basis, we will say that the function is
universally integrable (or shortly u-integrable).
3.3. Equivalent expressions and variants

Here are some equivalent expressions and a possible variation of the above
definition, which are worth to be remarked so that we can have a better
understanding of the nature of our noncausal integral.
(a) As a limit of the sequence of random Stieltjes integrals:
Given the pair (Wt,{ c p n } ) we introduce the sequence of approximation
processes W$(t) in the following way.
It is immediate to see that this gives a pathwise smooth approximation

of the Brownian motion W ( t , u ) .Moreover, by virtue of the famous
theorem due to K.It6 and M.Nishio,2 we know that for any choice of
as n -
the basis {cp,} the sequence { W z ( t ) } converges uniformly in t E [0,1]
co with probability one. Now we notice that our noncausal
integral can be expressed as the limit (in probability) of the sequence
of random Stieltjes integrals;
Proposition 3.1. It holds that,
1
1 1
f d p W t := l i m l f d W $ ( t ) (in probability).
(b) Riemannian definition:

Let us take the Haar functions { H n , i ( t ) } for basis {cpn}. This is a case
of special interest because we have the following,
Lemma 3.1 (1984, Ref. 21). Let us define the approximation pro-
cess W f ( t )for this case by the following formula,
2k-1 rt
T h e n each W:(t) i s the Cauchy polygonal approximation of the process

Wt taken over the set of dyadic points { k / 2 n ; 0 5 k 5 2 n } , that is,
To check this, we introduce the indicator function, xn,i(t) =

2 n 2 1 [ 2 - n i , 2 - n ( i + l ) ) ( t ) .It is immediate to see that
( X n , i ,Hm,k) = 0 for ( m ,k ) with m 2 n + 1,

here the symbol (., .) denotes the inner product in L2(0,1).
Therefore each xn,i should be represented as linear combination of the
membres {Hm,k,m 5 n } , say;
2(m- 1)- 1
Xn,i(t) = C ( ~ , JO,O)Ho,o
; + c c
l<m<n k=O
C ( n ,2; m, k ) H m , k ( t ) .
q n , i;m, k ) = ( X n , i ,H m d .
It is also easy to see that we have the following relation;
c
2n- 1
i=O
C ( n ,i; m, k ) C ( n ,i; l , j ) = 6m,&j.
Based on this relation we can get the following equality,
and this is what we want to see.

308 5'. Ogawa
Now applying this result to the expression given in the Proposition 3.1
of (a), we see that in this case the defining formula of the noncausal
integral is given as the Riemannian sum,
1' f d H W t = lim c .I'

2n-
n-+w i=o
2n
2-ni
2-"(i+l)
f (s)ds.{ w ( 2 - n (i+l))-w ( 2 - " 2 ) } ,
(14)
This type of definition can be found in recent publications of some
authors. But as we have seen in here ( 2 1 ) 1 this is merely a special case
of our integral.
(c) Let Dn(t,s ) be the kernel given by,
n
Dn(t,). = C Y k ( t ) P k ( S ) , ( t ,s E [0,11).
k=l
Then we have the following representation for the noncausal integral,
i'fd,W(t) lim
= n-co 1' dt f ( t ,w ) D n ( t , s ) d W s (limit in probabil-
ity).
For the case of trigonometric functions, the kernel Dn(t,s ) is the Dirich-
let kernel appearing in the theory of Fourier series.
(d) A generalization of the above view: Replace the kernels {D,(t, s)} in
the above interpretation by any 6- sequence say { K n ( t , s ) } then
, wc
will get a generalized formula for the noncausal integral:
3.4. Condition for the integrability - in the framework of

the Homogeneous Chaos theory
Let Ho be the totality of all random functions f ( t , w > E H such that,
1
If
( t ,w)12dt < 00. By Wiener-ItB's theory of Homogeneous Chaos, we
know that such function f E Ho can be decomposed into the slim of multiple
Wiener integrals, that is:
There exists a set of kernels, say { k L ( t ;t l , . + ., tn)}rZo,

such that
k: E L 2 ( [ 0,In+'
, ) with ~ n ! I I ] c ~ l l<~03,
+,
n
symmetric in n-parameters ( t l ,., tn) E [0,lIn and that,

M
n=O
where 11 . [In stands for the norm in L2([0,1In)space.
We will denote by H1 the totality of all Ho- functions f ( t ,w ) such that,

w
~ n n ! l / k ~ l l<~ 00.
+ lGiven a function f E H1 we introduce its stochastic
n=l
derivative V f by the following formula,
n=l
Since El1 i 1 ( V f ( t , s ) ) 2 d t d s= ~ n n ! / l k ~ / l ~we

+ l notice
, that the
stochastic derivative D f (t, s) is well defined for the f E H I . Now we can
state the condition for the p-integrability of the HI-class functions in the
following theorem that was obtained by the author in 1984.
Theorem 3.3 (Ref. 22). Let f E H1 and let { p n } be an arbitrary or-

thonormal basis. T h e n the necessary and suficient condition f o r the random
function f t o be cp-integrable is that the lim
n-+m
exists in probability.
3.5. Relation between symmetric and noncausal integrals

We call a random function f ( t , w ) semi martingale when it admits the
decomposition, f ( t , w ) = a ( t ,w ) + s' f d W s where f E M and a ( t ) is such
that almost every sample path is of bounded variation in t over [0,1].Notice
that if sup E l a ( t ) - a(s)I2 = o(h) then f is B-differentiable.
t,slt-sl<h
The followings are the basic results concerning the relation between the
symmetric integrals with the noncausal integral.
310 S. Ogawa
Theorem 3.4 (Ref. 18). Every causal B- differentiable function is inte-

grable in noncausal sense with respect to the system of Haar functions and
the sum coincides with that of the symmetric integrals:
We say that an orthonoral basis {cp,} is regular provided that it satisfies

the next condition:
Remark 3.3. Notice that this condition (15) is equivalent to the fact that,
1
w - lim u, = - (in L ~ )
n-cc 2
namely to the fact that, for any f ( t ) E L2(0,1) it holds the following,
Theorem 3.5 (Ref. 18). Every semi martingale (causal or not) becomes
cp-integrable, iff the basis (9,) is regular. I n this case the noncausal integral
coincides with the symmetric integrals.
Related to this result is a natural and interesting question asking whether

there can or can not be a basis {cp,} which is not regular. This question is
affirmatively answered by P.Mayer and M . M a n ~ i n o .We
~ can go on further.
The next result shows that a smoothness in Wt of the integrand ensures
the integrability with respect to any orthonormal basis.
Theorem 3.6 (Ref. 18). Every semi martingale that is twice B-

differentiable, namely the B-derivative f^ is again a semi martingale, is
u-integrable.
In their article^,^,^ of the authors M.Zakai and D.Nualart, the noncausal

integral was referred as the intrinsic Ogawa integral. We like to remark at
this stage that, precisely saying what they referred there was our noncausal
integral for the case of u-integrable functions,
Now let us finish this paragraph by giving another result on the way of
convergence of noncausal intgerals for semi-martingales, for its usefulness
in appications.
Proposition 3.2 (1985, Ref. 18).

If the B-derivative f of the semi-martingale f , df = fdWt + g(t,w)dt,
satisfies the following condition,
P{
I f 4 ( t ) d t < m} = 1.
Then the noncausal integral

verges uniformely in t E [O, 11.
I fd,Wt with respect to the regular basis con-
4. Applications to the Noncausal SDEs

Now we like to show basic results concerning the Cauchy problem of the
noncausal SDEs and the validity of the noncausal It6 formula, which will
give us at the same time the appropriate answers to those noncausal prob-
lems listed in the Paragraph 2.
4.1. Noncausal Cauchy problem

First notice that the SDE in (3) becomes meaningful in the framework of
the noncausal stochastic calculus. that is:
t E (O,T],
d X t = a(t,Xt,77(w))dt+b(t,Xt,ll(w))d,Wt,
(16)
Xo(w) = E ( w )
here the (9,) is a regular basis in L2(0,l),which we will fix throughout
the discussion.
We notice at this stage that when the parameters [(w),

~ ( ware
) not random
and the solution Xt can be supposed to be causal, then by virtue of the
Theorem 3.5 the SDE in (16) is reduced to the usual SDE with symmetric
integration,
dXt =a(t,Xt,77)dt+b(t,Xt,rl)dWt, t E (O,T],

(17)
XO(W) = E.
The Cauchy problem for the noncausal SDE was first studied by the
authorlg for such simple case where the parameter 77 is not random or
312 S. Ogawa
does not appear in a ( t lIC), b ( t , z) and only the initial data [ ( w ) arises as a
noncausal factor.
d X t = a ( t ,X t ) d t + b ( t ,X t ) d p W t , t E (O,T],
(18)
Xo(w) = [(w).
For this case, the results on the existence and a kind of uniqueness proper-
ties of the solution are proved under a milder assumption on the regularity
of the coefficients a ( . ) ,b(.) as follows:
Assumption 4.1. T h e coeficients a ( t ,IC), b ( t , x) are suficiently regular

in such sense that,
a2
(1) a ( t ,x ) , -b(t, x) are of Cl-class,
8x2
(2) a ( t , x),b ( t , x ) are suficiently regular in the sense that the causal Cauchy
problem 17 admits the unique strong solution X ( t ,w ; [) and that the
X ( t ,w ; <) i s continuous in ( t ,E ) with probability one.
Notice that under such conditions the composit,e
w1w ) = X(t1 w ; E ( w ) )
of the strong solution X ( t ,[, w ) of the (17) and the random variable [ ( w )
is well defined, which we expect to be a solution of the noncausal Cauchy
problem (18). In fact we have the following,
Theorem 4.1 (1985, Ref. 19). T h e composite X ( t , w ) i s a solution of
the noncausal Cauchy problem (18).
We have also found that this solution X ( t , w ) verifies the It6 formula of
noncausal type, that is:
Proposition 4.1 (1985, Ref. 17). For any function F ( z ) E C4 it holds
the equality,
d F ( X t ) = F(Xt){a(t,Xt)dt + b ( t ,X t ) d , W t } 05t51
As an application of this we can show the following result that concerns
the uniqueness of the solution for the noncausal problem (18),
Corollary 4.1 (Ref. 17). W h e n the b ( x ) # 0 , the composed function
X ( t ,w ) = X ( t ,w ; [(w))i s the unique solution among all random functions
verifying the It6 formula 4.1 of noncausal type.
The proof of this together with that of the previously presented Proposition
4.1 will be given in the next paragraph for a more general case.
4.2. Discussions f o r the more general cases

We are going to give in this paragraph the results on the Cauchy problem
for the more general case (16).
Assumption 4.2. W e suppose that the coeficients a ( t , x;r ] ) , b(t, x ; r ] ) are

suficiently regular in such sense that, for an arbitrary couple of parame-
ters (6,r ] ) the causal Cauchy problem 1'7 admits the unique strong solution
X ( t , w; E, r ] ) and that the X ( t , w; [, v) is continuous in ( t ,E, r ] ) with proba-
bility one.
Remark 4.1. The assumption is satisfied when, for example, the

a ( t , x;r ] ) , b(t, x ; r ] ) are of the C4-class in x , of C1-class in r] and all deriva-
tives are bounded on [0,1] x R1.
We also notice that under the Assumption 4.2 the composite
X ( t , w ) = X ( t , w; E(w),r](w))
of the strong solution X ( t , w ; [ , r ] ) of the (17) and the random variables
[ ( w ) , ~ ( wis) well defined, and as in the previous case we expect this com-
posite X ( t , w ) to be a solution of the noncausal Cauchy problem (16). In
fact we have the following,
Theorem 4.2. The X gives a noncausal solution of the noncausal Cauchy

problem 16.
For the verification of this, we need some preparations.
Proposition 4.2. Let f ( t , w; E , v) (5, v E [-A, A ] ) be a semi-martingale

such that for each fixed ( I ,r ] ) ,
d f ( t ,w ; E , 7 ) = d t ,w; E , r1)dt + h(t1w; 77)dWt
El (19)
where g(.), h(.) are causal random functions satisfying the following condi-
tion,
.rJ_A, 4 1; dr] 1 1 { g 2 ( t ,w; El r ] ) + h2(t,w; E , r])W<I . . =1
(i) Then for any regular basis {pn} in L2(0,l), the following equality holds,
(in probability)
314 S. Ogawa
(ii) Moreover i f the coeficient h(t,w ; <,r ] ) in the decomposition (19) again
becomes a semi-martingale satisfying the same condition as the f(.),
then the equality (20) still holds true for any basis {pn).
Proof. Put f = f1 + f2 where,
and
Hence with the help of the Theorem of Nishio-It6 we confirm that,
(in probability)
For the term f2 we have the decomposition,
where,
and, zn = 1
We are t o show that:
I
Pn(t)dW(t).
A
I&(c,r]) = 0 (in probability), (1 5 i 54).
Since for the quantities I,,, (i = 1 , 3 , 4 ) this could be easily done by a usual
routine, it would suffice to show the result only for the term I z , ~ .
By taking the Remark3.3 into account, we see that for each fixed (E, r ] ) we
have,
1
n+o3
h(t,w; E, r]){ 51 - u,(t)}dt = 0
On the other hand we have,
Now given the unique solution X ( t ,w ; <,'I) of the causal problem 17, we
introduce the sequence of random functions in the following way,
X,"(t, w; E , r ] ) = E+
Ita(s, X ( s ,w; E , 'I); r])ds+
It b(s, X ( s ,w; E , 77); v ) d W Z ( s )
(21)
where W$ is the approximate process of the Brownian motion introduced
in the previous paragraph.
We easily see by the Theorem 3.5 that for each fixed t , (I r],
) , we have
lim
n-+m
X g ( t , w ; <,77) = X ( t ,w ; E , 'I), (in probablity). Moreover we can see
that this convergence is uniform in ( E , r ] ) on every finite set CA = [-A, A] x
[-A, A].
Proposition 4.3. For an arbitrarily large A > 0 it holds the following
relation at each fixed t E [ O , 1 ] ,
lim sup I X z ( t ,w ; <,77) - X ( t ,w; <,'I)\ = 0 (in probability)
n-o3
(,V)ECA
Proof. Put
316 S. Ogawa
From equations (17),(21) we obtain the following:
&(t, w ; E , V ) =l b ( X ( s ;E, 71); rl){dW$(s)

On the other hand we have the following expression,
which implies that,
where
in probability as n -
We are to show that for each fixed t these J l ( n ) ,&(n),J 3 ( n ) tend to zero
00. Since a t this stage the parameters E , 7 remain
as deterministic constants, we notice that the X ( t ,w ; E, rl) is causal and
derivable in E, 77. In fact under the assumption (4.1) on the regularity of the
coefficients a ( . ) ,b(.) it is easy to verify that the derivatives,
IX 3 ( 0 ) =0
This combined with the expression (22) would imply that the quantity
A,@, w ; 7) is derivable in <,7 and that the order of the derivation in I ,7
and the integration is exchangeable. For example,
Hence by virtue of the Proposition 4.2 we only need to show that the
following quantities,
are semi-martingales satisfying the condition in that Proposition. Since this

can be verified by a simple routine work, we see that we are done. 0
Now we are going to give the proof for our Theorem,

318 S. Ogawa
5. Question of Uniqueness - Noncausal It8 Formula

The noncausal solution of the problem (16), a(t,w ) = X ( t ,w ; [ ( w ) , ~ ( w ) )
constructed in the Theorem 4.2, has a remarkable property as stated in the
next ,
Theorem 5.1 (Noncausal It6 formula). For any random variable [ ( w )
and any function F ( x ,y ) , which is differentiable in (x,y ) and of C4-class
in x with bounded derivatives, it holds the following equality:
d F ( X t , C(w)) = ( W w t ,CbJ))MXt;rl(w))dt+ @t; r l ( w ) ) ~ , W t ) (23)
Proof. Let X ( t ,w ; E , 17) be the unique solution of the causal SDE (17) with
deterministic parameters (c,
77). Then by the usual It6 formula for causal
functions, we have for each fixed deterministic parameters C), the (c,~,
following relation:
F ( X ( t ;E , r l ) , C) (24)
=F(t, C) + I' ( & F ) ( X ( S ,w ; t ,r l ) , C){a(Xs;77)ds + q x s ; rl)dWs)

Here the stochastic integral
s dWt stands for the causal symmetric integral.
Given this we introduce the approximation sequence as follows,
F Y t , w ; E , rl, C) (25)
Following the same argument as in the proof of Proposition 4.2, we would

easily verify that for each fixed t E [0,1] the sequence F n ( t , w ; c , ~ , C )
converges to F ( X ( t , w ; E , v ) , C )in probability as n + 00, uniformly in
(6,q, C) E C> on any finite set C> = [ - A , AI3. Hence we confirm, again fol-
lowing the same argument as in the proof of the Theorem 4.2, that for each
fixed t the sequence F n ( t ,w ; c ( w ) , ~ ( w )<,( w ) ) converges in probability to
the F ( X ( t , w ;[ ( w ) , ~ ( w )C(w))
, = F ( X ( t , w ) ,[ ( w ) ) . Now from the equation
(25) we see that the following limit,
lim
n-cc Lt(W(X(Sw
, ; E ( w ) , rl(w)),C(w))b(X(s,w ; E(wL rl(w));rl(w))dW$ (s)
should converge in probability to the limit,

Noncuusal Stochastic Calculus Revisited 319
by definition of the X ( t lw ) and by definition of the noncausal integral with

respect to the basis {cp,}.
Thus from this fact we get the desired equality (23 ), by letting n 4 00 on
both sides of the equality ( 2 5 ) . 0
As we have mentioned in the previous paragraph, this fact that the so-
lution X ( t , w ) = X ( t , w ; [ ( w ) , ~ ( w ) of
) the noncausal problem (16) satisfies
the It6 formula of noncausal type (23) would give us a partial answer to the
question of uniqueness of the solution of our noncausal problem. In fact we
have the following result that is valid for the case of 1-dimensional SDE.
Corollary 5.1. If the b ( t , z ; q ) does not depend on the t and b ( z ; q )> 0

(or < 0 ) for all ( t ,x,q ) , then the solution X ( t lw ) is unique among the all
random functions that werify the noncausal It6 formula (23). We will call
such solution the regular solution of the Cauchy problem.
Proof. Without loss of generality we suppose that b(.) > 0. Put Y ( t )=

F ( X ( t ) )where F ( z ) is as follows,
Then we have, k ( t , w ) = F - l ( Y ( t , w ) ) . By applying the noncausal It6

formula to the function Y ( t )we get,
Since this is merely a family of ordinary integral equations parametrized by

the w , we see the uniqueness of its solution Y ( t )for each fixed and hence
the uniqueness of the r?(tlw ) . This completes the proof. 0
References
1. P. Cheridito, Arbitrage in fractional Brownian motion models, (2002).
2. K. It6 and M. Nishio, Osaka J . Math. 5 , 35 (1968).
3. M. Mancino & L. Pratelli, Rendiconti Acad. Nat. delle Scienze detta dei X L ,
Memorie d i Matematica, 112 Vol. xviii, fasc 1, (1994) pp. 89-101
4. P. Mayer & M. Mancino, Se'minaire de Probabilite's xxxi, 198 (1995).
5. D. Nualart and M. Zakai Ann. Probab. 17,1536 (1989).
6. A . N. Shiryaev, On arbitrage and replication for fractional models, Research
Report 20, MaPhySto, (Dept.of Math.Sci., Univ. of Aarhus, 1998)
7. M.Zakai & D. Nualart, Proha. a n d Math. Statist. 3,37 (1987).
320 S. Ogawa
8. S. Ogawa & T. Miura, The fractional B-S model interpreted by the Ogawa
integral, presented at the poster session of the Congres on RISK Control
Theory, Firenze (2005)
9. S. Ogawa, Stochastic integral equations of Fredholm type, (to appear i n )
Harmonic, Wavelet and P-adic Analysis, (2005)
10. S . Ogawa, On noncausal Cauchy problem for the noncausal SDEs, In:
S.Watanabe et al. (Eds.): Stochastic Processes and Applications t o Mathe-
matical Finance, (World Scientific, Singapore, pp. 289-304, 2004).
11. S. Ogawa & M. E. Mancino, Nonlinear feedback effects by hedging strate-
gies, In: S.Watanabe et al. (Eds.): Stochastic Processes and Applications to
Mathematical Fznance, (World Scientific, Singapore, pp. 255-270, 2004)
12. S. Ogawa & M. Mancino, On a noncausal insider trading equilibrium model
of asset pricing (in preparation)
13. S . Ogawa, Rendiconti Acad. Nazionale delle Scienze detta dei XL,119, Vol.
xxv (2001)
14. S. Ogawa, On a stochastic integral equation for the random fields, Seminaire
de Proba. vol.xxv, (Springer Verlag, Berlin, pp. 324-339. 1991)
15. S. Ogawa, Topics in the theory of noncausal stochastic integral equations,
Diffusion Processes and Related Problems in Analysis, edt by M.Pinsky,
(Birkhauser Boston, 1990).
16. S. Ogawa, On the stochastic integral equation of Fredholm type, W a v e s and
Patterns (monograph) (Kinokuniya and North-Holland, pp. 597-605, 1986).
17. S. Ogawa & T. Sekiguchi, On the It& formula of noncausal type, Proc.Japan
Acad. Sci, (1985).
18. S . Ogawa, The stochastic integral of noncausal type as an extension of the
symmetric integrals, Japan J.Applied Math., (1985).
19. S. Ogawa, Sur la question dexistence de solution de lequation stochastique
differentielle du type noncausal, Kyoto J . of Math., (Kyoto Univ., 1985).
20. S. Ogawa, Quelques propriMs de lintegrale stochastique du type noncausal,
Japan J.Applied Math., (1984)
21. S . Ogawa, T 6 h o k u Math. J., 36,41 (1984).
22. S. Ogawa, Noncausal Problems in Stochastic Calculus (in japanese) in Proc.
of the Workshop o n Noncausal Calculus and T h e Related Problems held at
RIMS Kyoto Univ. in February 1984, (Suuri Kagaku Kokyuuroku (RIMS
Report) 527, 1984).
23. S. Ogawa, C.R. Acad. Sci., Paris, t.288, SBrie A, pp.359-362, (1979).
24. S. Ogawa, 2. W.verw. Geb., 28, 53 (1973).
25. S. Ogawa, Proc. Japan Acad. Sci., 70 (1970).
@ 2007 World Scientific Publishing Go.
$17. INFINITE-DIMENSIONAL STOCHASTIC ANALYSIS

AND FOUNDATIONS OF QUANTUM MECHANICS
ANDRE1 KHRENNIKOV
International Center for Mathematical Modeling in
Physics and Cognitive Sciences, MSI,
University of Vaxjo, S-35195, Sweden
E-mail: Andrei.Khrennikov@msi.vxu.se
We show t h a t by using an asymptotic expansion of Gaussian integrals of ana-

lytic functions on infinite-dimensional spaces it is possible to represent quan-
t u m mechanics as a projection of classical statistical mechanics on infinite-
dimensional space.
1. Introduction
The problem of reduction of quantum mechanics to classical statistical me-
chanics has been discussed from the first days of quantum mechanics, see,
e.g., Refs. 1-4. Now days this problem is known as the problem of hid-
den variables or completeness of quantum mechanics, see, e.g., Refs. 5-8
for debates. There is a rather common opinion that quantum mechanics
is complete and that it is impossible to introduce hidden variables pro-
viding more detailed description than quantum mechanics. In Ref. 9 it
was demonstrated that in the opposition to this opinion it is possible to
represent quantum mechanics as projection of classical statistical mechan-
ics on infinite-dimensional space. In this paper we present this approach
(which was called in Ref. 9 Prequantum Classical Statistical Field Theory
~ PCSFT) on the mathematical level of rigorousness; in particular, some
functional spaces introduced in Ref. 9 should be modified to obtain the cor-
rect results. In the present paper we also find connection of PCSFT with
background Gaussian random field on Hilbert space. Finally, we extend the
approach of Ref. 9 by considering unbounded operators, see Sec. 6. There
are also differences in interpretations of a small parameter of our asymp-
totic procedure of dequantization. In [9] this parameter was identified with
the Planck constant h. In this paper we introduce a new parameter a giving
322 A . Khrennikov
the dispersion of prequantum fluctuations, see Ref. 10 for more details on

physical interpretation. To simplify considerations, in this paper we con-
sider quantum formalism over the field of real numbers, see Ref. 10 for
dequantization of complex quantum mechanics. To exclude possible mis-
understanding, we emphasize from the very beginning that our paper is
n o t about d e f o r m a t i o n quantization for s y s t e m s w i t h t h e infinite n u m b e r of
degrees of freedom, see, e.g., Ref. 11, but about dequantization of conven-
tional quantum mechanics for systems with a finite number of degrees of
freedom by using analysis on infinite-dimensional space.
2. Infinite-dimensional analysis
Let H be a real Hilbert space and let A : H -+ H be a continuous self-
adjoint linear operator. The basic mathematical formula which will be used
in this paper is the formula for a Gaussian integral of a quadratic form
f($) fA("b) = (A'$,$1.
Let dp(+) be a a-additive Gaussian measure on the a-field F of Bore1
subsets of H , see Ref. 12. This measure is determined by its covariation
operator B : H + H and mean value m = mp E H . For example, B
and m determines the Fourier transform of p : P(y) = S H e i ( Y ' G ) d p ( $ ) =
e ~ ( B y ~ y ) t i ( m ~yY )E, H . In what follows we restrict our considerations
to Gaussian measures with zero mean value m = 0, where ( m , y ) =
S(,y, + ) d p ( $ ) = 0 for any y E H . Sometimes there will be used the symbol
p~ to denote the Gaussian measure with the covariation operator B and
m = 0. We recall that the covariation operator B = cov p is defined by
(By1,yz) = ~(~~,~CI)(Y~,$)~P(~CI),Y~,Y~ E H , and has the following proper-
ties: a). B > >
0, i.e., ( B y , y ) 0 , y H ; b). B is a self-adjoint operator,
B E C , ( H ) ; c). B is a trace-class operator and Tr B = sH
II$l12dp(+). This
is dispersion cr2(p) of the probability p. Thus a 2 ( p ) = Tr B.
We pay attention that the list of properties of the covariation operator
of a Gaussian measure differs from the list of properties of a von Neumann
density operator [4] only by one condition: Tr D = 1, for a density operator
D.
We can easily find the Gaussian integral of the quadratic form fA($) :
(1)
The differential calculus for maps f : H + R does not differ so much

from the differential calculus in the finite dimensional case, f : R" -+ R.
Instead of the norm on R", one should use the norm on H . We con-
Stochastic Analysis and Foundations of Quantum Mechanics 323
sider so called Frechet differentiability. Here a function f is differentiable

if it can be represented as f($o + A$) = f($o) +
f($o)(A$) +
o(A$),
where l i m ~ ~ ~ Ilo(A)tl ~ ~ +=o0. Here a t each point $ the derivative f($)
~IlA+ll
is a continuous linear functional on H ; so it can be identified with the
element f($) E H . Then we can define the second derivative as the
derivative of the map $ ---f f($) and so on. A map f is differentiable
n-times iff: f ( $ o + A$) = f($o) + +
f($o)(A$) +f($o)(A$,A$) +
+ +
... &f()($o)(A$,...)A$) on(A$), where f()(+o) is a symmetric
continuous n-linear form on H and limlla+ll-o = 0. For us it
is important that f($o) can be represented by a symmetric operator
u ,H
f($o)(ulu) = ( f ( $ ) o ) ~ , ~ ) , E u (this fact is well know in the finite
dimensional case: the matrice representing the second derivative of any two
times differentiable function f : R + R is symmetric). We remark that
+ + +
f($) = f ( 0 ) + f(O)($) $f(O)($, $1 .. $f(0)(7b1 . . . I $) on(+). +
For a real Hilbert space HI denote by the symbol H C its complexifica-
tion: H C = H @ iH.We recall that a function f : HC + C is analytic if it
can be expanded into the Taylor series:
which converges uniformly on any ball of H C
3 . Dequantization
3.1. Classical and quantum statistical models
We define classical statistical models in the following way, see Ref. 9 for
more detail (and even philosophic considerations):
a). Physical states w are represented by points of some set R (state
space).
b). Physical variables are represented by functions f : R + R belonging
to some functional space V(R).*
c). Statistical states are represented by probability measures on R be-
longing to some class S(R). +
* T h e choice of a concrete functional space V(R) depends on various physical and math-
ematical factors.
+ I t is assumed (by the Kolmogorov axiomatics) t h a t there is given a fixed u-field of
subsets of R denoted by F. Probabilities are defined on F It is, of course, assumed
t h a t physical variables are represented by random variables - measurable functions. T h e
choice of a concrete space of probability measures S ( R ) depends on various physical and
mathematical factors.
324 A . Khrennakov
d). The average of a physical variable (which is represented by a function

f V(R)) with respect to a statistical state (which is represented by a
E
probability measure p E S(R)) is given by
A classical statistical model is a pair M = (S(R), V(R)). We recall that

classical statistical mechanics on the phase space R2, = R" x Rn gives
an example of a classical statistical model. But we shall not be interested
in this example in our further considerations. We shall develop a classical
statistical model with an infinite-dimensional phase-space.
In real Hilbert space H a quantum statistical model is described in the
following way:
a). Physical observables are represented by operators A : H -+ H be-
-
longing to the class of continuous self-adjoint operators C, = C C , ( H ) .
b). Statistical states are represented by density operators, see Ref. 4.
The class of such operators is denoted by D D ( H ) .
d). The average of a physical observable (which is represented by the
operator A E L C , ( H )with
) respect to a statistical state (which is represented
by the density operator D E D ( H ) )is given by von Neumann's f ~ r m u l a : ~
< A >o- Tr DA (4)
= (D(H)L
The quantum statistical model is the pair Nquant , ,(H)).
3.2. Asymptotic Gaussian analysis

Let us consider a classical statistical model in that the state space R = H
(in physical applications H = L2(R3)is the space of classical fields on R3)
and the space of statistical states consists of Gaussian measures with zero
mean value and dispersion
where a > 0 is a small real parameter. Denote such a class of Gaussian

measures by the symbol S$(R). For p E S$(O), we have
Tr cov p = 0' (6)
We remark that any linear transformation (in particular, scaling) preserves
the class of Gaussian measures. Let us make the change of variables (scal-
ing) :
(we emphasize that this is a scaling not in the physical space R3, but in
the space of fields on it). To find the covariation operator D of the image
p~ of the Gaussian measure p ~ we, compute its Fourier transform:
Thus
D=-=%!!?.
B
a a
We shall use this formula later. We remark that by definition:
To make our further considerations mathematically rigorous, we should

attract the theory of analytic functions f : RC + C . Here RC = R @ ZR is
the complexification of the real Hilbert space R.
Let b, : RC x ... x RC 4 C be a continuous n-linear symmetric form.
We define its norm by llb,ll = supIIGlII1I&($, ...,$)I. Thus
Ibn(+r "'7 $)I 5 Ilbnll11+1In (9)
Let us consider the space analytic functions of the exponential growth:
If(+)/ 5 aebllQll,+ E oC. (10)

Here constants depend on f : a = a f , b = b f .
Lemma 3.1. l 3 The space of analytic functions of the exponential growth

coincides with the space of analytic functions such that:
~ ~ f ( n )5 ~ ,~n = 0,1,2, ...
( 0c )r n (11)
Here constants c = c f and r =rf depend on the function f .
Proof.
+
A). Let f have the exponential growth. For any E RC, we consider
the function of the complex variable z E C : g+(z) = f (z+). By the Cauchy
integral formula for gG(z) we have: g$)(O) = sJz,=R
g+ ( z ) z - ( " f l ) d z ,
where a t the moment R > 0 is a free parameter. Thus: Ig$'(O)l 5
n!RPn If (Reie$)l 5 afn!R-nebfRII+II. By choosing R = n and
observing that g$)(O) = f(")(O)($, ...,$) we obtain:
11 f(") (0)11 5 a;e-nn1/2ebfn.
326 A . Khrennikov
Thus the derivatives o f f satisfy the inequalities (11) with r f = e b f .

B). Let now derivatives of f satisfy the inequalities (11). Then
by the inequalities (9) we have If(+)l 5 C,"==, Ilf(")(0)IIII+lln/n!5
cf C;==,(rf11 ~ ~ + ~ ~ )5" /cfe'fII+II.
n,! Thus f has the exponential growth with
bf =r f . 0
We denote by the symbol V(R) the following space of functions f : R 4

R. Each f E V(R) takes the value zero a t the point $ = 0 and it can
be extended to the analytic function f : RC 4 C having the exponential
growth.
Example 3.1. In particular, any polynomial on R belongs to the space

V(R). For example, let A l l ..., AN be continuous linear operators. Then
function f ( + ) = C ~ s l ( A n + $, ) n belongs to the space V(R).
Any function f E V(R) is integrable with respect to any Gaussian mea-
sure on R. Let us consider the family of the classical statistical models
Ma = (S,%(R),V(R)).
Let a variable f E V(R) and let a statistical state p~ E S,%(R).Our
further aim is t o find an asymptotic expansion of the (classical) average
s,
< f > p s = f ( $ ) d p ~(+) with respect to the small parameter a.
Lemma 3.2. Let f E V(R) and let p E S,%(R). T h e n the following asymp-
totic equality holds:
a
< f >p= - Tr D f"(0)
2
+ o(Q), 4 0, (12)
where the operator D i s given by (8). Here
.(a) = a2R(a1f l P I , (13)
where IR(a,f , p)I Icf so erfll+lldpD($).

Proof. In the Gaussian integral /, f ( $ ) d p ( + ) we make the scaling (7):
where
We pay attention that
because the mean value of p (and, hence, of p ~ is) equal to zero. Since
p E Sz(R),we have Tr D = 1.
The change of variables in (14) can be considered as scaling of the
magnitude of statistical (Gaussian) fluctuations. Negligibly small random
fluctuations a ( p ) = 6 (where a is a small parameter) are considered in
the new scale as standard normal fluctuations. If we use the language of
probability theory and consider a Gaussian random variables <(A), then the
transformation (7) is nothing else than the standard normalization of this
random variable (which is used, for example, in the central limit theorem):
E(A)-EE (in our case E J = 0).
q(A) = JE(E(A)-EE)Z
We now estimate the rest term R(a,f , p). By using the inequality (11)
we have for a 5 1 :
By using the equality (1) we finally come the asymptotic equality (12). 0
We see that the classical average (computed in the model M a =

( S g (R), V(R)) by using the measure-theoretic approach) is coupled through
(12) to the quantum average (computed in the model N q u a n t = (D(R),
&(a))by the von Neumann trace-formula).
The equality (12) can be used as the motivation for defining the following
classical quantum map T from the classical statistical model M a =
---f
(S& V ) onto the quantum statistical model Nquant = (D, L S ):

cov p
T : Sg(R) 4 D(R), D = T ( p )= -
a
(the Gaussian measure p is represented by the density matrix D which is
equal to the covariation operator of this measure normalized by a ) ;
1
T : V(Q)-+ L ( o ) , Aquant = T(f) = ,f(O). (17)
Our previous considerations can be presented as
328 A . Khrennikov
Theorem 3.1. The map T : Sz(R) + D(R) i s one-to-one; the m a p T :

V(s2) + C,(R) is linear surjection and the classical and quantum averages
are coupled by the asymptotic equality (12).
We emphasize that the correspondence between physical variables f E

V(s2) and physical observables A E &(a)is not one-to-one.
Let again R = H . We consider a classical statistical model such that the
class of statistical states consists of Gaussian measures p on R having zero
mean value and unit dispersion
These are Gaussian measures having covariance operators with the unit
trace. Denote the class of such probabilities by the symbol SG = SG(R). In
this model we choose a class of physical variables consisting of quadratic
forms f A ( 2 ) = (Az,z).
We denote this class by Vquad. We remark that this
is a linear space (over R). We consider the following classical statistical
model:
As always in a statistical model, we are interested only in averages of phys-

ical variables f E Vquad with respect to statistical states p E &(a).w e
emphasize that by (1):
Let us consider the following map T from the classical statistical model
Mquad = (SG,Vquad) to the quantum statistical model Nquant= ( D ,L,) :
T : SG(R) + D ( n ) , T ( p )= cov p (19)
(the Gaussian measure p~ is represented by the density matrix D which is

equal to the covariation operator of this measure), and we define:
(thus a variable f E Vquad(fl) is represented by its second derivative).
Theorem 3.2. The map T provides one-to-one correspondence between the

classical statistical model M q u a d and the quantum model N q u a n t .
Stochastic Analysis and Foundations of Q u a n t u m Mechanics 329
4. G a u s s i a n u n d e r g r o u n d for pure states

In quantum mechanics a pure quantum state is given by a normalized vector
@ E H : llPll = 1. In our model such a state is not pure a t all (in the
sense that such a vector Q does not provide a description of an individual
system). Such a normalized vector Q is the label of a Gaussian statistical
mixture. The corresponding quantum statistical state is represented by the
density operator: Dq = 9 @ @. In particular, the von Neumann's trace-
formula for expectation has the form: Tr DqA = (A@,@). Let us consider
the correspondence map T for statistical states for the classical statistical
model M a = ( S & V ) , see (16). A pure quantum state 9 (i.e., the state with
the density operator Dq)is the image of the Gaussian statistical mixture
pq of states $J E H . We use the capital @ to denote a quantum pure state.
This is just the special system of labeling of the Gaussian measure pq by
the normalized vector @ of Hilbert space. Points of the sample space on
that this measure is defined we denote by the low $J. The measure pq has
the covariation operator Bq = a&. This means that the measure pq is
concentrated on the one-dimensional subspace H a = {x E H : x = s q , s E
R}. This is one-dimensional Gaussian distribution.
5. Pure states as one-dimensional projections of spatial

white-noise
In section 4 we showed that so called pure states of quantum mechanics have
the natural classical statistical interpretation as Gaussian measures concen-
trated on one-dimensional subspaces of the Hilbert space H . On the other
hand, it is well known that any Gaussian measure o n H i s determined by
i t s one-dimensional projections. To determine a Gaussian random variable
[ ( w ) E H , it is sufficient to determine all its one-dimensional projections:
[ ~ ( w= ) ( @ , [ ( w ) ) , Q E H . The covariation operator B of [ (having the
zero mean value) is defined by ( B @@) , = EJ:. We are interested in the
following problem:
Is it possible t o construct a Gaussian distribution o n H such that i t s one-
dimensional projections will give u s all pure quantum states, @ E H , 11@11 =
l?
We recall that in our approach a pure quantum state @ is just the label
for a Gaussian random variable [\I, such that E[; = 0111611~ = 01. Thus
the answer to our question is positive and pure quantum states can be
considered as one-dimensional projections of the &-scaling of the standard
Gaussian distribution on H . The standard Gaussian distribution p on H (so
330 A . Khrennikov
the average of p is equal to zero and cov p = I , where I is the unit operator)
is nothing else than the white noise on R3 (if one chooses H = L2(R3)),
see [12] for details. Thus pure quantum states are simply one-dimensional
projections of the spatial white noise. It is well known, see, e.g., [12], that
the p is not a-additive on the a-field of Bore1 subsets of H .
To escape mathematical difficulties and concentrate on the dequanti-
zation of quantum mechanics, we start with consideration of the finite-
dimensional case.
5.1. The finite-dimensional case

We consider the family of Gaussian random variables [Q, 9 E R,E& =
0 , E [ i = a119112.This family can be realized as [ ~ ( w )= (Q,[(w))where
[ ( w ) = &q(w) and q(w)E R is standard Gaussian random variable (so
Eq = 0,cov q = I ) . For any $ E R, we define the projection Pq to this
vector: Pqj(k) = (Q, k ) 9 .
Denote by the symbol V(R) the class of functions f : R -+ R such
that f(0) = 0 and f can be continued analytically onto C and this con-
tinuation f(z) has the exponential growth.
Proposition 5.1. Let f E V(R). Then we have:
E f ( W ( w ) )= 7 +
a9112j - ( f ( 0 ) 9 ,9) o ( a ) ,a + 0. (21)
Proof. By using the Taylor expansion o f f we obtain:

1
Ef(Pd(w))= p,
c ( w ) ) 2 ( f ( 0 ) Q*), 4a),+ + 0.
By setting into this asymptotic equality the dispersion of the random vari-
able P Q [ ( w we
) obtain (21).
If 11911 = 1 (a pure quantum state), then we get:
Ef(PqjJ(w))= %(f(O)Q, +
9) o ( a ) ,a -+ 0. (22)
Here A = f(0) is a symmetric linear operator. We quantized the classical
variable ~ ( z ) , zE R, by mapping it to the operator A = $f(O), see
Theorem 3.1. The Gaussian random variable cqj, I 19 I I = 1 is quantize by
mapping it into the pure quantum state 9.
Theorem 5.1. There exists a Kolmogorov probability space such that all
pure quantum states can be represented by Gaussian random variables on
this space. The correspondence Q 4 & ( w ) is linear:
where XI, A2 E R.
Proof. We choose R = R" as the space of elementary events, the a-field of

Bore1 subsets is the space of events and the standard Gaussian measure p
as the probability measure. Then for @ -+ &(w) = &(Q, w ) , w E R",we
have: XIEql(w)+ M i h 2 ( w ) = Al(@i,u) + L ( Q ' z , w ) = (Xi*1 + h Q 2 , w ) =
~Xl*'l+X2Q2 (w). 0
This theorem is rather surprising from the common viewpoint (by that
essentially nonclassical probabilistic features of quantum states are conse-
quences of the non-Kolmogorovian structure of the quantum probabilistic
model).
We pay attention that physical variables &,(w) = P q [ ( w ) , @ E R",
(one-dimensional projections of the scaling ( ( w ) of the standard Gaussian
random variable ~ ( w E) Rn)cannot be mapped onto nontrivial quantum ob-
serwables. Prequantum classical physical variables &,(w) = (9, w ) are linear
functionals of w . Therefore T(&) = tg(0) = 0. Nevertheless, quantum me-
chanics contains images of c,p given by quantum states @, but only for Q
with 11Ql1 = l!
We call ( ( w ) a background random field. All pure states could be ex-
tracted from the the background random field by projecting it to one di-
mensional subspaces. PCSFT explains the origin of the scalar product on
the set of pure quantum states. We consider the l/a-amplification of the
covariation of two Gaussian (prequantum) random variables <Q, (w)and
[u2 ( w ) . We have:
Conclusion. The Hilbert space structure of quantum mechanics is in-

duced by the (prequantum) Gaussian random field (the background field
[ ( w ) ) through the a 4 0 asymptotic.
At the moment we proved this only in the finite-dimensional case. In sub-
section 5.2 we shall do this in the infinite-dimensional case. Finally, we pay
attention to the fact that, for quadratic physical variables f(x) = +(Ax,x),
where A : H -+ H is a symmetric operator, the asymptotic equality (22)
is reduced t o the precise equality of averages. By considering directly the
332 A . Khrennikov
standard Gaussian random variable q(w) (instead of the background ran-

dom field t ( ~=)&q(w)) we come to the following classical probabilistic
representation of the quantum average: Ef(Pq,q(w))= ( A Q ,Q) =< A >q, .
5.2. Prequantum white noise field

To repeat consideration of susection 5.1 for the infinite-dimensional case,
we consider measures on the so called rigged Hilbert spaces. We apply
some rather abstract mathematical constructions. However, finally we shall
consider a simple concrete example which will be then used as the basis of
our prequantum classical statistical model.
Let R be a nuclear Frechett topological linear space and R its dual
space. Suppose that R is densely and continuously embedded into a Hilbert
space H , so R c H . Thus the dual space H is densely embedded into R.
By identifying H and H we obtain the rigged Hilbert space:
RcHcR (25)
In our final application we shall set R = S ( R 3 ) . This is the space of
Schwartz test functions on R3. Here R = S(R3)is the space of Schwartz
distributions. In this case we choose H = L2(R3) and we shall consider the
rigged Hilbert space:
S(R3)c L2(R3) c S(R3) (26)
Readers who are not so much interested in general theory of topological
linear spaces can consider this rigged Hilbert space throughout this section.
A Gaussian measure p on 52 is determined by its characteristic func-
tional (Fourier transform) which is defined on R : p(9) = e-ib(*>),
where b : R x R --f R is a continuous positively defined quadratic form.
By the well known theorem of Minlos-Sazonov, see e.g., [I, p is o-additive
on 0 and its covariation functional is equal to b. Here b ( Q 1 , Q2) =
Jo,(q5, @ I ) ( + , 92)dp(4), where * I , E R. This functional defines the co-
variation operator B : R --f R by ( B ~ Q2) I , = b ( * ~ , Q2). This operator is
self-adjoint in the following sense. The dual operator B : 0 0. But,
--f
since the topological linear space R is a nuclear Frechet space, it is reflex-

ive. Hence, R = R. Thus the operator B : R -+ 52. Thus it is meaningful
to speak about self-adjoint operators in this framework (by extending the
ordinary theory of self-adjoint operators in Hilbert space). We also pay
attention to the fact that the covariation operator B is positively defined.
$so complete metrizable and locally convex

Let us consider the standard Gaussian distribution p on H that is de-

fined by its covariation functional:
b p l , Q2) = (91,*2).
The corresponding covariation operator B = I : R R' is the canonical

.--)
embedding operator. Since the embedding R c H is continuous, b : R x R +

R is continuous and, hence, the measure p is a-additive on R'. Therefore
there is well defined the corresponding Gaussian random variable r ] ( $ ) E R'.
In the case of the rigged Hilbert space (26) the Gaussian random field
~ ( 4 E) S'(R3) is nothing else than the spatial white noise. We extend
this terminology and we shall call ~ ( 4 white
) noise even in the abstract
framework. Let us consider &-scaling of white noise
<($I =fir]($)
and its one-dimensional projections: &($) = (C($),P),@E R. We have

E<Q= 0,E<$ = allPl12.The<($) is the backgroundfieldinour prequantum
model (PCSFT).
For any P E R, we consider the one-dimensional projector Pq($) =
(4, a)*, q5 E R', and the R-valued random variable PQ<($) = < ~ ( $ )If9 .
((lkl(= 1, then the T-image of the corresponding Gaussian distribution p~
is nothing else than the pure state *.
This correspondence can be extended form the space R to the Hilbert
space H . If 9 E H, then <q,(q5) = (4, 9) is also well defined, but is is not
a continuous linear functional on the space R'. The <Q($) is defined as an
element of the space of square integrable functionals of the white noise:
Cq, E L2(R', d p ) . To define <*,we approximate P E H by elements P, of
R , 9 , + 9 in H (we recall that R is dense in H ) . Then (Q = limn+m <qn
in La(R', d p ) .
Lemma 5.1. Let f : H + R be a polynomial and let f(0) = 0. Then, f o r

any P E H , the asymptotic equality (21) holds.
Proof. Here the main difference from consideration in section 3 is that the
measure p is not concentrated on Hilbert space H on which the function
f is defined (and continuous). Therefore even the exponential growth o f f
on H would not help so much, because Jn,eal1411dp(+) = 00 (since even
,J, ll$lldp($) = 00). We have for a polynomial f:
334 A . Khrennikov
Since the sum is finite and derivatives of f are continuous forms on H , we

obtain (21). 0
We quantize f ( u )by mapping it into 9.

For quadratic functionals
f(u)= $ ( A u , u ) , AE L , ( H ) , we have the precise equality and we can
directly use the average with respect to the canonical Gaussian random
variable ~ ( w )Here
.
1
Ef(P*rl(w))= +f(O)*, *I.
6. Unbounded operators
Let f : R -+ R be a smooth function. Then, at any point $0 E R, f($o) :
R + 0. Therefore f(0) is in general unbounded operator in H .
Moreover, in this way (i.e., starting with PCSFT) we obtain the class of
linear operators (quantum observables) that is even essentially larger than
in the conventional quantum formalism. In general, A = f(0) maps R not
into H . but into R.
Example 6.1. Let us consider the rigged Hilbert space (26). We consider
the map f : S(R3)-+ R determined by a fixed point zo E R3 :
1
f($) = 2$2(zo).
(For example, the classical field $(z) = e- is mapped into the real number
e-:). Then ( f ( O ) $ l , 7,h)= $I(~o)$z(~o). Thus
1 1
A+(x)= -f(O)+(z) = -b(z - Z O ) $ ( X )
2 2
is the operator of multiplication by the &function d(z - 20).Hence
c L ~ ( R ~ ) .
f(o)(~(~~))
For any @ E S(R3),we have
1
) ) ;*(Q)
E f ( P q q ( ~= = (A*, @) =< A >* .
However, in general for 9 E L2(R3)the average < A > q is not well defined.
Stochastic Analysis and Foundations of Q u a n t u m Mechanics 335
We can consider not only pure states, but general density operators.
Let us now consider a Gaussian measure p E S E ( H ) which has the support
on the space 52. Thus p can be considered as a measure on R. For such a
measure p its covariation operator B : R' 4 R and its Fourier transform
p is defined on 52'. We denote this class of statistical states by the symbol
SE(52).We remark that SE(R) c S E ( H ) .
Let E be a complex locally convex topological linear space. We recall
that the topology of E can be determined by a system of semi-norms (the
notion of a semi-norm p generalizes the notion of a norm 11 . 11; the only
difference is that p ( $ ) can be equal to zero even for a nonzero vector $).
Let b, : E x ... x E -+ C be a continuous n-linear symmetric form. There
exits a continuous semi-norm p on E such that
llbnllp = SUP Ibn($, ..., $41 < 00

P(*)< 1
(here p zz pb, ) . Thus
lbn($, "'7 $>I 5 llbnll P Y $ > (27)

An analytic function, see, e.g., [Ill for details, f : E -+ C has the exponen-
tial growth if there exits a continuous semi-norm p on E on such that:
If($)l 5 aebp($),$ E E. (28)

Here the constants and the semi-norm depend on f :a = a j , b F b j ,p 3 p f .
Lemma 6.1. l 3 The space of analytic functions of the exponential growth
coincides with the space of analytic functions such that there exists a con-
tinuous semi-norm p = p f :
~ ~ f ( n ) ( 05 )c~rn,
~ pn = 0 , 1 , 2 , ... (29)
Here constants c = c f and r =rf depend on the function f .
Proof.
A). Let f have the exponential growth. For any $ E E l we consider the
function of the complex variable z E C : g $ ( z ) = f (zq9). As in Lemma 3.1,
we have: IgF'(0)I 5 n!R-nsupole12s I f(Reie$)I 5 a f n ! R P e b f R p ( * ) .We
obtain:
~ ~ ~ ( n )5( a;e-nnl/zebfn,
o)~~p
Thus the derivatives o f f satisfy the inequalities (29) with r f = ebf
336 A . Khrennikov
B). Let now derivatives of f satisfy the inequalities (29) for some con-
tinuous semi-norm p . Then by the inequalities (27) we have
lf(42 c
03
n=O
Ilf'"'(o)llpPn(lcl)l~! ICferfP(@).
Thus f has the exponential growth with b f =rf and the same continuous
semi-norm p as in (29).
We denote by RC the complexification of R : RC = R e i R . We denote by

V(R) the class of functions f : R + R, f(0) = 0, which can be analytically
continued onto RC and they have the exponential growth.
Lemma 6.2. L e t p E S$(R). T h e n , for a n y f u n c t i o n f E V(R), t h e follow-

ing asymptotic equality holds:
< f >,- s, f('$)dP('$) = ;J'

n
( f " ( O ) h ! b ) d P D ( U ) + 4aL a + 0 , (30)
where D = 7 .Here
.(a) = a2R(a,f l P I , (31)
where
IR(a,f,P)I 5 C f 1 n
eTfp(')dPD('$). (32)
T h e s e m i - n o r m p i s determined by t h e inequality (28).
The proof of this Theorem repeats the proof of Lemma 3.2. Instead of
Lemma 3.1, we apply its generalization to the case of an arbitrary locally
convex topological linear space, see Lemma 6.1.
We pay attention that D : R' -+ R, and A = : R + R',so

C = D A : R -+ R. In general, this operator can not be extended to a
continuous operator in H . We would like to obtain an analogue of the
formula (1) for linear continuous operators A : R + R' :
(A'$, '$)dpD (u)= Tr D A (33)
The main mathematical problem is that in general the operator C = D A is

not even continuous in H , so it is not a trace class operator in the Hilbert
space H . Nevertheless, we can introduce the notion of trace even in such a
framework.
We recall that systems of vectors { e j } Z 1 ,ej E 0, and {eg}jO=ll

eg E R,
are called biorthogonal topological bases in R and 0 if
j=1 j=1
where the series converge in R and R, respectively.

Definition 6.1. A linear continuous operator C : R --+ R is called trace-
class operator if, for any pair of biorthogonal topological bases, the series
M
j=1
converges and its sum does not depend on bases.
Lemma 6.3. Let p be a Gaussian measure on R and let A : R 4 R be a

continuous operator. Then the operator C = D A , where D = covp, belongs
to the trace class and the equality (53) holds.
As a consequence of Lemmas 6.2 and 6.3, we obtain:
Theorem 6.1. Let p E SE(R). Then, for any function f E V(R), the
following asymptotic equality holds:
< f >,= f ( $ ) d p ( + ) = Tr Df(0)/2 + o ( a ) ,a + 0, (34)
where D = y .
Thus our prequantum model, PCSFT, provides the motivation to extend
the set of quantum observables and consider all continuous operators A :
52 --+ 0. Operators should be self-adjoint in the ordinary sense: A = A .
We recall that here A : R + R, but 0 3 0, since R is a nuclear
Frechet space and hence it is reflexive. Denote the set of such operators by
the symbol C,(R,0).Denote the set of covariation operators of Gaussian
measures belonging the space S&(R) by the symbol D(R, 0).
Definition 6.2. A statistical quantum model corresponding to a rigged
Hilbert space 7 given by (25) is the pair N,.,,,,t(7) = (D(R, R), Cc,(R,a)).
A generalized density operators D E D ( W , 0) represents a statistical state;
a linear operator A E C,(R,R) represents a quantum observable. The
average of such an observable with respect to such a statistical state is
given by the following generalization of the von Neumann trace-formula:
<A >D= Tr DA (35)
338 A . Khrennikov
We choose the state space R - a nuclear F'rechet space. For a rigged

Hilbert space 7 given by (25), we consider the classical statistical model
s,
M " ( 7 ) = (S$(R),V ( 0 ) ) .Here as always < f >,= f ($)dp($).
The equality (34) can be used as the motivation for defining the fol-
lowing classical 4 quantum map T from the classical statistical model
M " ( 7 ) = (S,%(R),V ( 0 ) )onto the quantum statistical model Nquant(T)=
(D(R', R), LC,(R,0'))by (16), (17). Our previous considerations can be pre-
sented as
Theorem 6.2. The map T : S,%(R) -+ 2)(0',0)i s one-to-one; the map

T : V(R) -+ LC,(R,R') is linear surjection and the classical and quantum
averages are coupled by the asymptotic equality (34).
Example 6.2. The position operators P j , j = 1 , 2 , 3 can be obtained as

?,.-A
3 - 2 fXj(01, where
'I
Here the operator of multiplication Pj : S(R3) 4 S(R3),$ 4 xj$, is

continuous. Hence ij : S(R3)4 S'(R3)is also continuous. Thus, for any
measure p E S,%(S(R3)), we have
D = cov p/cr (here the trace of the composition DPj is well defined).
Example 6.3. Let ICO be a fixed point in R3. Let now A+(%)= 6(x -
q ) $ ( x ) , $ E S(R3).This operator does not belong to the domain of the
conventional quantum formalism. It could not be represented as an un-
bounded operator in H = L2(R3) with a dense domain of definition. Nev-
ertheless,
and the trace of the composition D A is well defined.
Example 6.4. The momentum operators & , j = 1 , 2 , 3 , can be obtained

as fjj = if;,
(0), where
Here t h e operator p j : S(R3)+ S(R3)is continuous. Hence, fij : S(R3)-+

S(R3)is also continuous. T h u s for any measure p E S a ( S ( R 3 x) S(R3)),
+
we have (for $(x) = q(x) ip(x)):
< fpj >p- -i / /

S(R3)xS(R3) dxj
% ( x ) m d x d p ( $ ) = cw~rD p j ,
where D = cov p / a . Here Dfij : S(R3)4 S(R3)is t h e trace class operator.

Similar considerations can be done for angular momentum operators.
This paper was partially supported by EU-network Q u a n t u m Proba-

bility a n d Applications.
References
1. D.Hilbert, J. von Neumann, L. Nordheim, Math. Ann., 98,1 (1927).
2. P. A. M. Dirac, T h e Principles of Q u a n t u m Mechanics, (Oxford Univ. Press,
1930).
3. W. Heisenberg, Physical principles of quantum theory, (Chicago Univ. Press,
1930).
4. J. von Neumann, Mathematical foundations of quantum mechanics, (Prince-
ton Univ. Press, Princeton, N. J., 1955).
5. A.Yu. Khrennikov (editor), Foundations of Probability and Physics, Q. Prob.
White Noise Anal., Vol. 13, (WSP, Singapore, 2001).
6. A. Yu. Khrennikov (editor), Q u a n t u m Theory: Reconsideration of Founda-
tions, Ser. Math. Modeling, 2 , (Viixjo Univ. Press, 2002).
7. A . Yu. Khrennikov (editor), Foundations of Probability and Physics-2, Ser.
Math. Modeling, 5, (Vaxjo Univ. Press, 2003).
8. A. Yu. Khrennikov (editor), Q u a n t u m Theory: Reconsideration of
Foundations-2, Ser. Math. Modeling, 10, (Vaxjo Univ. Press, 2004).
9. A. Yu. Khrennikov, A pre-quantum classical statistical model with infinite-
dimensional phase space. J . Phys. A : Math. Gen., 38,9051 (2005).
10. A. Yu. Khrennikov, Quantum mechanics as an asymptotic projection of sta-
tistical mechanics of classical fields: derivation of Schrodingers, Heisenbergs
and von Neumanns equations. http: //www. arxiv.org/abs/quant-ph/O511074
11. A. Yu. Khrennikov, Infinite-dimensional pseudo-differential operators. Dues-
tia Akademii Nauk U S S R , ser.Math., 51,46 (1987).
12. T. Hida, M. Hitsuda Gaussian Processes, Translations of Mathematical
Monographs, 120,(American Mathematical Society, 1993).
13. A. Yu. Khrennikov, Equations with infinite-dimensional pseudo-differential
operators. Dissertation for the degree of candidate of phys-math. sc., (Dept.
Mechanics-Mathematics, Moscow State University, Moscow, 1983).
$18.NONCOMMUTATIVE TRIGONOMETRY AND

QUANTUM MECHANICS
Karl Gustafson
Department of Mathematics,
University of Colorado, Boulder, Colorado
E-mail: gustafs@euclid. colorado. edu
A noncommutativeoperatortrigonometryoriginated by this author in 1966 will

be summarized. Some important applications t h a t have taken place during the
intervening years will be recalled. Then we will turn t o t h e recent application
of this noncommutative operator trigonometry t o Bells inequality in quantum
mechanics. Our geometrization of this celebrated inequality clarifies t h e proper
meaning of several physical issues in the Einstein-Podolsky-Rosen Paradox.
Those clarifications lead us t o look more closely at Von Neumanns formulation
of the collapse of the quantum mechanical wave packet. These investigations
give us a new clarification of the quantum Zen0 Paradox. We also announce
a fundamental new theorem on quantum mechanical reversibility and we take
note of some of its physical implications.
Keywords: noncommutative operators; antieigenvectors; Bell Inequality;

Einstein-Podolsky-Rosen Paradox; Zen0 Paradox; reversibility
1. Introduction, Background, and Summary

Starting from a purely operator-theoretic question which this author had
formulated in 1966 within the Hille-Yosida-Phillips-Lumer theory of ab-
stract operator semigroups, this author developed in the period 1966-1972
the essentials of a noncommutative operator trigonometry. In this begin-
ning period, I called this a theory of antieigenualues and antieigenuectors,
because intuitively it compares and contrasts in a natural manner to the
Rayleigh-Rita variational theory of eigenvalues and eigenvectors. From 1973
*This paper is a summary of the two lectures, Noncommutative Trigonometry presented

at The Second International Conference on Abstract and Applied Analysis 2005-Quy
Nhon, Vietnam, June 4-9, 2005, and Noncommutative Operator Algebra and Bell In-
equality presented a t the Institute of Mathematics, Vietnamese Academy of Science and
Technology, Hanoi, Vietnam, June 15, 2005. In addition, some new results are included.
342 K. Gustafson
to 1993, this author was chiefly involved with other research, including scat-
tering theory, mathematical physics, computational fluid dynamics, optical
computing systems, neural networks. However, a few further results for
the noncommutative trigonometry were obtained. Then from 1994 to the
present I returned to the noncommutative trigonometry and successfully
applied it to numerical linear algebra, wavelets and control theory, statisti-
cal estimation and efficiency, and quantum mechanics. In 1997 I published
two books'?' which contain chapters on this noncommutative trigonometry.
At that time I decided to call this theory operator trigonometry. This, be-
cause I was beginning to realize its importance beyond the original context
in which I had viewed it variationally, juxtaposed against the Rayleigh-Ritz
variational theory. More recently, I have settled upon the name noncommu-
tative trigonometry. This, because the noncommutative trigonometry now
seems to me to convey a spirit much like the theory of A. Connes, which
now carries the name noncommutative geometry. Suffice i t t o say that this
author has always viewed operator theory as primarily noncommutative. It
is time to emphasize that fact with the name noncommutative trigonometry
for the noncommutative operator trigonometry I will describe herein.
A bibliography of 62 papers written by this author that deal wholely or
in part with this noncommutative trigonometry may be found in the recent
~ u r v e yThus,
.~ the reader may find more bibliography, facts, and history in
Refs. 1-3. Although this noncommutative operator trigonometry has had
contact with related contributions by a few other notable mathematicians,
including M. G. Krein, H. Wielandt, V. Ptak, C. Davis, B. Mirman, L.
Kantorovich, C. R. Rao, for the most part it is my creation.
In Section 2 we quickly recall the key essentials of the noncommutative
trigonometry. In Section 3 we quickly recount some of its main application
to date. More details about these may be found in Ref. 1-3 and especially
in the 62 citations of Ref. 3. However, the recent survey3 could not present
the recent results of this author in which the noncommutative trigonometry
fundamentally clarifies key issues in the celebrated Bell and Zen0 problems
of quantum mechanics. So I go beyond3 here and in Sections 4 and 5, respec-
tively, I treat those two problems from the viewpoint of the noncommutative
trigonometry.
2. Essentials of the Noncommutative Trigonometry

To fix ideas, let us restrict attention to A a symmetric n x n invertible
positive definite (SPD) matrix. The theory extends to arbitrary positive
selfadjoint operators in a Hilbert space and to arbitrary invertible matrices
Noncommutative %gonometry and Quantum Mechanics 343
and even to accretive operators in a Banach space and we refer the reader to
Refs. 1-3 for those results. Let A have eigenvalues 0 < A 1 < A:! < * * < A n
9
where again for simplicity we assume them to be all distinct, although the
theory does not need that. Then the essential facts of the noncommutative
trigonometry depend upon the following fundamental four entities:
$(A) = the maximum turning angle of A (3)
x* =f (L)12zl
+ (A)12z,
A1 + A n
+An (4) A1
I named p1 the first antieigenvalue of A. The entity v1 is equally important
in applications. I called $(A) as defined from either p1 or vl, the angle of
A. I called the two vectors z* the first antieigenvectors of A. They are the
vectors (normalized t o norm 1 in (4)) most turned by A. The z1 and z,
A good example is A = [:p6].

are normalized eigenvectors for A1 and A, respectively.
Then $ ( A )
antieigenvectors are z+ = (4,3)/5 and 2 - = (-4,3)/5.
= 16.2602 and the two
Early on I realized that these entities form the basis of an (angular)

spectral theory as rich as the usual (dilation) spectral theory based upon
eigenvalues. Moreover I showed that the Euler equation
211Az11211z112(ReA)z- IIzl122Re(Ax,z)A*Az- IIAzlI2Re(Az,z)z = 0 (5)
generalizes the usual Rayleigh-Ritz variational theory, and moreover, con-
tains it: for A symmetric or normal, As antieigenvectors, and as well all
eigenvectors, satisfy (5). I defined the sequence of higher antieigenvalues and
corresponding higher antieigenvectors on successive reducing subspaces of
A, more or less in analogy with the Rayleigh-Ritz theory. All of that was
done in the period 1966-1972. See Refs. 1-3 for more details.
A couple of remarks here may be helpful. The two antieigenvectors may
not be linearly combined to form an antieigenspace. However, their span
is of course the reducing subspace sp{zl,z,} formed by the smallest and
largesteigenvectors. Thus the antiegenvectors zk are rather special vectors
contained within that spectral subspace. Specifically4 they have relative
344 K. Gustafson
angle (x+,x-)= -sin$(A), and thus the angle between them is always
4(A) + 7rP.
Second, in applications I often must use the operator trigonometry and
the usual trigonometry in conjunction. However, the operator trigonometry
generally does not satisfy the identities of the usual trigonometry. More-
over, it has some of its own interesting identities. A useful example is:5
sinqh(A1/2)= sin+(A)/[l + C O S ~ ( A )Here
] . A1/2 denotes the positive square
root of the operator A. Generally we have the operator angle inequality:
44AB) 5 $ ( A )+ + ( B ) .
Third, although I developed some rudiments of the noncommutative
operator trigonometry for operators A on a Banach space, it is better to stay
+
in Hilbcrt space, because the essential identity sin2 4(A) cos2 4(A) = 1
does not hold generally in Banach space.
3. Quick Summary of Applications to Date

Applications of this noncommutative trigonometry t o date include, roughly
and chronologically: perturbation of operator semigroups, positivity of op-
erator products, Markov processes, Rayleigh-Ritz theory, numerical range,
normal operators, convexity theory, minimum residual and conjugate gra-
dient solvers, Richardson and relaxation schemes, wavelets, domain decom-
position and multilevel methods, control theory, scattering theory, precon-
ditioning and condition number theory, statistical estimation and efficiency,
canonical correlations, Bells inequalities, quantum spin systems, and quan-
tum computing. Here is a quick summary.
In the initial development period 1966-1989 one could say that the re-
search emphasis was functional-analytic and theoretical. However, I had
turned to computational mathematics in the 1980s, working with several
engineering groups. An example is in Ref. 6. In particular, in such research
I had used some steepest descent and conjugate gradient scheme computa-
tions and had learned of the basic convergence rate bound
Here E ( x ) is the error ((x - z * ) , A ( x - 2*))/2, z*the true solution. This

bound also occurs in a key way in some parts of optimization theory. Imme-
diately from my operator trigonometry I knew that what (6) really meant,
geometrically, was trigonometric:
Noncommutative P i g o n o m e t y and Quantum Mechanics 345
Although (6) was known as a form of the Kantorovich inequality, nowhere

in the numerical nor in the optimization literature did I find this important
geometrical result: that gradient and conjugate gradient convergence rates
are fundamentally trigonometric. That is, they are a direct reflection of the
elemental fact of As maximum turning angle 4(A). The known conjugate
gradient convergence rate, by the way, becomes, trigonometrically written
now,
See Ref. 7 where I published this result (8) for conjugate gradient conver-
gence. And I first announced the steepest descent result (7) at the 1990
Dubrovnik conference.8
Many other important iterative computational linear solvers for large
problems Ax = b were shown by the noncommutative trigonometry to have
trigonometric convergence rates. I obtained these results in the period 1994-
2004. All were new. It is really quite astounding that these fundamentally
trigonometric convergence geometries had lain undiscovered all those years,
going back t o Richardson (1910), Kantorovich (1948), other notable com-
putational mathematicians. However, without the cos 4(A) and sin 4(A) of
the noncommutative trigonometry a t your fingertips, one cannot see the
geometry. See Ref. 9 to come somewhat up-to-date on these results.
Another important application of the noncommutative trigonometry
has been to statistics. For example, there is the famous Durbin-Watson-
Bloomfield (1955, 1975) lower bound for least squares relative efficiency
of estimators, namely,
where the As are the eigenvalues of the covariance matrix V. In (9) the
p* is the best possible estimator and X is an n x p regression design
matrix. Coincidentally, I had recently4 extended my operator trigonom-
etry to arbitrary matrices and there (also in earlier papers) I had stressed
the alternate definition of higher antieigenvalues and their corresponding
higher antieigenvectors as defined combinatorially in accordance with (1)-
(4) but with successively smaller operator turning angles now defined by
corresponding pairs of eigenvectors. It follows readily that the geometric
346 K. Gustafson
meaning of statistical efficiency then becomes clear: l o
RE(& 2 n
P
i=l
cos2 &(V) (11)
It is that of operator turning angles.

Other intimate connections between statistical efficiency, statistical es-
timation, and canonical correlations may be found in Ref. 10. For example,
LaGrange multiplier techniques used in statistics lead to equations for the
columns xi of the regression design matrices X to satisfy
v2xz
(VXi, X i )
+ (V-lxz,
Xi
Xi)
= 2vxi, 2 = 1 , .. . , p
I call this equation the Inefficiency Equation. In the antieigenvalue theory,

the Euler equation (5) becomes in the case when A is symmetric positive
definite,
The combined result is the following. All eigenvectors x j satisfy both the
Inefficiency Equation and the Euler Equation. The only other (normalized)
vectors satisfying the Inefficiency Equation are the inefficiency vectors
(my terminology)
The only other vectors satisfying the Euler equation are the antieigenvectors
Further recent results from thc application of the noncommutative

trigonometry t o matrix statistics may be found in Refs. 11,12.
For other applications of the noncommutative trigonometry, see the ci-
tations of Ref. 3. These include applications to quantum computing and to
elementary particle theory, which I will not discuss in this paper. In the
next two sections I will concentrate on the Bell and Zen0 problems from
quantum theory.
4. The Bell Inequalities of Quantum Theory

Addressing fundamental issues raised by Einstein, Podolsky, Rosen, l 3 con-
cerning the foundations of quantum mechanics of Von Neumann,14 Bell15
Noncommutative Trigonometry and Quantum Mechanics 347
presented his famous inequality
and exhibited quantum spin measurement configurations whose quantum

expectation values could not satisfy his inequality. Bells analysis assumed
that two measuring apparatuses could be regarded as physically totally sep-
arated, and free from any effects from the other. Thus his inequality could
provide a test which could be failed by measurements performed on corre-
lated quantum systems. It was therefore argued that local realistic hidden
variable theories could not hold in quantum mechanics if future physical
experiments would violate Bells inequality. Later Aspect et al. l6 indeed
demonstrated violation of Bells inequality in their laboratory experiments.
But the controversy about the Bell inequality and related inequalities to be
mentioned below and their physical implications, continues to this day.
Bells arguments in arriving a t his inequality were classical probabilistic
correlation arguments. However, it is known and easy to prove that this
inequality holds for any real numbers a , b, c in the interval [-1,1]: then
+
ab - bc ac 5 1. Here is a proof. From b2 5 1 and c2 5 1 we have
+ +
b2(1 - c 2 ) 5 1 - c2 and hence b2 c2 5 1 b2c2. Adding 2bc to both
sides and multiplying by a2 5 1 we therefore have a2(b2 c2 2bc) + +
+ + + + + +
b2 c2 2bc 5 1 b2c2 2bc, that is, a2(b c ) 5~ (1 bc)2. Taking the
+ + +
positive square root yields a ( b c ) 2 la1 Ib cI 5 1 bc.
Wigner17 presented his own version, making more clear the issues of
locality and socalled realism. Furthermore, Wigner was sure to use quan-
tum mechanical probabilistic correlations. His version of Bells theory then
becomes the inequality
where the Oi, are angles between spin directions w i and W k .

Another important Bell-type inequality is that of Clauser, Horne, Shi-
mony, Ho1t.l Let a,b, c, d be four arbitrary chosen unit vector directions
in plane orthogonal to the two beams produced by the source. Let vi(a)
and wi(d) be the hidden predetermined values f l of the spin components
along a and d, respectively, of particle 1 of the ith pair, similarly w i ( b ) and
wi(c) for particle 2 values along directions b and c . Then the average cor-
relation value for particle 1 spins measured along a and particle 2 spins
measaured along b is E(a,b ) = EL=,
vi(a)wi(b)/N. Taking into account
in the same way the average correlation values E(a,c ) , E ( d ,b), E(d,c )
348 K. Gustafson
and adding up all pairs, one arrives a t the CHSH inequality
In a series of papers starting with Ref. 19, see Ref. 3 for citations to
later papers, I placed all of the inequalities of the Bell theory, into my non-
commutative trigonometry. From my results, it may be argued that many
important issues in the Bell theory, about which there are a t times furi-
ous arguments about physical and probabilistic meaning among physicists,
are really better seen as new mathematical quantum geometry from the
noncommutative trigonometry.
As a first example, consider Wigners version (17) of the Bell inequality
discussed above. For the three coplanar directions, our corresponding in-
equality equality20 becomes in Wigners quantum spin setting terminology
sin2 (;el2) + (:ez3)

sin2 - sin2 (ie13)
= 2 cos (iell) (;el3)
[cos - cos (:el2) cos (:ezj)] (19)
Violation of the conventionally assumed quantum probability rule

I (u,TI)l2 = cos2 9u,vfor unit vectors u and v representing prepared state
u t o be measured as state v, is equivalent according to Wigner to the right
side of this identity being negative. This is his Bell violation test. How-
ever, from our point of view, there is no violation, there is just a quantum
trigonometric identity, valid for certain formulations of measurement of cer-
tain spin systems.
As a second example, let us consider the important CHSH inequality
(18) given above. Wishing now to preserve equality, we write
Squaring this expression and writing everything quantum trigonometrically,

la. b + a .c + d . b - d . cI2 = (2 + 2 cos ebc) cos2 e a , b f c
+ (2 2 cos 6bc) cos2ed,b-c)
-
+ 2(4 4 cos2 &)1/2 cos e a , b + c cos Od,b-c

-
= 4 cos2(8bc/2) cos2 6a,b+c

+ 4sin2(Obc/2)cos2 ed,b-c
+ 4 Sin ebc cos ea,b+c cos Od,b-c.
In the above we used two standard trigonometric half angle formulas. Now
substituting the double angle formula sin 6bc = 2 sin(Obc/2)cos(BbC/2)into
the above we arrive at
l a . b + a . c + d .b-d.CI2 = 4 [ c o s 8 ~ c / 2 ) C O S ~ a , b + c + S i n ( e b c / 2 ) C O S ~ d , b - c ] 2
and hence we have a new quantum CHSH equality
la. b + a .c +d .b - d . CI = 21 cOs(6)bC/2)COSea,b+c + Sin(&/2) COSed,b-cl
We may also write the righthand side as twice the absolute value of the
two-vector inner product
U1 u2 (cos(ebc/2), Sin(ebc/2)) (cos ea,b+c, cosed,b-c)
to arrive at the quantum trigonometric identity
l a . b + a . c f d . b - d . c l = 2 ( c 0 s ~ e ~ , fbc+o~s 2 ed,b-c)1/2/COS8,1,,,1
(20)
The right sides of these equalities isolate the classical limiting probability
factor 2 from the second factor, which may achieve its maximum fi.Fix
any directions b and c. Then choose a relative to b+c and choose d relative
to b - c so that cos2 Oa,b+c = 1 and cos2 6d,bPc = 1, respectively. Now we
may choose the free directions b and c to maximize the third factor to
COS~,,,,, = f l . But that means the two-vectors u1 and u2 are colinear
and hence
u1 = (COS(6bc/2), sin(ebC/2))= 2 ~ l / ~ ( c&,b+c,
os cosed,b-c)
- 2-1/2(&1, +l)
and thus the important angle 6bc is seen to be f7r/2. More to the point, our
identity allows one to exactly trace out the violation regions analytically
in terms of the trigonometric inner product condition 1 5 Iu1 . u21 5 fi.
Thus, and we have only presented two examples here, the noncommuta-
tive trigonometry applied to the celebrated inequalities of the Bell theory
350 K. Gustafson
has shown that those inequalities may be made special cases of identities
for the noncommutative trigonometry. This enables much clearer delin-
eations of physical, probabilistic, and geometrical implications of those in-
equalities. Further analysis leads to the opinion that one cannot claim
quantum locality or nonlocality on the basis of satisfaction or violation of
Bells inequalities. Rather, one is ledz0 to the assertion that the real issue
as concerns nonlocality is the Von Neumann projection rule.14 That rule,
under which measurements of quantum probabilities correspond t o projec-
tions on Hilbert space subspaces, is considered next, in the last section of
this paper.
5. The Zen0 Problem of Quantum Theory

The term Zen0 Paradox, i.e., a watched pot never boils, was introduced
in Misra, Sudarshanl to highlight certain fundamental issues in quantum
mechanical measurement theory. There is a long history to such quantum
measurement problems. For example, Von Neumann, l4 when he created his
Hilbert space model of quantum mechanics, proved that any given state q5 of
a quantum mechanical system can be steered into any other state in the
Hilbert space by an appropriate sequence of very frequent measurements.
Thus, in particular, you can freeze a quantum evolution in time by con-
tinually watching it. The allusion t o the greek Zen0 refers t o a debate
which he and Parmenides had with Socrates in Athens approximately 445
B.C.. Parmenides, an elderly philosopher, presented to Socrates an interest-
ing proposition t o start the debate: reality never changes, therefore motion
is not possible. Although Parmenides relied on the younger Zen0 for ar-
gumentive support, the originating thesis was that of Parmenides. Thus is
might be more accurate to call the quantum versions Parmenides Paradox
and Parmenides theory.
What is a t issue is not a matter of names, it is the fundamental phys-
ical question of what happens when a quantum mechanical wave function
$(t) is measured by some classical measuring instrument. The act of mea-
suring is seen as a collapse of the wave function, which according to the
Von Neumann postulates is implemented by an orthogonal projection. This
gives rise t o numerous mathematical issues. I was involved, going back to
1974, in some of these questions, and left them unresolved, and hence as
concerns my partial results, unpublished. Misra and Sudarshanl went from
the Schrodinger picture to the Heisenberg picture t o partially move beyond
these issues, but a close reading of Ref. 21 shows that the issues are not
removed, they are only assumed away. Recently I returned to some of these
questions and some of the following results are taken from Ref. 23. Because
there is now a large literature on the Zen0 Problem, I must refer the reader
to Ref. 23 for further bibliography. Also one should look at several articles
in the 2003 volume in which Ref. 20 appeared, a volume dedicated in sig-
nificant part to the Zen0 problem, and its relation to quantum computing.
At issue both physically and mathematically is the nature of the interac-
tion of the measuring instrument or subspace with the evolving wave packet
$t = U ~ + Owhere $0 was the initial state vector and where Ut = eitH is the
unitary evolution group generated by the governing self-adjoint Hamilto-
nian. One also needs t o look a t projected evolutions Zt = PUtP. Here P is
an orthogonal projection. It turns out that commutativity or noncommu-
tativity play critical roles in both the mathematical considerations and in
the physical outcomes.
The following was known early. Let U ( t ) be a n arbitrary unitary evo-
lution with selfadjoint infinitesimal generator H and let P be an arbitrary
selfadjoint bounded projection onto closed subspace M in Hilbert space 'H.
Let D ( T ) denote the domain of an operator T in 7-l, and R ( T ) its range.
We recall that P u t = UtP iff M is a reducing subspace for Ut, and more
generally P u t P = Ut P iff M is an invariant subspace for Ut . More general
yet, we have (see Refs. 24-27)
Lemma 5.1. The projected evolution Zt = PUtP is a semigroup for all t 2
0 iff M is a proper subspace without regeneration for Ut, i.e., PUsPLULP=
O for all t , s 2 0.
We remark that when M is a reducing subspace then Zt is a unitary
group, when M is an invariant subspace then Zt is a semigroup of partial
isometries, and if M fails t o meet the requirement of Lemma 5.1, the Zt
loses the semigroup property entirely.
The possibility of retaining a unitary, hence reversible, evolution on
socalled Zeno subspaces has been investigated in a number of recent
b o o k ~ / p a p e r s . ~For
~-~a ~better understanding of what is involved in such
formulations, we wish to take note of certain unbounded operator-theoretic
considerations. The following was clear but not published due to
the general incompleteness of my study of the Zen0 issues.
Lemma 5.2. Assume D ( H P ) is dense. Then P H P is symmetric in 'H,
and P H P is selfadjoint i f PH is closed in 'H.
Proof. Since D ( H P )= D ( P H P ) is dense, ( H P ) *and (PHP)' exist and

( P H P ) * = ( H P ) * P 2 ( P H ) P = P H P so P H P is symmetric. If PII
352 K. Gustafson
is closed, then P H = ( P H ) * * = ( H P ) * , so one obtains equality in the

previous sentence. 0
Before looking more closely a t the fundamental underlying issue of

D ( H P ) dense, we wish at this point to recall a few operator-theoretic facts
which we will use in the remainder of this paper. For more details see the
books [32,33]among others. In particular, the notion of (closed) invariant
subspace M for an unbounded operator T is better thought of in the more
general context as a decomposition of T by a direct sum IFI = M @ N of a
pair of subspaces with the requirements that the projection P on A4 map
D ( T ) into D ( T ) , T maps M into MI T maps N into N . Here P is the
(generally oblique) projection of M along N . Such decomposition of T is
equivalent to T commuting with P : PT c T P . Then T P = P T P = PT on
D ( T ) . When T is a selfadjoint operator H and P an orthogonal projection
and U, = eiHt, then when one says that A4 reduces H one is saying all of the
following: P H c H P , P : D ( H ) into D ( H ) , ( H - z I ) - ' P = P ( H - zI)-'
for all z with I m z # 0, P E ( s ) = E ( s ) P for all real s and all spectral family
projectors of H , and P u t = UtP for all real t. Our point-of-view is that
such a simplified situation is not that of the Zen0 issues. A measuring pro-
jection already within H's spectral calculus is a far oversimplified situation,
although certain proposed spin measurers may have the property.
A second set of facts to remember is that T* exists iff D ( T ) is dense,
that D ( T * ) need not be dense, but D ( T * ) is dense iff T is closeable, and
then its closure satisfies T = T**.For any two densely defined operators A
and B and if AB is densely defined, then (AB)' 2 B*A*, with equality if
A E B('H). Other conditions for equality when A and B are both unbounded
may be found in [34]. We also will use other unbounded operator theory,
such as the associativity TI(T2T3)= (TlT2)T3,without comment.
We may now sharpen Lemma 5.2 to further clarify the situation.
Theorem 5.1. D ( H P ) is dense iff P H i s closeable. T h e n (HP)' is defined

and ( H P ) ' = PH IIP H has domain at least as large as D ( H ) . Thus
generally ( P H P ) * = P H P whenever P H is closeable. Furthermore the
polar factors satisfy ImI= ( H P 2 H ) ' / ' and lHPl 2 ( P H 2 P ) l 1 2 .
Proof. Because P H is densely defined, its adjoint exists and is ( P H ) * =

H P . This operator is densely defined iff P H is a closeable operator.
Then H P is a closed densely defined operator and ( H P ) * = ( P H ) * * =
PH 3 P H . To obtain the polar factors we form (=)*PH = H P 2 H and
( H P ) * H P = =HP 3 PH2P.
Noncommutative Pigonometry and Quantum Mechanics 353
The key assumption that H P be densely defined is made through-

out the recent treatment [35] of quantum Zen0 dynamics. Their approach
of quadratic forms follows that of [36] and one is concerned with the
form J I H 1 / 2 P ~ 1 1with
2 form domain D ( H 1 l 2 P ) .The operator H p :=
(H1/2P)*(H1/2P is)associated with this form. It is noted in [35] that H p
may not be densely defined but they then assert that it will be a selfadjoint
operator in some closed subspace of H. From our analysis here we would
like to defer slightly, or a t least point out some ambiguity in such a conclu-
sion. When H 1 / 2 Pis not densely defined in H, one cannot even speak of
an operator ( H 1 / 2 P ) * ( H 1 / 2 POf
) . course one can then reduce ones con-
siderations to the smaller Hilbert space M . But if H 1 l 2 P was not densely
defined in HI then H 1 / 2will not be densely defined in M either. Moreover,
whatever its domain there, the range R ( H 1 / 21 ~ will ) generally fall a t least
partially outside of M .
The same reservation applies t o the analysis of Ref. 37. The interesting
formulation there combines continuous measurement with a coupling limit
t o force the system t o evolve in a set of orthogonal subspaces of the parent
Hilbert space. These socalled quantum Zen0 subspaces are the eigenspaces
of a Hamiltonian which is supposed to represent the interaction between the
evolving quantum dynamical system and the measurement apparatus. The
use of a superselection rule and an adiabatic theorem are assumed to deter-
mine the subspaces that the apparatus is able to distinguish. Thus the
physical description is now that of a dynamical evolution allowing chang-
ing Zeno subspaces. However, from our point of view, since the modelling
and analysis is carried out in the density matrix formulation, its underlying
rigorous validity, e.g., the denseness of the domains of the effective Hamilto-
nians in the individual Zen0 subspace evolutions, has not been considered.
A similar comment applies to the Von Neumann Subalgebras Zen0 formu-
lation [38], especially if one does not want the measuring projection E to
have t o be within the Hamiltonians functional calculus.
Finally, I would like to announce here a fundamental new theorem with
I believe fundamental implications to quantum mechanics in general. The
reversibility of a quantum mechanical Schrodinger evolution corresponds
just t o the unitarity of its semigroup eitH. The following result [23] grew out
of such thoughts combined with the considerations about domains elsewhere
in this paper.
Theorem 5.2. A unitary group Ut necessarily exactly preserves its in-

finitesimal generators domain D ( H ) . That is, Ut maps D ( H ) one to one
onto itself, for all -00 < t < 00.
354 K. Gustafson
Proof. For every -oo < t < oo the linear isometry IlUtzll = llzll property
and the commutativity property UtH c HUt guarantee that Ut maps D ( H )
one to one into D ( H ) . Is it onto D ( H ) ? Suppose not. Then for some t there
exists an z in D ( H ) which is not in the image U t ( D ( H ) ) Apply
. Url to z.
By the commutativity property again, we know z = U-tx is in D ( H ) . But
then Utz = x must have been in D ( H ) .
The result is quite evident and natural once one sees it. Therefore it may
exist elsewhere in the literature. But in a limited search, I did not find it. It
importantly distinguishes the special action of Ut on D ( H ) from its one to
one onto action on the whole Hilbert space. Remember that D ( H ) ,although
dense, is a Baire Category 1 subspace. In particular, a quantum mechanical
Schrodinger evolution Ut = eitH where H is the atomic Schrodinger partial
differential equation must take this Baire Category 1 subspace D ( H ) of all
prepared states exactly onto itself. The rest of the Hilbert space, a Baire
Category 2 set which is not a subspace, is separately taken onto itself. Thus
Theorem 5.2 expresses a global regularity of exact domain preservation.
Contrast that with the heat equation evolution Zt = e&*, which in-
stantly takes the whole Hilbert space (e.g., L z ( - m , GO)) into Coo(-oo,oo).
See for example the rigorous proof in Ref. 39, pp. 128-131. The Heat evolu-
tion operator Zt is an infinitely smoothing operator. Its smoothing action is
a local regularizing of each initial given heat distribution f E L z ( - m , oo),
whether f be in the domain W232(-oo,oo)of the Laplacian -A, or not.
Instantly, information in f is spread out and lost. In particular, the domain
W2i2is immediately lost. Not so in the Schrodinger evolution: its unitarity
requires that W2,2be mapped exactly onto W 2 >by 2 Ut at every instant t ,
according to Theorem 5.2.
To express our interpretation of Theorem 5.2 in dynamical system terms,
the Schrodinger partial differential equation initial value problem
$(t) = iH$(t), t > 0
$ ( O ) = $0 given
has two classes of states: $0 in the domain D ( H ) ,and the rest. Only those
$0 in D ( H ) are legally allowed wave packets from the differential opera-
tor theory, in which it is essential that the bounded evolution operator Ut
commute with the infinitesimal generator H . Stated in elementary terms:
otherwise you cannot substitute the solution back into the differential equa-
tion.
One can prove Theorem 5.2 in another way which is more useful when
one wants to consider also contraction semigroups Zt with infinitesimal
Noncommutative %gonometry and Quantum Mechanics 355
generator A. To that end we first establish two lemmas.
Lemma 5.3. Let T be a closed densely defined operator in a Hilbert space

FI and let Ut = eitH be a unitary evolution there which commutes with T .
Then Ut remains unitary o n the graph-norm Hilbert space FIT.
Proof. HT is the Hilbert space D ( T ) equipped with the inner product

+
(2, y ) r = (2, y) ( T z ,Ty). By the commutativity UtT C TUt we have the
isometry property retained,
Also the adjoint U,+ considered in HT is the same as the original adjoint,
using again the commutativity,
(UtG Y ) T = (Utx,Y) + (UtTx,TY)

= +
(x,U,*Y) ( T x ,TU,*y) = ( 2 ,U,*Y)T (23)
Thus U; = U r l in the original space carries over to HT. 0
In particular, Lemma 5.3 provides an alternate proof of Theorem 5.2.

Let T be H . Then Ut and U,+ are both onto HT = D ( H ) and the rest
follows easily from bounded operator theory.
Lemma 5.4. Under the conditions of Lemma 5.3, a contraction semigroup

evolution Zt = etA remains a contraction semigroup on F I T .
Proof. As above. Note that the assumed semigroup commutativity prop-

erty ZtT C T Z t is essential, as is its subproperty that Z, map D ( T )
into D ( T ) In particular, one may take T = A, the infinitesimal generator
of zt 0
Now we can provide a partial converse to Theorem 5.2 in the sense of

asking, suppose a contraction semigroup Zt exhibits the regularity preser-
vation property that it map the domain D ( A ) of its infinitesimal generator
one to one onto itself. What unitarity properties does Zt exhibit for all
t 2 O?
Proposition 5.1. Let Zt = etA be a contraction semigroup on a Hilbert

such that 2,maps D ( A ) one to one onto D ( A ) . Then Zt on HA
space H
can be extended to a group 2-t = 2;.
356 K. Gustafson
Proof. By Lemma 5.4 above with A = TI we know 2,remains a contrac-

tion semigroup on 'HA. Because Zt maps 'HA1-1 onto itself, we know ZF1 is
bounded and also maps 'HA 1-1 onto itself. By known results in semigroup
theory (e.g., see Ref. 33, p. 393), we may extend Zt to a group by defining
Z-t = Zcl. We mention that 2: remains the same in 'HA as it was in 'HI
so that one can relate the adjoint semigroups by (Z,")-' = (Z;')*, pursue
further converse statements for the original Hilbert space 'HI etc., which we
will not do here. One reason we present these results here is in hopes that
they be used elsewhere by those working in Zen0 theory. U
What Theorem 5.2 states within the context of unitary quantum me-
chanical evolutions is that one cannot losc any of the totality of detail
embodied in the totality of wave functions $ in D ( H ) . It is quite interest-
ing to think about how much 'mixing around' of D ( H ) the evolution Ut
can do. Perhaps one will investigate such questions elsewhere, e.g., from
the viewpoint of ergodic theory. However, just by itself, Theorem 5.2 says
that to have reversibility in quantum mechanics, you cannot lose a single
wave function from D ( H ) as you proceed forward in time. In other words,
you must be able to account for all evolving domain probabilities, future
and past, all of the time.
This finding also gives renewed importance to the role of preparation
of states in quantum mechanics. The ensemble ( $ 0 ) of experimental initial
states which are prepared to simulate a posteriori in an experiment what
you could expect a priori from all possibilities from D ( H ) ,must be suffi-
ciently extensive within D ( H ) in such a way that their trajectories continue
to predict properly during evolution. Ideally, you need a dense subset from
D ( H ) . Experimentally of course, one must settle for less.
To emphasize the above discussion, which to us seems important, we
may state, at the risk of sounding repetitive, the following.
Corollary 5.1. A quantum mechanical evolution Ut must and does contin-

uously and simultaneously account f o r all probabilities I$(t)12and all wave
functions +(t) in the domain of its Hamiltonian. No probabilities m a y be
lost.
It is important to note that Theorem 5.2 is of wider interest, i.e., it

will apply to any unitary evolution Ut = eitH in any context. For exam-
ple, even if one has "resorted" to the Heisenberg interaction picture and
to unitary evolutions p ( t ) = U,poU;, one knows from Theorem 5.2 that
the underlying Hamiltonian must have the domain regularity preservation
Noncommutative Pigonometry and Quantum Mechanics 357
property. Perhaps more interestingly, effective interaction Hamiltonians

such as those assumed in quantum dynamical d e ~ o u p l i n g ~ must~ ) ~ also
enjoy and respect the domain regularity preservation property of Theorem
5.2. Our point of view is that in such situations, this is a physical property
that one should try to understand mathematically, in order to obtain a
deeper understanding of the physics.
There is another way to place Theorem 5.2 in context. There is a school
of thought (see Refs. 41,42) in physics which views irreversibility as going
out of the Hilbert space. This point of view certainly has some merit and
goes back in physics to the notion of Rigged Hilbert space, among others.
However, as I pointed out in Ref. 43, one is already out of the Hilbert space
whcncver one needs to consider weak solutions in partial differential equa-
tions, a concept going back to the 1930s. Also, and in quantum mechanics,
the radiation eigenfunctions over the positive (scattering) spectrum of the
Schrodinger Hydrogen operator are already non- L2 functions. So we con-
struct L2 wave packets, or work in some Banach function space, or a Rigged
Hi1bert space.
What Theorem 5.2 says, within this context, is that you do not even
need to go out of the Hilbert space to get irreversibility. You need only
lose one wave function II,from the domain D ( H ) of the Hamiltonian. Then
the unitarity of the dynamical system has been lost. The wave function
II, can remain in the Hilbert space. Should Ut immediately reconstitute its
unitmity, the lost wave function 1c, would continue its trajectory under the
dynamical system Ut but it could not re-enter the eligible D ( H ) prepared
states until the evolution operator Ut again lost its unitarity for an in-
stant and somehow switched II, back into D ( H ) as another wave function
gets thrown out. I have not seen this conceptual picture anywhere in the
quantum theory literature. I would assert that going out of the operator
domain is a much sharper characterization of irreversibility.
Elsewhere recently44 I showed that a number of Nobel Prizes depended
in a fine analysis on an assumption of detailed balance a t a microscopic
physical level. This then leads to a selfadjoint operator and hence to an
implicit unitary evolution. The detailed balance was usually a t the level of
some, e.g., microscopic reaction operators. I may here interpret Theorem
5.2 as a more precise detailed balance statement. It is at the wave function
level and says that Ut must preserve the balance between any two wave
functions, and not lose any of them.
From the above considerations, which are of course important in them-
selves, there now enters the question of how to bring the operator trigonom-
358 K. Gustafson
etry to further enrich our understanding of the microscopic dynamics tak-

ing place when quantum measurements are performed. The very act of
a quantum measurement, except in the case when measuring instrument
and the dynamics commute, involves an interaction which is, intuitively,
irreversible. This must be, in my opinion, the key ansatz in the Zen0 prob-
lem, and we asserted it in our discussion a t the beginning of this Section.
On the other hand, Theorem 5.2 (and the results following) involve only
a commuting situation. That gave the important new results about micro-
irreversibility as the loss of a single wave function. Can the noncommutative
operator trigonometry now address the measurement (noncommutative)
dynamics? It should be remembered that to date, in my opinion, no-one
has successfully treated the quantum measurement interaction dynamics in
a mathematically rigorous way.
In Ref. 45 a beginning may have been made. There I showed that under a
multiplicative perturbation H 4B H , although the direction of time is not
changed, the speed of evolving dynamics has been altered. Moreover, there
always exists a small disturbance such that a unitary evolution U, becomes
a completely nonunitary contraction semigroup. In [46] a preliminary theory
of interaction antieigenvalues for interacting operators A and B has been
worked out. See Refs. 47,48 for some earlier discussion of the Bell and Zen0
problems.
Acknowledgements
The invitation to speak at the ICAAA 2005 and the warm hospitality shown
me in Quy Nhon by Professors Nguyen Minh Chuong and Nguyen Minh Tri
and the others there was greatly appreciated. The invitation to speak at
the Institute of Mathematics in Hanoi afterward and the warm hospitality
offered me there by Professor Le Tuan Hoa and his group was equally
appreciated.
References
1. K. Gustafson, Lectures on Computational Fluid Dynamics, Mathematical
Physics, and Linear Algebra, (World Scientific, Singapore, 1997).
2. K. Gustafson and D. Rao, Numerical Range: The Field of Values of Linear
Operators and Matrices, (Springer, Berlin, 1997).
3. K . Gustafson, Noncommutative Tbigonometry, Operator Theory: Advances
and Applications: Wavelets, Multiscale Systems, and Hypercomplex Analysis
(D. Alpay, ed.), Vol. 167 (2006), pp.127-155.
4. K. Gustafson, Lin. Alg. & Applic. 319,117 (2000).
Noncommutative Rgonometry and Quantum Mechanics 359
5. K. Gustafson, N u m . Lin Alg. with Applic. 34, 333 (1997).

6. K. Gustafson and N. Sobh, Computer Physics Communications, 65, 253
(1991).
7. K.Gustafson, Linear and Multilinear Algebra 37,(1994) 139.
8. K. Gustafson, Antieigenvalues in Analysis, in Fourth International Workshop
in Analysis and its Applications, Dubrovnik, Yugoslavia, June 1-10, 1990 (C.
Stanojevic and 0. Hadzic, eds.), Novi Sad, Yugoslavia (1991), pp. 57-69.
9. K. Gustafson, N u m . Lin Alg. with Applic. 10,291 (2003).
10. K. Gustafson, Lin. Alg. tY Applic. 354,151 (2002).
11. K. Gustafson, Research Letters I n f . Math. Sci. 8 , 105 (2005).
12. K. Gustafson, International Statistical Review, 74,187 (2006).
13. A. Einstein, B. Podolsky, and N. Rosen, Physical Review 47,777 (1935).
14. J . Von Neumann, Die Mathematische Grundlagen der Quantenmechanik,
(Springer, Berlin, 1932).
15. J. Bell, Physics 1, 195 (1964).
16. A. Aspect, J. Dalibard and G. Roger Physical Review Letters 49,1804 (1982).
17. E. Wigner, American Journal of Physics 38, 1005 (1970).
18. J. Clauser, M. Horne, A. Shimony, and R. Holt, Physical Review Letters 23,
880 (1969).
19. K. Gustafson, The Geometry of Q u a n t u m Probabilities, On Quanta, Mind,
and Matter: Hans Primas in Context, (H. Atmanspacher, A. Amann, U.
Mueller-Herold, eds.), Kluwer, Dordrecht (1999), pp. 151-164.
20. K. Gustafson, Bell's inequalities, in The Physics of Communication, Proceed-
ings of the X X I I Solvay Conference o n Physics, (I. Antoniou, V. Sadvnichy,
H. Walther, eds.), (World Scientific 2003), pp. 534-554.
21. B. Misra and G. Sudarshan, Journal of Mathematical Physics 18, 756 (1977).
22. W. Guthrie, A History of Greek Philosophy, V . 2: T h e PreSocratic Tradition
f r o m Parmenides t o Democritus, (Cambridge University Press, UK, 1965).
23. K. Gustafson, Reversibility and Regularity, to appear in International Jour-
nal of Theoretical Physics.
24. K. Sinha, Helvetica Physica Acta 45, 621 (1972).
25. K. Gustafson, On the "Counter Problem" of Q u a n t u m Mechanics, unpub-
lished (1974), pp. 14.
26. K. Gustafson, S o m e Open Operator Theory Problems in Q u a n t u m Mechanics,
Rocky Mountain Mathematics Consortium Summer School on C" Algebras,
Bozeman, Montana, August 1975. Unpublished notes, (1975), pp. 7.
27. K. Gustafson, Irreversiblity questions in chemistry, quantum-counting, and
time-delay, In Energy Storage and Redistribution in Molecules, (J. Hinze,
ed.), (Plenum Press, 1983), pp. 516-526.
28. M. Namiki, S. Pascazio, and H. Nakazato, Decoherence and Quantum Mea-
surements, (World Scientific, Singapore, 1997).
29. P. Facchi, V. Gorini, G. Marmo, S. Pascazio, and G. Sudarshan, Physics
Letters A 274,12 (2000),.
30. P. Facchi, S. Pascazio, A. Scardicchio, and L. Schulman Physical Review A
65,012108 (2001).
31. S. Tasaki, A. Tokuse, P. Facchi, and S. Pascazio, International Journal of
360 K. Gustafson
Quantum Chemistry 98, 160 (2004).

32. T. Kato, Perturbation Theory for Linear Operators, (Springer, Berlin, 1980).
33. Riesz, F. and Sz.-Nagy, B. Functional Analysis, (Ungar Publishing, New
York, 1955).
34. K. Gustafson, A Composition Adjoint Lemma, Stochastic Processes, Physics,
and Geometry: New Interplays. I1 (F. Gesztesy, H. Holden, J. Jost, S. Paycha,
M. Rockner, S. Scarlotti, eds.), American Mathematical Society, Providence,
2000, pp. 253-258.
35. P. Exner, and T . Ichinose, Product formula related to quantum zeno dy-
namics, in Proceedings of the X I V International Congress on Mathematical
Physics, (Lisbon, World Scientific, 2003). See also arXiv:math-ph/0302060.
36. C. Friedman, Indiana University Mathematics J . 21, 1001 (1972).
37. P. Facchi, S. Pascazio, Quantum zeno subspaces and dynamical superselec-
tion rules, in The Physics of Communication (I. Antoniou, V. Sadovnichy, H.
Walther, eds.), (World Scientific, 2003), pp. 251-286.
38. A. Schmidt, Journal of Physics A 35, 7817 (2002).
39. K. Gustafson, Partial Differential Equations and Halbert Space Methods, 3rd
Edition, Revised, (Dover Publication, N. Y . , 1999).
40. L. Viola, E. Knill, and S. Lloyd, Physical Review Letters 8 2 , 2417 (1999).
41. I. Prigogine, Is Future Given? (World Scientific, Singapore, 2003).
42. I. Antoniou, G. Lumer, (eds.) Generalized Functions, Operator Theory, and
Dynamical Systems, Chapman and Hall, CRC Research Notes in Mathemat-
ics 399, (London, 1998).
43. K . Gustafson, Rigged Spectral States: A Proclivity for Eigenvalues, in Gen-
eralized Functions, Operator Theory and Dynamical Systems (I. Antoniou,
G. Lumer, eds.), CRC Research Notes in Mathematics 399, London, 1998,
pp. 48-55.
44. K. Gustafson, Discrete Dynamics in Nature and Society 8 , 155 (2004).
45. K. Gustafson, Semigroups and antieigenvalues, in Irreversibility and Causal-
ity (A. Bohm, H. Doebner, P. Kielanowski, eds.), Lecture Notes in Physics
504, Springer, Berlin (1998), 379-384.
46. K. Gustafson, J . Math. Anal. A p p l . 299, 174 (2004).
47. K. Gustafson, The quantum zeno paradox and the counter problem, in Foun-
dations of Probability and Physics-2 ( A . Khrennikov, ed.), (Vkixjo University
Press, Sweden, 2003), pp. 225-236.
48. K. Gustafson, International Journal of Theoretical Physics 44, 1931 (2005).

Advances in Deterministic and Stochastic Analysis (2007)

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

Advances in Deterministic and Stochastic Analysis (2007)

Diunggah oleh

Hak Cipta:

Format Tersedia

Advances in

British Library Cataloguing-in-PublicationData

ADVANCES IN DETERMINISTIC AND STOCHASTIC ANALYSIS

Printed in Singapore by World Scientific Printers ( S ) Pte Ltd

and the possible appearance of shock waves to a class of equations gen-

Chapter I Integral and Pseudodifferential Operators

$1.Pseudodifferential Operators of Second Order with Sign-Changed

$2. A Semilinear Nonclassical Pseudodifferential Boundary Value

$3. Singular Integral Operators in Functional Spaces of Morrey

$4. Classification of Integral Transforms 43

Chapter I1 Partial Differential Equations

$5. Unified Minimax Methods 75

$6. Some Remarks on Single Conservation Laws 91

$7. Iterative Method for Solving a Mixed Boundary Value Problem

$8. Numerical Solution to a Non-Linear Parabolic Boundary

$9. A Class of Solutions to Maxwells Equations in Matter

510. On the Cauchy Problem for a Quasilinear Weakly Hyperbolic

511. Some Singular Perturbation Problems Related to the

Chapter I11 Geometric Analysis

$12. Monotone Invariants and Embeddings of Statistical

$13. Graded Cech Cohomology in Noncommutative Geometry 255

$14. Sobolev Spaces with Weight on Riemannian Manifolds 269

Chapter IV Stochastic and Infinite-Dimentional Analysis

$15. Stochastic Population Control and RSDE with Jumps 281

$16. Noncausal Stochastic Calculus Revisited - Around the

5 17. Infinite-Dimensional Stochastic Analysis and Foundations of

818. Noncommutative Trigonometry and Quantum Mechanics 341

51. PSEUDODIFFERENTIAL OPERATORS OF SECOND

NGUYEN MINH TRI

In this paper we prove hypoellipticity for pseudodifferential operators of second

In this paper we consider operator of the form

in a bounded domain R c R$' = IRE x @, where XI,. . . ,xk are proper

2000 Mathematics Subject Classification Primary 35H10; Secondary 35A08, 35B65,

(x,t ) -+ (21,t); 5 = 5 , t = 6t; 6 E (0,1].

For a function u E C r ( 0 ) we can define the function E C r ( 0 ) as

Note that when 6 = 1 we can identify the coordinates ( z , t )and ( d , t ) .

we can get easily the desired estimates.

Let Ps be the operator in the coordinates (d,

Lemma 2. Let the operator Pa have the form ( 2 ) . Then f o r arbitrary s E R

where 11 . ]Is, ( , ) s are taken in the (x,t) coordinates.

Proof. Put Bj,6 = 6tXj,s - (dtXj,s)*. Obviously, ord Bj = 0 (since Xj,a

Thus, Lemma 2 is proved. 0

Corollary 1. From the inequality (3) it implies that

where o([t,X j , & ]denotes

Lemma 4. Assume that X j satisfies the condition of Lemma 3. For every

is a homogeneous function of degree one in ((1, . . ., Jnl 7) we have the fol-

Putting in Eq. (5) (1 = . . . = Jn = 0 , =~1 we get

Note that if X j satisfy the condition (4) in Lemma 3, then, equivalently,

where 11 . / I s , (., .)s are taken in the coordinates (d,

Proof. Consider the expression

By Lemma 3, the operator P6 may be rewritten in the form

where ord (pj,a = 0. It follows that

where @a may be estimated by

Let us estimate the term -6

It remains to estimate the second term in Eq. (10)

Put Dj,3,6= [Xj,hlA'23-1]. Since ord Dj+,g = 2s - 1, the second term in

It remains to estimate the first term in Eq. (12)

where C is the constant appearing in Lemma 4. By choosing 60 =

Corollary 2. Under the hypothesis of Lemma 5 the following inequality

Corollary 3. Under the hypothesis of Lemma 5 the following estimate

for some constant C = C(s, K ) .

Proof. First choose 6 = 2 , g = f6 in Eq. (15), then apply Lemma 1 we

Corollary 4. Let the hypothesis of Lemma 5 be fulfilled and the algebra

for some constant C = C(s, K).

/(l+ It[+ IEnl)-dEn = (1 + ltl)-+l s,(1+ Itl)-*dt,