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
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
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
The Editors
vii
CONTENTS
Preface V
INTEGRAL AND
PSEUDODIFFERENTIAL OPERATORS
This page intentionally left blank
Advances in Deterministic and Stochastic Analysis 3
Eds. N. M. Chuong et al. (pp. 3-13)
@ 2007 World Scientific Publishing Co.
j = 1 , ..., k.
Xj(~,0,0,1)=0,
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,)
ug(z, t) = u ( 5 ,t6).
c k
j=1
ll6txj,6911,2 6 21(6t%71g>sl + C(~)lIgl1,2, vg E C,-(K), (3)
where
a k
Qs = -t:
at
+ btcs(x,t) - S2tXj,s(t)Xj,6; ord Qs = 1.
j=1
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
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)
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)
-6 C a
t'Xj,sg, -[Xj,a, (12)
j=1
k ( at'
j=1
We have
Consequently
By Lemma 4
for all g E C r ( K ) .
10 N . M. Tri
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
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.
Therefore
+IlXE * fll, G Ml
uniformly in E. Passing to the limit as E 4 0 we obtain f E Hq+'. By the
Schwarz inequality
j=1 j=1
P.d.0 of Second Order with Sign-changed Characteristic Form 13
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).
This page intentionally left blank
Advances in Deterministic and Stochastic Analysis 15
Eds. N. M. Chuong et al. (pp. 15-32)
@ 2007 World Scientific Publishing Co.
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
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 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
hence,
3. Pseudodifferential Operators in R
= (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
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
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
then
Semilinear Nonclassical P.d. BVP in Sobolev Spaces He,p 21
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
aA(<,4) =
bI+P=-m
qp<" /
yq
e-i(Zif)K,p(z)dz.
4. Pseudodifferential Operator in R y
Q = { z E Q'lyl 5 argz 5 7 2 ) .
24 N . M . Chuong and D. A . Than
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
It is easy to prove
Proof.
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
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
Q = { z E Clyi L argz I 7 2 } .
We now study the boundary value problem
It is obvious that
Ellipticity C o n d i t i o n
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
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
Semilinear Nonclassical P.d. BVP in Sobolev Spaces He,p 29
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).
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.
30 N . M . Chuong and D. A . Than
0, if x is outside of r d
hd(x) =
1, if x is in
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 .
(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
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).
Advances in Deterministic and Stochastic Analysis 33
Eds. N. M. Chuong et al. (pp. 33-42)
@ 2007 World Scientific Publishing Co.
LUBOMIRA SOFTOVA
Postal address: Politecnico d i Bari
Dipartimento d i Matematica
Via E. Orabona 4 ,
70125 Bari, Italy
E-mail: 1ubaQdm.uniba.it
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
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
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
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
.
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,
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.
for any integer 1 > 1. Replacing the kernel with its expansion, we get
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
s :dy)L4ds))
I'Hsm(Y - x) - 'Hsm(Y)l dY
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
(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
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
:= 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
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
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.
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).
Advances in Deterministic and Stochastic Analysis 43
Eds. N. M. Chuong et al. (pp. 43-72)
@ 2007 World Scientific Publishing Co.
VU KIM TUAN
Department of Mathematics,
University of West Georgia, Carrollton, GA 30118, USA
E-mail: vu@westga.edu
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
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.
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
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
f +g=k* f.
Depending on the convolution, convolution transforms can be divided into
following types:
d z ) = (Hf)(z)=- fody,
Y-2
&dz.
Y-X
dz)= k(zy)f(y)dy.
f(Y) = (Hvg)(Y) =/ 0
fiJv(zy)s(z)dz. (15)
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
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 Hankel functions HL')(z) and HL2)(z)in the kernels' that are
unitary on L2(-m, m).
Classification of Integral Transforms 49
0 < c < Ik*(S)Sal < c < 03, a > 0,W(s) = 1/2. (32)
with the confluent hypergeometric function 1F1 ( a ;c;x ) in the kernel, the
Saigo transform28
50 V. K. Tuan
Y X
F3(-a', n - a , -b', n - b; n - c; 1 - -, 1 - -)g(z)&, 0 < Rc < n,
X Y
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
with the modified Bessel function K,(z) and I,(z) in the kernels,' the
Meijer K , transform
ds) = 1 0
M
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
or of Laplace type
c*=p+q-2(m+n)>0. (61)
with the real variable inverse of the form
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.
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
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
Jo * 4
Classification of Integral Transforms 55
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
and therefore,
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
Then
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
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,
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+.
As examples we can take
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 ~ ' ~ ~ ~
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)
rcsinh2.rrxJg(x)12dx= 1 00
If(y)12dy. (94)
Classification of Integral Transforms 59
1 00
y1-2alf(y)12dy =
1 00
1 +
z s i n h 2 m p 2 ( a iz)g(z)12dz. (95)
k ( z , y) = -
27r Jm e-izt+iYV(t)dt (96)
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
Lebedev transforms.
60 V. K. Tuan
- If cp(t) = sinht + at, a > -1, we obtain the following pair of integral
transforms
f ( y) = - 1
Y
/
--m
00
as a special case.
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
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
also holds.
Here is the composite structure
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 > $, c1 and c2 are constants, which may be different for x --t +m
and x .+ -CQ.
To derive composition structure and a Plancherel-type theorem for the
Cherry transform, we rewrite the Cherry expansion (112) in the following
form:
and
and
Classification of Integral Transforms 63
into the formulas (7) and (8), and interchanging the order of integration
that is permissible if f ( y ) E S(R), the Schwartz space of all infinitely
differentiable functions which rapidly vanish at infinity together with all
derivatives, we arrive at the following composite representations for the
transforms D+, D-, 0; and DI1:
and
Here M and M- are the Mellin transform and its inverse, and F and F-
are the Fourier transform and its inverse. The composite representations
(13) and (14) of the transforms D+ and D- are key tools in studying the
Cherry expansion (6).
By L:(R;e*) (similarly, by L,(R;e*)) we denote a subspace of
&(R) consisting of all square-integrable functions f ( y ) such that the
Fourier transforms of f(y)e$ 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
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
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
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
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:
(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).
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
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 a bounded operator from L2(0,m) onto L2(-00, m) and has the inverse
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).
Classification of Integral Transforms 71
MARTIN SCHECHTER
Department of Mathematics University of California, Irvine
Irvine, C A 92697-3875
E-mail: mschechtOmath.uci.edu
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.
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
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)
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
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.
AnB=$ (20)
and
B n K # 4, K E K.
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.
bosa<co (23)
and (3) holds.
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.
Assume also that for each K E K there is a p > 0 such that G ( u )is Lipscluitz
continuous on
We present some linking methods which are special cases of linking with
respect to minimax systems.
We have
We also have
Assume that
a : = inf sup G ( r ( s ) u ) (34)
0<5-<1
uEA
Kc u
tE[o,Tl
a(t)AUa(T)K.
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
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=
$(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
We shall prove
+(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.
a 5 inf G.
B
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.
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
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
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
+
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
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
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
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
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).
Unified Minimax Methods 89
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
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
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-.
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
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).
Fig. 1.
Fig. 2.
Some Remarks on Single Conservation Laws 97
Fig. 3.
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.
and
which leads us to
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
This implies that the entropy condition is satisfied for t > T . Summing up
the above results, we have the following
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
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.
Advances in Deterministic and Stochastic Analysis 103
Eds. N. M. Chuong et al. (pp. 103-113)
@ 2007 World Scientific Publishing Co.
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.
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
and denote
B as follows:
B : w + Bw,
e)
where
w= , Bw = (i."Ir1)
'p+bu
,
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
Setting
BW= F. (23)
Property 1. B is symmetric in H
Proof. We have
Therefore,
( 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,
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:
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
Now , denote
+
(blavsl2 blzs12)dx. (39)
110 Q. A Dang and L. T. Son
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
i=l
where
(_l)N+l-iiN+l
"li =
z. ! ( N + 1 - i)! ' (41)
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
we obtain
where
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.
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).
This page intentionally left blank
Advances in Deterministic and Stochastic Analysis 115
Eds. N. M. Chuong et al. (pp. 115-129)
@ 2007 World Scientific Publishing Co.
where a: > 0 is given and y(z, t ) = y(z, t ;u ) is the solution to the non-linear
parabolic problem
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
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 .
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
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).
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 .
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.
Yk E ah(xk).
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)
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
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
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,
Numerical Solution to a Non-Linear Parabolic Boundary Control Problem 125
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.
R__mdn.rnWMn
(c) (4
Fig. 2. Reconstruction of the control function of example 1, with a = n/2
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.
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
Numerical Solution to a Non-Linear Parabolic Boundary Control Problem 129
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).
This page intentionally left blank
Advances in Deterministic and Stochastic Analysis 131
Eds. N. M. Chuong et al. (pp. 131-175)
@ 2007 World Scientific Publishing Co.
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.
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
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)
( V , B) = 0 ,
47T i ( H , d s ) = 4-7TI
(V,H)=-J,
C C
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
(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 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.
and
and
2
A=
a2(cosh 2u - cos 2v)
Solutions to Maxwell's Equations 139
KCyl: +
z = a2(cosh2u cos2v sinh2u sin2 v ) , 0 5 u 5 uo, 0 5 v 5 2n,
where
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.
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
= (Lr
f w
s_dd J ( X
d
- Ic)2
dzdx
+ ( y + !) + ( 2 -
dzdx
Z)2
By(5)= --
a9
dY
= -M 1: s_dd (e 2) - dzdx,
and
Solutions to Maxwells Equations 141
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),
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
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
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
Solutions t o Maxwells Equations 143
o Bx
-1
-2
Y
X
0 By
-1
-2
Y
x
-2
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.)
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
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
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
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,
In order to obtain formulae for the magnetic field BE, two cases need
to be considered.
CASE I: T > T O
Jo
The azimuthal component B,#,of BE is obtained in a similar fashion. As
B+ = (-1/r)(N/a4), the result is
Using the fact that d1l/dr = (1/2)[10 + 121 (Ref. 3, p. 397, 11.116) gives
*m
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
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
Proof. See Ref. 5 or Ref. 8, for instance. For definitions and results see
Ref. 1 (Ch. 8) or Ref. 3 (Ch. 12). 0
I sin#cos[m(+ - $)]dq5 =
ifm>l
.rrsin+ if m = 1.
Therefore, in the third term inside the bracket only the m = 1 summand
contributes to the sum over m. Thus,
150 P. Massopust
The integral over 8' can be exactly evaluated using the following approach.
Let u = cos 8'. Then
= 1' JsPt,l(-u)du
= ( - ~ ) ~ + l /d' 3 P e , l ( ~ ) d ~ .
0
Now,
(Ref. 3, p. 424, 12.34 and 12.35) and the moment integral is equal to
11uP*(u)du= { ~ ~ ~ + ~ ~
if[= 1
i f e+= 2A X)
2
if l = 2 X + 1,
(23)
(Ref. 1,8.14.15).In the equation above, r denotes the real Gamma function,
which is the extension of the factorial to real numbers: r(z)= (z+l)!.Using
the fact that (Ref. 1, 6.1.17 and 6.1.12)
Solutions to Maxwells Equations 151
S,, (X+1)(2X)!!
sin2 epe,l(cos e)de =
(I
Therefore one arrives a t the following result.
ife=1
if C = 2X + 1.
(24)
5
+
X=l
(-1)X+(2X - 3)!! p;x
(2X+2)!!
2x+1p2X,1(cos0) sin 4
p,
1
where
(-1)x+1(2X - 3)!!
Nx =
(2X + 2)!! .
B, =coseB,-sinBBe.
4.1. Semi-spherical V
To complete the discussion of spherical V s , an exact expression for the
magnetic flux leakage field BE generated by a semi-spherical V is derived.
where
and
(e- m)! p:
e=i m=l
x
x/2
/x
2x
sin2d'sin 4'cos[m(4 - 4')]Pem(cosO')d$'dO' Pem(cos0)
154 P. Massopust
where
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
Solutions to Maxwell's Equations 155
and
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.
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.
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
and
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
+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
with E< = min{E, E l } and E> = max{E, E}. Finally, the following result is
obtained.
Solutions to Maxwells Equations 159
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,
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).
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
-
-
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
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.
Solutions to Maxwells Equations 163
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
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.
+ cos[k(z - ~)]~e~~+l(~)~e2~+1(~)9,]
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
= -4T/C.
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
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 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
+ 12*
luo cos [Z - :(cash 2.11 - cos 2 4
1
x [sinhusin3v - sinh3usinvl se2,+1(vl k)se2,+1(v, k )
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)
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
and, in particular,
(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
one obtains
and
C O S ~ ~ + ~ sin 3v
( ~ ~ sin(2v
) + 1)vdv
( ~ ~2(v
C O S ~ ~ + ~ [cos ) + 2)v - cos 2 ( -~ l)w] dv (44)
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
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
] 12Tk:[
sin k) du
cos 2v] sin v~e2~+1(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)
- 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
Proof. The formula for Gzn+l is essentially given by Szn+1 with sehzn+1
replaced by gekhzncl.
In summary,
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
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
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).
Solutions t o Maxwells Equations 175
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
0,Vai /
where the coefficients A,B,C are calculated at
'--\-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
ax,
L aa2
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
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 ) ) .
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
The Cauchy problem for system (14) consists in looking for U ( a 1 ,a2) E
C 1 such that
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
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
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.
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
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
Pl = Pl,
(22)
P
2 + a12(X,2,P ) X l + a22(X,2,PIX2 =p2
where
and
Proof. Set
{
fl
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
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
Cauchy Problem for a Quasilinear Weakly Hyperbolic 187
- -
-&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)
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)
Cauchy Problem for a Quasilinear Weakly Hyperbolic 189
+ ( - =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=
Since (X, p ) are sufficiently closed to (X', Po),from (41) and (21) it follows
that d # 0. The assertion 2) follows now.
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
Cauchy Problem for a Quasilinear Weakly Hyperbolic 191
P2 -
192 H. T. Ngoan and N. T. Nga
and
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
XIT =o,
A- & =O.
dP2
6. Examples
AX2 - P2)-
aPl + ( 4 x 1-
aa1
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).
+
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.
(-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
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
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
'-
at
- EAUE + (U".V) u~ + v p " = f, in R,,
\ uEl,,o = u o .
198 M . Hamouda and R. T e m a m
'
EAU"- UR3v" + (v'.V) ue + Vp" = f,
aVE
-_ in R,,
at
div u" = 0, in R,,
\ uEl*=O= uo.
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
1 -U +
D ~ P 0p0 = j , in R,,
I~30div u0 = 0, in R,,
I 0
u =O,
= 0, on
on
ro,
rh,
(4)
{ 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)
avo
-- UD~U+
O Vpo = f,
at (7)
and, for j >_ 1:
-
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
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.
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.
202 M . Hamouda and R. T e m a m
-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)
div v 1 = 0, in R,,
{
v31 = 0, on ro,
v1 = O , on F h ,
on rourh.
As at order zero, we instead construct an approximation for 0'1' as follows:
@IE = +
e.s.t., with a tangential component
Singular Perturbation Problems Related to the Navier-Stokes Equations 203
?)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,
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.
is given by
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)
206 M . Hamouda and R. T e m a m
at
d
p s = --(a;) cN
j=O
&J+1 ( j
.j+l
+ I)! U 2 N + 1 - j
---UZ/E-
we deduce that
where
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.
+ 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
208 M . Hamouda and R. T e m a m
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
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.
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
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
+ 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.
in ,a,
in R,,
IvE=O,
'JI
on ,,?I
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.
v0 = 0 , on Fh,
Remark 5.1. The existence of a smooth solution for the system (53) is
proved in Ref. 8.
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
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)
First, we infer from (56) the equations satisfied by the mode w1 at order
one. These are described by the following system:
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,
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 5.5. 1. The approximation "3: of O:+ satisfies exactly the same
equation (60) as @:YE.
-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),
rh.
rh, (65)
FlyEare
216 M . Hamouda and R. T e m a m
Proof. Let
Before writing the equations satisfied by w ' ~ ~we, start by expanding the
nonlinear term:
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
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
Singular Perturbation Problems Related to the Navier-Stokes Equations 217
{
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
obtain
218 M. Hamouda and R. T e m a m
where,
k3i' =- U D ~ W. ~wli"dR
~ ' = 0,
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)
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
The same estimates are also valid for the tangential derivatives of ( p l l E .
~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
{
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
Let us now recall the statement of the Hille-Yosida theorem (for more details
see for example Refs. 1,2,5 and 13):
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.
+
(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
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.
(89)
where X is chosen such that X 6 {Ulkl, k E Z2}.
Lemma 0.1. We have p E L2(R).
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:
(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:
Let us state and prove the following Lemma in order 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.
226 M. Hamouda and R. T e m a m
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)).
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)
Singular Perturbation Problems Related t o the Navier-Stokes Equations 227
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).
This page intentionally left blank
Chapter I11
GEOMETRIC ANALYSIS
This page intentionally left blank
Advances in Deterministic and Stochastic Analysis 231
Eds. N. M. Chuong et al. (pp. 231-253)
@ 2007 World Scientific Publishing Co.
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.
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.
l p ( z , w ) d w = l'dz E M .
g F ( V ,W ) X = /R
(8vlnp(z, W ) ) ( a wlnp(z,W ) ) P ( Z ,w ) . (3)
T ( X , Y ,2)=
I (axlnp(~,w))(dyInp(~,w))(~zln~(~,w))p(~,w).
(7)
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
Monotone Invariants and Embeddings of Statistical Manifolds 237
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.
Clearly we have
0 I M1(T)I M 2 ( T )I M 3 ( T )I IITII.
+ + + + + +
- 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.
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
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.
A ~ ( T:=
) max T ( x ,Y, 2)
Ix)=ly~=lzl=1,<z,y>=<y,z>=<z,x>=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 .
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.
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
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
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.
i=l l<i,j,k
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).
L(v2):= -/Tl
(/F, /?-, -/?). (16)
Monotone Invariants and Embeddings of Statistical Manifolds 243
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.
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
4.1. 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
M 1 ( T ) o= SUP M 1 ( T ( z ) ) .
XEM
here z E R.
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 ) .
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)
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.
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.)
Proof. We put
L m + l ( x l , . . . ,xm+1) = XI),... f"+'(xrn+l))
7
as follows.
f3(2) = A - l . f i ( x )@ 0f2).
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).
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
+
. 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
Monotone Invariants and Embeddings of Statistical Manifolds 251
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
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
10 = I1 x * * * x L + l
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).
This page intentionally left blank
Advances in Deterministic and Stochastic Analysis 255
Eds. N. M. Chuong et al. (pp. 255-268)
@ 2007 World Scientific Publishing Co.
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
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
j=1 p=l
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
- --
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
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.
HomD,(M, O X ) ~ +
Y H o m ~ , ( M y OY)
,
i s an isomorphism.
A ( U ) = l%A/In(U)
n
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:
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 ) .
Graded Cech Cohomology in Noncommutative Geometry 261
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
and therefore
H q ( G ) ( X )= l@ H q ( R , G )
R-X
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
where Ciij :=
(il
n
,..., i k ) E l k
M(Ui, x . . . x U i k ) and f = ev (even) or od (odd).
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.
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
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.
[sine case]
cos 0 - sin 0
-I=[:;].
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.
Advances in Deterministic and Stochastic Analysis 269
Eds. N. M. Chuong et al. (pp. 269-278)
@ 2007 World Scientific Publishing Co.
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.
iJ=l
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)$, s E R, belongs to x(Rn).
If w E K ( R n ) then the functions:
e=o
Sobolev Spaces with Weight on Riemannian Manifolds 271
Ilfllp,w = IbJfllP.
As in Ref. 2 we can get Theorems 2.1 and 2.2 below:
Theorem 2.1. cp(Rn)is dense in Hp,k+,(Wn), that is Hi,k,w(Rn)=
HP,k,W(R").
f,(O = f(< - 7 ) .
We have:
so:
Hence:
Note that:
e=o
Hence the sequence {fi} and { IAfil} are uniformly bounded on K , so the
sequence {fi} is uniformly bounded and is equicontinuous on K . Since
Rn is separable and countable a t infinity, the sequence {fi} possesses a
subsequence converging uniformly on every compact subset of Rn. This
subsequence, may be denoted by {fi}. Given any E > 0, by wz([)/w1(<) 40
as [ -+ 00,there exists a large enough closed ball S such that:
where C = supwz
- > 0. Since llfi - fjllp,2,wl I2 and the sequence {fi}
S
uniformly converges on S we get:
Remark 2.1.
(1) With the assumptions of the Theorem 2.5, it is obvious that the in-
clusion operator CF(Rn) n Hp,k,wl(Rn)-+ Lp,wz (Rn)is completely
continuous, for all integers k 2 2.
(2) When Mn is a Riemannian manifold, it is also a metric space that is
separable and countable at infinity. But a ball S in M,, is not necessary
to be a Riemannian submanifold with boundary, so we cannot apply
the Sobolev embedding theorem as above to S. However, the above the-
orem holds true for the class of Cartan-Hadamard manifold, that is, a
simply connected complete Riemannian manifold with non-positive sec-
tional curvature. By the exponential map at any point such a manifold
is diffeomorphic to an Euclidean space of the same dimension. Conse-
quently a closed ball in a Cartan-Hadamard manifold is a Riemannian
submanifold. So we get:
where E is any positive number (see Ref. 1, pages 31-33). By the Minkowski
inequality, we obtain:
Wn\G ' G
Since G and supp f are compact sets, for u large enough, we have x - <,, q!
supp f , V x E G. So:
-(/w$(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
Proof. Since Mn is compact, there exists p > 0 such that V P E M,, the
polar coordinates exist on the closed ball:
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
Sobolev Spaces with Weight on Riemannian Manifolds 277
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:
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
STOCHASTIC AND
INFINITE-DIMENTIONAL ANALYSIS
This page intentionally left blank
Advances in Deterministic and Stochastic Analysis 281
Eds. N. M. Chuong et al. (pp. 281-296)
@ 2007 World Scientific Publishing Co.
SITU RONG
Department of Mathematics, Zhongshan University,
Guangzhou 510275, China
E-mail: mcsstr@zsu.edu. cn
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$t of the second term is the
amount of persons who will arrive a t age interval [l,2) in the next year tt 1.
Similar physical meaning can be made for Ax:, 2 5 i 5 r,. If one considers
that the time variable changes continuously, and writes the corresponding
282 S. Rong
Bt= [ 0 .. 0 bL1
0 .... ..
.. b y 0 .. 0
1
.. .. .... 0 . (4)
+t = 1
;4 s ) d I& >
,n(t)E Nz,,
as ~t E aR$,
Stochastic Population Control and RSDE with Jumps 283
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 ,
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.
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.
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:
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 3.2. 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
Stochastic Population Control and RSDE with Jumps 287
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 ~
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
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$(x) is
bounded and jointly continuous such that
IPt(x) - Pt(y)I I ~ N , T ( ~ ) P N , T ( -
I XY I ' ) ,
as 1x1, lyl 5 N , t E [O,T],W'<00;
where kN,T(t) and P N , T ( U ) have the same properties as those in the con-
dition 4" of Theorem 3.1. Then the population SDE, i.e. (6) with (7) and
with the coeficient (8), has a pathwise unique strong solution.
i u(t,x ) = co + ( c p x , .. . , Cid)x)
+
c ( t , x , z ) = (CO G z ) f ( z ) ,
(9)
where Co, Cii),i = 1,* . . ,d and 7 7 1 are constant d x d matrices, and 770
is a constant Rd-vector, (i.e. all elements in the matrices or vectors are
constants), and , in addition, all elements in 770 and 771 are non-negative,
i.e. ?$,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.
+ 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
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
+ s, p ( S , z, z ) - c y s , 5 , .)I 2
T(dZ)]dS}.
290 S. Rong
+s, (c"(t,J:?-,
+ 44: -43.
2 , ~-
) Co(t,J:y-, z , w ) ) f i k ( d t ,d z )
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
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).
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 ) '
Stochastic Population Control and RSDE with Jumps 295
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
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).
Advances in Deterministic and Stochastic Analysis 297
Eds. N. M. Chuong et al. (pp. 297-320)
@ 2007 World Scientific Publishing Co.
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
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,
-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.
LZ(0,l).
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.
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.
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 ) ;
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)
Noncausal Stochastic Calculus Revisited 303
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 ) .
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).
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;
1
1 1
f d p W t := l i m l f d W $ ( t ) (in probability).
Noncausal Stochastic Calculus Revisited 307
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
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.
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,
(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
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:
k: E L 2 ( [ 0,In+'
, ) with ~ n ! I I ] c ~ l l<~03,
+,
n
Noncausal Stochastic Calculus Revisited 309
n=O
n=l
The followings are the basic results concerning the relation between the
symmetric integrals with the noncausal integral.
310 S. Ogawa
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.
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.
Noncausal Stochastic Calculus Revisited 311
P{
I f 4 ( t ) d t < m} = 1.
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.
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:
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,
(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).
and
(in probability)
where,
Noncausal Stochastic Calculus Revisited 315
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
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,
Noncausal Stochastic Calculus Revisited 317
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,
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 Y t , w ; E , rl, C) (25)
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)
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.
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).
Advances in Deterministic and Stochastic Analysis 321
Eds. N. M. Chuong et al. (pp. 321-339)
@ 2007 World Scientific Publishing Go.
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
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
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)
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
-
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)).
(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:
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
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
Stochastic Analysis and Foundations of Quantum Mechanics 327
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
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:
Let us consider the following map T from the classical statistical model
Mquad = (SG,Vquad) to the quantum statistical model Nquant= ( D ,L,) :
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.
E f ( W ( w ) )= 7 +
a9112j - ( f ( 0 ) 9 ,9) o ( a ) ,a + 0. (21)
By setting into this asymptotic equality the dispersion of the random vari-
able P Q [ ( w we
) obtain (21).
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
Stochastic Analysis and Foundations of Quantum Mechanics 331
where XI, A2 E R.
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:
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
b p l , Q2) = (91,*2).
<($I =fir]($)
Proof. Here the main difference from consideration in section 3 is that the
measure p is not concentrated on Hilbert space H on which the function
f is defined (and continuous). Therefore even the exponential growth o f f
on H would not help so much, because Jn,eal1411dp(+) = 00 (since even
,J, ll$lldp($) = 00). We have for a polynomial f:
334 A . Khrennikov
Example 6.1. Let us consider the rigged Hilbert space (26). We consider
the map f : S(R3)-+ R determined by a fixed point zo E R3 :
1
f($) = 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
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$)I 5 a f n ! R P e b f R p ( * ) .We
obtain:
~ ~ ~ ( n )5( a;e-nnl/zebfn,
o)~~p
Thus the derivatives o f f satisfy the inequalities (29) with r f = ebf
336 A . Khrennikov
B). Let now derivatives of f satisfy the inequalities (29) for some con-
tinuous semi-norm p . Then by the inequalities (27) we have
lf(42 c
03
n=O
Ilf'"'(o)llpPn(lcl)l~! ICferfP(@).
Thus f has the exponential growth with b f =rf and the same continuous
semi-norm p as in (29).
where D = 7 .Here
.(a) = a2R(a,f l P I , (31)
where
IR(a,f,P)I 5 C f 1 n
eTfp(')dPD('$). (32)
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.
j=1 j=1
j=1
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
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,
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).
This page intentionally left blank
Advances in Deterministic and Stochastic Analysis 341
Eds. N. M. Chuong et al. (pp. 341-360)
@ 2007 World Scientific Publishing Co.
Karl Gustafson
Department of Mathematics,
University of Colorado, Boulder, Colorado
E-mail: gustafs@euclid. colorado. edu
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.
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:
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 $(A) as defined from either p1 or vl, the angle of
A. I called the two vectors z* the first antieigenvectors of A. They are the
vectors (normalized t o norm 1 in (4)) most turned by A. The z1 and z,
angle (x+,x-)= -sin$(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.
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
RE(& 2 n
P
i=l
cos2 &(V) (11)
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
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
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.
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. 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
Also the adjoint U,+ considered in HT is the same as the original adjoint,
using again the commutativity,
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.
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