Proceedings of the Third German-JapaneseSymposium
Infinite Dimensional Harmonic Analysis 111
This page intentionally left blank
Proceedings of the Third German-Japanese Symposium
Infinite Dimensional Harmonic Analysis Ill University of Tubingen, Germany 15 - 20 September 2003
editors
Herbert Heyer University of Tubingen, Germany
Takeshi Hirai Kyoto Universitx Japan
Takeshi Kawazoe Keio University,Japan
Kimiaki Sait6 Meuo University, Japan
NEW JERSEY
-
World L O N D O N * SINGAPORE
BElJlNG
-
Scientific SHANGHAI
-
HONG KONG
TAIPEI
-
CHENNAI
Published by
World Scientific Publishing Co. Re. Ltd. 5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601 UK ofice: 57 Shelton Street, Covent Garden, London WC2H 9HE
British Library Cataloguing-in-PublicationData A catalogue record for this book is available from the British Library.
Cover design by Katrin Heyer, Catodesign, Munich, Germany.
Proceedings of the Third German-Japanese Symposium INFINITE DIMENSIONAL HARMONIC ANALYSIS I11 Copyright 0 2005 by World Scientific Publishing Co. Re. Ltd.
All rights reserved. This book, or parts thereof; may not be reproduced in any form or by any means, electronic or mechanical, includingphotocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.
ISBN 981-256-593-0
Printed in Singapore by World Scientific Printers (S) Pte Ltd
PREFACE In continuation of previous meetings on the same topic the 3rd GermanJapanese Symposium on Infinite Dimensional Harmonic Analysis was held from September 15th to September 20th 2003 at the Department of Mathematics of the University of Tuebingen. As in previous years the successful collaboration between Japanese and German mathematicians will be manifested in a volume of Proceedings which the organizers and additional editors are offering to the mathematical public with great satisfaction.
It is hoped that successive symposia on new developments in the traditional field of research will continue to reveal what can be achieved by productive exchange between Japanese and German scientists. Naturally, the success of an international meeting is a consequence of the achievements of its renowned contributors. Their one-hour talks on recent discoveries in their domains of research formed the central part of the symposium. There was a wide range of topics presented to the participants of the meeting, the connections and interactions between the various subjects being given by structure-oriented thinking. Three sections of research constituted the program: Stochastic processes in groups and related structures Structural aspects of quantum stochastics Infinite dimensional stochastic analysis
As in previous years (Tuebingen 1995, Kyoto 1999) the 2003 meeting has been organized under the auspices of the “German-Japanese Cooperative Science Promotion Program” established in 1992 by the Japan Society for the Promotion of Science (JSPS) an the German Research Society (DFG). Unfortunately only DFG decided on the granting along the cooperative program. But even without the support of JSPS the Japanese researchers managed to participate at the symposium. Their individual sources together with the generous financial help of DFG deserve our great appreciation. We are also
vi thankful to the German-Eastasian Science Forum at Tuebingen for spiritual and physical support, also to secretaries and students of the host Department for their practical involvement. All contributions to these Proceedings have been refereed. The editors are grateful to the referees, for their efficient help. It also seems in order to mention heartfelt thanks to be extended to Mrs. Megumi Mori, Tuebingen, for her translation of Professor Heyer’s necrology in memory of Professor Shozo Koshi and t o Ms Katrin Heyer, Muenchen, for providing the cover design of the present book. Last not least we express our deep appreciation to World Scientific Publishers, in particular to Ms E. H. Chionh as their acting representative for a remarkably pleasant cooperation. For the local organizers Herbert Heyer, Tuebingen Burkhard Kuemmerer, Darmstadt For the editors Herbert Heyer, Tuebingen Takeshi Hirai, Kyoto Takeshi Kawazoe, Tokyo Kimiaki Sait6, Nagoya February 2005
CONTENTS
Preface
V
Shigoto nakama de ari yujin de atta kare no omoide
ix
In Memory of a Colleague and Friend H. Heyer
xi
In Memory of Professor Shozo Koshi Y. Takahashi
xiii
Recent Developments on Harmonic Forms and L2-Betti Numbers of Infinite Configuration Spaces with Poisson Measures S. Albeverio and A. Daletskii
1
Une R6ciprocit6 de Frobenius H. Fujiwara
17
Markov Property of Monotone L6vy Processes U. Franz and N . Muraki
37
Geometric Properties of the Set of Extensions of a Stochastic Matrix R. Gohm
59
Heat Kernel Analysis on Infinite Dimensional Groups M. Gordina
71
On Mehler Semigroups, Stable Hemigroups and Selfdecomposability W. Hazod
83
On Infinite Divisibility and Embedding of Probability Measures on a Locally Compact Abelian Group H. Heyer and G. Pap
99
Character Formula for Wreath Products of Finite Groups with the Infinite Symmetric Group T. Hirai and E. Hirai
119
Remark on Biane’s Character Formula and Concentration Phenomenon in Asymptotic Representation Theory A . Hora
141
Real Hardy Spaces on Real Rank 1 Semisimple Lie Groups T. Kawazoe
161
viii
White Noise Analysis, Filtering Equation and the Index Theorem for Families R. Le'andre
177
Laplace Approximations for Diffusion Processes S. Liang
187
A Note on Afine Quotients and Equivariant Double Fibrations K. Nishiyama
197
Admissible White Noise Operators and Their Quantum White Noise Derivatives U. C. Ji and N. Obata
213
PDE Approach to Invariant and Gibbs Measures with Applications M. Rockner
233
Deformations of Convolution Semigroups on Commutative Hypergroups M. Rosler and M. Voit
249
An Infinite Dimensional Laplacian Acting on Multiple Wiener Integrals by Some LCvy Processes K. Suit6
265
L6vy Processes on Deformations of Hopf Algebras M . Schurmann
277
Unitary Representations of the Group of Diffeomorphisms via Restricted Product Measures with Infinite Mass H. Shimomura (joint work with T. Hirai)
289
An Application of the Method of Moments in Random Matrix Theory 313 M. Stolz Isotropy Representation for Harish-Chandra Modules H. Yamashita
325
X
In memory of a colleague and friend On July 11 th, 2003 Professor Shozo Koshi passed away. He was our friend, whose helpfulness and affection will remain fundamental to my wife and myself in experiencing and understanding Japan and its culture. It had been a cold winter day in Sapporo in February 1985. During the passed night a strong storm had swept violent masses of snow into the town which were removed by convoys of trucks in order to make a minimal traffic circulation possible. The man's - high column of the bus stop in front of the guest house of Hokkaido University was completely covered with snow. Only with some effort I found my way to the Mathematics Department, where my host, Professor Yasunori Okabe, intended to introduce me to his colleagues. On the very day I met Professor Koshi for the first time. His name was known to me, since prior to my departure for Japan Professor Edwin Hewitt from the University of Washington at Seattle had written in a letter to me that I would find Professor Koshi a "great gentleman" and a "fine collaborator". The meeting with Professor Koshi was very cordial. He obviously enjoyed remembering his time of study with Professor Gottfried Kothe in Heidelberg during the years 1963 through 1965. From today's point of view it is clear to me that already at our first meeting a genuine liking took its start growing rapidly into a friendship between the two of us which later extended also to our spouses. Along with our common mathematical interests this relationship provided us with a deeper understanding of each other and of our all too different cultures. Numerous mutual visits within Japan and Germany make still an excellent memory of a most fruitful and pleasant time. I just mention the sightseeing tours starting from Tubingen (Stuttgart, Berlin, the Rheingau, the Reichenau, Bad Wildbad) and the wellness activities in Japan (Awara onsen, Jozankei, Yumoto Spa, Ito, Ogoto onsen, Haruna ko). There were contemplative hours during which we discussed the interplay or clash of our cultures. The Koshis tried to explain to us the esthetic significance of sansui (the combination of mountain and water in drawings and paintings) - on our joint trips in Japan we always stayed over night with a view over a lake towards a mountain -, and my wife and myself made efforts to clarify the misunderstanding among Japanese people of the notion of romanticism. But there were also cheerful rounds: in Susukino with shochti no bancha - wari, and in the Kloster Eberbach with Riesling wines of various ages. An unforgettable event deserves special mention. On the occasion of a meeting of the Japan Mathematical Society at Tokyo Toritsu Daigaku we enjoyed a most exciting closing party and became witnesses of the traditional tapping of the Sake barrel, a technically rather demanding action incumbent upon the great Professor Kiyosi Ito.
xii Clearly, while our family communities developed, we two mathematicians remained scientifically active. Professor Koshi was one of my coorganizers of the first Japanese German Symposium on “Infinite dimensional harmonic analysis” held in Tubingen in 1995. He also participated actively at the second Symposium in Kyoto in 1999. To our great regret we had lost him prior to the third Symposium in Tubingen last year. Professor Koshi and myself collaborated also on the analysis on hypergroups; our set of Lecture Notes on “Harmonic analysis on the disk hypergroup” appeared since 1993 in two editions. In Kyoto (1990), Sapporo (1990) and Berlin (1998) we met for discussions on measures on groups and hypergroups on which also some of our former students participated. Professor Koshi was a highly cultured man whose openness and capacity of constructive criticism have been exemplary for the exchange of ideas between East and West. To me he was always a partner of sensitive proximity who taught me essentials on dealing with the Japanese world. He gave me the chance to test my modest attempts to speak some Japanese by patiently listening and correcting efficiently, always with a sense of humour. His attitude in conversations to add to his statements “to omoimasu” or ”may be” or “possibly” impressed me, since these suffixes seemed to indicate responsible thinking and talking. With Professor Koshi’s death an academic researcher and teacher disappeared from this earth who spread a comprehensive intellectuality. His example has been profitable for all those who crossed his way. I myself consider it a privilege to have met this fine man and to have enjoyed over an interval of 20 years a significant human affinity.
Herbert Heyer
In Memory of Professor Shozo Koshi Too suddenly to me, Professor Shozo Koshi, Professor Emeritus of Hokkaido University, passed away on July 11, 2003, in Sapporo, Japan. When I met him about three months before, he had looked so well that the news of his death really surprised and shocked me. To him I was just one of his many students but to me Professor Koshi was indispensable for my research life. I would like to express my heartful sorrow and laments here and try to call back his life and warm personality. Professor Koshi was born on November 18, 1928, in Otaru, a seaport city in Hokkaido, Japan, and grew up there. He studied mathematics at Hokkaido University. After graduating, he was appointed Assistant to the Department of Mathematics, Hokkaido University in June 1951. He was promoted to Lecturer in May 1958 and to Assistant Professor in April 1960. Earlier in his career he had studied in the area of semi-ordered linear spaces and convexity. After a research stay in West Germany, under Professor G. Kothe, a world famous mathematician, from 1963 to 1965, Professor Koshi turned his research interest to nuclear spaces and measures on linear topological spaces. I first met Professor Koshi in April 1967, when he was appointed Professor of Mathematics for the Faculty of Science, Okayama University, as a student of the Department of Mathematics. He always impressed us deeply with a great deal of preparation, which made his lectures very understandable and enjoyable. The young Professor Koshi in those days was full of vitality and confidence. Then I entered the graduate school and continued to be his student. As he was interested in nuclear spaces and measures on linear topological spaces, I followed him naturally. We were always impressed with his rich knowledge and deep insight. One day after a seminar he told me that perseverance was important above all for a mathematics student. As I had little perseverance for anything, his words struck me with a shock. I owe a great deal to him if I have some perseverance now. Although in October 1970, Professor Koshi was appointed Professor for the Faculty of Science, Hokkaido University, he continued to stay in Okayama for six months more. I followed Professor Koshi in April of the following year, when he moved to Sapporo, as Assistant of Hokkaido University. This was the start of my researcher’s life. He looked more stable and confident. I stayed in Sapporo for eight years, during which time I learned a great deal from Professor Koshi. At first he seemed too great for me and I thought it was far beyond my power t o understand his deep insight. One
xiv
day I was invited to his house for dinner. As I had heard Mrs. Koshi was a daughter of Professor Kiyoshi Noshiro, a great mathematician, I felt so nervous. But when I was first introduced to her, all my anticipation was banished. She was such a friendly and pleasant lady. I had a really good time with the family, finding what a nice cook Mrs. Koshi was. After I myself had a family, we were often invited to his house. I soon discovered what a loving nice Daddy he was. He was so generous and broadminded. I am especially grateful to him for allowing me to research freely and to my heart’s content. I have no memory of being scolded by Professor Koshi. He never tried to find faults with anyone else. He always tried to extract merits from anyone. Professor Koshi was elected Dean of the Faculty of Science in September 1985 and proved himself so able in his job. By then I had already left Sapporo and met him much less frequently than before. But whenever I met him at a conference or a seminar from time to time, he did not seem painful at all under such a burden. He even seemed to be enjoying himself. He was so cooperative, generous, and friendly that naturally he had so many friends and acquaintances both in Japan and overseas. Professor Koshi retired from Hokkaido University in March, 1991, but still continued to do his research with perseverance in mathematics. At the same time, it is said, he used to keep early hours in order to listen to the language programmes on the radio and television, such as English, German, French, etc. I feel sure Professor Koshi owed his great success as a scholar and educator to his perseverance and generosity. He seemed to have made it a rule to show himself as an example instead of compelling or forcing students to do something. In that sense too, needless to say, I owe a great deal to Professor Koshi. In recent years I met him much less frequently than before, but whenever we met, he always spoke to me with a broad smile. He had so many topics and time flew while I talked with him. I felt strangely relaxed whenever I met and talked with him. I last met Professor Koshi on the occasion of the Annual Meeting of the Mathematical Society of Japan held at the University of Tokyo. That was only three months before he died. Remembering how he talked joyously and enjoyably about his old students and looked so well, I still cannot believe this sad news. I am really grateful to Professor Koshi for teaching me so many things. May his soul rest peacefully! Yasuji Takahashi
Infinite Dimensional Harmonic Analysis I11 (pp. 1-15) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
Recent developments on harmonic forms and L2-Betti numbers of infinite configuration spaces with Poisson measures Sergio Albeverio Inst. Ang. Math., Universitat Bonn; BiBoS; SFB 611;
IZKS; CER.FIM (Locarno); Acc. Arch. (Mendrisio)
Alexei Daletskii School of Computing and Technology The Nottingham Trent University’
U.K.
Abstract A review of recent results on harmonic forms, L2 cohomology and L2-Betti numbers of infinite configuration spaces equipped with Poisson measures is given.
1 Introduction The space rx of infinite configurations in a Riemannian manifold X is defined as the space of locally finite subsets (configurations) of X . Configuration spaces of such type appear in a natural way in different parts of mathematics and mathematical and theoretical physics (statistical mechanics, quantum field theory, representation theory, probability). The growing interest in the theory of configuration spaces as developed in recent years can be explained by various important applications, as well as by the rich and interesting intrinsic structure of rx. Let us remark that rx,in contrast to the spaces of finite configurations, does not possess any proper structure of an (infinite dimensional) manifold. However, many important geometrical objects, like differential forms, connections and the de Rham complex over rx can be introduced in a specific way. A crucial role here is played by a probability measure on rx (in particular, a Poisson or Gibbs measure), which is quasi-invariant with respect to the component-wise action of the group Diff,-,(X) of diffeomorphisms of X . This
2
property of Poisson measures is used in the pioneering works [26] and [42] for the construction of operators of quantum field theory and irreducible representations of Diffo(X) respectively. In fact, the existence of a probability measure which is quasi-invariant with respect to a rich class of transformations, replaces in some sense a manifold structure. This philosophy, inspired by I261 and [42], has been initiated and developed in 191, [lo] and has lead to many interesting and important results in the field of stochastic analysis on configuration spaces and its applications. In particular, many questions related to spectral properties of Dirichlet operators of Poisson and Gibbs measures, corresponding stochastic dynamics, description of measures, and their applications have been studied, see a review in [39]. Further step in the development of the "manifold" approach to rx has been done in the works [4], [5], where the spaces of differential forms on r x are studied. In the present review, we give to a short description of the results of these works and more recent results on the L2-cohomology of r x equipped with the Poisson measure T . In Section 2, we define and study the de Rham complex of .Ir-squaie-integrabledifferential forms over rx and the corresponding Laplacian acting in this complex. We describe the structure of the spaces of harmonic forms. In Section 3 we discuss the concept of von Neumann dimension of a symmetric tensor power of Hilbert spaces, in view of applications in the next section. In Section 4, we consider the case where X is an infinite covering of a compact manifold, and compute the L2-Betti numbers of rx. That is, we introduce a natural von Neumann algebra containing projections onto the spaces of harmonic forms, and compute their traces. Further, we introduce and compute a regularized index of the Dirac operator associated with the de Rham differential on T'X. For a more detailed exposition, see [l],[23], [22] and references given therein. Let us remark that the spaces of finite configurations, which unlike rx possess a natural manifold structure, have been actively studied by geometers and topologists, see e.g. [B],[25] and references given therein. The relationship between these works and our L2-theory,which is relevant for the spaces of finite configurations too, is not clear yet. The situation changes dramatically if the Poisson measure .~ris replaced by a different measure (for instance a Gibbs measure). From the physical point of view, this describes a passage from a system of particles without interaction (free gas) to an interacting particle system, see [lo] and references within. For a wide class of measures, including Gibbs measures of Ruelle type and Gibbs measures in low activity-high temperature regime, the de Rham complex has been introduced and studied in [6]. The structure of the corresponding Laplacian is much more complicated in this case, and the spaces of harmonic forms have not been studied yet. Our long-term goal is to find invariants which reflect the complicated interplay between the properties of measures on configuration spaces and the topology and geometry of the underlying manifolds.
3
Acknowledgment. The first named author is very grateful to Professor Herbert Heyer for his very kind invitation to a most stimulating workshop.
2
De Rham complex over a configuration space
Let X be a complete, connected, oriented, Riemannian manifold of infinite volume, with a lower bounded curvature. We fix the inner product (., .)x in each tangent space T x X , x E X . Let Vx stand for the gradient on X . The configuration space l?x over X is defined as the set of all locally finite subsets (configurations) in X :
rx := {y c X I 17 n A / < ca for each compact A c X } .
(1)
Here, IAl denotes the cardinality of a set A. We can identify any y E l?x with the positive, integer-valued Radon measure
CE"c M P ) ,
(2)
XEY
where
E,
is the Dirac measure with mass at x ,
CXEB&, :=zero
measure, and
M ( X ) denotes the set of all positive Radon measures on the Borel a-algebra B ( X ) . The space rx is endowed with the relative topology as a subset of the space M ( X ) with the vague topology, i.e., the weakest topology on rx with respect to which all maps (3) are continuous. Here, f E Co(X)(:=the set of all continuous functions on X with compact support). Let B ( r x ) denote the corresponding Borel g-algebra. Let dx denote the volume measure on X , and let 7r denote the Poisson measure on rx with intensity dx. This measure is characterized by its Laplace transform
is a probability measure. On the other hand, it has some properties of the Riemannian volume. In particular, if a function F on rx is integrable with respect to 7r and local, i.e., F ( y ) = F ( ~ Afor ) some compact A c X , then T
4
Following [42], [9], we define the tangent space to r x at a point y as the Hilbert space
Remark 1 The configuration space rx does not possess any proper structure of a Banach manifold (or some other type of infinite dimensional manifold). Thus T,rx is n o t a tangent space in a traditional sense. I n fact, the definition of T J X i s motivated by the existence of point-wise (diagonal) action of the group D i f f o ( X ) o n configurations, @(y) = {..., @(x), @(y), @(z),...}, where y = {...,x,y,z,...} , q5 E D i f f o ( X ) . Here D i f f o ( X ) i s the group of compactly supported (trivial outside of a compact set) diffeomorphisms of X . Under a differential form W of order n over r x , we will understand the mapping
rx 3 Y ,+ ~
( 7E )(
~,r~)"~.
(7)
Let y E r x and x E y. By O,,,we will denote an arbitrary open neighborhood of x in X such that OT,x n (y \ {x}) = 0 . We define the mapping
This is a section of the Hilbert bundle
(T,Yrx)An Y E ++
u-,,x.
(9)
The Levi-Civita connection on T X generates in a natural way a connection on this bundle. We denote by V;, the corresponding covariant derivative and use the notation
V r W ( d := (V$, Wx(r, x))xe,.
(10)
We say that W is differentiable if VrW(y) exists, and
v r w ( ~E)T,rx
@ (T,rX)hn.
(11)
Higher order differentiability can be defined in the similar way. A form W : r x -+ (Tl?x)"n is called local if there exists a compact A = A(W) in X such that W(y) = W(y*) for each y E F X . Let F P denote the set of all local, infinitely differentiable forms W : r x + ( T l ? X ) A n which together with all their derivatives are polynomially bounded, i.e., for each W E 332'' and each m E Z+, there exists a function p E C o ( X ) and k E N such that
5
where V(')W := W . We define on the set FO" the L2-scalar product with respect to the Poisson measure:
The integral on the right hand side of (13) is finite, since the Poisson measure has all moments finite. Moreover, (W,W ) p p > 0 if W is not identically zero. Hence, we can define a Hilbert space L 2 ( r x -+ A n ( T r X ) ;T ) as the completion of FOn with respect to the norm generated by the scalar product (13). We denote by L:On the complexification of L2(I'x -+ An(Trx);T ) . We will now give an isomorphic description of the space Laon via the space L20° of 7r-square integrable functions on I'x and spaces L 2 0 n ( X m )of squareintegrable complex forms on X m , m = 1,. . . , n. We have n
where
T P ' x ~:=
... A ( T , , , , x ) A for ~ ~ any finite
~ l <-k l , . . . , k , ~ d ( T ~ l ~ ) A k ' kl+ k,=n
...+
configuration r] = { x l , ..,,x m } . We denote by W m ( y ;r ] ) the projection of W(y) E (TTI'X)An onto the subspace Te)Xm. Theorem 1 [5] Setting f o r m = 1,. . . ,n
one gets the isometry
Remark 2 Actually, formula (15) makes sense only for where
-
xm:= {(XI,.. . , x m ) E x mI xz # xj However, since the set X" lead to a contradiction.
\ 5 is
if i
(XI,. .
#j
. , x,)
} .
-
E Xm,
(17)
of zero d x l . . . dx, measure, this does n o t
Proof. The result follows from the generalized Mecke identity
6
where f is a measurable function on l?x x X m for which at least one of the integrals in (18) exists. This formula can be proved (see [38])by a repeated application of the well-known Mecke identity
Remark 3 T h e corresponding statement in [5]is more refined (the explicit description of the image of In is given). T h e exact formulation is cumbersome, and we give its version o n the level of the spaces of cohomologies (formula (27)). We define the de Rham differential d,: L;nn -+ Lion+' by the formula
+
ASn+l(VrW(y)),
(dnW)(y) := (n
(20)
on the complexiiication of Fan, where ASn+l: (T,rx)@(n+l) (T,I'x) A ( n + l ) is the anti-symmetrization (remark that Vr is the analog of the Levi-Civita connection). Let d; be the adjoint operator. -+
Theorem 2 [4],[5] d; : L;Sln+'
-+
L;Qn is a densely defined operator.
Proof. Let y E r x and z E 7 be fixed. Let C"(U,,, -+ An(T,rx)) denote the space of smooth sections of the Hilbert bundle (9). We define the corresponding exterior differential dz,n : ~ " ( o , , , An(T,rx)) -+
-+
c"(Q,,,
-+
A ~ + ' ( T ~ ~ X ) ) (21)
whose action, in local coordinates on the manifold X I is given as follows: d,,, 4(y) hi A . . . A h,
=
(n
+ 1)1'2Vx4(y) A hi A . . . A h,,
(22)
4 E C"(O,,, -+ R), h k E T,,X, Z k E y,k = 1,.. . ,n. It easily follows from the definition of d, and Vr that, for W E F W , (dnW)(y)=
(23)
dZ,nW~(7~ XE7
The proof follows now from (5), and the corresponding statement for "finite dimensional" operators d,,,, see [5]for more details. Thus the operator d, is closable. We denote its closure by d, and introduce an infinite Hilbert complex
3. . .
'~1
5 L:-p+1
.. .
(24)
7
(for the definition and discussion of properties of Hilbert complexes see [21]). Let us define in a standard way
HP) := dndEPl+ dEd,,
(25)
It will be called the Hodge-deRham which'is a self-adjoint operator in La.". Laplacian of the Poisson measure 7 r . Standard operator arguments (see [40], [13])show that the Hilbert spaces 7-l: := Ker d,/clo{Im d n - l } and Ker HP) are canonically isomorphic, and we identify them. Theorem 3 [5] 1) Let H$?, be the Hodge-deRham Laplacian in L 2 R n ( X m ) . Then:
2) T h e isometry I" generates the unitary isomorphism of Hilbert spaces 7-l;
@
2: Sl,..
. ,sd=o,1,2...
(H'(X))k"1 @ . . . @ ( 7 - l d ( x ) ) : s d l
(27)
Sif2Sz+...+dSd=Tl m
where 'HI"(X) := KerHim), m = 1 , 2 , ..., d =: dimX - 1, and o s means the symmetric tensor product i f m is even and the skew-symmetric tensor product i f m is odd.
Proof. 1)Formula (26) can be checked directly on the set DR" := DnLZR", which is dense in LZR". Here 2) is spanned by the forms W such that ( I " W ) ( y ,5 1 , . . . Q) := ]
0, k # m , F(y)w(xl1 . . . z m ) , k
=m
'
with F a smooth local function on rx and w a smooth n-form with compact support on X". It is known that H$i is essentially self-adjoint on the set of smooth forms with compact support. Hio) is the Dirichlet operator of the Poisson measure 7r studied in [9]. In particular] it has been shown there that HP) is essentially self-adjoint on the set of smooth local functions. From this and the properties of operators with separable variables [17]it follows that DR" is a domain of essential self-adjointness for HPI1which implies the result. 2) It has been shown in [9] that KerH;') consists of constant functions] which together with (26) and the Kiinneth formula (see e.g. [19]) implies that I" generates an isometry
7-l;
N
@
('HIl(X))@S1 € 3 . . . €3 (7-ld(X))@-.
,...,Sd=O,1,2.. . S1+2S2 +...+dsd=n S1
(29)
8
In order to describe its image we need to factorize the right-hand side with respect to the commutation relation
w2 = (-1)PlP2w2 €3 w1,
(30)
where w1 E ‘FIP1(X), w2 E 7-lP2(X),which gives ( 2 7 ) .
0
w1 €3
Remark 4 Formula (27) also holds f o r spaces of finite configurations, see [22]. I n fact, 7-l:
21 Har”
(B?)),
(31)
where Har” ( B g ) )is the space of L2-harmonic n-forms o n the space B?) of configurations of n o more than n points. Let us remark that B g ) = Ug=, x k / s k , where X k i s defined by (l7), and (27) i s in this case a symmetric version of the Kunneth formula.
-
-
We see from ( 2 7 ) that all spaces n E N,are finite dimensional provided the spaces 7-lm(X), m = 1,..., d , are so. In this case, S 1 , ...,Sd=0,1,2... S1+2Sz +...fdsd=n
where /3f’
of)
:=
:= 1. Here
( ),
if k is odd, ,Bf) :=
, if k is even, and
,& := dimHk(X).
Example 1 [5]LetX be a manifold with a cylindrical end (that is, X = M u (N x ) :XE f o r some compact manifold M with boundary N ) . I t i s proven in [16] that 7-lm(X) i s isomorphic to the image of the canonical map H t ( X ) 4 Hk(X), where Ht(X) resp. H k ( X ) i s the space of the compactly supported de R h a m cohomologies resp. all de R h a m cohomologies of X. B y e.g. [19], the spaces Hk(X) are finite-dimensional. Thus, all H k ( X ) are finite-dimensional and, in general, non-trivial, and hence so are all spaces 7-l:. For a bigger class of examples of manifolds X with finite-dimensional spaces 7-lk(X) see [34]. In Section 4 we will consider a special case where the spaces Hk(X) are infinite dimensional. This requires however the development of specific techniques, which we discuss in the next section.
3 Von Neumann dimensions of symmetric tensor powers of Hilbert spaces This section is independent of the theory of configuration spaces. However, the results obtained here will be used in Section 4 , which is concerned with L2Betti numbers of configuration spaces. Here, we study the following problem,
9
which can be formulated in quite a general form, see [23], [l],[22], and [20], [41] for general notions of the theory of von Neumann algebras. Let us consider a d-dimensional complex Hilbert space 7-1 = C d . It is easy to compute the dimensions of the symmetric and antisymmetric n-th tensor powers and 3-1"" of 7-1 respectively. We have obviously dim?@'"
dimx"'"
=
=
(
d+n-I
(t)
)
d ( d + l ) ...(d + n - l ) n!
=
d(d - 1)...(d - n = n!
1
(33)
+ 1)
Let 7-1 be a subspace 2f some Hilbert space XI which may in general be infinite dimensional. Then 'H@"and 7-1"" are subspaces of X@". Let
(34)
P:X+3-1,
p,'"' ; X@"
--f
3-16"
(35)
and
pp) : x@n 7-1"" +
(36)
be the corresponding orthogonal projections. Then we have T r P = d, and formulae (33) can be rewritten in the form
np,'")
=
Tr P(Tr P + 1)...(TrP + n - 1) n!
I
(37)
npp)
=
T r P ( T r P - I)...(T r P - n + 1) n!
If the space 7-1 is infinite dimensional, T r P = cm and formulae (37) have no sense. There are however situations where P belongs to a von Neumann algebra A (different from the algebra of all bounded operators in X), equipped with a trace %A, and Trd p < 03. It is interesting to ask whether analogues of formulae (37)involving P hold. The following general statement is proved in [22]. Let M be a 11, factor and H be a separable M-module, and S, 3 g t)U, be the natural action of the symmetric group S, in H@" by permutations. Then the orthogonal projections P, ; H@" + H6", Pa ; H@" + HA" have the form
(38)
10
We set M P ) := {MBn,Ps}",M p ) := {MBn,Pa}". In what follows, we denote by T ~ N the faithful normal semifinite trace on a II, factor N. Theorem 4 [22] M F ) and M P ) are 11, factors, and f o r any A E M
Proof. Let W*(M@,,S,) be the crossproduct of the factor M@'"and the symmetric group S, with respect to the action U, (see e.g. [41]). It is known that W*(M@'nlS,)is a II, factor (because S, is a finite discrete group, see [28]). It is shown in [22] by methods of the general theory of factors that the von Neumann algebras M P ) and M p ) are isomorphic to W*(M@,,S,). Formula (40) follows now from the well-known expression for the trace on a cross-product (see e.g. [41]). 0
Remark 5 A proof of the analogous statement in the particular situation of the next section, which i s based o n the explicit structure of von N e u m a n n algebras appearing there rather than o n the general theory of factors, is given in [23] (the case of n = 2) and [l](the general case).
4
L2-Betti numbers of configuration spaces of coverings
An important example of a manifold X with infinite dimensional spaces "(X) is given by an infinite covering of a compact Riemannian manifold (say M ) . In this case, an infinite discrete group G acts freely by isometries on X and consequently on all spaces L 2 R P ( X ) ,and X / G = M . The orthogonal projection
P, : L 2 W ( X ) -+ W ( X )
(41)
commutes with the action of G and thus belongs to the commutant d, of this action, which is a semifinite von Neumann algebra. The corresponding von Neumann trace b, := Tr~,p,gives a regularized dimension of the space W ( X )and is called the L2-Betti number of X (or M ) . L2-Betti numbers were introduced in [15] and have been studied by many authors (see e.g. [33] and references given therein). It is known [15] that b, < 00. It is natural to ask whether the notion of L2-Betti numbers can be extended to configuration spaces over infinite coverings. It particular, is formula (32) valid in this case (with p k replaced by bk)? In order to answer this question, we need to construct a natural von Neumann algebra containing the orthogonal projection
P, : L
p
4
7t;
(42)
11
and compute its von Neumann trace. There are two natural group actions which we can try to use. One is the diagonal action of G on rx: g{...x,Y,Z,
...I= {...9~,9Y19z, ...I ,
(43)
where g E GI {...x, y, z , ...} E l?x (cf. Remark 1). This action commutes with H(P) and thus Pp belongs to its cornmutant. Unfortunately, the factor-space r x / G is very big (certainly not compact) and the corresponding trace of Pp is therefore either zero or infinite. On the other hand, we know that 7-l: is isomorphic t o the space HarP(B$') of L2-harmonic pforms on the finite configuration space B$) = xk/sk (see Remark 4). Thus
u;=,
-
k
and we can employ the natural actions of the product groups G x ... x G on the spaces X k . The corresponding commutant
-
rI
A(P) := S1, S1
Aysl@I ... @I AFsd
(45)
..., Sd = 0,1,2...
+ 2S2 + ... + d S d = p
does not contain Pp. However, the structure of the von Neumann algebra A@):= { A(p),Pp}"can be described using the results of the previous section. We assume that G is an ICC group (that is, all non-trivial classes of conjugate elements are infinite). Under this condition, all von Neumann algebras d k are 11, factors (see e.g. [35]). Theorem 5 [l]A(p) is a 11, factor, and
where b l , ..., bd are the L2-Betti numbers of X .
Proof. For each k = 1,..., d we define the operator
P,""Ps, k i s even, P,""Pa, k is odd
(47)
12
and the von Neumann algebra
{
dr,P S } " , k is e v e n ,
(48)
{dp,Pa}",k is odd,
generated by the von Neumann algebra df"and projections P, and Pa respec-
8n
tively. Thus, P p ) is the orthogonal projection (L2R"(X))@"+ ('Ftk(X)) ,n = 1,2, ... . Obviously, P f ) E A?). It follows from Theorem 4 that d?) = W * ( A pSn). , We will use the convention dp) = C'. Further, we have @
... @
dtd).
(49)
..., Sd = 0,1,2... 5 1 + 252 + ... + dSd = p S1,
Since all algebras dfk)are IIm-factors, so is A(p), with the trace given by the 0 product of the traces in At'). Formula (46) follows now from (40). We will use the notation b, := Tr,(,,P(P) and call bp the p t h L2 Betti number of rx.
Example 2 Let X = Wd, the hyperbolic space of dimension d . It is known [12\ that the only non-zem L2-Betti number of Wdis b d p (provided d i s even). T h e n
Let us introduce a regularized index indr, (d + d*) of the Dirac operator associated with the deRham differential of the configuration space setting co
indrx (d+d*): = x ( - l ) k b k .
(51)
k=O
We will use the convention bo
=
1.
Theorem 6 [l]T h e series o n the right hand side of (51) converges absolutely, and
indr, (d
+ d*)=
(52)
where x ( M ) is the Euler characteristic of M . The proof can be obtained by a direct computation using formulae (46) and (51).
Corollary 1 The L2-cohomo10gy of F x i s infinite provided x ( M ) # 0.
13
References [l]S. Albeverio, A. Daletskii, L2-Betti numbers of Poisson configuration
spaces, The Nottingham Trent University, Mathematics and Statistics Research Report Series, 22/03, 2003. [2] S. Albeverio, A. Daletskii, and Yu. Kondratiev, Stochastic analysis on product manifolds: Dirichlet operators on differential forms, J. Funct. Anal. 176 (2000), 280-316. [3] S. Albeverio, A. Daletskii, and Yu. Kondratiev, De Rham complex over product manifolds: Dirichlet forms and stochastic dynamics, “Festschrift of L. Streit” (eds. S. Albeverio et al.), World Scientific, Singapore 2000. [4] S. Albeverio, A. Daletskii, and E. Lytvynov, Laplace operators on differential forms over configuration spaces, J. Geom. Phys. 37 (2001), 1 5 4 6 .
[5] S. Albeverio, A. Daletskii, and E. Lytvynov, de %am cohomology of configuration spaces with Poisson measure, J. Funct. Anal. 185 (2001), 240-273. [6] S. Albeverio, A. Daletskii, Yu. Kondratiev and E. Lytvynov, Laplace o p erators in deFtham complexes associated with measures on configuration spaces, J. Geom. Phys. 47 (2003), 259-302. [7] S. Albeverio, Yu. G. Kondratiev, and M. Rockner, Differential geometry of Poisson spaces, C. R. Acad. Sci. Paris 323 (1996), 1129-1134.
[8] S. Albeverio, Yu. G. Kondratiev, and M. Rockner, Canonical Dirichlet operator and distorted Brownian motion on Poisson spaces, C. R. Acad. Sci. Paris 323 (1996), 1179-1184. [9] S. Albeverio, Yu. Kondratiev, and M. Riickner, Analysis and geometry on configuration spaces, J. Funct. Anal. 154 (1998), 444-500.
[lo] S. Albeverio, Yu. Kondratiev, and M. Riickner, Analysis and geometry on configuration spaces: The Gibbsian case, J. Funct. Anal. 157 (1998), 242291. [ll]S. Albeverio, Yu. Kondratiev, and M. Rockner, Diffeomorphism groups and current algebras: Configuration spaces analysis in quantum theory, Rev. Math. Phys. 11 (1999), 1-23.
[12] M. Anderson, L2 harmonic forms on complete Riemannian manifolds, Geometry and analysis on manifolds (Katata/Kyoto, 1981;), Lecture Notes in Math. 1339 (1988), 1-19. [13] A. Arai, A general class of infinite dimensional Dirac operators and path integral representation of their index, J. Funct. Anal. 105 (1992), 342-408.
14
[14] A. Arai and I. Mitoma, De Rham-Hodge-Kodaira decomposition in dimensions, Math. Ann. 291 (1991), 51-73.
00-
[15] M. F. Atiyah, Elliptic operators, discrete groups and von Neumann algebras. Colloque ”Analyse et Topologie” en 1’Honneur d e Henri Cartan (Orsay, 1974), pp. 43-72. Asterzsque, No. 32-33, SOC.Math. France, Paris, 1976. [16] M. F. Atiyah, V. K. Patodi, and I. M. Singer, Spectral assymetry and Riemannian geometry. I, Math. Proc. Cambridge Philos. SOC.77 (1975), 43-69. [17] Yu. M. Beresansky, “Selfadjoint Operators in Spaces of Functions of Infinitely Many Variables,” Amer. Math. SOC.,Providence, R.I., 1986. [18] C. F. Bodigheimer, F. Cohen, and L. Taylor, On the homology of configuration spaces. Topology 28 (1989), 111-123. [19] R. Bott and L. W. Tu, Differential Forms in Algebraic Topology, SpringerVerlag, New York, Heidelberg, Berlin, 1982. [20] 0. Bratteli, D. W. Robinson, ”Operator algebras and quantum statistical mechanics I”, Springer-Vertag, New York, Heidelberg, Berlin, 1979. [21] J. Briining and M. Lesch, Hilbert complexes, J. Funct. Anal. 108 (1992), 88132. 1221 A. Daletskii, A. A. Kalyuzhnyi, Permutations in tensor products of factors and L2 Betti numbers of configuration spaces, Proceedings of V International Conference ” Symmetry in Nonlinear Mathematical Physics” (Kiev, June 2003), eds. A.G. Nikitin, V.M. Boyko, R.O. Popovych, I.A.Yehorchenko, Proceedings of Institute of Mathematics of NAS of Ukraine 50 (2004), 1025-1528. [23] A. Daletskii, Yu. Samoilenko, Von Neumann dimensions of symmetric and antisymmetric tensor products, Methods of Functional Analysis and Topology 9 (2003), No. 2, 123-133. [24] K. D. Elworthy, X.-M. Li, and S. Rosenberg, Bounded and L2 harmonic forms on universal covers, Georn. Funct. Anal. 8 (1998), 283-303. [25] E. R. Fadell, S. Y. Husseini, ”Geometry and Topology of Configuration Spaces”, Springer Verlag, Berlin/Neidelberg, 2001. [26] G. A. Goldin, J. Grodnik, R. T. Powers, and D. H. Sharp, Nonrelativistic current algebra in the N / V limit, J. Math. Phys. 15 (1974), 88-100. [27] R. S. Ismagilov, “Representations of Infinite-Dimensional Groups,’’ AMS, Providence, R. I., 1996.
15
[28] V. Jones, S. Sunder, ”Introduction to subfactors” . London Math. SOC.Lecture Note Series 234,Cambridge Univ. Press. 1291 J. Kerstan, K . Matthes, and J . Mecke, “Infinite Divisible Point Processes,” Akademie-Verlag, Berlin, 1978. [30] J. F. C. Kingman, “Poisson Processes,” Clarendon Press, Oxford, 1993. [31] S. Kusuoka, de Rham cohomology of Wiener-Riemannian manifolds, Proceedings of the International Congress of Mathematicians, Vol. 1,II (Kyoto, 1990), 1075-1082, Math. SOC.Japan, Tokyo, 1991. 1321 V. Lipscher, Integration by parts formulae for point processes, in ”Festschrift of S . Albeverio”, Vol. 2, eds. F. Gesztesy et al., Canadian Math. SOC.Conf. Proc. 29,AMS, Providence, 2000. [33] V. Mathai, L2 invariants of covering spaces, in “Geometric analysis and Lie theory in mathematics and physics,” Austral. Math. SOC.Lect. Ser., Vol. 11, pp. 209-242, Cambridge Univ. Press, Cambridge, 1998. [34] W. Muller, On the L2-index of Dirac operators on manifolds with corners of codimension two. I, J. Diflerential Geometry 44 (1996), 97-177. [35] M. A. Naimark, ”Normed algebras”, Groningen : Wolters-Noordhoff, 1972. [36] P. Pansu, Introduction to L 2 Betti numbers, in “Riemannian geometry” (Waterloo, ON, 1993), Fields Inst. Monogr., Vol. 4,pp. 53-86, Amer. Math. SOC.,Providence, RI, 1996. [37] N. Privault, Connections and curvature in the Riemannian geometry of configuration spaces, J. Funct. Anal. 185 (2001), 367-403. [38] A. L. Rebenko and G. V. Shchepan’uk, The convergence of the cluster expansion for continuous systems with many-body interaction, J. Stat. Phys. 88 (1997), 665-689. [39] M. Rockner, Stochastic analysis on configuration spaces: Basic ideas and recent results, in “New Directions in Dirichlet Forms” (eds. J. Jost et al.), Studies in Advanced Mathematics, Vol. 8,pp. 157-232, American Math. SOC.,1998, [40] I. Shigekawa, De Rham-Hodge-Kodaira’s decomposition on an abstract Wiener space, J. Math. Kyoto Univ. 26 (1986), 191-202. [41] M. Takesaki, ”Theory of operator algebras”, Springer-Verlag, New York, 1979. [42] A. M. Vershik, I. M. Gel’fand, and M. I. Graev, Representations of the group of diffeomorphisms, Russian Math. Surv. 30 (1975), 1-50.
This page intentionally left blank
Infinite Dimensional Harmonic Analysis I11 (pp. 17-35)
Eds. H. Heyer e t al. @ 2005 World Scientific Publishing Co.
UNE RECIPROCITE DE FROBENIUS HIDENORI FUJIWARA RBsuMB. Soient G un groupe de Lie reel nilpotent connexe et simplement connexe, H un sou-groupe ferm6 connexe de G et x un caractere unitaire de H. Nous 6tudions la representation induite indg x et sa d6sintCtgrationcentrale canonique T = J$ m(?r)?rdp(?r) afin d’6tablir une sorte de reciprocite de Frobenius. A savoir, nous montrons que la multiplicit6 m(?r)s’obtient par la dimension de l’espace des vecteurs gh8ralis6s H-semiinvariants de K E 6.
1.
INTRODUCTION E T NOTATIONS
Soit G = expg un groupe de Lie rbel nilpotent connexe et simplement connexe d’algebre de Lie g. On note G le dual unitaire de G , c.-bd. l’ensemble des classes d’bquivalence des reprksentations unitaires et irrbductibles de G. Etant donnks un sous-groupe analytique H = exp lj de G, ayant l’algebre de Lie fj, et son caractere unitaire x,construisons la reprksentation induite r = indg x de G par translation B gauche dans l’espace ‘H, des fonctions mesurables q5 : G + C telles que q5(gh)= X(h)-l+(g)(g E G, h E H ) et que ScIH 1q5(g)I2dg < 00 pour une mesure invariante dg sur G/H. On sait ( [ 5 ] ,[ll])que la dbsintbgration centrale canonique de r s’bcrit 7- r?
l@
m ( K ) m ( K ) ,
(1)
oh la multiplicitk m ( ~ et ) la mesure p sur G se dkcrivent comme suit en termes de la mkthode des orbites (cf. [3], [13], [17]). Soit g” l’espace vectoriel dual de g. G agit sur g* par la reprksentation coadjointe et G se rkalise comme l’espace g*/G des orbites coadjointes au moyen de la bijection de Kirillov. Pour x E G on note O ( K ) = OG(K) l’orbite associbe. Le caractere x s’kcrit X(expX) = = xf(expX) pour une certaine forme linbaire f E g* sur g telle que f ( [ l j , fj]) = (0). Alors, la mesure p est l’image par l’application de Kirillov d’une mesure finie sur g* bquivalente h la mesure de Lebesgue sur le sous-espace affine rT = f g’- de g*, ici b’- = g’J* = {l E g* : llf~ = 0). Et la ) bgale au nombre des H-orbites contenues dans O ( K )n rT multiplicitb m ( ~est Ceci ktant, nous nous trouvons dans l’alternative suivante : soit il existe un borne ) m ( ~=) 00 pour K E G quelconque. Conformbment a ces uniforme pour toutes m ( ~soit deux Cventualitbs, nous dirons que r est h multiplicitbs finies ou infinies.
+
Date: October 30, 2003. 1991 Mathematics Subject Classification. 22327, 43A85. K e y words and phrases. Orbite coadjointe, Representation unitaire, Reciprocit6 de Frobenius.
18 Pour une reprksentation unitaire p de G, on note H,,Hp” et Hp” respectivement l’espace de p, celui des vecteurs diffkrentiables de p et l’antidual de ‘Hp” (cf. [2], [IS]). Pour a E H:” et b E H;”,on note (a,b) l’image de b par a, de sorte que (a,b) = < b, a >. Posons ( H , y ’ X = { a E H,W;p(h)a = x(h)a,Vh E H} . Dans ces circonstances, il est bien connu [15] que la multiplicitk m(n) a la dksintkgration (1) vkrifie m(n) 5 dim
presque partout pour /I et on peut se demander ([7], [8], [14]) une variante de la rkciprocitk de Frobenius : la multiplicitk m(n) s’obtient-elle par la dimension de I’espace des vecteurs gknkralisks H-semi-invariants de n ? Le but de la prksente note est de rkpondre affirmativement B cette question, qui remonte pour l’auteur jusqu’au Benoist 111. Nous donnons kgalement de simples exemples variks de notions et d’outils utilisks pour l’ktude des reprksentations monomiales. 2. U N LEMME CLEF
Commenqons par nous prkparer un lemme qui nous servira essentiellement pour dkmontrer la rkciprocitk. Soit
s:(0)
=go
c g 1 c ... c gn-l c gn = g
une suite d’idkaux de g telle que dimgk = k (0 5 k 5 n). Soit
z= { i l <
< .. . < i d ) l’ensemble d’indices i (1 5 i 5 n) tels que IJ n gi # IJ n gi-l,et posons J’ = (jl < j , < . . . < jp}= { 1,2,. . . , n} \ Z ( p = dim (g/IJ)). En posant t o = IJ et t,. = IJ + gj, pour 1 5 r 5 p , on obtient une suite de sous-algebres 22
de g :
IJ = t o c t1 c . . . c tp-l c tp = g, dim (t,./t,.-l) = 1. IJ, = IJ f l gi, (1 5 s 5 d ) nous fournissent
D’autre part, les sous-algebres d’idkaux de IJ :
(0) = Co
c h i c ... c hd-1 c f)d
une suite
= 6, dimIJ, = S.
En extrayant Y,E KJ, \ IJs-l pour 1 5 s 5 d et X,. E t,. \ t,-l pour 1 5 r 5 p , on forme une base de Malcev de g relative B KJ. Considkrons l’algebre D7(G/ H) des opkrateurs diffkrentiels G-invariants sur le fibrk b base G / H associk aux donnkes (H, x f ) ,et l’exprimons comme d’habitude en termes de l’algebre enveloppante U ( g ) de gc. Soient d
s=l
19 U ( g ) a , l’idkal a gauche de U ( g ) engendrk par a, et U ( 8 ,7)= {X E W ) [;X ,YI E U ( e ) a , ,w E 5 ) . Alors, nous avons [4] que &(G/H)
U ( g l7 ) / U ( g ) a , .
Nous allons rappeler le travail de Corwin-Greenleaf ([4], [la]) dans le cas oii 7 est B multiplicitks finies. Au passage de tj-1 B tj, notre alg6bre D,(G/H) s’agrandit si et seulement si les H-orbites dans r j = {C E ( t j ) * ; e l b= fib} sont gknkriquement nonsaturkes dans la direction &.-I)*. Cela signifie qu’on trouve un nouveau gknkrateur rationel de DT(G/H) si et seulement si tj ng(C) # ng ( e ) pour e E rT gknkrique, g ( e ) dksignant le radical de la forme bilinkaire alternke Be sur g x g dkfinie par B e ( X ,Y ) = [ ( [ X ,Y ] ) .Ceci posB, nous nous demandons ce qui se passe lorsque cette 6largissement tj n g ( e ) # tj-1 n g(C) s’effectue dans Ij elle-mCme. Soit e E g * , et reprenons la suite d’idkaux S. Pour 1 5 k 5 n, notons e k ( C ) la dimension de G-orbit dans (&)* passant la restriction t ) g k de C A g k . Posons e(C) = (el(C),. . . ,en([)). Soit e = ( e l , e2,. . . en) l’unique n-uple d’entiers non-n’egatifs tel que la couche Ue = {C E g * ; e(e) = e } (ou une sous-couche de Ue) donne des 61kments e-centraux dans un ouvert de Zariski de r,. Soit T ( e ) = { m ;em = em-l} = im1 < m2 < . . . < m:}. Allons utiliser la base de Malcev {Xj}F1 donnee au dkbut, i.e. X j = X, si j = j , et X j = Y, si j = is. D’aprBs Corwin-Greenleaf [4], chaque mk E T ( e )nous fournit un klkment e-central (Tk = & X m k Ak, ou Ak E U(gmk-l)et oii & est un polynbme des u1,. . . U k - 1 vkrifiant rt(&)# 0 pour e E rTgknkrique, 18 re dksigne la reprksentation unitaire irrkductible de G construite au point e dans la thkorie des orbites. Supposons que mk = is pour un certain s (1 5 s 5 d ) , d’oii g k = &Xmk Ak = &kys Ak, et que b, n g(C) n g ( e ) pour e E rTgknkriques. Alors nous avons un lemme qui sert de la clef pour attaquer la conjecture de commutativitk. Notons 0 l’anti-automorphisme principal de U ( g ) .
+
+
+
+
Lemme [9]. Il existe un polyndme P de k variables (certaines variables peuvent y &re manqukes) tel qu’on ait p (O(‘Ji), O(gz), . . . avec a,-1 =
o(gk-i)l
O(Ak)) 0 modU(gmr-i)as-i
x:~lC ( ~ +Z f l j ( v ) ) ,oi apparatt efiectivement O(Ak).
Remarque 1. Ici dans ce lemme il n’est pas question que ou infinies.
7 soit
b multiplicitks finies
3. U N E RECIPROCITE DE FROBENIUS
Nous allons appliquer le lemme-clef B l’ktude des vecteurs gknkralisbs H-semi-invariants de r E GI i.e. des klkments de (7i;m)H’Xf et montrons une rBciprocit6 de Frobenius. Plus prkciskment nous nous proposons de dkmontrer le :
20
Theoreme (Rkciprocitk de Frobenius). I1 s’e‘tablit p-presque partout que m(7r)= dim
(Z;M)H’Xf.
Avant de commencer la dkmonstration du thkoreme, rappelons certains rksultats obtenus dans [7]. Pour l E g*, nous rkalisons la reprksentation unitaire irrkductible correspondante Te a l’aide d’une polarisation b en e, i.e. comme representation monomiale re = indgxe avec B = exp 6. Bien que la notation dissimule le choix de 6, cela ne causera pas de confusion dans ce qui suit. Lorsque e appartient a r7,on peut fabriquer le vecteur de Frobenius ae E (Z;F)H’Xf par la formule
b e , 4) =
/
HIHnB
4(h)Xr(h)dh
pour tout 4 E ZE. Le cas essentiel pour aboutjr au thkoreme est celui oh T est a multiplicitks finies. La, pour p-presque toutes 7r E G, l’intersection R(7r) n r7 consiste en m ( r ) composantes connexes w 1 , . . . , ~ ~ ( dont ~ 1 chacune , est une seule H-orbite. Fixons maintenant e dans une telle orbite E R(n) et rkalisons 7r comme 7r = re = indgxe. En transportant a wk cette realisation par l’opkrateur d’entrelacement et en y appliquant ensuite le vecteur de Frobenius, nous pouvons construire des klkments de (Z;”)H’Xf en nombre m(7r).A savoir, pour 1 5 k 5 m(7r),soient g k E G tels que g k . e E wk, et dkfinissons a: E (z;”)~”’ par la formule
pour tout 4 E Z,”, dkh ktant une mesure invariante sur l’espace homog6ne HfHngkBg;’. Alors, on sait [7] que a;, . . . ,a$n) sont linkairement indkpendants dans (Z;”) H’Xf S’il . en est ainsi, il nous s’agit de montrer que
presque partout pour p si T est
a multiplicitks finies
Demonstration du theor8me. I1 est kvident que l’kgalitk a montrer n’a un sens que p-presque partout. Tout d’abord nous pouvons supposer que Ij contient le centre 3 de g. Puis, en passant au quotient g/(tjna nker f ) et en prockdant par recurrence sur dimg, on se r a m h e au cas oh dim2 = 1,fla # 0. Introduisant comme d’habitude X,Y,Z et g’ tels Y ]= Z et que g’ dksigne le centralisateur de Y dans que 3 = RZ, f(2)= 1, f ( Y )= 0, [X, g. D’oh, g = RX g’. Pour l E g*, on considere la reprksentation unitaire irrkductible x de G qui correspond A l’orbite G . l c g*. Si el2 # fla, on constate immkdiatement que (7-l;cO)H,Xf = (0). Soit donc l ( Z ) = 1. En prennant une polarisation b en e contenue dans g’, on rkalise 7r = indg Xi, oh B = exp 6, et pose G’ = exp g’, 7r‘ = indg Xe de sorte
+
21
qu’on a 7r = indg, d.Lorsqu’on kcrit tout g E G comme g = exp(xX).g’ avec x E R et g’ E GI, il vient que
H;
= S(Rm) = S(R)6S(]Wm-1) = S(R)63’H?,
oh m = dim(g/b) et oh S dksigne l’espace de Schwartz. Maintenant soit a E (H;m)H’Xf.Lorsque g $ g’, on peut prendre X dans b n kerf, et la semi-invariance de a sous l’action de ht = exp(tX) E H (Vt E R) implique aussit6t
que
h 4 W=
(@w)(a’,+) (4
E S(R),$ E 7.m
avec un certain vecteur a’ E ( H ; , W ) ~ ”pour ~ ~ H’ = exp b’, g~ = tj ng’. Cela nous permet de descendre au sous-groupe G’. Supposons dksormais que g C 8’. Si Y E b, la semi-invariance de a sous l’action de ht = exp(tY) E H n B (Vt E R) entraine
ou 4 E S(R),$ E HF.On en dkduit que a n’est autre que la mesure de Dirac pour la premiere coordonnke x en z = l(Y) et que (a,@&J) = rn(4 $) avec un certain vecteur a’ E ( ~ ; m ) ~ , oh ~ iH, = exp(-e(Y)X)Hexp(e(Y)X) et oh .f E g* est dkfinie par .f = exp(-e(Y)X).f. Ainsi, on se retrouve encore une fois dans G’. Maintenant allons examiner le cas essentiel oh Y @ ti. Comme l’inkgalitk m(n)5 dim (H;M)H’Xf
est bien connue, nous pouvons supposer que T est A multiplicitks finies. Alors, puisque toutes les G-orbites rencontrant rT sont saturkes dans la direction (g’)l, de m&mepour H-orbites gknkriques. Considkrons l’indice k (1 5 k 5 d ) tel que gknkriquement dans rT les Hk-1-orbites soient non-saturkes dans la direction (8‘)’ mais que les Hk-orbites y soient saturkes. D’apres le lemme-clef, q ktant le nombre des Clkments de l’ensemble { i E T ( e ) ; i5 ik}, il existe (cf. [9]) des polynBmes Qr (0 5 r 5 ml) des images sous l’anti-automorphisme 0 des Clkments e-centraux u1, . . . , uq et de Yk, oh uq = AYk 8 avec un polyname A de u1,. . . , uq-l et un certain 8 E U(gik-l), vkrifiant
+
m‘
C QrYr
0
mod U(gik)ak.
r=O
Choisissons cette relation (2) de faGon que m’ > 0 soit le plus petit que possible. Appliquons le conjuguk complexe de la relation (2) A notre vecteur gknkralisk a, ce qui nous donne F ) . = O .
(3)
22 C'est-a-dire que, pour p-presque toute
T,
oh x dksigne la premiere coordonee dans l'espace de T = re, et oh les constantes c, (0 5 T 5 m') sont donnkes par Qra = q a . Remarquons que toutes ces constantes q ne dkpendent que de l'orbite O ( T ) = G . C. D'oh le support de a, comme distribution
tempkrke sur Wm, se concentre pour la premiere coordonnke x sur l'ensemble de points finis, i.e. sur l'ensemble des racines de l'kquation
c m'
G . ( G ) ' ( [ ( Y) x ) r = 0.
(5)
r=O
Par un changement de variable on peut supposer que x = 0 est une racine de l'kquation ( 5 ) , dans un voisinage de laquelle on Ccrit a comme une somme de dCrivkes de la mesure de Dirac 6 E S(W)' au point x = 0 :
avec des distributions Dj E S(Wm-')' (0 5 j 5 q) vkrifiant D, # 0. Comme b c g', [b, Y ] = (0). D'oh rksulte que xa = ( G Y C(Y))a appartient a (.H;m)H'Xf. Compte tenu de (6), un simple calcul mene
+
comme distribution. En rkpktant ce prockdk, nous observons que
xqa = (-l),q!D, appartient ( 3 - 1 ; " ) ~ ' X f . Rksumons ce qui precede. Si x = X est la valeur projetke sur la premiere coordonnke du support d'un a E (H;m)H'Xfnon nul, alors il s'ensuit que 6 ~ 6 D dkfini , par
46$J)= r n ( D q ,11) (4 E SP),$J E S(Wm-l)
(~A6Dq,
nous fournit un k1Cment non nu1 de (.H;")H'Xf.
= .H?)
Ceci &ant, il en rCsulte que
D,
E
et oh f E g* est dCfinie par f = exp(-XX).f. C'est ainsi que, si l'on s'intkresse a montrer en passant que, a &ant considkrke comme distribution sur G,le support de a est contenu dans
(H,;")H'xf, oh H
= exp(-XX)Hexp(XX)
S = {g
E G;g. ( C f 6')
n (f + b') # 0},
23 on a d’apr8s l’hypoth6se de recurrence que le support de a est localement contenu dans exp(XX)S’, oh
s’ = { g / E G!;g f .
(el
+ b13(g’)*) n (f” + k ~ g ’ ) * )#
avec C’ = Clg’, f” = flg’, f’ = flg’ et oii k exp(XX)S‘ c S. En somme, on ktablit la :
=
0)
Ad(exp(-XX))(b). Mais il est clair que
Proposition. Soit r = indz x f = se8 m(r)xdp(.rr)comme plus haut et supposons que r est a multiplicitb finies. Pour p-presque toute .rr, on a l’assertion suivante. Soit a E (H;”)H’Xf. Rialisons x = indg X e et regardons a comme distribution bi-semi-invariante sur G, i i gauche par H et droite par B. Alors le support de a est contenu dans S = ( 9 E G ; g .(e+ b)l n.(f bL) # O}.
+
Revenons a la-suite de la dkmonstration du thkor8me. On aimerait voir pour p-presque toutes les x E G qu’B l’expression (6) de a on a q = 0. Pour cela, il suffit de noter que l’kquation ( 5 ) n’a pas de racine multiple. Restreignons nos attentions aux rkalisation .rr = .rrt pour C E rT ghkriques. Alors x = 0 est une racine de 1’Cquation m‘
C C , ( G(ep-) ~)~ - x)T
= 0.
(5)
T=O
Si cette racine ktait multiple sur un sous-ensemble non-nkgligeable de rT,ce qui revient au meme de dire que si le terme au premier degri? pour x dans le membre gauche de ( 5 ) ktait identiquement nu1 pour C E rT car il en dCpend polynomialement [lo], on aurait que m’ 2 2 et la dkrivke par rapport B z du membre gauche de ( 5 ) va annuler le vecteur de F’robenius a = ae :
pour C E r e gknkriques, ou encore
pour C E r7 gknkriques. Par conskquent, nous en dkduisons d’apr8s la formule de Plancherel pour r (cf. [7], [15]) que fd-1
(r + l ) Q T + l Y T
=0
(mod U(gik)adl
r=O
ce qui contredit la minimalitk de m‘. En somme, les C E rTgknkriques tels que x = 0 soit une racine multiple de 1’Cquation ( 5 ) forment un ensemble semi-algkbrique A de r7 qui est nkgligeable pour la mesure de
24 Lebesgue. Comme r est A multiplicitks bornkes, il en rCsulte que (G . A) n rT est aussi nkgligeable. S'il en est ainsi, en revenant a 1'Cquation (5), nous supposons comme prCcCdemment que x = 0 est une racine simple de ( 5 ) . Compte tenu de (4),nous avons pour si E (H;")H'Xf quelconque
(&ix+=o j=1
(7)
avec certains coefficients E j E C (1 5 j 5 m') vkrifiant E l # 0. Reprenons l'expression (6) de a E (HH,")H'Xf dans un voisinage de x = 0, et montrons que q = 0. En effet, si q 2 1, nous savons que
appartient L (H;w)H'Xf.Par conskquent, nous dCduisons de (7) que
ce qui se trouve contradictoire. Nous concluons de tout ce qui precede que pour p-presque toute 7r E G chaque ClCment a E (H;")H'Xf n'est qu'une mesure de Dirac par rapport a la premiere coordonnCe, ce qui nous permet de descendre au sous-groupe G' auquel s'applique l'hypothese de recurrence pour achever la demonstration du thCor8me. 0 Compte tenu de ce qu'on a not6 avant la dkmonstration du thCor&me,nous avons le :
Corollaire. Supposons que r est ci multiplicit6s Jinies. Pour p-presque toutes les G, il vient que
17
E
4. EXEMPLES Nous allons observer certains exemples concernant notre sujet.
Exemple 1. Soit g = ( x l ,x 2 ,y, zl, zZjwavec les crochets non nuls : [ X I , y] = z , [xz,y] = Alors les ClCments zl, 2 2 , z ~ x l - z l x z sont centraux dans U ( g ) .Mais, Ctant donnCes f = z1* et g = Rxz Rzl, on obtient r = inds x j a multiplicitCs finies et les gCnCrateurs de U ( g ,r ) modulo U ( g ) a ,sont plut6t z2, XI. De plus,si l'onconstruit le vecteur de F'robenius a! en e E rT gCnCrique, on voit facilement que x1ae = xl(il)ae. Soient T E G et l E R(7r). Si ( H H , " ) ~#' ~{0}, ~ C(zl)= f(zl) = 1. Ecrivons z2.
+
e = ( Y I X l * + CYZXZ* + yy* + z; -tBzz*
25
avec cq, az, y, 6’E R. Quand 0 = 0, rkalisons re dans L2(R) B l’aide d’une polarisation b = ( 2 2 , y, zl, Z Z ) ~ i.e. , identifions 4 E 7-1, avec @ E L2(R) dkfini par @(t) = +(exp(tzl)) pour t E R. Alors l’invariance de 0 # a E (7-1;F)H’Xf sous l’action de xz E t, exige az = 0 et par consequent que O(T) n rT # 0. Quand 6’ # 0, on constate facilement que n(n) n rT # 0 en faisant bouger e par exp(Ry). En somme, (7-1;03)H’Xf = (0) si
n(T)n rT= 0.
Soient maintenant f = 0 et t, = Rzz. I1 en vient que T = i n d z x f est a multiplicitks infinies et que 2 2 , y, ~ 2 ~ engendrent x 1 U(g, T ) modulo U(g)aT.Construisons ici encore le vecteur de Frobenius at en e E rTgCnBrique. Sous toute reprksentation unitaire irrCz2 agit scalairement et on a zzae = zz(il)at. ductible de G dont l’orbite rencontre D’autre part les actions de y, 21 sur dkpendent du choix de la polarisation utiliske. Si nous utilisons la polarisation de Vergne pour la suite de Jordan-Holder
rr,
Rzz
c (Z2,Zl)R c (ZZ,Zl,Y), c ( ~ Z , ~ l , Y , Q ) g c g -
de g, il est immkdiat que yae = y(ie)at. Finalement on considere la sous-algebre I = (zzry, 21, z& qui est de codimension un dans g. Alors gknkriquement dans rT,les Gorbites sont saturkes par rapport B I, mais les H-orbites ne le sont pas. Conformkment, l’klkment central w = zzzl - zlzz tombe dans U(I) U(g)a, et, p : S + U ( g ) ktant = -zl(i!)zz(ie) pour e E rT. l’application symktrisation, on a p-’(w)(i[)
+
Exemple 2. Soit g = ( X I , 2 2 , b, d, u , y,zl, Z Z ) ~avec les crochets non nuls : [XI,y] = z1, [zz, y] = 22, [zl,2 2 1 = -b, [zl,u] = [Q, dj = y, [b,d] = -21, [b,u] = 2 2 . On vkrifie dans ce casque l’klkment w = .z2z1-z1zZ-by est central dans U(g). Soient f = 0, t, = Rzl+Rd et
e = c(zz)zz*+ qb)v+ e(v)u*+ e(y)y*
+
+ e(Zz)z2*E rT.
I1 rksulte d’un simple calcul que
+
g ( 4 = (l(zz)z1 - e(Z1)zz - !(y)b - W ) y , l(zz)d l(z1)u - l(y)y, 21, zz)w, Comme G = exp(Rz1) exp(Rb) exp(Rv) exp(Ry)G(l)
rT
avec G(e) = exp(g(l)), il suffit de voir n exp(Rb) exp(Rw) exp(Ry) . e afin d’examiner l’intersection r,nG.e. Soit g = exp(tb) exp(uw) exp(wy) avec t,u,w E W tel que g-i? E rT. L’kgalitk (g . e)(d) = 0 entraine t = 0 et (9.C)(z1)= 0 donne ue(y) d ( z l ) = 0 D’oh
+
lorsque e(z1) # 0. On en dkduit que T est sans multiplicitk. Rkalisons la reprksentation unitaire irrkductible re au moyen de la polarisation b = (e(zz)zi - !(zi)zc,
- e(y)b, d, 21,
Y, zi, 2 z ) R .
26 Supposons ~ ( z z # ) 0 et identifions 7& (resp. 7-l:) avec L2(R2) (resp. S(Rz)) par l’application @(s, t ) = q5 (exp(sz2)exp(tb)) (s,t E R) pour q5 E IdTl. I1 rCsulte de calculs :
La deuxicime relation montre que tout a E (H;;OO)H’Xfest une distribution sur la sous-variCtC (s,t ) E B2;t = C(Z$ de R2 si l(zl)
- 2c(zl)
# 0. Posons
Alors la premi6re relation notCe ci-dessus implique
d’ou nous avons
(a,a) = c/
~(s)e”9(22)+$$P(b))dy
w avec une certaine constante c E C D’autre part, on calcule le vecteur de Frobenius at : (at, 4) =
s, q5(exp(szl))ds. Compte
on vkrifie que
En somme, (‘H;;OO)H’Xf= Cat et dim (‘H;F)H’Xf = 1. Posons a prCsent g = R q Rb Rzl. I1 entraine que w E U(g)cl,.Des calculs directs en l E r7 gCn6rique nous fournissent que dim G . C = 4,
+ +
g(l) = (C(zz)z, - C(Y)b, %)d
- C(Y)Y,
Z1,ZZ)R
27 et que b = (%I,22, Y, 21, d, C(z2)zl - e(Y)b)R est une polarisation. D’oa, T est A multiplicitCs infinies car fj g(C) n’est pas lagrangien (cf. IS]). Fabriquons A l’aide de cette b la representation irrkductible re associke A l’orbite G.C. Remarquons qu’il en vient C(zz)zl-C(y)b E fjflg(C) dans cette couche principale de rT.S’il en est ainsi, c’est une bonne occasion de se procurer d’un exemple du lemme-clef. Pour cela nous allons constater que a = z2zl - yb = O(O),0 designant comme avant l’antiautomorphisme principal de U ( g ) ,est un element e-central dans cette couche. En effet, soit 9 ( s , t ) = $ (exp(sz2)exp(tb)) E 7ig g S(R2). On calcul :
+
et puis (ybWs,t) = -2
a9
M Y ) - 4s))K’
D’oh
(9) = P-’
(iwa
(O(0))
Tout compte fait, O(0)agit scalairement dont les valeurs constituent encore une fonction polynomiale sur r7.Finalement O(a)- O(z1)zl E 0 modulo U(g)a2, oh az = Cb Czl, et il existe modulo U(g)a2 une relation polynomiale entre O(a),O(z1),2 1 comme pr6vu. De mbme, soient maintenant f E g* tel que f(zl) = 0 et fj = Wd Wy Wz1. Gkneriquement dans on vkrifie que dim G . C = 4,
+
+ +
rr
~(e= ) (C(z2)zi - C ( Y ) ~- V ) y , C(z2)d - ~(Y)Y,z1, z2)R et que b = (C(z~)zl - C(Y)~ - e(b)y, d, 21, Y,z1, z2)w est une polarisation en C. Fabriquons a l’aide de cette b la representation irrkductible re associee A l’orbite G . C. I1 en rksulte que T = indgxf est a multiplicites infinies. Remarquons que C(zz)d - C(y)y E fj n g(C) genkriquement dans Voyons que a = 2zzd-y2 = O(a)est un Clement e-central dans la couche en question. En effet, on constate aiskment que o repr6sente un 616ment central dans l’algkbre enveloppante de g/Rzl. Autrement, si nous proc6dons tout comme avant, soit 9 ( s , t ) = $(exp(szz) exp(tb)) E ‘Hg g S(R2)).Nous avons
rr.
(
(2~2dQ)(s,t)= 2il(~2)i C(d) - sC(Y) et (Y2@)(S,
1
+ TC(zz)
9
t ) = [i M Y ) - Sqz2))l29.
D’oh,
(00)9 = p-l (Oa)(iC)Q. Nous voyons ainsi que O(a)-20(~2)d-f(y)~= 0 modulo U(g)a2, oh a2 = C (y if(y))+ Czl et qu’il existe modulo U(g)a2 une relation polynomiale entre O(o),O(zz), d comme pr6vu.
+
28 Exemple 3. Soit g = (51, z ~tl, , t z , y1, yz, z1, z ~avec ) les ~ crochets non nuls : [zI, tl] = Y i , [zi,tz]= YZ, [zz, tl] = 21, [ ~ , t 2 = ] 22. Si nous partons de la suite de Jordan-Holder formCe selon la base de g dkcrite plus haut mais ayant ses BlCments rang& dans l’ordre inverse, alors pour le 8-uple e = (0, 0, 0, 0 , 1 , 1 , 2 , 2 ) d’entiers de dimension des G-orbites nous avons la couche
ue = {e E g* : ~ ( Z Z )# 0, e(~l)e(zz)- t(y2)e(z1) = 0) Plus prkcisi:ment, nous venons de faire appel
la suite de Jordan-Holder
~ : { ~ } = ~ O c ~ 1 ~ ~ ’ ~ c ~ 7 c ~ 8 = ~ , OIl
81 = Rzz, 8 2 = 81 85 = 84
+ RZl, 83 = + RYZ, 84 = 8 3 + RYl, 82
+ RtZ, 8 6 = 85 + Rtl, 8 7 = 8 6 + Rz2.
Cela entraine que l’klkment W = ylzz
-
y2z1, qui est central dans U ( g ) ,satisfait a
re(W) = ~i(ie)zz(ie)- yz(it)zi(ie) = ~2(C)zl(e)- yl(e)zz(!) = O pour e E U,quelconque. D’autre part, p = {X E g ; e ( X )= 0 , V l E U,}= (0) et W $ U ( g ) p . On constate que le centre de g est a = (z2,z1, y2, y& et que, pour e E U,,
B ( l ) = a + ( W z 1 - q Y z ) z 2 , W t l - e(zl)tZ),. I1 s’ensuit que b = g(!) Rtz est une polarisation en e E U,,et on r6alise r e comme la reprksentation induite moyennant B = exp 6, savoir re = indg X e . I1 se voit que, si V E U ( g ) satisfait a rg(V) = 0 pour e E U,quelconque, V s’kcrit comme V = TW avec un certain T E U ( g ) et donc que, p dksignant la symktrisation, P-’(V)(il) = 0 pour tout
+
e E u,.
Pour un 8-uple d’entiers non-nkgatifs e = (ej)ljj58 de dimension, soit
T(e) = (1 5 j 5 8; ej = ej-l}, c’est-&-dire que T(e) dksigne, lorsque ej = d i m G . ( l l g j ) , l’ensemble des indices oh la dimension des G-orbites n’augmente pas. Pour notre e mentionni: plus haut,
T(e) = { 1 , 2 , 3 , 4 , 6 , 8 }. Quand on identifie l’espace de re avec L2(R) sous l’application $ $(exp(uzz)) pour tout u E R, on a
= -4 d
ds
(cxp(uz2
+ suz1) exp (-s
(t1-
-t2))
exp
(-s-t2))
ct
@, oh @(u)=
I s=o
29 De meme,
(.e(tz)@) (u)= i (W - ~ ( z z )@(u). ) Par consequent, Te(zzt1 - z1tz)@ = (e(zl)e(tz) - t(Z,)t(tl)) a. C’est ainsi qu’on vkrifie que us = zztl-zltz est un klkment e-central de Corwin-Greenleaf correspondant A 6 E T(e). Faisons des calculs similaires. Nous avons d d (dzi)@) (u)= z$(exp(--sn) exp(uz2)) = Z ~ ( ~ X P ( exd-szi)) ~Q) ls=o
En Somme,
Te(Zzz1 - W Z ) @ = (e(Yz)e(zz) - e(zz)e(zl))@, ce qui montre bien que = zzzl - yzzz est un klkment e-central de Corwin-Greenleaf correspondant a 8 E T ( e ) . Ainsi, on trouve sur la couche U, des 616ments e-centraux u1 = z 2 , 0 2 = z1, u3 = y2, a4 = y1 et u5,0 6 introduits ci-dessus. Soient b = ( t z , y1, y& et f = zz* E g*. I1 rksulte de ce qui prkcdde que r = i n d g x f est a multiplicitks finies car I?, = f b’ c U,. I1 est immkdiat que tl, z1 appartiennent a U ( g ,r ) ,et, ae dksignant le vecteur de Frobenius a e E r,, on a tlat = iC(tl)ae,zlae = il(z1)ae.D’ou nous pouvons directement constater que la conjecture polynomiale (cf. 141) s’ktablit dans ce cas-ci, i.e. l’algdbre D,(G/H) est isomorphe a l’algdbre C[r,JHdes fonctions polynomiales H-invariantes sur r,. Soient maintenant b = ( t z , y1, ~ 1 et) f ~= zz* E g*. Ici aussi r = indg xf devient de m6me a multiplicitks finies. Tout comme avant, on a t l E U ( g , r ) et tlae = iC(t1)at. Ensuite, si w E U ( g ,r ) , on kcrit
+
w
= xlnwn + xln-lwn-l + . .
’
modulo le noyau U ( g ) a , avec certains polyndmes wj (0 5 j 5 n) de zz, yz, tl, 2 2 . I1 s’avdre aussitBt que w, est indkpendant de 5 2 . Puis, en calculant [w, t z ] E U(g)aT,on montre l zz, yz, t l tels que qu’il existe certains polyndmes G,,?-Zl,-de w, = G,z2, w,-1 = -nzzGnyz etc. D’ou, il vient que G, = 2i),zzn-’ avec un certain polyndme 6,de
w -(
2 1~zzY2)ncn ~
E U ( g ,r )
ZZ, yz, tl
et que
30 est de degre inferieur ou Bgal A n - 1 par rapport A zl. Ainsi, l’emploi d’une recurrence sur le degre en question nous assure que 20 se met en polyn6me de z2, y2, tl, xlzz - x2y2, ce qui Btablit sans encombre la conjecture polynomiale dans le cas envisage. ) ~ f = cltl* c2t2’ E g* avec cl,cZ E R. rT n U, etant Soient b = ( t ~ , t z , y l , y zet semi-algebrique dans rT,1’6galit6 dim(b g(C)) = 7 pour C E rTn U, nous assure que 7 est b multiplicites finies. Soit b’ = b n g5. Alors, b n g(C) # b’ n g(C) pour C E rT n U,. Si l’on h i t a5 = z2tl - zlt2 = ultl AS, on a
+
+
+
O(&)
= A5 = -2lt2
ZE ic2.21 = ~ c Z U = ~ Z Q ~ ( U ~ )
+
+
+
modulo U(g5)a’’ oii a’ = C(t2 i f ( t 2 ) ) Cyl Cyz. A savoir, il existe une relation polynomiale O(&) - ic2O(u2) = 0 modulo U(g5)a’entre O(ul), O(u2), ()(us),O ( U ~ )O(A5). , Soient b = (XI, 2 2 , y1, ~ 1 et )f =~ C X Z * E g* avec c E R. Comme precedemment 7 est A multiplicites finies. Ici, on voit generiquement dans rTque b n g(C) # b’ n g(C) avec b’ = b n 87. Pourvu que nous Ccrivions 0 6 = zzzl - yZzz = u1z1 &, alors
+
O(&)
= A6 = -y2z2
+
=
+
= Z C O ~ = icO(u3)
+
modulo U(g7)a’’ O~ a’ = C(z2 if(%)) Cyl Cz1. En bref, il existe aussi dans cette situation une relation polynomiale O(A6)- icO(u3) = 0 modulo U(g7)a’ entre O(ui), O ( ~ Z 0(03), ), 0 ( 0 4 ) , O(u5)1()(As). Exemple 4. Soit g = ( z ~ , ~ ~ , ~ ~ l t ~ , t ~ ,avec y ~ ,les y ~crochets , z ~ , znon ~ ) nuls R : = Y2, b 2 , tll = Zll b 2 1 t21 = z2, [ X o , z11 = 2 2 , b o , Y11 = 21, [zo,Y21 = z2. Si Yon considere la base de Jordan-Holder de g construite da la msme facon que dans l’exemple precedent, on obtient une couche b 1 7 tll = Y11 [21’t21
ue = {C
E g*;[(z2)
# O,C(~l)l(z2)- C ( Y ~ ) [ ( ~ I ) = 0)
pour le 9-uple e = (0, 0,1,1,2,2,3,3,4) d’entiers de dimension. L’61Bment W = ylz2-y2zl est central dans U ( g )et satisfait a re(W) = 0 pour tout C E U,. Nous avons deja remarque que p = {X E g ; C ( X )= 0,VC E U,} = (0) et que W 4 U(g)p. D’autre part, si V E U ( g ) satisfait B r l ( V ) = 0 pour C E U, quelconque, V s’ecrit comme V = TW avec un certain T E U ( g ) et, /3 designant la symetrisation, on a P-’(V)(iC) = 0. Mais cette derniere constatation depend essentiellement de la forme de 1’61Cment central W et il serait trop optimiste de 1’6sperer dans le cas gkneral. Le centre de g est 3 = (z2,zl)R et, pour C E U,,
fl(4= b + (@2).l
- C(y2)zz - -e(ZZ)YZ, C(Z2)tl - C(zl)t2,e(a)Yl - qz1)YZ)R.
+
I1 s’ensuit que 6 = g(C) (yz,t& est une polarisation en C E U,, et nous realisons re comme la representation induite moyennant B = exp 6, A savoir re = indg x!. I1 s’avere que T(e)= {1,2,4,6,8). Identifions l’espace de re avec L2(R2)sous l’application $ ++ Q, oii Q(u, IJ) = 1c, (exp(uzo)exp(vz2)) pour (u,v) E R2 arbitraire. Nous avons d‘abord rdzzt1
-Z
l W = (e(zl)C(tz) - c(zz)e(tl))Q
par les calculs similaires B ceux effectues dans le cas precedent.
31
Puis, des calculs similaires a ceux qu'on afaits ci-desssus nous donnent
Tout compte fait, nous arrivons a re(zzzi - yzzz)Q = (l(yz)t(zz) - l(zz)l(zi)) Q. C’est ainsi qu’on trouve sur la couche U, des klkments e-centraux de Corwin-Greenleaf u1 = z2, uz = z1, u3 = zzy1 - zlyz, u4 = zzt1 - Z l t 2 , a5 = Z2Xl - yzxz.
Soient b = (20,t z , y1, ~ 1et )f =~ ZZ* E g*. I1 rksulte de ce qui prkcede que r = indg xf est A multiplicitks finies car rT = f b’ c U,. Comme [zo,z2z1- yZxz] = 0, nous pouvons prockder pareillement au deuxieme cas de l’exemple prkckdent pour aboutir b l’ktabliisement de la conjecture polynomiale. Soient b = (yl, y2, ~ 1et )f =~ cyz* E g*. Alors, il rksulte du fait que r7n U, est semialgkbrique et dense dans rTet de l’inkgalitk dim (b + g ( l ) ) = 6 < 7 pour l € rTn U, que r est b multiplicitks infinies. En effet, les H-orbites gknkriques dans rTsont saturkes dans la direction (g8)’ et donc l’alg8bre D,(G/H) coi’ncideavec D7,(G8/H), oh 7’= i n d 2 xf. Puisque est contenue dans le centre de g8, il est trivial que l’algebre D7,(G8/H)n’est pas commutative. Soit t E r7rl U,. On a b n g ( l ) # b’ f l g ( l ) avec b’ = b n g3. Ecrivons 03 = zzyl - zlyz = ulyl As. Donc,
+
+
O(A3)= A3 = -zlyz
+
= iczl = icuz = icO(uz)
+
modulo U(g3)a‘ avec a’ = C(y2 if(y2)) Czl. C’est qu’il existe modulo U(g3)a’ une relation polynomiale O(&) - icO(u2) = O entre O(ul), O(UZ),O(A3).
32 Soient b = (21, ~ Z , Y I z& , et f = m1* E g* avec c E R. I1 rCsulte du fait que r,nU, est semi-algkbrique et dense dans r, et de l’inCgalit6 dim (b g(C)) # 7 pour C E rTnu, que T est a multiplicitCs infinies. En effet, les H-orbites gCnCriques dans r, sont non-saturCes dans la direction (&)’ quoique les G-orbites gCnCriques y soient saturkes. I1 est pourtant vrai que dim (b & ( e l ) ) = 7 pour C’ = C(g8 et par suite que T’ = i n d z xf,Gg = exp 8 8 , est multiplicitks finies. Encore, 20,yz E U ( g ,T ) et [ZO, yz] = z2 $ U(g)%, ce qui montre que D,(G/H) n’est pas commutative. Soit d E rTn U,. On a b n g(d) # b’ n g(d) avec b’ = b n 87. Ecrivons u5 = zzxl - ~ 2 x 2= ulzl + As. Donc, O(A5)= As = -yzxZ 0 modulo U(g7)a’ avec a’ = Cz2 Cyl Czl. Ce n’est autre que la forme rCduite de notre relation polynomiale modulo U(g7)a’ entre O(ul),O(uz),O(ug),O ( u 4 ) , O(A5). Finalement, soient b = (q,2 2 , y1, yz, ~ 1 et )f = ~ bx2* cyz* E g* avec b, c E R. Tout comme dans le cas prBcCdent, pour d E rTn U,,on a b ng(d) # b’ng(d) avec b’ = b ng7. Puis a5 = ~ 1 x 1 A5 avec O(A5)= A5 = -yzxz = ibyz = bc, i.e. O(A5)- bc = 0 modulo U(g5)a’ avec a’ = C (x2 Zf(x2)) Cy, C (y2 if(y2)) CZ1.
+
+
+
+
+
+
+
+
+
+
+
Enfin, avant de terminer cette note, calculons dans un cas typique des ClCments ecentraux de Corwin-Greenleaf, qui ne sont autre que des ClCments centraux de U ( g )dans notre cas, et voyons qu’ici le lemme clef devient facile B comprendre.
Exemple 5. Soit g = (Xl,.. . , X& avec les crochets non-nuls : [X,, Xk] = Xk-l, (2 5 .k 5 n - 1). D’oh le centre de g est 3 = RXl et b = (XI,.. . , Xn-l)R est un ideal abklien de codimension 1 dans g. Les G- orbites coadjointes ont gCnCriquement la dimension 2 et 18 b se qualifie pour polarisation commune de g. Ceci posC, notre reprksentation monomiale 7 = indg xf est B multiplicites finies m6me si dim i~ = 1, sauf le cas ou l~ = 3 et oh f l 3 # 0. En posant g j = (Xi,.. . ,Xj)Rpour 1 5 j 4 n, on obtient une suite de Jordan-Holder de g : (0) = go
c 81 c ... c gn-1 c g n = g.
Nous allons chercher tout simplement des BlCments centraux de U ( g ) .I1 y en a 01
= XI, 0 2 = 2X1X3 -
xz2,0 3 = Xl2x4- xlXzX3 +
3 etc.. Soient, par exemple, n = 5,b = (X3,rX4)get f(X3) = X,f(X4) = K. Alors, T est 5, multiplicitks finies et 1’algGbre DT(G/H) coi‘ncide avec l’algkbre des polyn6mes de X1,Xz. Donc, cs3 ne contribue pas a la formation de D,(G/H). Posons
+ if(X3))+ C(X4+ if(X4))= C(X3+ i X ) + C(X4+ et a1 = C(X3 + Zf(X3)) = C(X3+ ZX). On a = -2ZXu1 - Xz2 modulo U(g1)al et (53 = u12~4 + ( i ~ u l +?1. ~ 2 2 ) ~2 = u12x4+ a, = C(X3
ZK)
cs2
33 D'oh, si l'on prend b' = RX4, u3 contribuerait effectivement ind$ xf avec H' = exp b', mais 9(u3 - uTX4)'
former D+(G/H'), 7' =
+ + 2iXu1)(u2 - i X ~ 1=) ~0 modLI(g2)a, (UZ
et en particulier
+
+ + 2i~u1)(u2- i~u1)'= o m o d ~ ( g 4 ) %
9 ( ~ina12)2
(u2
Nous pouvons continuer ainsi de suite. En effet, pour n 2 6, a 4 = 2X1X5-2XzX4+X32 est central dans U ( g ) ,et si l'on part de b = (X3, X4, X S )avec ~ f(X5) = C, on obtient 04 =
2~1(-i<)- 2Xz(-in)
+ ( - 2 ~ ) ~= -2i
2inxZ - X2
modulo U(g3)%.D'oh, (U4
+
22
f-
-4K2x22
4K2(U2
+ 2iXU1)
modulo U(g5)%. Soit maintenant n 2 7. Considkrons u5 =
1 1 1 x14x6 - X , ~ X , X+ ~ -xl2xZ2x4 - -x1xZ3x3 + -xZ5 2 6 30
qui est encore central dans U ( g ) . Prenons b = (X3,X4, X 5 ,x6)~ pour que assez grand en l E r7 gknkrique. Soit f(&) = y. Alors,
n g(!)
soit
modulo U(&)a,. Tout indique que u1, u2, par exemple, engendrent algkbriquement l'algebre des elements r,-centraux. Soient maintenant b = (X3,X5)Ret f(X3) = A, f(X5) = C comme avant. On a
34 modulo U ( g 3 ) ~Donc, .
= -2icu13
+ -A31
2
u1
2ix 2 + -u1u2 + -u22 3 3
- 2xZu3
modulo U(gs)a,. Finalement, (3Ui2ff4
-2
~
+
+
-22 i X~ ~ l u 2 6iCui3 - X2u12)2 36(u2
+2 i X ~ l ) u = 3 ~0
modulo U(g5)a,. Si l’on jette un coup d’ceil au cas oh n = 6, D,(G/H) s’identifie avec 1’algBbre des polyntimes B trois variables X I ,X2,X4. Enfin, soient = (X4,X5), et f(X4) = IC, f ( X 5 ) = comme prkckdemment. Cette fois, U4 = 2Ul(-iC) - 2X2(-iIC) + x32= -2i
<
+
+ +
2 3 u2’~+ 2u2Xz2 ~ X ~+ x24 4 ~u 14 - 8 i C ~ 1 8 i 1 ~ u 1 ~ X z(u2 Xz2)’ = -8i
+ 8i
( 4 ~ 1 ~ ~2~ ~ 4
X U ~ ) ~
-
modulo U ( g 5 ) ~Tout . comme avant, l’algkbre D,(G/H) est 1’algBbre des polyntimes a trois variables X I ,X Z ,X3 lorsque n = 6. 5. REMERCIEMENTS Cette note est un sous-produit des collaborations [9], [lo] sur la conjecture de commutativitk due B Duflo (61 et B Corwin-Greenleaf [4]. L’auteur remercie vivement ses collaborateurs Gkrard Lion, Bernard Magneron et Salah Mehdi, ainsi que l’universiti! Paris 13 qui a rendu possible ces collaborations en l’invitant pour un skjour d’un mois en juin 1999. Ce travail a kti! partiellement financi! par le Ministkre de 1’Education du Japon qui a attribuk B l’auteur la subvention N”11640189.
REFERENCES [l] Y . BENOIST, Analyse hamonique sur les espaces symitriques nilpotents, J. Func. Anal., 59 (1984), 211-253.
35 [2] P . CARTIER,Vecteurs diffkrentiables dons les reprksentations unitaires des groupes de Lie, Lect. Notes. Math. Springer 514 (1975), 20-34.
Representations of nilpotent Lie groups and their applications, [3] L. CORWINAND F. GREENLEAF, Part I : Basic theory and examples, Cambridge University Press, 1990. [4] L. CORWIN AND F . P . GREENLEAF, Commutativity of invariant differential operators on nilpotent homogeneous spaces with finite multiplicity, Comm. Pure Appl. Math., 45 (1992), 681-748. [5] L. CORWIN,F . P . GREENLEAF AND G. GRELAUD, Direct integral decompositions and multiplicities for induced representations of nilpotent Lie groups, Trans. Amer. Math. SOC.,304 (1987), 549-583. [6] M. DUFLO,Open problems in representation theory of Lie groups, edited by T. Oshima, Katata in Japan 1986, 1-5.
Reprdsentations monomiales des groupes de Lie nilpotents, Pacific J. Math., 127 [7] H. FUJIWARA, (1987), 329-351. [8]
H. FUJIWARA, Analyse harmonique pour certaines reprt?ssentations induites d’un groupe de Lie nilpotent, J. Math. SOC.Japan, 50 (1998), 753-766.
G. LION,B. MAGNERON AND S.MEHDI,Commutatioity criterion for certain alge[9] H. FUJIWARA, bras of invariant differential operators on nilpotent homogeneous spaces, Math. Ann., 327 (2003), 513-544. G. LION ET B. MAGNERON Algdbres de fonctions associkes a m reprksentations [lo] H. FUJIWARA, monomiales des groupes de Lie nilpotents, Prepub. Math. Univ. Paris 13, 2002-2,2002. 1111 G. GRELAUD,Sur les reprksentations des groupes de Lie rksolubles, Th&e, Univ. Poitiers, 1984. [12] F. P . GREENLEAF, Harmonic analysis on nilpotent homogeneous spaces, Contemporary Math., 177 (1994), 1-26. [13] A. A. KIRILLOV, Reprisentations unitaires des groupes de Lie nilpotents, Uspekhi Math. Nauk., 17 (1962), 57-110. [14] R. LIPSMAN, Attributes and applications of the Cowin-Greenleaf multiplicity function, Contemporary Math., 177 (1994), 27-46. 1151 R. PENNEY, Abstract Plancherel theorem and a Fmbenius reciprocity theorem, J. Func. Anal., 18 (1975), 177-190. [16] N. S. POULSEN, On Cm-vectors and intertwining bilinear forms for representations of Lie groups, J. Func. Anal., 9 (1972), 87-120. [17] L. PUKANSZKY, LeGon sur les reprisentations des groupes, Dunod, Paris 1967.
HIDENORI FUJIWARA, FACULTE DE TECHNOLOGIE A KYUSHU,UNIVERSITE DE KINKI,IIZUKA 8208555, JAPON. E-mail address: fujiwaraOfuk.kindai.ac. j p
This page intentionally left blank
Infinite Dimensional Harmonic Analysis I11 (pp. 37-57) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
MARKOV PROPERTY OF MONOTONE LEVY PROCESSES
UWE FRANZ Institut fir Mathematik und Informatik Emst-Mon’tz-Amdt- Universitat Greifswald Jahnstrajle 15a, D-17487 Greifswald, Germany Ernail: jkznzOuni-greifswald.de NAOFUMI MURAKI Mathematics Laboratory Iwate Prefectural University Takizawa, Iwate 020-0193, Japan Ernail: rnurakiOiwate-pu. ac.jp Monotone LBvy processes with additive increments are defined and studied. It is shown that these processes have natural Markw structure and their Markw transition semigroups are characterized using the monotone LBvy-Khintchine f0rmu1a.l~ Monotone L6vy processes turn out to be related t o classical L h y processes via Attal’s “remarkable transformation.” A monotone analogue of the family of e x p e nential martingales associated to a classical L h y process is also defined.
1. Introduction
One of the remarkable features of quantum probability is the existence of several different notions of independence. The most prominent examples are certainly tensor independence and free independence. Tensor independence is based on the tensor product of algebras and linear functionals. It generalizes the notion of independence used in classical probability and was used to develop the quantum stochastic calculus on the symmetric Fock space, 6.Ref. 13,22. Free independence can be motivated by the free product of groups. Even though it is incompatible with classical probability there exist no non-trivial examples of classical random variables that are freely independent - it has been used to develop the secalled free probability theory, which has many parallels to classical probability, cf. Ref. 26, 27. Another independence for which limit theorems, infinite-divisibility, convolution semigroups, etc. have been studied is boolean independence,
38
see Ref. 25. Monotone independence18J7 arose from a non-commutative Brownian motion and an interacting Fock space that was introduced in Ref. 15, 16, 14. Recently, Accardi, Ben Ghorbal, and Obata’ have discovered that monotone independence also appears in the study of the comb product of graphs. Ranzg has shown that monotone independence leads to an ass* ciative product of quantum probability spaces and can be used to develop a theory of quantum Levy processes analogous t o the ones for tensor, free, and boolean independence in Ref. 24. The monotone product in not commutative, unlike the tensor, free, or boolean product. MurakiZ0Jghas proven that there exist only five universal notions of independence in quantum probability. These are tensor, free, boolean, monotone, and anti-monotone independence. Ranz’O subsequently found a construction that reduces boolean, monotone, and anti-monotone independence to tensor independence. As an application the theories of quantum LBvy processes with boolean, monotone, and anti-monotone increments can be reduced to the theory of Levy processes on involutive b ialg eb r~.’~ This actually implies that for any quantum LBvy processes with monotone increments their exists a natural family of conditional expectations with respect to which it is Markovian, cf. Corollary 4.4 in Ref. 10. But the construction is algebraic and does not give a direct expression for the semigroup of Markovian transition kernel. In the present article we give a description of the semigroups of Markovian transition operators and kernels of quantum LCvy processes with additive monotonically independent increments. This is similar t o the program carried out by Biane5 for quantum LBvy processes with additive or multiplicative freely independent increments. But unlike the free product, the monotone product does not preserve traces, see Remark 2.2. Therefore the existence of the conditional expectations does not follow from the theory of von Neumann algebras, but relies on the explicit construction given in Ref. 10. It turns out that monotone LBvy processes can be obtained from the Fock space realization of classical LBvy processes via Attal’s “remarkable transformation,”2 see the proof of Theorem 4.1. The Fock space realization of a classical LBvy process defined in Equation (6) is a quantum semi-martingale in the sense of Ref. 2 only if the LBvy measure has compact support. In this case it turns out that the associated monotone LBvy process is even bounded for all t 2 0. We also associate a family of martingales t o a monotone Levy process that are analogous to the exponential martingales of a classical LBvy process, see Theorem 7.1.
39
2. Monotone Independence In this section we present the definition of monotone independence and its main properties.
Definition 2.1. Let 3t be a Hilbert space, R E 3t a unit vector, and define a state @ : B(3t) + C on the algebra of bounded operators on 3t by
+(X) = (R,XR),
for X E B(3t).
Subalgebras d1,... , 5 D(3t) are called monotonically independent w.r.t. @, if the following two conditions are satisfied.
(a) For all X E A, Y E dj,2 E dk with j > max(i,k), we have
XYZ = @(Y)XZ. (b) For all XI E A,,. .. ,X, E din,Y E dj, 21 E d k l , ...,2 , Edkm with il > i2 > -.. > in, kl > k2 > . . . > km, and j c min(i,, IFm), we have
@(XI-**X,YZ,. . . Z , ) = @(x~)~~~@(x,)~(Y)@(z* Operators XI,. .. ,X, E B(3t) are called monotonically independent w.r.t. 0, if the subalgebras = alg(Xi) = span{Xtlk = 1,2,.. .}, i = 1 , . . . , T , they generate are monotonically independent.
Remark 2.1. We do not require the subalgebras t o be unital. If the ith algebra di contains the unit, then @ ~ I A ~ is a homomorphism for k > i, since
@(XlX2)= @(XllX2)= @(1)@(Xl)@(X2) = @(Xl)@(X2) for d l XI,x2 E Ak. We will call a triple (d,3t, 0)consisting of a Hilbert space 3t, a unit vector R E 3t, and a subalgebra d B(3t) a quantum probability space. By an operator process we shall mean an indexed family (Xt)tEIof elements of some quantum probability space. A quantum random variable is a homomorphism j : B + d from some *-algebra into a quantum probability space and a qzlantum stochastic process is an indexed family (j,)tEr of quantum random variables.
Proposition 2.1. Let (di,Nj,Rj))i = 1,2, be two quantum probability spaces, and denote the states associated to R1 and 0 2 by 01 and @2, respectivel y .
40
Then thew exists a quantum probability space (A, 3c, R) and two injective state-preserving homomorphisms Jj : dj + d,i = 1,2, such that the images 51(d1)and Jz(d2) are monotonically independent w.r.t. R . Proof. We set 3c = 311 8% and R = R1 €302. Denote by P2 the orthogonal projection on cCn2 C_ 3c2. We define the embeddings Jj : d i + B(3c) by
Jl(X) = x €3 P2, J2(X) = 1 @X,
for X E d1, for X E d2.
For d we take the subalgebra generated by Jl(d1) and Jz(d2). It is clear that J1 and J2 are injective, state-preserving homomorphism. A simple calculation shows that Jl(d1) and Jz(d2) are monotonically independent w.r.t. R. E.g., for products of the form J~(X~)JZ(Y)JI(X~), X1,X2E d1,Y E d2,we get Jl(Xl)J2(Y)Jl(X2)
= (Xl€3P2)(l €3Y)(X1€3P2) = (XlX2)€3P2YP2 = Qi (J2 (Y))51(XlM(X2).
On the other hand, for Qi(J2(Yl)Jl(X)J2(Y2))
J~(YI)J~(X)J~(Y~), X E d1,Yl,Y2E d2,we get = (Rl €3 R2,(1€3 Yl)(X €3 P2)(1€3 Y2)Ol €3 R2) = (01 €3 R2,X€3 (YlPY2)Rl8 02) = Qil(X)Qi2(Yl)Qi2(&) = 9(52 (Yl)) Qi (J1
(XI) Qi (J2 (Y2)).
We will call the quantum probability space (d,%,R) constructed and in the previous proposition the monotone product of (di,?fi,Ri) (d2,3c2,R2). When there is no danger of confusion, we shall identify the algebras d1and d2 with their images Jl(d1) and J2(d2), respectively.
Remark 2.2. (a) The monotone product is associative and can be extended t o more than two factors, see also Fief. 9. But it is not commutative. (b) The state Qi on d = alg(d1, d2)depends only on 9 1 and Qi2l~~. (c) The embedding J1 : d1+ A is not unital. It is a consequence of Remark 2.1 that it is impossible to get a unital embedding of the first algebra, if the state on the second algebra is not a homomorphism.
41
(d) The product is not tracepreserving. If equal to zero, then the calculation
@1I.+
is not identically
@l(X)@2(KY2) = @(XYlY2)= @(&Xy1)= @l(X)@2(Yl)@2(Y2) for all X E d1, Y1 ,Y2 E d2 shows that @ can only be a trace on d, if @ 2 1 is~ a~ homomorphism. Note that we identified the elements of d1 and d2 with their images under the embedding J1 and J2 to simplify the notation. We now recall several results from Ref. 17, which will be important for the following sections. By cC+ = { z E CISa > 0) we denote the upper half plane. Denote by R x ( z ) = ( z - X)-l the resolvent of an operator X. Let X E B(31)now be self-adjoint. We will need the reciprocal Cauchy transform
of the spectral measure of X evaluated in the state @. We use the same notation for the reciprocal Cauchy transform
of a probability measure v on R Theorem 2.1. (Theorem 3.1 in Ref. 17) Let XI,... ,X,, E B(X)be monotonically independent self-adjoint operators. Then
Hx,+...+x,( 2 ) = H X l ( * * - ( H x , (2))
*
--)
forzEC+. This theorem suggests to define the monotone convolution of probability measures as follows. Definition 2.2. (Definition 3.2 and Theorem 3.5 in Ref. 17) For any pair of probability measures p, v on W there exists a unique probability measure X on R, whose reciprocal Cauchy transform is given by
HA = H,(H,(z)), z E @. We will call X the monotone convolution of p and u and denote it by X = p 0 u.
42
In the proof of Theorem 3.5 in Ref. 17, Muraki gives the following formula for the monotone convolution, r
where the measures p,, y E R, are given by their reciprocal Cauchy transforms E C?. HV,,(z)= H V ( z )- y , The measures v,, y E R, can also be defined by v, = S, D v. The monotone convolution is associative, f i n e in the first argument, and weakly continuous in both arguments. Note that it is not commutative. A monotone convolution semigroup is a weakly continuous family (pt)t?o of probability measures on P such that p,+t = ,usD pt for all s,t 2 0 and po = SO. Monotone convolution semigroups are in oneto- of reciprocal Cauchy one correspondence with continuous families (Ht)t>O transforms that form a semigroup w.r.t. composition. These semigroups are even differentiable with respect to t and the derivative 1 a E C? A ( z ) = lim - ( H t ( z ) - z ) ,
t is called their generator. Muraki has classified all monotone convolution semigroups in terms of their generators, see Theorem 4.5 in Ref. 17. Since we will only consider monotone convolution semigroups of probability measures with compact support in this paper, we will need Theorem 5.1 from Ref. 17. t\o
Theorem 2.2. (Ldvy-Khintchine formula for the monotone convolution) Let (pt)t>o be a weakly continuousfamily of probability measures and assume pt # Sofor some t > 0. Then the following are equivalent.
(a) (pt)t>o is a monotone convolution semigroup of compactly supported measures on B. (b) There exists a unique pair (a,p) # (0,O) of a real number a and a compactly supported finite measure p on P such that the Pick function
generates H,,, an the sense that for z E C?, H P t ( s ) = w i s the unique solution w E @ of the equation
43
Remark 2.3. Composition semigroups of holomorphic functions on a half plane or on the disc and their generators were also studied in Ref. 7, 12. We will call the pair (a,p) the chamcteristic pair of the monotone convolution semigroup (pt)t>o, -a its drift coeficient, p((O}) its digusion coeficient, and 3 ( p - p({6})60) its Ldvy measure.
3. Conditional Expectations We will first introduce “conditionalexpectations” for the monotone product of two quantum probability spaces.
Proposition 3.1. Let (Ai,Xi, a,), i = 1,2, be two quantum probability spaces and denote by (A,31, 0) their monotone product. Then there exists a linear map E1 : A + A1 such that
(a) (b) (c) (d) (e)
E I ( J I ( X ) Y J ~ ( Z=)XEl(Y)Z ) for all X , Z E AI, Y E A, @1 0
E1 = a,
El 0 J1 = id, El is completely positive, if A1 and A contain units 11 and 1, then we have El(1) = 11.
We will call El the conditional expectation from (A,31,R) onto (A1,3 1 1 , 01). Proof. Let P denote the orthogonal projection onto P = 1 €3 P2. Since
3 1 1 C3 0 2
C 31, i.e.
P ( X €3 Y ) P = x €3 P2YP2 = @ ( Y ) X€3 P2, for X E B(311), Y E B(’Hz), we see that the image of the map from A to B(31) defined by Z I+ PZP lies in Jl(A1). Therefore we can define El : A+ A1 by E l ( Z ) = Jc’(PZP). It is straight-forward to verify that El satisfies all the properties listed in the proposition.
Remark 3.1. The map Z I+ PZP from B(31)t o B(311)€3 P2 is basically the conditional expectation for the tensor product of quantum probability spaces. Note that it is impossible to define a conditional expectation onto A2 in the same way, since the image of the map Z I+ P’ZP’, where P’ is the orthogonal expectation onto 0 1 €3 3 1 2 E 31, does not lie inside Jz(A2).
44
Proposition 3.2. Let X1 and X2 be two self-adjoint operators on 5% and fi. unit vectors fli E X i , i = 1,2. Denote by A, the C'-algebras Ai = {f(Xi)lfE C@)} generated by X1 and Xz,respectively. Let (A,X,fl) be the monotone product of (Ail?tili,a,), i = 112,and denote the images of XI and X2 in A again by X1 and XZ. Then we have 1 1
for I E
e.
(2 - (X,+X2J
= HX,(.)
-x1
Proof. The idea of the proof is the same as for Proposition 3.2 in Fief. 5. Denote again by
Rx(z) = .( - x>-' the resolvent of an operator X. Then we have RX,+X,(I) = Rx,(z)(1 - X1RX2(4)-1 As in Ref. 5 we can expand this expression into a norm convergent series
c RX2(.)(X1Rxz(4)k 00
Rx,+xz(.) =
k=O
for 1.1
> llXlll + IlX~ll.Therefore
and 1
El (RXl+XZ(4) = Hxz(z)- Xl. By uniqueness of analytic continuation follows that the identity holds for all z E C\[-IIXl II - IIX2IL IIXiII + 11X2111~ Corollary 3.1. For all f E C@), we have
+
(f(x1 x2))
where the operator T is defined by
= (Tf)(X1),
45
with the measures ~
2
determined , ~ by their reciprocal Cauchy transforms,
HP2,.( z ) = Hx2( z ) - z,
for z E C+.
R e m a r k 3.2. We can also prove this formula using only the fact that T does not depend on the distribution of XI and Formula (1). Denote the conditional expectation of f(X1 X2) by Tf(X1) = El (f(X,+ X2)). Since the conditional expectation preserves expectations, we get
+
@ ( W X d )= @(f(Xl+X2)). Denoting by p1 and p2 the distributions of XI and X2 w.r.t. R, this becomes
Substituting p1 D p2 with Formula (l),we obtain
Since this has to hold for all probability measures p1, we get the desired result. 4. Monotone LQvyProcesses
Definition 4.1. A family of self-adjoint operators (Xt)t>o 2 B(31)is called monotone Ldvy process w.r.t. R E 31, llRll = 1, if the following conditions are satisfied.
(a) (Independence of increments) For all n E N and 0 5 tl 5 the operators xt, ,xt2
- xt,7 - * * 7 Xt,
-.- 5 t,,
- XL-1,
are monotonically independent w.r.t. R. (b) (Stationarity) The distribution of an increment X t - X, w.r.t. to the state = (0, R) depends only on t - s, i.e. @(a)
-
@(f(Xt- xs>> = @(f(&+h - Xs+h))
for d l 0 5 s 5 t , h 2 0 , and f E cb(R). (c) (Weak continuity) limtlo @(f(Xt))= f(o) for all f E cb(R).
46
We will call two monotone LBvy processes (Xt)t>o and (Xi)t>o,defined on (R, 0 ) and (R’, Q’), equivalent if their marginal.distributio& coincide, i.e.
0%f ( X t - XS)Q) = (a!,f(&-
X W )
for all 0 5 s 5 t , f E cb(R). Due to the monotone independence of the increments this implies that all finite joint distributions also coincide. Denote by C[z] the algebra of polynomials generated by one self-adjoint indeterminate z = z*. It is a dual group with the comultiplication A : q x ]+ C[z] C[z] E C[z’, z”]defined by b ( z ) = x’ x”, see Ref. 9, 10. The conditions in Definition 4.1 imply that the homomorphisms jst : qz]+ B(R)defined by
+
for 0 5 s 5 t and k = 1,2,. .. form a monotone LBvy process on qz]in the sense of Definition 2.5 in Ref. 10. In Ref. 9, 10 it was shown that such a processes is uniquely characterized by its convolution semigroup (p&oof states on C[z] given by pt(zk)= ( n , x ; Q ) ,
t 2 0, k 2 0.
Conversely, a monotone LBvy processes (jst)O<s
xt = j o t ( z ) ,
t 1 0,
if these operators are bounded (and hence self-adjoint) for all t 2 0. Since we are only considering bounded operators, the marginal distributions ( , u t ) t l o defined by
and are for t 2 0 have compact support contained in [-IlXtll,IlXt11] uniquely determined by their moments. Therefore we obtain the following classification of monotone LBvy processes.
Proposition 4.1. We have a one-to-one correspondence between monotone Ldvy processes (Xt)t>o (modulo equivalence) and monotone convolution semigroups ( h ) t >-o of compactly supported probability measures.
47
We will now apply the results of Ref. 9, 10 to show how the monotone LCvy process associated to a monotone convolution semigroup of compactly supported measures with characteristic pair (a,p) can be constructed. Let p be a compactly supported probability measure on R, suppp 5 [-M, MI, with M > 0, and define q,,: C[x]+ C by
Cp(P>=
J P(4dP(Z),
p E C[4.
R
Then q,,is uniquely determined by the generating function 00
1 -dp(z)
H,, (4
z-x
k=O
for 1.1
1
=-
> M . It follows that the generator
> o , qt = q,,* for t 2 0, assoof the convolution semigroup of states ( q t ) t ciated to a monotone convolution semigroup (pt)t>o, - can be characterized by the generating function
for sufficiently large .1
Therefore we get
0 L(xk) = { - a JR
if k = 0, if k = l , xk-2dp(z) if k = 2,3,. ...
Note that L : C[x] 3 C extends to a unique continuous functional on C,"(R), which can be given by
for f E C,"(R). Recall that a SchGnnann triple (a,q, L ) on a *-algebra B with a character E : B + C over some (pre-) Hilbert space H consists of a 0 0
a *-representation a of B on H , a n-e-cocycle q, i.e. a linear map q : B
d a b ) = a(.)rl(b) for all a, b E
D,and
+H
such that
+7 ) ( 4 4 b )
(3)
48 0
+ C such that L(ab) = 4 a ) L ( b )+ M a * ) , d b ) )+ L ( 4 4 b )
a hermitian linear functional L : B
(4)
holds for all a, b E B. Lemma 4.1. Let H = L2(H$p) and set n(f) = operator on L2(Ihp,p),M f ( g ) = fg) and
Mf (the multiplication
for f E C,”(R). Then (n,q, L ) with L as in Equation ( 2 ) defines a Schdrmann triple on C,”(R) with the character ~ ( f=) f(0) for f E C,”(R).
For 0 5 s 5 t , let qs,t[= : L 2 ( R + , H ) + L 2 ( R + , H )be the orthogopf. nal projection onto L~(Is,t[“,H ) L2(R+,H ) , i.e. qsrtpf = ~ l ~ , ~We , ~J?(I)s,tp) ~ E Z3(I‘(L2(R+,H))), 0 5 s 5 t , the secdenote by J ? I ~ = ond quantization of Ijs,t[c. For the case t = 00, we introduce the shorter and rS] =r ~ ~ , ~ p . notation P,] = qs,oo[c
49
Theorem 4.1. Let a be a real number, p a compactly supported finite measure on R, and H = L2($p). Denote by Ma the operotor H 3 f I+ xf E H and by 1~ E H the constant function with value one. Then the quantum stochastic diflerential Equation %t
=
/
t
+ dA,(M,) + d A s ( 1 ~+) dA,(lR)
(-t,dA,(id)
- ad7)
(5)
8
has a unique bounded solution. Furthermore, (Xt)t>o - with
xt = Zotrtl,
t 2 0,
is a monotone Lduy process w.r.t. the vacuum vector s2. The monotone convolution semigroup associated to (Xt)t>o - has characteristic pair (a,p). If (X&O- is another monotone Ldvy processes whose convolution semi-
p u p has characteristic pair (a,p), then (Xt)t>o - and (Xi)tzo are equivalent. Proof. Denote by B involutive algebra generated freely (as an algebra) by two self-adjoint generators x and p , with the coalgebra structure defined by
A(z) =z@p+l@z, E(Z)
= 0,
A@) = p @ p , E(p)
= 1.
It follows from Section 3.2 in Ref. 10 and Schiirmann's representation theorem23 that Equation ( 5 ) has a solution on some dense invariant subspace of I'(L2(R+, H))and that j&)
=28t
and
j88tb) = q e , t [ "
defines a LoIvy process ( j 8 t ) o l s l t on the involutive bialgebra B w.r.t. t o the vacuum state in the sense of Ref. 23. The operator process (yt)t>o - defined by
yt = At(M,)
+ A z ( 1 ~+) A t ( l R ) - atid,
t E B+,
is a quantum stochastic integral process with bounded coefficients and belongs therefore to the algebra S' of possibly unbounded quantum semimartingales introduced by Attal? see also Ref. 4. The operator process (&t)t>o- satisfies the quantum stochastic differential equation
1 t
20t= yt
-
%osdA8(id),
t E R+.
(6)
50
By Proposition 9 from Ref. 2 and Proposition 37 from Ref. 4 the solution of this quantum stochastic differential equation is unique and equal to the image DY of (K)t>ounder Attal’s “remarkable transformation” D,see also Ref, 4 for the generalisation to a Fock space with more than one degree of freedom. Therefore (&,t)t>o belongs to the algebra S of bounded quantum semimartingales in the sense of Ref. 2,4. The operators X,t can be obtained fiom &-,by a timeshift and are therefore also bounded. By Theorem 3.7 from Ref. 10, j,t(z) = X,, = for 0 5 s 5 t defines a monotone LCvy process on 44 w.r.t. to the vacuum state in the sense of Definition 2.5 in Ref. 10. Since 2,t and rt]are symmetric, bounded, and commute for all 0 5 s 5 t , the operators Xt = &trt] are also symmetric and bounded, hence self-adjoint. This implies that (Xt)t?o is a monotone LBvy process in the sense of Definition 4.1. The coefficients Mz = n(z),1~ = ~ ( z )and , -a = L ( z ) in Equation (5) correspond exactly to the Schiirmann triple associated to the characteristic pair ( a , p ) in Lemma 4.1. Therefore (Xt)t>o has the correct monotone convolution semigroup.
Remark 4.1. It follows from Ref. 11 that R is cyclic for (Xt)t>o. - But R is not separating, except for p = 0 6.e. the pure drift process, see Subsection 6.1). Set Zt = X t - 9(Xt)rtl for t 2 0. Then we have 9 ((X,(Xt - X,))*X,(X,- X)) = 9 ((% - X,)x:(Xt- X,)) =
and therefore
- X,(Xt
but
(qxt- xs))2@(x:) =0
- Z,)R
= 0,
(a,( X t - X,))’)= 9 (X,(Xt - x,)2z)
9 (X,(X, - X,)
=9
- -
(q ((X,- X,)”> 9
proves that X , ( X t - X,)# 0 for 0 < s < t, unless xt = 0 for all t 2 0.
5. The Markov Semigroup of a Monotone LQvy Process Let (a,p), (a,p) # (O,O), be a non-trivial characteristic pair. In this section we will always assume that (Xt)t>o - is the monotone LBvy processes on I‘(L2(& ,L2($ p ) ) constructed in Theorem 4.1 for (a,p).
)
51
As “conditional expectations” we will use the linear maps Et f? (r (Wb.2 H ) ) ) + f? (L2(w,H I ) )
:
(r
E t ( X )= r t p t ] ,
x E f?(r(L2(%H))),
(7)
for t 2 0. Denote the image of Et by
At] = E t ( B ( r ( L Y R + , H ) ) ) ) . These are exactly the operators on the Fock space which are R-adapted in the sense of B e l t ~ n The . ~ ~algebra ~ dtj consists of all bounded oper) R C r ( L 2 ( [ 0 , t ] , H )&I ) ators that leave the subspace r ( L 2 ( [ 0 , t ] , H )&I r ( L 2 ( [ t , w [ , H ) FZ ) r ( L 2 ( & , H ) ) invariant and vanish on its orthogonal complement. The conditional expectations have the following properties.
Lemma 5.1.
(a) (b) (c) (d)
E, o Et = E, = Et o E,, for 0 5 s 5 t , and in particular E; = Et. Et i s completely positive for all t 2 0. E t ( X Y 2 ) = X E t ( Y ) Z f o r a l l X , Z E .At], Y Ef?(I‘(L2(R+,H))). Et(id) = rt]for all t 2 0.
Let (kt)t>O - be the quantum stochastic process on 421 defined by
Using functional calculus we extend (kt)t>o to cb(R) by k t ( f ) = f(xt)rt]. Denote by dt the algebra generated by I$ and Xt, i.e. dt = k t ( q z ] ) and by at =
{rt]f(xt)lf E cb(R))
the image of the extension of kt to cb(]w). We have of course
-48
2 d, 2 .At]
for all 0 <_ s <_ t .
Theorem 5.1. The monotone Ldvy process is Markovian, i.e. we have
&(kdP))
E
and
&(kt(f)) E
for all 0 5 s 5 t , P E C[4, and f E cb(R).
&I
52
The semigroup of Markovian transition opemtors Tt : Cb(R) + Cb(R) with & ( k t ( f ) ) = k,(Tt-s(f)) for 0 5 s 5 t , f E Cb(R) is given by T t f ( 4=
f(Y)dPt,&),
(8)
- maps for t 1 0, f E Cb(R), where pt,= = 6, D pt. The semigroup (Tt)t>o Cb (R) to itself and polynomials to polynomials. Proof. The computation of Tt-, is the same as in the proof of Proposition 3.2 and Corollary 3.1, just write X t as a sum X t = X, ( X t - X , ) of two monotonically independent self-adjoint operators. The formula in Proposition 3.2 can be interpreted as a generating function and shows that polynomials are mapped t o polynomials. On then other hand, equation ( 8 ) implies that Ttf is again in Cb(R),due to the continuity CI of the monotone convolution.
+
As in Ref. 5, 6, the Markov property implies the existence of a classical version. Corollary 5.1. There exists a cZasstcal Markov process (At)t>o - on P that
has the same time-ordered joint expectations as (Xt)t>o, - i.e.
E( f l ( X t , ) for all n E N,0 5 tl 5
* * *
fn(&
5 t,,
1) = q
f l
f l , . ..,fn
( X t 1) * * * f n ( X t n ) )
E Ca(P).
We have the following expression for the generator of the semigroup
(Tt)t>o. Proposition 5.1. Let (X&o be the monotone Ldvy process whose monotone convolution semigroup has characteristic pair (a,p ) . Then the generator
of the associated semigroup of transition opemtors is given by
for f E C,"(P).
53
Proof. We use f z ( x ) =
as a generating function. We have
and therefore
Using a partial fraction decomposition, this becomes
For sufficiently large 121, the series converge uniformly on a bounded interval containing suppp. Therefore we can interchange summation and integration and deduce that the formula given in the proposition holds for polynomials. Since p has compact support, the formula extends to functions in C,Z(R) by a StoneWeierstrass type approximation. 0 6. Examples 6.1. Pure drift process
The simplest case is the monotone LBvy process associated to the characteristic pair (a,O),a E I[$. We get A ( z ) = a, H t ( z ) = z at, and pt = -at. The associated monotone LBvy process is just Xt = -aid.
-
6.2. Monotone Brownian motion
Consider now (0,do). The we get A(z) = -$, H t ( z ) = d m 2 and , pi is absolutely continuous w.r.t. to Lebesgue measure, with density
n dl m1
2 O.
] - a , a [ 9
The generator of the semigroup (Tt)t>O - is given by f (2) -f(O)
Lf (4 = ) f ” ( O ) z ’ on f E C,”(R).
-2
f’( 2 )
if x # 0, if x = 0,
54
The process 2 0 t is equal to the quantum AzBma rnartingale2l with parameter q = 0, cf. Remark 4.9 in Ref. 10. The classical version is the classical AzBma martingale.8 The monotone LBvy process with characteristic pair (a,&), a E R, is the monotone analogue of a Brownian motion with drift -a. We get A(z) = a - $. The reciprocal Cauchy transform is given as the unique solution of H&) = w in C+ of aw - 1 a(w - a ) +In -= a2t. aa - 1 The generator of the semigroup (Tt)t?o is given by f(z)-f(O)-z(l+oz)f'(z~
if x # 0, if x = 0,
on f E C,"(R).
6.3. Monotone Poisson process Let now (a,p) = (-$,$&) with A > 0. Then we have A(z) = $& and &(a) = w is the unique solution in C+ of - - x( w - z ) - - l n -x- w - 1 - t. 2 2 z-1 The corresponding probability measures where determined in Ref.18, where they arose as limit distributions in a Poisson-type limit theorem, see also Example 4.4.(3) in Ref. 17. The monotone LBvy process associated to this characteristic pair is the monotone Poisson process. The generator of its Markov semigroup (Tt)t>O - is given by
for,f E q ( R ) . 7. Martingales
In this section we show how one can construct a family of martingales from a monotone LBvy process that is analogous t o the family of exponential martingales of a classical LBvy process. If (Xt)t>o - is a classical Ldvy process with characteristic functions cpt(u) = E(eiuXt),then for any u E R, the - with process (MF)t>o
M,U = exp (iuXt - logcpt(u)),
for t
2 0,
55
is a martingale w.r.t. the filtration of (Xt)t>o. - An analogous family of martingales for free increment processes has been defined by Biane, see Section 4.3 in Ref. 5. Definition 7.1. Let H be a Hilbert space, r ( L 2 ( R + , H ) the ) Fock space over L2(R+, H ) , and (Et)tlo the family of conditional expectations introduced in (7). We call a family ( M t ) t E I of operators on I'(L2(Et+,H ) ) indexed by an interval I C R+ a martingale, if EB(Mt) = ME
holds for all s , t E I with s 5 t. L e m m a 7.1. Let (pt)tlo be a monotone convolution semigroup of probability measures with reciprocal Cauchy transforms ( H t ) t l o . Then the Ht are injective on UY for aZZ t 2 0. Proof. If pt = a0 for all t 2 0, then Ht = id for all t 2 0 and the lemma is true. Assume now that (pt ) is a non-trivial monotone convolution semigroup. Let t 2 0 and 2 1 , E~ Ht(U?) ~ with Ht(z1) = Ht(z2). Then we have
where w = Ht(zj), i = 1,2,and A ( z ) denotes the generator of (Ht)t>O,see Theorem 4.7 in Ref. 17 (or also Theorem 2.2 for the case where the-pt are compactly supported). This implies
and 21 = Ho(z1)= 2 2 by the uniqueness of the solution in Theorem 4.7.(2) in Ref. 17 or Theorem 2.2(b). 0 Theorem 7.1. Let T > 0, and let (pt)t>o be a monotone convolution semigroup of compactly supported probability measures with reciprocal Cauchy tmnsforms (Ht)t>O and monotone Lkvy process (Xt)tzo. Then for any x E HT(UY), theperator process ( M t ) o l t l T with
is a martingale.
56
Proof. M[ is well-defined, since we have H T ( C + )E Ht(C+) for 0 5 t 5 T by the semigroup property of (Ht)t>o - and since the Ht are injective by Lemma 7.1. Let now 0 5 s 5 t 5 T, then we have
6.Proposition 3.2. References 1. L. Accardi, A. Ben Ghorbal, and N. Obata. Monotone independence, comb graphs and BoseEinstein condensation. EMAU Greifkwald Preprint-Reihe Mathematik 3712003, 2003. 2. S. Attal. The structure of the quantum semimartingale algebras. J. Operator Theory, 46(2):391410, 2001. 3. A.C.R. Belton. Quantum hemimartingales and stochastic evolutions. J. f i n c t . Anal., 187(1):94-109, 2001. 4. A.C.R. Belton. An isomorphism of quantum semimartingale algebras. preprint, 2003. 5. P. Biane. Processes with free increments. Math. Z., 227(1):143-174, 1998. 6. M. Boiejko, B. Kiimmerer, and R. Speicher. q-Gaussian processes: Noncommutative and classical aspects. Commun. Math. Phys., 185(1):129-154, 1997. 7 . E. Berkson and H. Porta. Semigroups of analytic funct.ions and composition operators. Michigan Math. J., 25(1):101-115, 1978. 8. M. Emery. On the Az6ma martingales. In Sdminaire de Pmbabilitds XXIII, Lecture Notes in Math., Vol. 1372, Springer-Verlag, Berlin, 1989. 9. U. Franz. Monotone independence is associative. Infin. Damens. Anal. Quantum Probab. Relat. Top., 4(3):401407, 2001. 10. U. Franz. Unification of boolean, monotone, anti-monotone, and tensor independence and L6vy process. Math. Z., 243(4):779-816, 2003. 11. U. Franz, M. Schiirmann, and M. Skeide, 2003. in preparation. 12. V. V. Goryahov. Fractional iteration of probability-generating functions and the embedding of discrete branching processes into continuous ones. Mat. Sb., 184( 5):55-74,1993. 13. R. L.Hudson and K. R. Parthasarathy. Quantum Ito's formula and stochastic evolutions. Comm. Math. Phys., 93(3):301-323, 1984. 14. Y.G. Lu. An interacting free Fock space and the ascsine law. Pmbab. Math. Statist., 17(1): 149-166, 1997.
57
15. N. Muraki. A new example of noncommutative "de Moivre-Laplace theorem". In Probability theory and mathematical statistics (Tokyo, 1995), pages 353362. World Sci. Publishing, River Edge, NJ, 1996. 16. N. Muraki. Noncommutative Brownian motion in monotone Fock space. Comm. Math. Phys., 183(3):557-570, 1997. 17. N. Muraki. Monotonic convolution and monotone LBvy-HinEin formula. preprint, 2000. 18. N. Muraki. Monotone independence, monotone central limit theorem and monotonic law of small numbers. Inf. Dim. Anal., quant. probab. and rel. fields, 4(1):39-58, 2001. 19. N. Muraki.The five independences as natural products. EMAU Greifswald Preprint-Reihe Mathematik 3/2002,2002. 20. N. Muraki. The five independences as quasi-universal products. Inf. Dim. Anal., quant. probab. and rel. fields, 5(1):113-134, 2002. 21. K.R. Parthasarathy. Adma martingales and quantum stochastic calculus. In R.R. Bahadur, editor, Proc. R.C. Bose Memorial Symposium, pages 551-569. Wiley Eastern, 1990. 22. K.R. Parthasarathy. An Introduction to Quantum Stochastic Calculus. Birkhauser, 1992. 23. M. Schiirmann. White Noise on Bialgebms, Lecture Notes in Math., Vol. 1544, Springer-Verlag, Berlin, 1993. 24. M. Schiirmann. Non-commutative probability on algebraic structures. In H. Heyer, editor, Proceedings of X I Oberwolfach Conference on Probability Measures on Groups and Related Structures, pages 332-356. World Scientific, 1995. 25. R. Speicher and R. Woroudi. Boolean convolution. In D. Voiculescu, editor, R.ee probability theory. Papers from a workshop on random matrices and opemtor algebm free products, Toronto, Canada, March 1995, Fields Inst. Commun., Vol. 12, pages 267-279. American Mathematical Society, Providence, RI, 1997. 26. D. Voiculescu, K.Dykema, and A. Nica. Rve Random Variables. American Mathematical Society, Providence, RI, 1992. 27. D. Voiculescu, editor. R.ee probability theory. Papers from a workshop on mndom matrices and opemtor algebra free products, Toronto, Canada, March 1995, Fields Inst. Commun., Vol. 12, American Mathematical Society, Providence, RI, 1997.
This page intentionally left blank
Infinite Dimensional Harmonic Analysis I11 (pp. 59-70) Eds. H. Heyer et ol. @ 2005 World Scientific Publishing Co.
GEOMETRIC PROPERTIES OF THE SET OF EXTENSIONS . OF A STOCHASTIC MATRIX
ROLF GOHM EMA U Greifswald, Institut fir Mathematik und Informatik Jahnstr. 154 17489 Greifswald, Germany E-mail: gohmOuni- greifsvdd.de
We consider stochastic matrices as restrictions of unital completely positive maps to diagonal subalgebras. The corresponding extensions of a stochastic matrix are classified by certain arrays of functionals and by positive dehite kernels.
Introduction The subject of extensions of completely positive maps given on operator algebras to the algebra of all bounded operators on a Hilbert space was first studied by Arveson l. For a survey see also the corresponding chapter in Effros/Ruan 4 . Recent work of Gohm establishes some applications in quantum probability and indicates a way to more concrete descriptions by a duality with dilation theory. Many interesting questions arise at this point, and in this paper we provide some playground by explicitly describing the most elementary case, namely that of stochastic matrices. This is also of interest for its own sake because embeddings of classical Markov chains into noncommutative ones are a natural topic for quantum probabilists. Surprisingly it turns out that a sort of geometric reasoning is appropiate here which has a natural starting point in Arveson's notion of a metric operator space We review it in Section 1 and give an alternative formulation by representing functionals. In Section 2 we define realizations of stochastic matrices and show that realizations are representing functionals which classify the extensions to normal unital completely positive maps on all bounded operators. The geometric character becomes explicit by a bijective correspondence with certain positive definite kernels. In Section 3 we include measures and states into the extension procedure. This is the setting of Gohm 5i6 and we show how the duality theory 596
'.
60
between extensions and dilations can be described very concretely in terms of realizations. Again there is a geometric picture given by certain positive definite kernels. A remarkable conclusion tells us that the extension set essentially only depends on the pattern of zeroes of the stochastic matrix. Details on that can be found in the remarks at the end of the paper. The following notation will be used invariably throughout the paper. P, G,%! denote (complex, separable) Hilbert spaces with ONBs { ~ k } f = , , {bj G }jn=,, {bp}gl (the superscripts will be omitted from now on). d, n, m may be finite or 00. Operators are identified with matrices relative to these ONBs. Inner products are linear in the second variable.
1. Metric operator spaces and their duals
We start with the objects given at the end of the introduction. Now assume that { a k } b l c B(G,%!)(bounded linear maps from G to %!) satisfy ~ ~ u <~00
xi=,
xi=,
xi=,
xi=,
xi=,
aij :
P -b c,
11c112
<
I+
<
(&,U(<)bj).
When the inner product of P is transferred from P to P" the dual basis
61
of {Ek)k=l becomes an ONB of P*. With a; := u i j ( e k ) = (6i,a&) we get the formula {Ek):=l
A
xi=,
For d = 00, note that Ila;ijlz < 00 implies that (a?):=, E Z2, i.e., the sum above is convergent. The inner product (a", ai'i') in P* is nothing i'j' d but the standard inner product of the vectors (a:):=, and (ak )k=l in Cd or P. The linear independence of the a k implies that the aij (i = 1 , . . . ,m and j = 1 , . . .,n ) are total in P*. In fact, if b = Cf=,&ek is orthogonal to all aij then
xk=l
i.e., d Pkak = 0 and thus P k = 0 for all k and finally b = 0. The argument can be reversed. For d < 00 we can start with a Hilbert space P and any total array aij E P* (i = 1,. .. , m and j = 1,. .. , n ) and construct backwards a metric operator space from that. For d = 00 the property Ila;cllz < 00 is an additional assumption. This motivates the following
'&
Definition 1.1. An array (aij)i,j of vectors in a Hilbert space P* is said to represent a metric operator space or the corresponding map if it is total in P* and if for an ONB { c k ) in the dual P the formula (&,a&) := a i j ( e k ) defines operators {uk}$=, C B ( G , X ) such that C$, lla;c112 < 00 for all E 3t.
c
One can check that this does not depend on the choice of the ONB { e k } in P. A representing array always represents a normal completely positive map 2 : B(G) + a(%) via the corresponding minimal Kraus representation. From the uniqueness result for the a k quoted above we conclude immediately that two arrays ( a i j )and ( b i j ) represent the same normal completely positive map 2 : B(G) + B(3t) if and only if there is a unitary u E B(P*) such that bii = uaij for all i , j . If we say that a geometric property is something that does not change under unitary transformations then an interesting way t o think about normal completely positive maps geometrically consists in looking at a representing array ( a i j )in P*. This tool has been used ad hoc and in special cases ( n = m = 2) by Kummerer and by Gohm and will be used more systematically in the following.
62
2. Realizations of stochastic matrices
By a stochastic m x n-matrix we mean a matrix with nonnegative entries with all row sums equal to one (convergent if n = 00). The following definition prepares the study of extensions of stochastic matrices.
Definition 2.1. If an array ( d j ) (i = 1 , . . . ,m and j = 1,. ..,n) of vectors in a Hilbert space P* satisfies (a) (b)
Tij .ij
= lluij112
1ai'j if i # if
for all i , j , then we say that T = (Tij).
(dj)
is a realization of the m x n-matrix
Realizations always exist. For example, we may choose all vectors uij orthogonal to each other with length according t o (a). Note that if u is unitary or antiunitary on P*,then (a") realizes the same matrix as (ua"). To classify realizations of a matrix T geometrically we can proceed as follows. Given a matrix T with nonnegative entries let Q j (with j = 1,...,n) be a Hilbert space spanned by vectors &j (i = 1 , . ..,m) satisfying Tij = ll
for all (C)with finite support. See for example Constantinescu for details on positive definite kernels. The properties of the &j are chosen in such a way that
extends to an isometry and we define K ( r ,s) := v,' v,. Then K is positive definite and satisfies K ( r , r ) = I for all r. If u E B(P*)is unitary then ( u d ) has the same associated kernel, i.e., we have indeed constructed a geometric property. We call K the correlation kernel of the realization (a"i).
63
Theorem 2.1. Let T be a stochastic m x n-matrix. Every realization of T represents Q normal unital completely positive map. There is a bijective cor~espondencebetween (1) unitary equivalence classes of realizations of T
(2) normal unital completely positive maps Z : B(G) + B(3t) which map the diagonal into the diagonal such that the restricted map is given by T (3) positive definite kernels K with K ( r ,s) E B(Q8,Q T )and K(r,r ) = I I f o r r , s E { l , ..., n}. If ( d j ) is a realization of T , then Z is the map represented by it and K is the associated correlation kernel.
Proof Suppose that (a") is a realization of T in P*. Choose an ONB { ~ k } g = of ~ P and define a? := a i j ( e k ) . On = CzlXiSi E 3t (with (Xi) E Z2 if m = 00) we can define bounded operators
<
(in particular
(%,akSj)
k
i,j
k
j
= a t ) . In fact, we have k,j
i
i,j
j,k
i
i
i
i.e., llai<112= 11<112 for all <,which characterizes a metric operator space corresponding t o a normal unital completely positive map, see Section 1. Let Z : B(G) + a(%)be a normal completely positive map, given by Z ( x )= akxa; with {ak};=, c B ( 6 , X ) . We compute its action on a diagonal matrix x with diagonal entries z1,.. .,zn:
Ck
k=l
k,j
j
Here we used the description of Z by the representing array (a") in P*. Thus (ai'j, a") = 0 for i # 'i is equivalent to the fact that 2 maps diagonal matrices into diagonal matrices. In this case the restricted map is given ~ ~ different ~ 2 . arrays by a matrix T with entries Tij = (aij,,aij)= ~ ~ u iThat represent the same map 2 iff they are unitarily equivalent has been noted
64
in the previous section. We conclude that (1) and (2) are related by a bijection. Given (l),it has already been shown above how to associate the correlation kernel K in (3) which only depends on the unitary equivalence class. Conversely, given a positive definite kernel K as in (3), there is a minimal Kolmogorov decomposition for it, i.e., there exists a Hilbert space Q and v j E B ( Q j , Q) (for j = 1,. .,n) such that K ( r ,s) := v: v8. Because K ( r ,r ) = 1 for all T the v j are isometries. Then putting aij := v j (&) for all i,j we can identify Q with P* and obtain a realization (aij) of T with correlation kernel K . A minimal Kolmogorov decomposition is unique up to unitary equivalence which implies that two arrays ( a i j ) and (bij) obtained from the same kernel are unitarily equivalent. Thus we have also a bijection between (1) and (3). 0
.
Theorem 2.1 characterizes a certain set in three different ways. It shows that realizations of stochastic matrices classify certain extensions of them and that this can be mapped bijectively onto a well known set of kernels. We mention two structures on this set which are now immediate. First, with the natural affine structure from (2) or (3), it is a convex set. Second, observe that antiunitary transformations on P* in general change the unitary equivalence class in (1). Explicitly, if aij = Ct==, a f e k for an Eyek (complex conjugation) may be called a ONB {ek} then ?$ = conjugate realization of T . Because two antiunitaries differ by a unitary, the unitary equivalence class obtained in this way does not depend on the choice of the ONB {ek} and thus we have a canonical antilinear inversion on our convex set which we call conjugation. The conjugate extension is where on the r.h.s. we have complex conjugation of given by z(z)= all matrix entries. Similarly we also have a conjugate kernel
xi=,
z
m.
r.
3. Extension of states included For probabilistic applications it is essential to include states into the extension procedure of the previous section, see Gohm Assume now that T is a stochastic m x n-matrix and let v = ( ~ j ) j ” =and ~ p = (pj)fl be probability measures on (1,. .. ,n } and (1,. ..,m} such that all U j and pi are nonzero and we have p o T = Y . Such probability measures can be found iff T has at least one nonzero entry in each column. So this is an implicit assumption on T from now on. We have a dual stochastic n x m-matrix
‘.
65
TI = (T;J given by the equation piTij = ujTji for all i, j .
Then Y o TI = p and (TI)' = T . We can think of the Hilbert spaces 4 and X as the GNS- Hilbert spaces for the (commutative von Neumann) algebras A and B of bounded functions on { 1, .. .,n } and { 1, .. .,m} for the states induced by Y and p . Note that T : A + B and TI : B + A. Let us write R, E 4 and R, E X for the corresponding cyclic vectors. Explicitly, we have R, = C;=,f i 6 j and RP -- ELl &&, where the basis vectors & , S j in X , 4 are realized by characteristic functions of single points which are also elements in B,A and denoted as such in the same way. This should not cause confusion. Now let (a") (i = 1 , . ..,m and j = 1 , . . . , n ) be a realization of T and let Z : B ( 4 ) + B(X)with Z(z) = a k z a ; be the corresponding completely positive extension, according to Theorem 2.1.
xi=,
Lemma 3.1. The following assertions are equivalent: (1) (a,, f l u ) = (aP, Z(z) a,) for all z E B ( 4 ) (2) a: R, = z k R, for some w 1 , . . , w d E C.
.
(3) There exists a unit vector R* =
wk
ek E P" satisfying
Proof: These elementary equivalences are shown in a more general context in Gohm ', A.5. To relate it to our setting we prove the equality d
n
-
for all j , where the w k satisfy (2). In fact, from n
we obtain
and hence k
i,k
66
For the following considerations it is convenient to introduce vectors aij E P by
(aij,<) = a”(<) for all E P, i.e., aij = ‘&ii?~k. Let us say that (dj)is a p-realization of T if the conditions in Lemma 3.1 are satisfied. From now on we assume this. Then we also have a unit vector R E P given by $2:= Cy’l
6.j.
Proposition 3.1. Let (a”) be a p-realization of T . Then the map
J :B
+A@B(P),
is a normal *-homomorphism such that ( 1 d@ ~ (R, * R))
0
J = TI.
We have used von Neumann tensor products and Dirac notation for rank one operators. In the language of Gohm Proposition 3.1 tells us that J is a weak tensor dilation of TI. Proof: Because aij I ai’j for i # i‘, we find that are onedimensional projections which are orthogonal for different i . Thus, writing J(b) = C;=, S j €3 Jj(b), we see that the Jj are representations of B on P. We conclude that J is a normal *-homomorphism. Further we get
i.e., ( 1 d ~ (R,.R)) @ o J = TI. 0 Proposition 3.1 is a concrete instance of a much more general phenomenon revealed in Gohm compare also Gohm/Skeide ’. Namely, whenever T is a normal unital completely positive map between von Neumann algebras A and 23, there is a correspondence between extensions of T respecting some previously fixed states and weak tensor dilations of the dual map TI. In our setting, where T is a stochastic matrix, starting from the weak tensor dilation J above we can construct an associated isometry 596,
: 31 3
6 @P, bRp C) J(b)(R, @ 0) (b B),
67
explicitly: G v ( 6 i ) = v(diR,) = simple formula
'&f i d j
€9 & a i j 7
i.e., we get the
n
j=1
With
aij
=
cf=,Zfek
this can be written as d
v6i = c
d j j,k
@afek = c a f d j €9 Ek = ca;(di)
c;=,
j,k
i.e. v 5 = a;(E) C 3 € k for all 2 of T can be written as
@ek,
k= 1
E 3t.
Thus the corresponding extension
c d
Z(2) = v*
2 €9
12)=
U k 2 a;,
k=l
and we have reconstructed the extension from the dilation. Our emphasis here is on the fact that both extensions and dilations of stochastic matrices are very conveniently described in terms of realizations. With a suitable equivalence relation on weak tensor dilations discussed in Gohm 6 , 1.4, this correspondence between extensions of T (respecting fixed states) and (equivalence classes of) weak tensor dilations of T' becomes a bijection. It was noted in Gohm 5*6 that the set 2 characterized in this way is convex and closed in suitable topologies. In our setting, with T a stochastic matrix, we can combine Theorem 2.1 and Lemma 3.1 to see that the set 2 = Z ( T , p ) can be described geometrically as the set of unitary equivalence classes of -realizations of T , i.e., realizations ( a i j ) of T in P* for which Cr=l&a"(= a*) does not depend on j . We notice that on 2 there is a canonical conjugation, inherited from the one discussed in Section 2. A description of 2 by kernels can be obtained as follows. We start with the Hilbert spaces Q j ( j = 1 , . . ., n ) defined in Section 2. They are spanned by vectors &j (i = 1,...,m)satisfying Tij = ll&j112 and &j I< i j j if i # i'. Since T satisfies POT= v,there are unit vectors R j E Q j given by R j := ELl &&j. Because now (a") is a p-realization, the isometries V j : Q j + P*,
68
Then KO is a positive definite kernel with K"(r,r) = I for all r which we call the restricted correlation kernel. If we take another p-realization ( b i j ) then because bij = uuij for a unitary u E B(P*)the restricted correlation kernel is not changed and it represents therefore a geometric property. Proposition 3.2. There is a bijective correspondence between the set 2 = 2 ( T ,p ) and the set of positive definite kernels KO with Ko(r,8 ) E B(Q:, Q:) and Ko(r,r) = 1 for all r , s E (1,.. .,n}. It is given by the restricted correlation kernel.
Proof: We only have to modify the proof of Theorem 2.1 in a suitable way. We have already seen how to produce a restricted correlation kernel KO from a p-realization corresponding to 2 E 2. Conversely, g'wen a kernel KO, we use it to agglutinate the spaces Q j in such a way that all the Rj become a single vector R*.
In detail, using a minimal Kolmogorov decomposition for the kernel KO, we get a Hilbert space Q" and isometries v! : f?(QY,Qo) such that Ko(r,8 ) = (v,")*vf for all r, s E (1,. ..,n}. We extend them to isometries vj E f?(Qj, Q0 @ CR*)by v j R j ;= R* for all j,where R* is a unit vector. Then it is easily checked that ua3:= v j t j for all i , j and a corresponding identification of Qo @ CO* with P* defines a p-realization of T , and KO is
69
the corresponding restricted correlation kernel. By the uniqueness property of minimal Kolmogorov decompositions two p-realizations arising from the same kernel are unitarily equivalent and belong to the same 2 E 2. Hence 2 I+ KO is a bijection. 0 Let us finish with some remarks which should help to put these results into perspective. To determine the kernels in Proposition 3.2 the only relevant information about the spaces QY is their dimension which equals the number of nonzero entries in the j-th column of T minus one. Of course, in the case where this dimension is zero the corresponding entries of KO also have to be zero. There is always a distinguished solution given by a kernel KO with K 0 ( r , s )= 0 for T # s. When in T there are less than two columns with two or more nonzero entries then this is the only solution. The fact that 2 is always nonempty has been observed in a more general context in Gohm/Skeide 7. The correspondence established in Proposition 3.2 yields the surprising fact that the set 2 depends only rather slightly on the details of T and T'. More precisely, though obviously the extensions 2 : B(G) + a(%)depend on all entries Tij and on the measures p and Y (or on T and T'), we have shown that whenever for two stochastic m x n-matrices TI and T2 the number of nonzero entries in the j-th column coincides for each j then there is a canonical bijection between the extension sets 21 and 2 2 , via the correspondence to kernels given in Proposition 3.2. Interpreting 2 as the set of (equivalence classes of) weak tensor dilations of T', see Proposition 3.1, we can give the following interpretation of the restricted correlation kernel. Using the notation introduced in Proposition 3.1 and its proof, we can consider the relative position of the commutative algebras J j ( B ) in B ( P ) ( j = 1,. . ., n ) by looking at the corresponding subspaces Jj(B)R c P. The formula : Si I+ Cy=,Sj @ aij for the associated isometry makes clear that it is this relative position that is described by the (restricted) correlation kernel.
It is very instructive to check all these observations in the simple example of 2 x 2-stochastic matrices with nonzero entries. Then n = m = 2
70
and the kernels KO are C-valued and parametrized by the off-diagonal entry c := K0(2,1), i.e., by a complex parameter c with Icl 5 1. This parametrization of Z has already been observed in G o b 6, where also the corresponding extensions 2 : M2 -+ M2 and dilations J : C2+ C218 B(P) are explicitly computed. References 1. W. Arveson, Subalgebras of C*-algebras. Acta Math., 123, 141-224 (1969) 2. W. Arveson, Noncommutative Dynamics and E!-Semigroups. Springer Monographs in Mathematics (2003) 3. T. Constantinescu, Schur Parameters, Factorization and Dilation Problems. Birkhauser (1996) 4. E.G. Efhos, Z.J.Ruan, Operator spaces. London Mathematical Society Monographs, New Series 23, Oxford Clarendon Press (2000) 5. R. Gohm, A duality between extension and dilation, in: Advances in Quantum Dynamics, AMS, ed.: G. Price (2004) 6. R. Gohm, Elements of a spatial theory for non-commutative stationary processes with discrete time index. Habilitation Thesis (2002). A revised version will appear under the title ‘Noncommutative Stationary Processes’ in the Lect. Notes Math.-Series of the Springer-Verlag. 7. R. Gohm, M. Skeide, Normal CP-maps admit weak tensor dilations. Preprint Greifswald (2003), submitted to: Inf. Dim. Analysis, Quantum Probability and Related Topics 8. B.Kiimmerer, Markou dilations on the 2 x 2 matrices. Operator algebras and their connections with topology and ergodic theory, Proc. Conf., Busteni/Rom. 1983, Lect. Notes Math. 1132, 312-323 (1985).
Infinite Dimensional Harmonic Analysis I11 (pp. 71-81) Heyer et al. @ 2005 World Scientific Publishing Co. Eds. H.
H E A T K E R N E L ANALYSIS O N INFINITE D I M E N S I O N A L GROUPS MARIA GORDINA ABSTRACT.The paper gives an overview of our previous results concerning heat kernel measures in infinite dimensions. We give a history of the subject first, and then describe the construction of heat kernel measure for a class of infinite-dimensional groups. The main tool we use is the theory of stochastic differential equations in infinite dimensions. We provide examples of groups to which our results can be applied. The case of finite-dimensional matrix groups is included as a particular case.
1. MOTIVATION AND
HISTORY
In this paper we review our results in [7],[8],[9],[ll] and show how they fit into a broader picture. In order to see the challenges this study presents, we first review what is known in finite dimensions and for the flat infinite-dimensional case. Our main results concern analogues of heat kernel (Gaussian) measures on infinitedimensional groups. 1.1. Finite-dimensional noncommutative case. The initial motivation for our work was the following result for finite-dimensional Lie groups. It has a long history which we address later in this section. Let G be a finite-dimensional connected complex Lie group with a Lie algebra 0 . The Lie algebra is a complex Hilbert space an orthonormal basis of g as a real with a norm I . I, and we denote by vector space. By identifying 0 with left-invariant vector fields at the identity e we can define the derivative
and the Laplacian
af = Ca:f. The heat kernel measure ,LL~ has the heat kernel as its density with respect to a Haar measure dz. The heat kernel is the fundamental solution to the heat equation --
at
- a A p t ( z ) , t> 0,z E G,
'Research supported by the NSF Grant DMS-0306468. Date: January 30, 2004. Key word9 and phrases. Heat kernel analysis, infinitedimensional Lie goups and Lie algebras, von Neumann algebras, noncommutative LP-space, Fock space.
72 We can define holomorphic functions on G as functions whose derivatives are complex linear. The kth derivative Dkf(e) is defined as the tensor (@f(g),Jil 8 ... @ti,) = (aia*..aikf)(g)* Denote by X L 2 ( G , p t ) the space of holomorphic functions which are squareintegrable with respect to pt. It is known to be a closed subspace. The following identity represents an isometry between X L 2 ( G ,p i ) and a certain space of tensors over the Lie algebra g.
This identity was first proved by B. Driver, L. Gross for connected complex Lie groups in [5). In particular, if G = Cn,then the tensors Dkf are symmetric, and with this norm they form a bosonic Fock space (see V. Bargmann, I. Segal in [l], P I 1 W1). Our goal is to describe similar results for some infinite-dimensional groups. As we will see later, the approach described above is not applicable in that situation, but we will make use of the following probabilistic approach. Assume that G is a matrix group. Then we can make sense of multiplying elements from G and g. Instead of using the heat equation to define the heat kernel measure, we can view this measure pt as the transition probability of the Wiener process X t on G satisfying the stochast,ic different,ial equation d X t = XtdWt,Xo = e, 2n
where Wt =
C W,"&is the Wiener
process on g and W i are independent one-
i=l
dimensional Wiener processes. 1.2. Finite-dimensional commutative case. The identity in Equation 1.1has a long history. It appeared in the commutative setting (G = Cn)as follows. Consider a system of n classical harmonic oscillators. They satisfy Newton's equation in Wn m.x'! 3 3 = -kjxj, j = 1, ...n.
The corresponding quantum state space for the system is L2(Wn,dz) with the Hamiltonian 1 1 -at2 1~1~ 2 Then the ground state for this Hamiltonian is 4j =
--A
e-Ix12/4t
lp=- ( 2 4 4 4
and the heat kernel pt satisfies
+
. n
39 = p
73 This works if n < 00, but not in infinite dimensions where there is no analogue of the Lebesgue measure. However, if the Lebesgue measure is replaced by the Gaussian measure, we can extend the setting to infinite dimensions. Let us first explain how it can be done in the finite dimensional commutative case. -I.IZ/t Let dpt = ptdx, dpt = +dz be the Gaussian measures respectively on W n denote a bosonic Fock space, where H = Cn is and C". By 5, = a state space of a single particle. As before 3CL2(Cn,pt) is the space of square integrable holomorphic functions. In this situation we can describe the isometries between the three spaces as follows. Let Hk(x) be an Hermite polynomial, then Hkl (xjl) . ... . Hk, (xj,) form an orthonormal basis of L2(Rn,p t ) , and thus
B (Hkl (Zjl). ... . Hkn(Zjn)) = Z;; 8,... 8 s xcjn kn St (Hkl (Zj,). ... . Hk,
T (z;:
+
(Zcj,))
... . z;:)
. ... . Z;: =z ; : 8,...@I, xjn kn
= 2;:
The Segal-Bargmann map St has been studied recently in different settings ([14], [15], [IS]). We concentrate on the Taylor map T since the Segal-Bargmann transform for a finite-dimensional Lie group K@requires the inner product on the Lie algebra to be AdK-invariant. The following diagram illustrates the maps described above including the ground state transformation, the Segal-Bargmann transform and the Taylor map.
L2(W", dZ)
L2(W",Pt)
The Taylor map can be alternatively described as follows. First, we can identify the kth derivative D kof a function f with symmetric tensors by
where {e,} is an 0.n. basis of Cn.Then Tf = satisfies the identity
ckD k f ( 0 )is in Zs,is surjective and
This identity is a commutative version of Equation 1.1.
74 1.3. Infinite-dimensional commutative case. Let ( B ,H C M ,p ) be a complex Wiener space, where B is a complex separable Banach space, HCM is the CameronMartin subspace (a complex separable Hilbert space which is densely and continuously embedded in B ) , and p is a Gaussian measure defined by its characteristic functional
I'
e w ( z ) d p ( z ) = e-+llqll&M
B
, 'p E B* C HGM.
By using weak Cauchy-Riemann equations 1.Shigekawa and HSugita in [19], [22], [23] introduced holomorphic functions on ( B , p ) . In this case the space of holomorphic L2( B ,p ) is a collection of L 2 ( B ,p)-equivalency classes rather than functions as it was for finite-dimensional Lie groups. Still for each f E X L 2 ( B ,p ) there exists a uniquely defined regular version (holomorphic skeleton) off on H C M . The following pointwise estimate explains why the holomorphic skeleton is defined only on the Cameron-Martin subspace HCM which is of p-measure 0.
lf(h)I2 < h E HCM. We will have similar pointwise estimates for infinite-dimensional groups in Section 5 . 1.4. N e w features in the infinite-dimensional noncommutative case. Let us list the problems that arise when we try to prove analogous results for infinitedimensional groups. First of all, we need a new construction to define a heat kernel measure. For the groups we consider there is no Haar measure. Moreover, the PDE theory is not developed enough to introduce a heat kernel as the fundamental solution to the heat equation. To deal with this problem we use the probabilistic approach described for finite-dimensional Lie groups earlier in this section and the theory of stochastic different,ial equations in infinite dimensions. This allows us to construct a heat kernel measure. Another problem is how to define holomorphic functions. As we showed in [7], [8]and [9], there might be no nonconstant square integrable holomorphic functions. This question is related to another interesting problem which is new in infinite dimensions. We have a choice of natural Lie algebras, and this choice depends on the norm on this infinite-dimensional Lie algebra. By changing the norm, we change the Lie algebra as a set making it larger or smaller. At the same time, the Lie algebra determines the Laplacian, which in turn defines the heat kernel measure. Note that the Lie algebra is also the set of directions of differentiation for holomorphic functions. Thus if we choose a larger Lie algebra, the space of holomorphic functions becomes smaller, and in some cases the space of holomorphic square-integrable functions is trivial. We also introduce an analogue of the Cameron-Martin subspace and holomorphic skeletons. The Cameron-Martin group can be described as the image of the infinitedimensional Lie algebra under the exponential map. This image has measure 0, but this is the set where holomorphic skeletons live.
2. NON-COMMUTATIVE LP-SPACES
B ( H ) is a von Neumann algebra on a (separable) Hilbert space Suppose M H, that is, M is a C*-algebra which is weakly closed and contains the identity operator I . Let T be a faithful normal semifinite trace on M . This means that 7
75 is a functional on the positive elements of M satisfying the following properties for any A, B E M + (1) T(A B ) = T(A) T ( B ) , (2) T(cA) = CT(A)for any 0 < c < 00, T(AB) = T(BA), (3) T is faithful means that T(A) = 0 implies A = 0, (4) T is normal means that T(A,) + T(A) for any increasing net A, converging to A in the strong operator topology, (5) T is semifinite means that, T(A) = s u p { ~ ( B ): B E M + ,B < A, T(B) < m}.
+
+
Example 2.1. M = B ( H ) , T(A) = TrA is the standard trace. This is a faithful normal semifinite trace. Example 2.2. Abelian von Neumann algebra M = Lw(X,R,m), where m is a probability measure, H = L 2 ( X ,R, m). Let ~ ( f=) fdv, where v is a semifinite X
localizable measure on ( X ,R). Then
T
is a faithful normal semifinite trace.
Example 2.3. A hyperfinite 111-factor (as the weak closure of a subalgebra of the CAR-algebra). Definition 2.1. Noncommutative Lp((M,T ) , l
< 00,
llAllco = IlAll, the operator norm. The elements of L p ( M ) may be identified with (possibly unbounded) operators on H . It is well known that 11 . [ I p is a norm and LP(M) is a normed vector space. This norm has the properties
<
IIA’llP = IllAlllP = IlAllm IIABII, IIAIIPIIBII,A E L P ( M ) , BE M . (2.1) Example 2.4. M = B ( H ) , T(A) = TrA. Then LP(M, 7)is the pth Schatten class, in particular, L 2 ( M ,T ) is the space of Hilbert-Schmidt operators and L’(M, T) is the space of trace-class operators on H .
Example 2.5. Abelian von Neumann algebra M = L w ( X , R, m), where m is a ) f d v , where v is a semifinite probability measure, H = L 2 ( X ,R, m). Let ~ ( f = X
localizable measure on ( X ,R). Then LP(M,T)
LP(X, ill v).
3. STOCHASTIC DIFFERENTIAL EQUATIONS
ON
L2(M)
Let Q be a bounded linear symmetric nonnegative operator on L 2 ( M ) . Denote by M Q the subspace Q’I2L2(M)with the inner product (A, B ) = (Q-’A, B ) L z ( M ) M . Denote by Wt a and the corresponding norm by I . I. Assume that MQ Wiener process in L 2 ( M ) with the covariance operator Q, that is,
76 where Wt are one-dimensional independent real Wiener processes and an orthonormal basis of M Q as a real space. m
W
j=1
j-1
W
Let B,F : L 2 ( M ,T )
-+
{<j}jOo=,
is
C t*<.,C tit;, C tj”are
L 2 ( M ,T ) . We assume that
j=1
bounded operators with the series convergent in the operator norm.
Theorem 3.1 (SDEs [ l l ] ) . Suppose that Q is a trace-class operator. Then (1) Suppose F arid B are Lapschitz coriti7iuou~on L 2 ( M ) . Then the stochastic differential equation, dXt = F ( X t ) d t + d W t B ( X t ) ,
xo = t , t E L 2 ( M , T ) has n un,ique solution in,L 2 ( M ) . ( 2 ) The function v(t,X ) = Pt(p(X) is a unique solution for the parabolic type equatiori
Example 3.1. M = B ( H ) , 7 ( A )= T r A . Let ez, be the matrix with 1 at the i j t h place and 0 at all other places. Define Qe,, = A1,eaJ, A,, 3 0. Then ta3= &el., is an orthonormal basis of MQ = Q1I2L2(M)and Wt = where Wi3
&Wi3Ea,,
13
are one-dimensional independent real Wiener processes. In this case Q is a trace class operator if and only if A,, < 00. Assume in addition that A,, = AJ, (e.g. 23
A , = ra+3,O
C
~ 3 t a 3=
a,j=1
W
00
03
W
a,p1
a.j=1
a3=1
a= 1
C tajJ:J = C ~1~3eaa7C ti = C -Yela
Example 3.2. Abelian von Neumann algebra L 2 ( M ,7) = L2(0,1) with the Lebesgue measure. Then en = fisin(n.rrz) is an orthonormal basis of L2(0,1). Suppose Q
c tzt,, = 00
is diagonal in this basis &en = Anen, Jn = &en.
c <,&7’,
n=l
00
n=1
Define g(x) =
00
=
C <.:
If Q is a trace-class operator, then
71=1
0 < g(z) < 2trQI E B(L’(0,l)). In particular, if Q-l = -A, then
dz)= G(z, XI, where G(z, y) is Green’s function for the Dirichlet problem on [0,1].
4. HEAT KERNEL Let
MEASURES
77
B ( X )=x, F ( X )= f
ct;x. J=1
Then Kolmogorov’s backward equation becomes the (group) heat equation.
Definition 4.1. The heat kernel measure pt on L 2 ( M )is the transition probability of the stochastic process X t . The next theorem address the issue of whether the heat kernel measure actually lives in the group. This result shows that the process determining the measure has a one-sided inverse, which is a double-sided inverse for the case of Hilbert-Schmidt groups (see Section 6).
Theorem 4.2 (Inverse [ll]).Denote
j=1 Then the solutions of the following stochastic dzfferential equations
+ +
dXt = TXtdt dWtXt, dZt = ZtTdt ZtdWt,
xo = 2 , zo = 2, satisfy
x,z E P ( A 4 , T )
ztx, = z x
with probability 1 for any t > 0 . 5. CAMERON-MARTIN GROUP
AND ISOMETRIES.
The following definition was first given in [8].
Definition 5.1. The Cameron-Martin group is GCM = { x E M X ,d ( x , I ) < 00)) where d is the Riemannian metric induced by I . 1:
1 1
d ( x , y ) = inf rr(0)==
lg(s)-lg’(s)lds
s(l)=u 0
Let iFcL2(pt)be the closure in L 2 ( p t ) of holomorphic polynomials 3cT on L 2 ( M ) .
Theorem 5.2 (Holomorphic skeletons [8]). For any f E 3CL2(pt) there is a holomorphic function f on GCM such that f o r any x E GCM and p , E XT L2(w* )
if Pm f , then p m ( x ) is given b y the formula +
This skeleton
f
f(x).
-+
where D k f is the k t h deriuutiue of function f , and c(s) = g(s)-’g’(s) smooth path g ( s ) from I to x .
for any
78 The Cameron-Martin group can be also described by a certain finite-dimensional approximation. Let M,, be finite dimensional von Neumann subalgebras, G, are the groups of invertible elements of M, with Lie algebras M Ic MZ c ... c M , c ... c M Q . Assume that all M,, are invariant subspaces of Q.
Notation 5.3. G , = UGn and g = u M n .
<
Remark 5.4. If I[X,Y]J clXJJYJ, then G, is dense in GCM in the Riemannian metric induced by I . I. This is the case for some Hilbert-Schmidt complex groups. We also introduce one more (larger) space of holomorphic functions.
Definition 5.5. ~ ' ( G c Mis )the space of holomorphic functions on the CameronMartin group GCM such that
Ilfllt,oo= JLE Ilfllt.n= lim n-m
J
lf(z)I2dp:(z) < 00.
G,
Theorem 5.6 (Pointwise estimates [7], [8], 191). For any f E V ( G C M )g, E G C M , O<s
Theorem 5.7 (Isometries [7], [8], [9]). (1) The skeleton map is an isometry f r o m 3CL2(pt) to !K'(GCM), that is, the restriction map on holomorphic polynomials KP extends to an isometry between the spaces X L 2 ( p t ) and xt( G c M ) . (2) If Q is u trace cluss operator, then !KL2(pt) is a n infinite-dimensional Halbert space. M
(3) The Taylor map f
c-)
generalized Fock space g with the norm
2D k f ( I ) is an isometry from ~ ' ( G c Mto) the
a-=n
&(i), a subspace of the dual of the tensor algebra of
Theorem 5.7 can be illustrated by the following diagram.
skeleton
- - inclusion map
&'restriction" map
Taylor map
XL2(pt) WGCM) zt(e) The generalized bosonic Fock space zt(g) is a Hilbert space in the dual of the universal enveloping algebra of the Lie algebra g defined as follows. Let T(g)be the tensor algebra over g , J be the two-sided ideal in T(g)generated by {t@ v - @ t -
xP
bo
[c,171;<,17 E g}. The algebraic dual of the tensor algebra T(g)is T'(g)= k=O C (gWk)*. Then J o is the annihilator of J in the dual space T'(g). Define
79
Then the generalized bosonic Fock space is
&(g) = {a E
J0 : 1 1a11:
< .I.
This is the space of Taylor coefficients of functions from the space ~ ' ( G c M ) . 6. EXAMPLES OF GROUPS
In this section we describe examples of infinite-dimensional groups for which our construction works. Roughly speaking, we would like to consider L2(M,T ) as the Lie algebra g, and the multiplicative group of M as the infinite-dimensional (Lie) group. As we have seen g actually has to be chosen smaller than the whole space L'(M,T) to ensure that the Brownian measure lives in the group. 6.1. Hilbert-Schmidt groups. Let M = B ( H ) be the space of bounded linear operators on a complex Hilbert space H . Then the group is M X = GL(H), invertible elements of B ( H ) . As before let T ( A )= T r A be the standard trace, then L2(M,7)= H S is the space of Hilbert-Schmidt operators on H with the inner product (., . ) H s . Suppose g H S is an infinite dimensional Lie algebra with a Hermitian inner product (., .) which in general is different from the Hilbert-Schmidt inner product. Note that we can view elements of H S as infinite matrices, namely, matrices { a i j } such that C laijI2 < 00. Let as before e i j be the matrix with 1 at the i j t h place
s
i:j
and 0 at all other places. Assume that Qeij = Xije,j,Xij 2 0. Then (ij = &eij is an orthonormal basis of M u , and the Wiener process Wt = C W y & j . Note that
Q is a trace class operator if and only of
C X i j < 00.
i7 j
i:j
n(l+
Example 6.1. The Hilbert-Schmidt general group G L H = ~ GL(H) H S ) with the Lie algebra g = Q1/'HS equipped with the norm IAl = IQ-1/2AI~s. Example 6.2. The Hilbert-Schmidt orthogonal group S O H is~ the connected component of { B : B - I E H S , B T B = BBT = I}. Let 5 0 ~ s= { A : A E H S , AT = -A}, then the Lie algebra for S O H is~ g = Q 1 / ' 5 o ~ sequipped with the norm IAl = IQ-'/'AIH~. Example 6.3. The Hilbert-Schmidt symplectic group S P H=~ { X : X
H S , X T J X = J } , where J =
(;
;I).
Let
5pHs
={X :
-I E
x E HS,XTJ+JX
=
0). Then the Lie algebra is g = Q'/'5p~swith the norm IAl = IQ-'/'AJHs.
Suppose the operator Q is as above. Theorem 6.1. Let Q : H S -+ H S (or ~ O H Sor 5 p H S ) . (1) If Q is trace class, then the heut kernel measure lives in GLHS (or S O H or ~ SPHS), and ?Ct(Gcb,)is a n infinite dimewional Halbert space. (2) I f the covariance operator Q is the identity operator, then W ( G C M contains ) only constant functions.
80
Remark 6.2. If the underlying Hilbert space H is finite-dimensional, then these are just usual groups GL, SO and Sp. Thus the finite-dimensional groups can be viewed as a particular case of our setting. R e m a r k 6.3. It is known that the Riemannian geometry plays an important role in understanding of how the heat kernel measure behaves. It is a much more complicated issue in infinite dimensions. We address this issue in [12].
6.2. A hyperfinite 111-factor (as the weak closure of a subalgebra of the CAR-algebra). In the following description we follow L. Gross [13] and I. E. Segal [20]. Let K be a complex Hilbert space with a Hermitian inner product (., .)K. Denote by A n ( K ) the space of skew symmetric tensors of rank n over K, and A o ( K )= C, H = A ( K ) = @T=oAn(K).For any z in K there exists a bounded creation operator C, on H = A ( K ) such that
+
C,u = (n 1)&zA u,u € A”(K), where z A u is the antisymmetric projection of z @ u. Then the annihilation operator is A , = C: and
Bz
def
=
C, + Aar.
Then
BZB, + B,B, = 2(2,2JY)KI. The von Neumann algebra M is the smallest weakly closed algebra of operators on A ( K ) containing all the operators B 2 , z E K , that is, M E B ( A ( K ) )= B ( H ) . The trace is defined by ~ ( u=) (uS2,52),where the “bare” vacuum C2 = 1 is an element in A o ( K ) C A ( K ) . It is known that the space M with this trace is a hyperfinite 111-factor [20]. As always, the trace determines a Hermitian inner product on M by ( A , B ) 2 M ) = T ( B * A )and Lp(M) is the completion of M in the norm “1
(
llAllp = ( ~ ( 1 A l p ) ) Ffor l
Theorem 6.5 ( [ 9 ] ) .If Q is the identity operator, then F C t ( G ~ ~contains ) only functions of the f o ~ m
81 REFERENCES [l] V.Bargmann, O n a Hilbert space of analytic functions and an associated integral transform,
Part I, Communications of Pure and Applied Mathematics, 24, 1961, 187-214. [2] V.Bargmann, Remarks on a Hilbert space of analytic functions, Proc. of the National Academy of Sciences, 48, 1962, 199-204. [3] G.DaPrato and J.Zabczyk, Stochastic Equations in Infinite Dimensions, Cambridge University Press,Cambridge,l992. [4] B.Driver, O n the Kakutani-ltb-Segal-Grossand Segal-Bargmann-Hall isomorphisms, J. of Funct. Anal., 133,1995,69-128. [5] B.Driver and L.Groas, Halbert spaces of holomorphic functions on w m p l e x Lie groups, Proceedings of the 1994 Taniguchi International Workshop (K. D. Elworthy, S. Kusuoka, I. Shigekawa, Eds.), World Scientific Publishing Co. Pte. Ltd, Singapore/New Jersey/hndon/Hong Kong, 1997. [6] E. B. Dynkin, Calculation of the coeficients in the Campbell-Hausdorfl formula, Doklady Akad. Nauk SSSR (N.S.)57, 1947, 323-326. [7] M. Gordina, Holomorphic functions and the heat kernel measure on an infinite-dimensional complez orthogonal group, Potential Analysis Volume, 12, 2000, pp. 325-357. [8] M. Gordina, Heat kernel analysis and Cameron-Martin subgroup for infinite-dimensional groups, J. Func. Anal., 171, 2000, pp. 192-232. [9] M. Gordina, Taylor map o n groups associated with Q III-factor, 2002, Infin. Dimens. Anal. Quantum Probab. Relat. Top., 5, 93-111. [lo] M. Gordina, Quasi-invariance for the pinned Brownian motion on a Lie group, 2002, to appear in Stochastic Processes and Their Applications. [ll] M. Gordina, Stochastic diflerential equations on noncommutative L2,2002, to appear in Contemp. Math. book series of the Amer. Math. SOC. (proceedings of the Conference on Analysis on Infinite-Dimensional Spaces in honor of Leonard Gross), H-H. Kuo and A.N. Sengupta, Eds. [12] M. Gordina, Hilbert-Schmidt groups: Rzemannian geometry, Lie algebras etc., 2004, preprint. [13] L. Gross, Existence and uniqueness of physical ground states, J. Funct. Anal., 10, 1972, 52-109. [14] L. Gross and P. Malliavin, Hall’s transform and the Segal-Bargmann map, It& Stochastic Calculus and Probability Theory, (M.Fukushima, N. Ikeda, H. Kunita and S. Watanabe, Eds.), Springer-Verlag, New York/Berlin, 1996. [15] B.Hall, Quantum Mechanics in Phase Space, Proc. of the Summer Research Conf. on Quantization, (L. Coburn and M. Rieffel, Eds.), 1997. [16] B.Hall, The Segal-Bargmann ‘coherent state’ transform for compact Lie groups, J. of Funct. Anal., 122, 1994, 103-151. [17] B. Hall and A. Sengupta, The Segal-Bargmann transform for path-groups, J. of Funct. Anal., 152, 1998, 220-254. (181 Robert H. Martin, Jr., Nonlinear operators and differential equations in Banach spaces, Pure and Applied Mathematics. Wiley-Interscience [John Wiley & Sons], New York-hndonSydney, 1976. [19] I. Shigekawa, It& Wiener ezpansions of holomorphic functions on the wmplex Wiener space, in Stochastic Analysis, (E. Mayer et al, Eds.), Academic Press, San Diego, 1991, 459-473. [20] I. E. Segal, Tensor algebras over Hilbert spaces. II, Ann. of Math., 63, 1956, 160-175. [21] I. E. Segal, Mathematical problems of relativistic physics, Lectures in Applied Mathematics (proceedings of the Summer Seminar, Boulder, Colorado, 1960), Vol. I1 American Mathematical Society, Providence, R.I. [22] HSugita, Properties of holomorphic Wienerfunctions-skeleton,contraction, and local Taylor expansion, Probab. Theory Related Fields, 100, 1994, 117-130. [23] HSugita, Regular version of holomorphic Wiener function, J. Math. Kyoto Univ., 34, 1994, 849-857. DEPARTMENT OF MATHEMATICS, UNIVERSITY E-mail address: gordinahath. uconn. edu
OF
CONNECTICUT, STORRS, C T 06269, U S A .
This page intentionally left blank
Infinite Dimensional Harmonic Analysis I11 (pp. 83-97) Eds. H. Heyer et 01. @ 2005 World Scientific Publishing Co.
On Mehler semigroups, stable hemigroups and selfdecomposability Wilfried Hazod, Dortmund
Introduction. Let V denote a Banach space - in most cases finite dimensional - and let (T, : t E R) be a continuous one-parameter group of linear operators acting on V, or, let N denote a simply connected nilpotent Lie group, and let ( T ~: t E W) denote a continuous one-parameter group of automorphisms. (In most examples it will turn out to be natural to assume that resp. ( T t ) is contractive, i.e. lirn,,, Ttx = 0 for x E v, resp. Ttx+ e.) In the last decade there exist parallel and independent investigations, for vector spaces as well as for groups, from different points of view and into different directions, concerning (operator) self-decomposabality on V (and N), cocycle equations - the solutions will be called M-semagroups in the sequel (operator) stable hemigroups, generalizing the concept of stable convolution semigroups, generalized Ornstein Uhlenbeck processes with or without invariant measures, and Mehler semigroups, the semigroups of transition kernels, and the corresponding background driving (Livy) processes. A few references: [17], [l],[13], $2.14, [8], [4],[31]. (For precise definitions see Definition 1.1 below). One of the aims of this paper i s to point out that thcse objects of investigations are closely related, in particular in the finite dimensional case: The missing links are obtained considering special representations of the space-time building as a semidirect extension G of the state space V (resp. N)- hence endowed with a non-Abelzan group structure - and the above-mentioned objects may be embedded into this enlarged state space. In fact, it turns out that for the finite-dimensional setup the means to develop a satisfactory theory are more or less already available, for vector spaces V as well as for groups N.For infinite dimension the situation is different. Nevertheless, it seems to be useful to begin with a general model and to point out where finite-dimensionality turns out to be essential. In $1we give precise definitions, discuss the models and assumptions and formulate the main results, in $2 we collect and - if nccessary - develop some results from the investigations of the embedding problem for probabilities on mostly locally compact - groups, which are applied in $3 to prove the before mentioned Theorems B and C.
(z)
Q 1 M-semigroups, stable hemigroups and self-decomposability We start with a series of definitions. 1.1Definition. a) A family (p(t))tEB+ n ' ( V ) is called M-semigroup if the mapping t H p(t) is continuous and fulfils the cocycle equation p ( t 3) = p ( t ) * Z ( P ( S ) ) t , s 2 0.
+
84 b) A probability measure v is called (operator-)self-decomposable if for all t 1 0 there exist cofactors p ( t ) E mZ'(V) such that v = p ( t ) * Tt(v). c ) Generalizing the concept of (operator-)stable laws, a family (pLs,t)Ols
01s < t. Th(ps,t)= ps+h,t+h d) A probability measure p E mZ1(V)is called M-znfinztely dzvaszble if for any n E N there exist p(n)E mZ1(V)such that p = P(n)* Tl/n(p(n))* . . . * qn-l)/n(P(n)). p ~ is ~called ) n-th M-root of p . d) For a group N and automorphisms (Tt) the definitions a) - c) are modified in an obvious way. It is well known that the above-defined objects are closely related. It is the aim of this paper to point out the interdependence of these different research objects. First we note that under mild conditions we obtain immediately relations between the objects in a)-c): Proposition A. a) If (p(t))t,Ois a M-semigroup then := T s ( p ( t- s)))s
85
In [25] the authors investigate continuity propertics of M-semigroups (on Banach spaces). This motivates investigations of embeddability of M-infinitely divisible laws into M-semigroups. In 53 we shall prove for finite dimensional V resp. N
Theorem B. p is M-infinitely divisible iff p is embeddable into a (continuous) M-semigroup (p(t))t>o such that p = p(1). Furthermore, we shall prove in this situation using the Lie-Trotter product formda for convolution semigroups on locally compact groups:
Theorem C. There exists a bijection between convolution semigroups (pt : t E R) (on V resp. on N ) on the one hand and M-semigroups ( p ( t ) : t E R) on the other, hence between Lkvy processes - called background driving processes and generalized Ornstein Uhlenbeck processes. Furthermore, this yields a bijection between L6vy processes with finite logarithmic moments and selfdecomposable laws v.
-
The proofs rely on the fact that the underlying state space is embedded into a (non-Abelian) group 9, and the above-mentioned objects, M-roots, Msemigroup etc. correspond to roots, convolution semigroups etc. on this group. Thus we may apply the well-known methods and results developed for embedding problems and convolution semigroups to the locally compact group 6 to obtain proofs of Theorems B and C above (for V resp. N ) . For infinite dimensional V comparable results are still considerably poor. This is due to the fact that for infinite dimensional (non locally compact) groups sufficient investigations into the structure of convolution semigroups and the embedding problem are still not available. On the other hand, as mentioned before, there exist independent profound investigations of these convolution structures ([4], [8], [31]) and of the corresponding operator semigroups, called Mehler semigroups. That is why it seems worth to formulate the problems in a general setup and to point out where finite dimensionality turns out to be essential. Furthermore, it should be mentioned that in the infinite dimensional situation the above-defined model is a special case of the afore-mentioned investigations, cf. in particular [8], (41, (311: There (Tt),," is assumed to bc a semi-group of operators, not necessarily (continuously] invertible. However, in most familiar examples - e.g. operator stable or self-decomposable laws on Banach spaces V - (Tt),?, may be extended to a group. Moreover, in most examples (Tt),?, on V (as well as ( T ~ ) ~ >on " N)is assumed to act contractively, i.e. limt+coTtx= 0 for x E V, rcspr.rtx -+ e. In this case obviously -1 belongs to the resolvent set of Tt for all t E R, a condition which will turn out to be crucial in $2 and 3. In fact, stable laws are concentrated on the contractible part {x : Ttx -+ 0}, and concerning selfdecomposable laws we refer to invcstigations in [19], [as], [30],showing in the more general situation of a discretc group of operators that there exist decompositions of an operator-decomposablc law into an invariant
86
measure and a measure on which the operators act contractively. In [17] contractivity is always assumed. Hence the spectral condition -1 $ Spec(Tt), t E R, in the following 52 and 3 is not a serious restriction. Note that the technical problems arising in the finite- and infinite-dimensional case are quite different. In particular, for an infinite dimensional Hilbert space only a reduced version of Bochner's theorem is available. Hence, if the convolution structure is investigated via Fourier transforms the existence of corresponding M-semigroups of probabilities and and of corresponding Ornstein Uhlenbeck processes is a priori not guaranteed (cf. e.g. [4], [S]).Therefore, if Sazonov continuity is not assumed, a suitable (Hilbert-Schmidt) embedding into a larger Hilbert space is necessary.
52 Remarks on root-compactness and embeddability Let G denote a topological Hausdorff group and p E fm'(G). p is called embeddable if there exists a continuous convolution semigroup (pt : t 2 0) with p1 = p. A challenging problem during the last decades was to obtain a descrip tion of embeddable measures within the class of infinitely divisible measures. (For locally compact groups the reader is referred to the surveys [21], [15], see also [7], and the literature mentioned there.) The following definition, originally due to W. Boge [3], is crucial for investigations of embedding problems: 2.1 Definition (Cf. [14], 3.1.1, [28], Section 2 ). A topological group G is called proot compact ( p E N) if for any compact C C 0 there exists a compact F G such that for any sequence { X I , . . . , x P } C 9 with x p = e and x, . C . xk . C n xi+k . C # 8 for a k 5 p we have { x , } C F . Q is root compact if 8 is proot compact for any p E N. G is strongly root compact if G is root compact and if the compact set F may be chosen independently of p E N.In a similar way strong proot compactness is defined.
+
2.2 Examples (Cf. [14], [28], [5], [6], [32]). 1. Banach spaces, morc generally, quasi complete locally convex topological vector spaces are strongly root compact. ([28], 6., Satz 4 ). In particular, the proof shows that for any compact C contained in a ball BR(O)with centre 0 and radius R we may assume F BQR(0).Hence, for (Cf. also [14], 3.1.3 finite dimensional vector spaces V we can put F = BQR(O) for dim(V) < 03). 2. (Cf. [5], [6], [28], 6., Satz 3, [14], 3.1.17). Connected (finite dimensional) nilpotent Lie groups are strongly root compact. 3. (Cf. [32], [14], 3.1.23). Let G be a connected solvable Lie group with Lie algebra g. If for all X E g the spectrum Spec(adx) is real, then G is strongly root compact. 4. However, there are negative results: Let Gz denotc the group C A W with ( x , t ) ( y , s ) := ( x + e t z . y , t + s ) , wherez := a+i.b E C x isfixed. Gz isasolvable connected real Lie group. Gz is strongly root compact if b = S z = 0 (according t o 3. above). If a = X z = 0 then G,.b is not 2-root compact (cf. [32], examples)
87 and, if a = !Rz # 0 and b = S z # 0, then Qz is not %root compact. (Cf. [32], , examples, [14], 3.1.26: There this result is proved for z = 1 - z . ~ / l o g 2 but the proof is easily extended.) 2.3 In the following, in particular in $3, we will be interested in topological groups of the following types: (a)Let V denote a real Banach space and let (Tt : t E R)denote a continuous one-parameter group of linear operators acting on V with infinitesimal generator A. Let Q denote the semidirect extension Q = V M R with group operation (2, t)(Y, s) := .( Tt(l/)l t + s). (b) Of particular interest is the case dim(V) < co. (Note that B is locally compact iff dim(V) < cm.)In this case, Q is a connectcd solvablc Lie group of dimension dim(V) 1, and we have Tt = exp(t . A ) for some A E GL(V). ( c ) Generalizing (b), let N be a simply connected (finite dimensional) nilpotent Aut(N) denote a continuous Lie group with Lie algebra n, let (Tt : t E R) group. Let again Q = N M Rwith group operation (z, t)(y, s) := ( z . T t ( y ) , t + s ) . In this case, rt = exp(t . A) for some derivation A E Der(n) GL(n).
+
+
The following results, complementary to 2.2, are probably known, however I was unable to find a reference. The proofs are straightforward and similar to the proofs of the standard results mentioned in 2.2. 2.4 Proposition As immediately seen, all groups mcntioned in 2.3 are solvable. In fact, [Q, Q] C (0, V) resp. C (O,N), hence Abelian rcsp. nilpotent. a) Let Q as in 2.3 a). Assume -1 $ Spec(Tt) for all t E R,equivalcntly, z . I[$ n Spec(A) (0). Then Q is 2-root compact. b) In particular, in the finite dimensional case, Q is a solvable 2-root compact Lie group. Note that the spectral condition is fulfilled if (Tt),?, is a contractive semigroup, i.e. if lim T,(z)= 0 for all z E V. t+w
c ) Let N, (q : t E W) and Q be as before in 2.3 c). Then, with the notations introduced there, we have: If i . R n Spec(A) {0}, G is a solvable connected 2-root compact Lie group. Again, this is in particular thc case, if ( T t ) acts contractively on N. d) Assume in a) - c) that PSpec(A) R then Q contains a closed subgroup which is not root compact, hence is not root compact. (Here PSpec denotes the point spectrum.) Proof: a) Let - w.1.o.g. - the compact set C Q be of the form C = r x [ - E , €1, with compact 'I C V. To prove 2-root compactness, let z1 := ( Z , t ) E Q, z2 = e = (0' ,0) the unit element and assume 5 1 . C . C n C # 0. For the second component t we obtain: Let T : Q + R denote the canonical projection with kernel ker T = V.Then t [-t, E] t + [ - E , E] n [ - E , E] # 0. R is (strongly) root compact, therefore for some a = a ( € )> 0 we obtain t E [--a, a]. (E.g., we may put a = 9 . E according to 2.2.1.) Considering the first cornpone$ Z we obser_ve: Therc cxist d' E I? such that o?= Z+Tt(a')+Tt+y(Z)++5+t+t(b)for some a', b E r, and JsI 5 t, It1 5-a ( €-)Thc. refore, z + T ~ + ~= ( zo?-Tt(Z) ) -Tzt+y(C) F for some compact set r = r(r,€1
+
c
+
88
2 V.Put u -
:= t+s thcn we obtain
2 E {(I
+ TJ1 (F)
:uE
[-a - E , a
+ el}
=: F , a compact set in V,since we supposed -1 $ Spcc(T,) for all u,hence the resolvents { ( I TU)-’} are uniformly bounded for u E [-a - E , a + €1. Hence a) and b) are proved. To prove d) for the vector space case, assume X = a z b, b # 0 and let 2 (in the complex vector space V i . V) be an eigenvector corresponding to A. Since X and 5; are eigenvalues there exists a (Tt)-invariant subspace W V, W E! R2 such that Ttlw = e t a . Ut for a rotation Ut E SO(2) (depending on b). Put 3-1 := W x E% defined by the action of (T,lw) on W = R2.This subgroup of 6 is isomorphic to Yuan’s example mentioned in 2.2 (4),thcrcfore 3-1 and hence G can not be root compact. We prove c) by induction on the dimension on N. For d = dim(N) 5 2 the group JVis a vector space, hence the result follows by stcp a), b). Y1,. . . ,YK}for n such that N K = Choose a vector space basis { X I ,. . . ,XN, d, { X I , . . . , XN}span n/a, with j := Cent(n), and {Yl,. . . ,YK}span a. Then N is parameterized b y C x , . X , + C y , . Y , =: H exp(Z)exp($ =: 4(%)$(ij) with central $($. 2 := Cent(N) is characteristic in N, hence 2 := ( 0 , 2 ) aG. For := x R with := N/2 the assertion is true by the induction
+
+
+
I
+
%:+a
hypothesis. Let x 1 = ( d ( Z ) $ @ ) , t ) E 6 , assume C = C x [ - E , E ] to bc compact in 8. W.1.o.g. assume C = 4(C1)$(C2) for compact C1 G n/a and C, 2 8 respectivedenote the canonical projection. With our parameterization ly. Let 7r : 4 -+ we obtain 7r (d(Z)$(Q)) = d(Z). Assume z1. (?. x1 .(? n # 0. Then 7r (zl) . 7r (5).7r (xl).7r C n7r C # 0.
(-> - ( 2
Whence by induction hypothesis we conclude that It1 5 a ( € )Z, E F = F(t,C) for some CY = a ( € )> 0 and some compact P C ii. Hence we obtain (4(2)$(9,
with It1
t ) .(4(q+ml s ) . ( d ( w W , t ) (. 4 ( U O $ ( V ’ ) > s‘) = ( d ( + m l t’) E
5 a, 4(Z)
#)(%I . $(Y)
E F , IsI, ls’l,It’l
. Tt ($(a))
’
Tt
5 E , U,U’, E
($J(3))’ Tt+s ’Tt+s+t
(+(.)I
’
E
x
C1, V, V’, E Cz. Thus
Tt+s
($(Y))
(#)(GI)) . Tt+s+t
($(a))
Therefore, the central elements $(@). rt+, -
c
(dj(,q)
= #)(El ’
$0) Ec
bclong to a compact set E C Cent(N). In other words, with u := t + s we concludc that $(y) . T ($(y)) ~ belongs to some compact set E’ C Cent(N). But 2 = Cent(N) is a vector space and by assumption -1 $ Spec(.r,lz), hence we obtain as in step a) that also $(g) belongs to some compact set F’ = F’(E’,e) C 2. Whence the assertion follows. To prove the assertion d) for the case of a group N, notc that from the existence of a non-real eigenvalue it follows as before that the complexification n@contains a (Tt)-invariant subalgebra tu generated a 2-dimcnsional subspacc 0,such that the action on the quotient space 7r (tu n n) is again of thc above-mentioned
89 form e t a . U,. Here 7r dcnotes the projection tu + tu/[tu, tu]. Therefore it follows easily that the (closed) subgroup exp (tu n n) and hcnce N can not be root compact.
2.5 Remarks. a) If in 2.3 we assume that Spec(A) & R then - at least for finite dimension - B is strongly root compact according to 2.2.c). b) In contrast, it could not be proved in 2.4 a)-.) that Q is strongly 2-root compact. This problem remains open. This property would guarantee continuous embedding, even in the infinite dimensional case. However, it turns out that for dim(V) < 00 this property will not be needed in the sequel. c ) The groups considered in 2.4 are connected and aperzodzc, i.e. possess only trivial compact subgroups. This follows immediately since the underlying vector space V resp. the group N are aperiodic and Q/V rcsp. Q / Nare isomorphic to R, hence aperiodic. In fact, the assumption on Spec(A) implies that t H Tt is injective and hence even {T, : t E R} is aperiodic, hcnce 2 R. The following results are folklore for investigations of embedding problems: 2.6 Proposition Let Q be a topological group. a) Let Q be 2-root compact. Let p E !311(Q). Then the sets of 2n-th roots { X E !311(B) : such that A’” = p } of p are relatively compact, n E N. b) Therefore in a) an infinitely divisible probability measure p E !311(Q)is embeddable into a convolution semigroup (p,. : r E D+), where D denotes the submonogeneous group of dyadic numbers. c ) A submonogeneous - in particular a dyadic - semigroup of probabilities on a connected solvable Lic group is locally tight, i.e. {p,. : 0 5 r 5 1,r E D+} is relatively compact. d) Therefore in c) there exists a compact subgroup K & Q such that a suitably shifted semigroup (A, := p,. * E,, : yi, E K ,r 2 0 ) has an extension to a continuous real convolution semigroup (A, : t E R+). e) In particular, for aperiodic, connected and solvable 2-root compact Lie groups Q we conclude: An infinitely divisible probability measure on Q is continuously embeddable. Proof. a)b) See e.g. [14], 3.1.4, 3.1.32; see also [28], 2., Satz 2. For c) see [21], Theorem 2. For d) Sec [28], 5., Satz 1, [14], 3.5.4 (for rational semigroups), the proof for general submonogeneous semigroups is similar (see e.g. [lo], 2.4, [23]). Finally, e) follows by c),d) since for aperiodic groups obviously K = { e } in step d).
In 53 below we shall apply the results of the abovc discussion in the condensed form:
2.7 Corollary Let 6 be as in 2.4 a)-.) and let p be infinitely divisible. Then p is embeddable into a dyadic convolution semigroup. In the finite dimensional cases b), c) p is continuously embeddable.
90
93 Space-time- and background driving processes Let ( X t )be a - not ncccssarily stationary - Markov process on a statc space := E x W (thc space-time-building). E with transition kernels (Ps,t)s
Psdz, t ) := J, g(Y, t + s)Pt,t+s(z,dY)
-
for measurable, bounded functions g : E + W. Thc semigroup of ker_nels (Ps)s20defines a stationary (i.e. time-homogeneous) Markov process on E. For a hemigroup (ps,t : s 5 t ) - defined as in the introduction - on E := V (resp. on E := N ), which is stable w.r.t (Tt : t E R) (resp. (Tt : t E W)) let (Ps,t),,tbe defined by the corresponding convolution kernels f H ps,t* f,0 5 s 5 t. In this case we havc
PsdzC,t ) := J, g b + Y, t + .)dPt,t+s(Y) =
s, g(x + Y, t + s)dTt (Po,,) (Y) =
J,S(a:+Tt(Y),t+s)dPo,s(Y) = S,g(z+Tt(y),t+s)dl.L(s)(y) (1) where ( p ( t ) := po,t)t,O - denotes the corresponding M-semigroup (cf. $1, Proposition A a)). In the following we shall use slightly different space-time processes
E g ( x , t ) := J , d Y + T s ( x ) , t + s ) 4 4 s ) ( y ) (In case of a group N wc have to replace y T,(x) by y . T,(x)). In particular, for functions g : (2, t ) H g(x,t) = 4 ( x ) . $(t) we obtain:
( 1')
+
.I,
ma:, t ) = dJ(s+ t ) 4(x + TS(Y))&cL(S)(Y) This yields for the projection T : B -+E defined by (x,t)
Os4(4
:= n-
(2) H
x:
(E ( 4 ( 4 W ) ) = J, 4b + Ts(Y))dP(s)(Y)
Note that (2') defines a semigroup of operators
(2')
(0,) 5 2 0 on a suitable func-
tion space, e.g. on L2 ( E )w.r.t. Haw measure. This semigroup is usually called (generalized) Mehler semzgroup. See e.g. [8], [4], [31], and the references mentioned there. This justifies the notation M-semzgroup for the family of measures ( p ( t ) ) ,though this family is not a semigroup. A further justification is given in 3.2 below. In the above situation, - the space-time-building in (1') may be represented as a semidirect product E = 6 = V x R (resp. = N x R) with W acting on V (resp. on N)by (z)(resp. by ( T t ) ) . 1.e. i? turns out to have the shape of the groups E considered in $2. Put A, := p(s) @ E, IE rnn'(G), s 1 0. Then (1') resp. (2) may be reformulated in terms of convolution operators
k 7 ( x ,t ) = (A, * 9)(x,t) = JG 9 ((Y, . (5, t ) )d (PL(S) @ EY) (Y,). with * denoting convolution on the (non-Abelian) group 9.
(3)
91 3.1 Definition. Put m:(Q) := {v 8 E~ : t E R, v E m'(V)} (rcsp. v E !%R1(N)), a closed convolution subsemigroup of 9JT' (8). Note that convolution in m:(Q) is given by ( p @ E t ) * (v 8 E s ) = ( p * Tt(v))8 E t + s . (4)
Here again
* and * denote convolution on V (N) and on G respectively.
A straightforward calculation yields: 3.2 Proposition. Let (At := p ( t ) 8~~E m:(Q) : t 2 0) with p ( t ) E 9JT1(V) (E m'(N))for t E R+. Then ( X t ) t 2 0 is a continuous convolution semigroup ) ~a > continuous ~ M-semigroup in ml(V) on D w.r.t. convolution * iff ( ~ ( t ) is w.r.t (Tt)(resp. in !JJ?l(N)w.r.t. and with convolution *. For a Proof see [ll],$3, [13], 2.14.12. See also the hints and references thcrc.
(~~7)
let
-
Let (At := p ( t ) 8 E ~ ) ~ > , ,be a continuous convolution scmigroup in E:(Q), be a corresponding LCvy process on Q. Thcn, as immediately seen,
(xt)t>o
X t = ( X i , t ) ,t 2 0, where ( X t ) is a process taking values in V (resp. in N) with distributions p ( t ) . Hence 3.1 yields (with the notations defined in $1): 3.3 Proposition. If ( y t ) = ( t ,X t ) is a LCvy process on the enlarged group Q, i.e. c&dl&gwith stationary independent increments on 4 , then ( X t )is a generalized Ornstein Uhlenbeck process, hence cAdl&gwith independent increments on the original state space V (resp. N), and vice versa.
Convolution on Q as described in (4)yields furthermore (for fixed automorphism groups (Ti)resp ( T t ) ) that M-roots p ~ (of~probabilitics ) p on V (rcsp. on N)correspond to roots = p ~ 8&lin ( ~ ) - in the usual scnse - of probabilities X = p 8 ~1 on Q. Thercfore we observe
3.4 Proposition. Let p E 9X1(V) (resp. 9JT'(N))be M-infinitely divisible (cf. ~ divisible on Q with roots in 9JT:(Q), Definition 1.1) then X = p @ is ~infinitely and vice versa. Now we have prepared the means for a Proof of Thcorem B: Let p be M-infinitely divisible, hence X = p @ c l is infinitely divisible in 9JTi(Q) (Proposition 3.4). According to Corollary 2.7, an infinitely divisible law X is embeddable into a dyadic semigroup {A, : T E D+} which bclongs to 9JT:(Q) if all roots belong to this subsemigroup of !T?'(G). In fact, (again according to 2.7) in the finite dimensional case, X is even embeddable into a continuous convolution semigroup. Now p is embeddable into a continuous M-semigroup by Proposition 3.2. The converse is proved in the same way. Hence Theorem B is proved. In the following we concentrate again on the finite dimensional situation.
3.5 Theorem. Let (At := p ( t ) @ Et E m:(Q): t 2 0) bc a continuous convolution semigroup in 9JT:(G) with corresponding M-semigroup ( p ( t ) : t E R+) in mm"(V) (resp. m'(N)).
92
a) Then any f E E := Cp(G)belongs to the domain of the generating functional A of ( A t ) , i.e. (A,f ) := $ ( A t , f ) l t = o exists. b) According to the particular shape of At, i.e., since At E M ; ( s ) , t 2 0 , we obtain a decomposition of the generating functional A = $Atlt=o
A = EG@ P
+B 8
EO
=:
P +B
(5)
glt=o
where P is the primitive form P = and B is defined as B = $p(t)It=o. (The latter derivative exists since the first ones, A and P exist.) For N replace EG by E,. c ) There exists a uniquely determined continuous convolution semigroup (pt)tGa, C ml(V) (resp. !Dl'(N)) with generating functional B such that
B = ddt+p( t )lt=o
d+
= xptIt=o'
(Note that in case of a non-trivial action of (T,) we havc pt # p ( t ) ,t > 0.) The LBvy process (Y,) on V (resp. on N) corresponding to that convolution semigroup (pJ is called background driving process. d) To enlighten the situation in a)b) we represent the group in the following equivalent form:
B 2 { ( x , T , ): t E R , z E V (resp. EN)}, a subgroup of the affine group V x Aut(V) (resp. N x Aut(N)). In this situation the decomposition of the generating functional has the following form:
A = 6 6 8 P + B 8 E I , ~= P
+ B,
(Id denoting the unit in Aut(V)) with
(P,dJ)= i d ( % ) It=O
= $(ETt,d)It=O
for dJ E cp(Aut(V))
and for $ E C,-(V).
( B , $ ) = %(P(t),$)lt=o
(The modifications for N are obvious.) Proof: a) follows by the Ldvy Khinchin formula for (non-Abelian) Lie groups (cf. e.g. [14],4.5, [13], $2.0.1). b)d) follow by Theorem 2.14.13 in [13]: According to the assumption on the spectrum Spec(A) the group (resp. ( T t ) ) is closed in Aut(V) (resp. Aut(N)) and isomorphic to (R, +). Hence the above-mentioned theorem 2.14.13 in [13] applies and yields the desired result. c) B = gp(t)It=o is an almost positive and normalized functional, hence by a result of E. Siebert ( [27],Satz 5 , [14], 4.4.18), B is the generating functional of a continuous convolution semigroup (pt) in !Dl'(V) (resp. in %V1(N)). Therefore, in particular we observe
(z)
B = $p(t)It=o
= xd+P t l t = 0 .
The above decomposition of the generating functional A = EC 8 P + L? enables us to prove Theorem C. First we notc the obvious
=p
+B 8
EO
93
3.6 Proposition. P := @ P and 23 := B @ EO are generating functionals of convolution semigroups on G, both belonging to !lXk(G): a) P generates the continuous convolution semigroup (pt := EG @ Et : t 2 0), in fact, this semigroup extends to a group, and (p-t := EB @ E-t)t>O- is generated by the functional -P. b) B generates the continuous convolution semigroup (vt := pt @ EO : t 2 0). c ) Therefore the Lze Trotter product formula [ see e.g. [9],I, 52 ] for continuous convolution semigroups on locally compact groups yields for all t 2 0 : At = and
ii% (pt/n * vt/n)"
(7)
* At/n)"
(8)
vt = n-w lim (P-t/n
Proof of Theorem C. The representation of At and vt via the Lie Trotter formula - with the previous introduced notations - yields for the first components:
~ ( t=)n-w lim
*
(pLt/n Tt/n (pt/n)
* . . * Tt (n-l)/n)
(9)
as well as the converse
*
*
pt = n-w lim p ( t / n ) . . . p ( t / n )= n+m lim (p(t/n))"
(10)
Therefore, by (9) and (10) we obtain that ( ~ ( t ) H ) ~(pt)t>o , ~ is bijective, mapping continuous M-semigroups on V (resp. on to continuous convolution semigroups, and vice versa. Self-decomposable laws v are limits of M-semigroups: v = limt,, p ( t ) . Moreover, these limits exist iff logarithmic moments exist. [ For vector spaces this result is well known, cf. e.g. [17], 3.6, [24]. For the the corresponding result in the group case see e.g. [13], 52.14, [12], and the references mentioned there. For further discussions of self-decomposable laws on non-Abelian Lie groups see [20], [26]. ] Theorem C is proved.
a)
3.7 Remarks. a) According to the Lie Trotter formula, convergence in (7) (10) is uniform on compact subsets of R+, whence replacing n by the subsequence [nt] we obtain c.g. for t 2 0 At =
iil(Pl/n * vl/n)
[ntl
(7')
and therefore
~ ( t=),l~l (pl/n * Tl/n (PI/.)
*
* . . * T([ntl-l)/n ( ~ 1 / n ) )
= n-w lim 1 9 9 4 ?J-l)/n (pl/n)
(90
b) The stable hemigroup (pL,,t)05sst defined in Proposition A, a) is immediately extended to the whole time axis by pLB,t := T, ( p ( t - s ) ) , --OO < s 5 t < co
94
On the other hand, put Tt:=T-t, t 2 0 and consider thc convolution semigroup generatcd by B - P (cf. notations of 3.6). Then wc have for * . t 2 0: At = P (t)C3 E - t . Comparing (9’) for (At) and resp. ( p ( t ) )and
(it)t2o
(lt)
(L
L
* Tl/n (P1jn) * ( t ) ) yields (t)= n+m lim T-t ( q n t ] - l ) / n If {T, ( p t ) } commute, in particular for vector spaces, we have
. .. * P l / J .
L ( t )= p-t,o =: p (-t) = T-t ( p ( t ) )
(11) 3.8 Example. Let ( p t ) bc a strictly (operator) stable continuous convolution i.e., S, ( p t ) = pr t, where (S,) denotes a semigroup on V Rd (rcsp. on N), onc-parameter group of automorphisms with multiplicativc parameterization, S,S,, = S,.,,, r,r‘ > 0 Let Tt := Set denote the corresponding additively parameterized group. Then by a simple calculation wc obtain in (9’) and (11): p
( t )= pet-l, p (-t) =
for t 2 0 and p8,t =
for s 5 t.
I.e., the distributions of the corresponding Ornstein-Uhlenbeck process coincide up to a deterministic time transformation - with the distributions of the stable background driving process. If ( p t ) is not strictly stable, S, ( p t ) = p, t * E,(,.,t) - in case of the group N we assume in addition that the shifts commute with thc probabilities ps - then obviously, a(r,t ) = t a(.) for some continuous function r H a(.) E V (rcsp. N). (In fact, a(.) is defined by a one-parameter group of affine transformations). In this case we obtain as above -
p ( t ) = petpl
* qt), with shift term b@) =
a (e‘) ds, t 2 0.
Similar representations are obtained for p(-t) and for ps,t, s 5 t . 3.9 Remarks. a) For vector spaces 3.6 ff is well known. The proof relies on random integral representatzons, cf. [17], 3.6. (For Banach spaces see e.g. [18].) In this situation, it can be shown via Fourier transforms that (9), (9’) follow by this random intcgral representation. The converse (lo), and hence the bijection between background driving processes and Ornstein Uhlcnbeck processes is also included in [17],cf. 3.6.10, however formulated in a different way. For the real line sce [16]. b) For non Abelian groups N , up to now there is no random integral representation available. Therefore we presented a proof for both - V and N - relying on the Lie Trotter formula. Note that (9), i.e. the mapping ( p t ) H ( p ( t ) ) ,is already observed in [ll],[13], the converse observation (10) seems to be new (at least for groups). c ) For finite dimensional V,[a] contains a generalization, a random integral representation for semi-selfsimilar processes (with a discrete group of operators). Up to now it was not possible to obtain a counterpart for groups N: A Lie Trotter formula is not applicable and on the othcr hand thcrc docs not exist any suitable random integral representation. This problcm remains open.
95
It is worth to note that the idea of this space-timc representation may be useful in different situations. We mention shortly the following example of a new limit theorem: 3.10 Proposition. Let ( p ( t ) )be a M-semigroup (on V or N ) with corresponding background driving process (pt) and continuous convolution semigroup (At = p ( t ) 8 Et) on G‘.Let again A and B denote the generating functionals of (A,) and (pt) respectively. and k, 00, a, > 0 with a,. kn --+ 1. Let vn E !7Jt1(V) (resp. nl(n/)) Then the following assertions are equivalent:
(i) vn * Ta,(vn) * ’ ’ ‘ * T((kn.t]-l).a,(vn) p(t), (ii) k,. (v, - E,) + B (iii) vih” 41 p t , t 2 0 +
20
+
(iv) (v, 8 tl xt, t 2 o (v) kn . (vn 8 &an - Ee,O) A Proof: Convergence of discrete convolution semigroups is equivalent with convergence of generating functionals, see e.g. [13],2.0.14. Therefore, applying this to Q and V (resp. Id)we obtain (ii)* (iii) and (iv)%(v). But (i) is nothing but the first component in (iv), and the second component follows by assumption since a, [k, . t] + t . Whence ( i ) H (iv). Furthermore, we have (v)+ (ii), and conversely, as easily seen, (ii) and a, . [k, . t] + t imply (v). --+
-+
Note that 3.10 shows that we not only have 1-1-correspondences between M-semigroups ( p ( t ) ) and semigroups ( p t ) but - as (i) @(iv) shows - also a 11-correspondence between limit laws. It seems worth to reformulate this result in terms of random variables and processes, e.g. for thc vector space case : 3.11 Proposition. With the notations introduced in 3.10, let (&) denote a generalized Ornstein Uhlenbeck process corresponding to ( p ( t ) ) and and (qt) a background driving LBvy process corresponding to ( p t ) , both taking values in v. Let furthermore x n , k : 1 5 k 5 k,, be an infinitesimal array of row-wise iid random variables and let v, denote the distribution of the variables Xn,k in the nthrow. Then
(Reformulation of ‘(i)
* (iii)’ in 3.10.) References
[11 Becker-Kern, P.: Stable and semistable hemigroups: Domains of attraction and selfdecomposability. J. Theoret. Probab. 16,573-598 (2003).
96
[2] Becker-Kern, P.:Random integral representation of operator-semi-self-similar processes with independent increments. Stoch. Proc. Appl. 109,327 - 344 (2004). [3] Boge, W.: Zur Charakterisierung sukzessiv unendlich teilbarer Wahrscheinlichkeitsverteilungen auf lokalkompakten Gruppen. Z. Wahrscheinlichkeitstheorie venv. Geb. 2, 380 - 394 (1964). [4] Bogachov,V.I., Rockner, M., Schmuland, B.: Generalized Mehler semigroups and applications. Probab. Theory Rel. Fields 105, 193 - 225 (1996). [5] Burrell, Q., McCrudden, M.: Infinitely divisible distributions on connected nilpotent Lie groups I. J. London M.S. 11. Ser. 7, 584 - 588 (1974). [6] Burrell, Q., McCrudden, M.: Infinitely divisible distributions on connected nilpotent Lie groups 11. J. London M.S. 11. Scr. 9, 193 - 196 (1974). Dani, S. G., McCrudden, M.: Embeddability of infinitely divisible distri[7] butions on linear Lie groups. Invent. Math. 110, 237 - 261 (1992). [8] Fuhrmann, M., Rockner, M.: Generalized Mehler semigroups: The nonGaussian case. 3. Potential Analysis 12, 1 - 47 (2000). Hazod, W.: Stetzge Faltungshalbgruppen von Wahrscheanlichlceitsmaj?enund [9] erzeugende Distributionen. Lecture Notes in Mathematics 595 (1977). [lo] Hazod, W.: Probabilities on groups: Submonogeneous embedding and semistability. In: Contnbutaons to Stochastzcs. W Sendler ed. Physica Verlag, Heidelberg, 164-174 (1987). [ll] Hazod, W.: On some convolution semi-and hemigroups appearing as limit distributions of normalized products of group-valued random variables. In: Analysas on znfinate-damenszonal Lze groups, Marsezlle (1997), H. Heyer, J. Marion ed. 104 - 121. World Scientific (1998). [12] Hazod, W., Scheffler, H-P.: Strongly 7-decomposable and selfdecomposable laws on simply connected nilpotent Lie groups. Mh. Math. 128, 269 - 282 (1999). [13] Hazod, W., Siebert, E.: Stable Probabalaty Measures on Euclzdean Spaces and on Locally Compact Groups. Structural Propertzes and Lamzt Theorems. Mathematics and its Applications vol. 531. Kluwcr A P. (2001). [14] Heyer, H.: Probabzlaty Measures on Locally Compact Groups. Berlin-Heidelberg-New York. Springer (1977). [15] Heyer, H.: Reccnt contributions to the embedding problem for probability measures on a locally compact group. J. Mult. Analysis 19,119 - 131 (1986). [16] Jeanblanc, M., Pitman, J., Yor, M.: Selfsimilar processes with independent increments associated with LBvy and Bessel processes. Stoch. Process. Appl. 100, 223 - 231 (2002). [17] Jurek, Zb., Mason, J.D.: Operator Lzmat Dastnbutaons an Probabzlaty Theory. J. Wiley, New York (1993). [18] Jurek, Zb., .Vervaat, W.: An integral representation for self-decomposable Banch space valued random variables. Z. Wahrsch. vcrw. geb. 62,247 - 262 (1983). [19] Krakowiak, W.: Zcro-one laws for A-decomposablc measures on Banach spaces. Bull. Pol. Acad. Sci. Math. 33, 1-2 (1985).
97 [20] Kunita, H.: Stochastic processes with independcnt increments on a Lie group and their selfsimilar properties. In: Stochastic d
W. Hazod, Department of Mathematics, University of Dortmund, D-44221 Dortmund, Germany e-mail: wi1fried.hazodQmath.uni-dortmund.de
This page intentionally left blank
Infinite Dimensional Harmonic Analysis I11 (pp. 99-118) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
On infinite divisibility and embedding of probability measures on a locally compact Abelian group By HERBERTHEYER(Tubingen) and GYULAPAP (Debrecen)
1
Introduction
In the present note the authors supplement significant properties of infinitely divisible and embeddable probability measures on a locally compact Abelian group G. There are at least two versions of infinite divisibility appearing in the literature which deserve special attention, and the problem of embedding those measure leads directly to the study of continuous convolution semigroups on G. It is evident from the classical setup of G = Rd for d 2 1 that in this context Gaussian semigroups and measures play a favorite r61e. The main result of Section 3 concerns the representation of Gaussian measures in terms of their characteristics and their relationship to Gaussian measures in the sense of Parthasarathy. Section 6 and 7 deal with the embedding of infinitely divisible probability measures in the weak or strong sense respectively. We present direct proofs to more or less known statements, but stress the irreplacable hypothesis that the dual 6 of G is arcwise connected. In a concluding Section 8 we initiate the study of Gaussian and diffusion hemigroups on G and their analysis, especially for 1-dimensional connected Abelian groups.
2
Preliminaries
Let G be a locally compact group with identity element e. Let m'(G) denote the semigroup of probability measures on G. For every z E G, E, denotes the Dirac measure in z. For a compact subgroup H of G, W H denotes the normalized Haar measure on H. 2.1 Definition. Let H be a compact subgroup of G. A family (pt)t>o in m'(G) (indexed by positive real numbers) is called an H-continuous convolution semigroup in fnz'(G) if ps * pt = p,+t for all s, t > 0 and limtp pt = W E .
A family (pt)t>o in ?M'(G) is called a continuous convolution semigroup in fnz'(G) if ps * pt = for all s, t > 0 and the limit po := limao pt exists in m'(G). The limit measure p~ is an idempotent element of m'(G) and hence (&>o is an H-continuous convolution semigroup in fnz'(G) for some compact subgroup H of G. This research has been supported by the Deutsche Forschungsgemeinschaft and the Hungarian Academy of Sciences.
100 2.2 Definition. A continuous convolution semigroup ( u , ) , , ~ i n m ' ( G )
Gaussian semigroup if
is called a
I lim-v,(G\N)=O
t for all Borel neighbourhoods N of e. A measure u E %T'(G) is said to be a Gaussian measure if there ezists a Gaussian semigroup (v&o in m'(G) with v1 = v. tl0
The class of Gaussian measures in m l ( G ) will be abbreviated by B(G), and the subclass of its symmetric elements b y B'(G). If (v&o is a Gaussian semigroup in ?JJZ1(G) then limGo vt(G \ N) = 0 for all Borel neighbourhoods N of e, hence vo(N) = 1 for all Borel neighbourhoods N of e with vo(aN) = 0. If G is metrizable then this implies vo = E,, thus (vt),>o is an { e}-continuous convolution semigroup. The above definition slightly differs from the Definition 6.2.1 in Heyer [6], since here a belonging to a continuous Gaussian semigroup may consists of Dirac measures (e&O one-parameter subsemigroup (x&o of G. It is known that for a Gaussian semigroup (vJt>o in m'(G) we have supp(vt) C Go for every t > 0, where Go denotes the connected component of the identity e (see Heyer [6, Theorem 6.2.31). Moreover, if G is a locally compact group and G # { e } then B(G) \ D(G) # 6, where D(G) := {e, : x E G} (see Heyer [6, Theorem 6.2.81). Consequently, if G is not totally disconnected and G # { e } then E(G) \ D(G) # 0, and if G is totally disconnected then B(G) = { e e } . 2.3 Definition. A measure p E m ' ( G ) is called infinitely divisible if for all n E M there exists p~ E Im'(G) such that p = (PA)" (and hence p i can be considered as an The set of all infinitely divisible measures in DI'(G) will be denoted b y n-th root of
/I").
Z(G). 2.4 Definition. A measure p E m ' ( G )
is called continuously embeddable if there exist a continuous convolution semigroup (pt),>o in m l ( G ) such that p = p l . The set of all continuously embeddable measures in ?YJT'(G) will be denoted by &(G). Clearly B'(G)
3
c 5(G) c &(G) c Z(G).
Gaussian measures
Let G be a second countable locally compact Abelian group. The dual group of G will be denoted by We will define Gaussian measures in the sense of Parthasarathy by their Fourier transforms.
e.
3.1 Definition. The Fourier transform jl :
defined by
--t
C of a measure
p E ml(G)
is
101 The mapping p ++ jl from EX1(G) into the set of mappings 3.2 Definition. A continuous function $ : 6
if $(XlX2)
+ $(XlXZ')
The set of all quadratic forms on
= 2($(x1)
6
+ R+
6+C
is injective.
is called a quadratic form on
+NXZ))
f o r all
x1,
6
xz E 6.
will be denoted b y q+(6).
3.3 Definition. A measure v E m'(G) is said to be a Gaussian measure in the sense of Parthasamthy if there exist an element m E G and a quadratic form $ E q+(6) such that
G(x) = x(m) exp{-$(x)}
holds for all
x E 6.
The class of Gaussian measures in the sense of Parthasarathy i n m'(G) will be abbreviated b y Bp(G), and the subclass of its symmetric elements by 5i(G). 3.4 Definition. A measure p E m'(G) is called weakly infinitely divisible i f for all n E N there exist a measure pn E im'(G) and an element x , E G such that p = &*c,,. The collection of all weakly infinitely divisible measures in m'(G) will be denoted by &(G).
Obviously Z(G)
c &(G).
3.5 Remark. For a bounded positive measure m'(G) with exponent T is defined by
T
on G, the Poisson measure eT-T(G)EeE
A measure v E m'(G) is a Gaussian measure in the sense of Parthasarathy if and only if it *X is weakly infinitely divisible and if for every factorisation of v of the form v = eT--T(G)EC with a bounded positive measure r on G and a weakly infinitely divisible probability measure X one has r = ace for some a E R+. If v E Bp(G) then an element m E G and a quadratic e m $ E q+(6) with the property such that C(x) = x(m) exp{-$(X)} holds for all x E G are uniquely determined. Moreover, if m E G and $ E q+(6) then there exists v E Bp(G) such that C(x) = x(m) exp{-$o()} holds for all x E 6. (See Theorems 5.2.7 and 5.2.8 in Heyer [S].)
For m E G and $ E q+(6) let vm,$ E m'(G) be defined by
Then
102 3.6 Definition. An element x E G is called continuously embeddable i f there exists a continuous one-parameter subsemigroup (x&>o in G such that X I = x.
T h e set of continuously embeddable elements in G wall be denoted by GE. Since G is a locally compact Abelian group, GE is a subgroup of G. 3.7 Theorem. Let G be a second countable locally compact Abelian group. T h e n = {vrn,$ : m E GE,
Consequently, GS(G)C B:(G) n G(G) and
II, E q + ( Q ) .
c B;(G)
U G(G) C Bp(G) C &(G).
Proof. If v E B(G) then there exists a Gaussian semigroup (vt)t>~ in ??Jtm'(G) with v1 = v. Consequently, for all t 2 0 we have that vt = v : / ~and lim,+,m nvt/,(G \ N ) = 0 for all Borel neighbourhood N of e. By Theorem 5.4.3 in Heyer [6],we obtain that vt E Gp(G) for all t 2 0. Thus there exist mt E G and $t E q+(@ such that
-vt(x)
for all
= x(mt) exp{-$&))
xE
and t 2 0. A
The semigroup property v,+~ = v,*vt for s , t thus
2 0 implies cs+t(x) = F,,(x)ct(x) for x E G,
A
vs+t(x) = x(ms+t)exp{-$,+t(x)} = x ( m J exp{-$Jx))x(mt) exp{-$t(x)) = x ( m 4 ~xP{-($,(x)
+ $t(x))).
Consequently m,+t = msmt for all s, t 2 0. Obviously mo = e. Moreover
n
is a continuous mapping from R+ into CC for all x E G, hence (mt)t)o is a continuous oneparameter subsemigroup in G. Consequently v = vml,+l and ml E GE. Suppose now that v = vm,+ with m E GE and $ E q+(z). Then there exist a continuous oneparameter subsemigroup (mt)t)o in G such that ml = m. For all t 2 0 there exists a measure vt E Gp(G) such that ct(x) = x(mt)exp{-t$(x))
for all
x E E.
Clearly is an {e}-continuous convolution semigroup in m'(G) such that vl = v. Consider the related continuous convolution hemigroup ps,t := vt-,, 0 s t. Let (Xt)t)o be an associated (timehomogeneous) increment process in G. By Theorem 2.6 in Bingham and Heyer [4] we conclude that (Xt)t)o has a. s. continuous paths. By Corollary 2 of Theorem 2 in Siebert [15] we obtain that (ps,t)o<s
< <
for all Borel neighbourhood N of e and for all T > 0. This implies that (vt)t>o is a Gaussian semigroup. Indeed, for all E > 0 there exists b > 0 such that C;==, p7h-l,Tk(G\ N ) E for all decomposition Z := (0 = TO < 7 1 < ... < 7, = 1) of [0,1] with 1Z1 < 6. Hence nvt(G \ N ) < E if t < 6 and nt 1, which clearly implies that limfvt(G \ N ) = 0. t+O Consequently v E G(G). 0
<
<
103
Continuously embeddable elements
4
Let again G be a second countable locally compact Abelian group. First observe that an element of G is continuously embeddable if and only if the related Dirac measure is a continuously embeddable measure, i.e., z E GE if and only if E, E &(G). Consequently, GE = {z E G : E, E &(G)} and {E, : z E G E } C B(G). Moreover, by Theorem 3.7, B(G) = Bp(G) holds if and only if {E, : z E G} c &(G). Further, note that each continuous one-parameter subsemigroup (zt)t20 of G can be extended to a continuous one-parameter subgroup (z&a of G by zt := (z-t)-I,
t < 0.
Hence GE = U{LP(R): LP E Hom(B, GI)
C
Go,
where Horn@, G) denotes the set of continuous homomorphisms from the additive group W into G. Consequently, if G is not connected then GE # G, and hence B(G) # Bp(G). Moreover, GE is a dense subgroup of Go, and GE equals to the union of the arcs of G which contain e, i.e., GE is the arc-component of e. Combining Theorem 3.7 with the results of Dixmier [5] we conclude the following theorem. 4.1 Theorem. Let G be a second countable locally compact Abelian group. Then the following statements are equivalent:
(i) B(G) = BP(G); (ii) GE = G; (iii)
{ E :~ z
E G} c E(G);
(iv) G is the union of all continuous one-parameter subgroups of G; (v) G is arcwise connected; (vi) G = W" x TI, where n E Z+ and I is an at most countable set.
C,
Note that each solenoidal group C, is a compact connected Abelian group with (C,)& # (see Example 5.4).
5 5.1
Examples Gaussian measures on Bd
Consider the group G = Rd, where d E N. The character group (Wd)' is topologically isomorphic with Rd. Every continuous character of Rd has the form xY(z) = ei((ZBY), z E Rd for some y E Wd. Every quadratic form II, on (Rd)"E Rd has the form
104 l(lB(xY)= (By, y) for some symmetric positive semidefinite matrix B E Rdxd. Clearly (Rd)&= Rd,hence
G(lRd) = Gp(Rd) = {v,,B : a E Rd,B E Rdxd with BT = B and B 2 0 } ,
G'(lRd) = ($(ad) = {VO,B : B where
V,,B
BT= B and B 2 0 } ,
E Rdxd with
E ?3JX1(Rd) is defined by
for all y E R~ 2 (R~ )".
, 5.2
Gaussian measures on Td
<
Consider the group G = T d , where T := {eZair: 0 z < 1) denotes the torus group. The character group (Td)" is topologically isomorphic with Zd. Every continuous character of Td has the form x m ( z )= nf=,zT, z = (XI,.. . ,zd) E Td for some m = (ml,.. .,md) E Zd. Every quadratic form l(l on (Td)" z Zd has the form $JB(x,) = (Bm,m) for some = Td, hence symmetric positive semidefinite matrix B E Rdxd. Clearly (p)&
G(Td)= Gp(Td) = {v,,B G'(Td) = where
V,,B E
:a E
Td, B
E Rdxd with BT = B ,
B 2 0},
= {v,,B : a E Td with a; E (1, -l}, B E Rdxd with BT = B , B
2 0},
?3JX'(Td)is defined by
. . ,eZniad) with Writing the element a = ( a l , . . . ,ad) E T d in the form a = (ezffioll,. 0 a k < 1, k = 1 , . . . ,d, we obtain x,(a) = e2ai(ol,m)with a := ( a l , .. . , a d ) , and
<
, 5.3
for all m = (ml,. . . ,md) E Zd z
(T~)".
Gaussian measures on Tm
Consider the group T" := n E I G k , where Gk := T for all k E N. The character group (T")" is topologically isomorphic with idm)consisting of the elements of Zoo having only finitely many nonzero coordinates. Every continuous character of Tm has the form xm(z)= z = (zl,zz,...) E Tm for some m = (ml,mz,...) E ~ ( " 1 . ~ ) bj,kmjmk Every quadratic form l(l on (T")" Z(") has the form l ( l , ~ (=~C;=, for some symmetric positive semidefinite matrix B = ( b l , k ) j , k = 1 , 2 , . . . E RmX" (i.e., Ed := (bj,k)j,+],,,,,dE Rdxd is positive semidefinite for all d E N). Clearly (T")E = IP", hence
nZlzT,
G(Tm) = Gp(Tm) = {Y,,B GS(Tm)= QE(T") = {v,,B
:a
E T", B E RmXm with BT= B , B 2 0 } ,
:a E
T"
with ai E (1, -l}, B E RmXmwith BT= B , B 2 O } ,
105 where v , , ~ E ml(Trn) is defined by
for all m = (rnl,mz,.. .) E Z(rn) (Ti'")". . .) with 0 in the form a = (e2zia1,e2*iaz,. exp ( 2 x i CEl f f k m k } , and
for all m = ( m l ,m2,. . .) E
5.4
Writing the element a = ( a l , a 2 , .. .) E Trn k = 1 , 2 , . . ., we obtain x,(a) =
< f f k < 1,
z ( ~2)(T")".
Gaussian measures on solenoidal groups
Let c = ( c l ,c 2 , . . .) E Nm with a subgroup of T", namely,
ck
2 2 for k
C, := {x = (21,22,. . .)
E N.
The solenoid C, can be considered as
E Ti'm : Z k = Xp+l
for all k E N}.
l for j < k , where Thus for x = (z1,x2,. . .) E C, we have x 3. -- z ~ c J + l " ~ c-k -xCli7*) cb,k) := c j c j + l . . . Ck-1 for j < k and c[k,k) := 1. In fact, C, is the projective limit of the projective sequence (Gk,T j , k ) , 1 < j < k , where Gk := T for all k E N,and xjvk : Gk -+ G j is the mapping z e z C U , k ) . It follows that C, is a second countable compact connected
group. It is well known that C, is not locally connected and not arcwise connected. The arc-component of the identity e = (1,1,. . .) of C, (i.e., the subgroup of continuously embeddable elements of C,) is
(Cc )&
--
{ (eznix, e2rrixlcl
If m = (ml,m 2 , . . .) E Z(co)2 (T")" j > k, hence k
e2zix/(clcz), . . .) : x E "} .
then there exists k E N such that
mj
= O for all
k
Consequently, the character group group
{
2,is topologically isomorphic with the (discrete) additive e y = -: l E Z, k E M cIl,k)
I
of rational numbers. Every continuous character of C, has the form x.,(x) = x:, x = (x1,x2,.. .) E C, for some y = C E Z, k E N. Every quadratic form $J on C, has the form $Jb(y) = by2 for some b 2 0. Hence
&,
106 where
va,b E
mZ'(C,) is defined by
e
-
Note that there exists an element a E C, with ai E {I, -1) and a # e if and only if there are only at most finitely many even numbers among {ck : k E N}. Moreover, if a E (CJE with a, E (1, -1) then a = e.
If u E ( C c ) ~then_there exists X E R such that a = (ezrix, $ n i X / c l , e2niX/(Clc2),".), hence for y = -C E C, we have x7(a)= atk -- eznixt/c[l.k) = e2n*X7. Consequently, c[l,k)
Fa,b(XT) = exp
{
2 n i ~y -b2 2l l
,
for all y E 5,.
<
Now let a = ( a l l az, . . .) E C. Write ak E T in the form a k = eZaiak with 0 a k < 1. Since a1 = a;, there exists an integer j , E Z such that a1 = c l a z - j , , hence az e2Ti(e1+j1)/c1. In a similar way one obtains ak = ~ z n i ~ ~ l + ~ l + ~ ~ ~ + C [ l , r - ~ ) ~ k - l ~ / C [for l,k) all k E N with some j = ( j l , j z , . . .) E Zm. Hence for y = -4E C, we have Cll.k) x7(a) = a: = ezri(al+~l+"'+c[l,k-l)~k-l)'/C[l,k) = e2?ri(oll+jl+...+C[l , k - l ) j k - 1 ) 7 , Consequently, 1
,.
va,b(x')
{
= eXP 27TZ(Cr1 -k j i
+ . . . + C[l,k-l)jk-i)Y - -by 2
9
,
for all y E
5,.
Note that each Gaussian measure on C, in the sense of Parthasarathy has full support and is singular with respect to a Haar measure on C,. Moreover, two Gaussian measures va,b, v,',b' E Gp(2,) in the sense of Parthasarathy are mutually absolutely continuous if a-la' E (C)c, otherwise they are singular with respect to each other. Particularly, any two Gaussian measures VaJ,, va!,b! E p(Cc) are mutually absolutely continuous. (See Bendikov and Saloff-Coste [2, 5.21.)
6
Weakly infinitely divisible measures
The purpose of this section is to characterize the set of locally compact Abelian groups G for which each weakly infinitely divisible measure is embeddable, i.e., G(G) = &(G).
6.1 Theorem. Let G be a second countable locally compact Abelian group. Then the following statements are equivalent: (i) G is arcwise connected; (ii) &(G) = Z(G) = &(G).
Proof. (i) ==+ (ii). Let us consider an arbitrary measure /I E &(G). We are going to show that /I E &(G). Let g : G x 6 + R be a local inner product (see Definition 5.1.7 in Heyer [S]). Then by Corollary IV.7.1 in Parthasarathy [13], the Fourier transform F admits a representation
107 for all x E 6,where H is a compact subgroup of G, m E G, 11 E q+(6) and q is a LBvy measure on G, i.e., q is a positive Borel measure on G such that q ( { e } ) = 0, q(G \ N ) < 0;) for all Borel neighbourhood N of e, and (1 - ReX(z))q(dz) < 0;) for all x E G. Since G is arcwise connected, there exists a continuous oneparameter subsemigroup (mt)t)oin G such that ml = m (see Theorem 4.1). Clearly tll, E q+(6) and tq is a Levy measure on G for all t 2 0. By Theorem IV.7.1 in Parthasarathy [13], for all t 3 0 there exists a measure pt E Z(G) such that
s,
A
k ( x )= $ H ( X ) X ( m t ) for all
{
exp -t'$(X)
-b t
/
G
( x ( z )- - ig(z,
x))7)(dz)
x E 6.We have 1 if x ( z ) = 1 for all z E H, =
0 otherwise.
Hence (In fact, the set { x E 6 : x ( z ) = 1 for all z E H } is the annihilator of H in 6.) we have GH(x)' = DH(x) for all x E 6.Moreover, for each s,t > 0 and x E G, we have x(ms)x(mt)= x(m,mt) = x(m,+t),Athuswe can conclude iis(x)Pt(x)= Ps+t(x). Further, limt$oj&(x)= ~ H ( x ) for all x E G implies limtlopt = w H . Consequently (pt)t>O is a H-continuous convolution semigroup in ER'(G) with 111 = p, and we obtain p E E(G). A
(ii) + (i). If G is not arcwise connected then consider an element z E G such that 0 E, E Zo(G) but E, # E(G), hence &(G) # E(G).
z $? GE (see Theorem 4.1). Then
7 Embedding property A locally compact group G is said to satisfy the embedding property if Z(G) = E(G). If G is a locally compact Abelian group such that any p E Z(G) is root compact (i.e., the root set R ( p ) := U n E ~ { v r n: v E ER'(G) with vn = v, 1 < m < n } is relatively compact in ER'(G)) then by Theorem 3.5.12 in Heyer [6] the following statements are equivalent: (i) Go is locally arcwise connected; (ii) Z(G) = E(G). We note that by Theorem 3.5.12 in Heyer [6] for any locally compact Abelian group G the following statements are equivalent: (i) any 11E Z(G) is root compact; (ii) the set of all compact elements of G is compact, and the set of divisible elements of G equals Go. In view of the above results one might conjecture that a locally compact Abelian group satisfies the embedding property if and only if any p E Z(G) is root compact and Go is locally arcwise connected, but such a result is false, as it is shown in Heyer [6, 3.5.211.
108 We can give a sufficient condition for a locally compact Abelian group in order to satisfy the embedding property in the following way. One can extend the statement of Theorem 4.1 along Dixmier [5] (see Heyer [6, Theorem B]). From this source follows that the class of second countable locally compact Abelian groups G with arcwise connected dual G consists of Abelian aperiodic groups which can be represented as closed subgroups of a locally convex vector space, that is as products R" x
z'
with a countable set Z (of cardinality IIl), embeddable into Rn+lIl
(Note that a second countable locally compact Abelian group G is aperiodic if and only if G = Rn x K with n 0 and with a connected and compact K such that k is torsion-free, see Heyer [6, 3.5.15, 3.5.181.) From Siebert's paper [14, 6, Satz 41 follows that this closed subgroup of the vector space is strongly root compact. In fact, if 111 is infinite, then there exists a closed, discrete, free subgroup N of Z' which is uniformly root compact such that G = N x R".
>
Now, apply Theorem 3.1.11 in Heyer [6], and one gets the uniform root compactness of G. From this it follows by Theorem 3.5.8 (or Corollary 3.5.9) of Heyer [6] that for a second countable locally compact Abelian group G with arcwise connected dual we have Z(G) = &(G), and for each p E Z(G) there exists a unique {e}-continuous convolution semigroup (pt)t>o with p1 = p . The aim of this section is to give another proof of the above statement.
7.1 Lemma. (Existence and uniqueness of the logarithm) (i) Let S be a locally arcwise connected topological space, and let (f n ) n ) l be a sequence of continuous mappings f n : S t C \ ( 0 ) satisfying f,"+l= f n f o r all n 1. T h e n there exists a continuous function h : S t C with f l = exp h.
>
(ii) Let S be a connected topological space, a E S and f : S t C \ ( 0 ) a continuous mapping with f ( a ) = 1. T h e n there exists at m o s t one continuous function h : S + C satisfying f = exph and h ( a ) = 0.
Proof. (i). Since every connected component of S is open, we may assume without loss of generality that S is connected. We fix a E S and assume that f n ( a ) = 1 for all n 1. Let H := { z E C :R e z > 0 )
>
and let Un denote the connected component of {s E
s : f n ( s )E H )
which contains a. Then is an increasing sequence of connected open subsets of S. For every continuous arc y : [0,1] t S with y(0) = a and sufficiently large n we have
109
u,>l
y([O, 11) c U, or S = U,. Indeed, there exists a continuous function 'p : [0,1] -+ C with f l o y = exp 'p. But then, for large n and 1/, := 2'-"'p the function f n o y = exp I) takes values in H . Now let log denote the principle branch of the logarithm on H and define h, on U, bY hn(S) := 2"-llog f n ( s )
for each n 2 1. The function h, is continuous on U, and coincides with h,+l on U,. Thus there exists a function h : S + C satisfying Resu,h = h, for all n 2 1. The function h is the desired logarithm. (ii). Let h and h' be two continuous functions S -+ C satisfying
f = exph = exph' and h(a) = h'(a) = 0. Then the continuous function h' - h takes on only values in 2 r i Z c3 and is therefore constant. Thus h' = h, since h'(a) - h(a) = 0. 7.2 Corollary. Let S be a connected and locally arcwise connected topological space, and let a E S . Moreover, let f : S -+ a2 \ ( 0 ) be a continuous mapping with f ( a ) = 1. W e assume that for every n 2 1 there exists a continuous mapping gn : S + C \ (0) such that 9," = f . Then there exists exactly one continuous function h on S with h ( a ) = 0 satisfying f = exp h. 7.3 Theorem. Let G be a second countable locally compact Abelian group with arcwise connected dual 6. Then
Z(G) = E(G). Moreover, for each p E Z(G) there exists a unique {e}-continuous convolution semigroup (Pt)t>O with P1 = P. Proof. Let us consider an arbitrary measure p E Z(G). Since 6 is supposed to be arcwise connected, 6 is connected, hence its dual G is aperiodic (see Hewitt and Ross [ll, (24.19)]). The aperiodicity of G implies that the only compact subgroup of G is { e } . By the representation (6.2) we conclude that G(x)# 0 for all x E 6. From the hypotheses of the theorem it is clear that Lemma 7.1 applies and that therefore a continuous branch of logp with log G( 1) = 0 exists, where 1 denotes the identity element of the dual group which is the constant function 1. Consequently we have
(ZgX)), thus
= X X ) = exp {logG(x)},
x E G,
110 Hence the n-th roots p i of p are uniquely determined. For each r = let
E
Q
with r > 0
hence the notation pT := p: is justified. Clearly
*
= iir(X)ii+(x) = exp{(r
( ~ r prl)-(x)
+ r’) logii(x)) = iir+rr(x),
xE6
for all r, r’ E Q with r, r’ > 0, hence pr * p+ = pr++ for all r, r’ E Q with r, T’ Defining po := eer we obtain a rational convolution semigroup (P,),~Q+ in m’(G). Next we show that the mapping T
from Q+ into m’(G) is with
7”- and
lim
n+m
> 0.
* Pr
.r,-continuous. Note that for pnrv E m’(G) (n E N)
iinM
x
= P(X),
E
z,
the continuity theorem implies that 7”-lim pn = v. n-tm
Let (rn)+1 be a sequence in Grit.,(x)
Q+ with r, -+ r E Q+.Then
= ~ X {Pr n log P(x)) + ~ X {P r log iiM) = P ~ ( x )
uniformly on compact subsets of 7”-lim n+m
6.This implies
pr, = pr,
hence
T ~ -lim n+co
pr,, = pr.
Finally we prove that extends to a unique {e)-continuous convolution semigroup (p&w+ in m’(G), and hence p = p1 E E(G). Let T + t E W+ with r E Q+. Then
i i r M -+ exp {tlogii(X)),
x E 6.
The limiting function
x
b-+
exp {t 1%
ii(x))
from 6 into C equals 1 at x = 1 and is a continuous positive definite function. Hence for every t E R+ there exists a unique measure pt E m’(G) such that iitit(x)= exp {tlog XX))
1
x E 6.
111 Clearly
for all t ,t' E R+. The uniqueness of a continuous for all t, t' E R+, hence pt * pt, = embedding semigroup (&R+ follows from the T,-continuity of the mapping
r
pr
from Q+ into 9J11(G). 7.4 Theorem. Let p be an infinitely divisible probability measure on a locally compact Then there exists a unique Abelian group G with arcwise connected dual group c?. continuous negative definite function 'p 5 + CC such that 'p(1) = 0 and
XX) = exp{-cp(x)},
x E (2.
d x ) = - logii(x),
x E 2,
In fact, and there exists an element m E GE, a quadratic form 1/1 E q+(@ on G such that
d x ) = - logx(m) + 44x1 -
s,
(x(4
-
and a Le'vy measure 77
1 - i d x , x)) rl(dz),
x E G^.
7.5 Remark. The converse of this statement is obvious.
By Schoenberg's theorem (see Berg and Forst [3, p.49, Theorem 8.31) a family 1) is an {e}-continuous convolution semigroup if and only if there exists a continuous negative definite function 'p : -+ CC such that xE6 X x ) = exP{-t'p()o),
Proof.
(pt)t>O of measures in 9J11(G) ( p 2 0, 11p11
<
for all t E R+. For t = 1 this gives us P(x) = exp{-cp(x)), But $(l) = 1 implies cp(1) = 0. Since
'p
x E (3.
is continuous, satisfies
XX) = exp{-cp(x)),
'p(1)
= 0 as well as
xEG
we see that 'p is the continuous branch of - log6 constructed above. The last statement 0 follows from Theorem IV.10.1 in Parthasarathy [13].
8
Gaussian and diffusion hemigroups
Let G be a locally compact group.
112 8.1 Definition. A family (ps,t)o+Gt in m ' ( G ) is called a continuous convolution hemigroup (briefly hemigroup) in E?X'(G) if ps,r * py,t = ps,t for all 0 s < r t , pt,t = E, for all t 2 0 , and if the mapping ( s , t ) ++ ps,t from 8 into m'(G) is continuous.
<
If (&o
0
<s
and
<
is an {e}-continuous convolution semigroup in m ' ( G ) then ps,t := vt-,, pt,t := t 2 0 define a continuous convolution hemigroup.
Now we recall the definition of Gaussian hemigroups (see Heyer and Pap [lo]).
A continuous convolution hemigroup ( p s , t ) o ~ 8 in ~ t m ' ( G ) is called a Gaussian hemigroup if
8.2 Definition.
n
for all Bore1 neighbourhoods N of e and for all T > 0. Corollary 2 of Theorem 2 in Siebert [15]implies that a convolution hemigroup (ps,t)oGsGt on a second countable locally compact group is a Gaussian hemigroup if and only if each associated left additive process has continuous paths with probability one. If G is a locally compact group Abelian group with countable basis of its topology then a family (ps,t)o+Ct in tm'(G) is a Gaussian hemigroup if and only if there exist a continuous function t tf mt from R+ into G with rno = e and a function t ++$t from R+ into q+(@ with Go = 0 such that t tf $&) is continuous for all x E G, $t - $s E q+(z) for all 0 s t , and such that
< <
iis,t(x)
1 = x(m,'mt) exp {-#(XI
-
M)}
(See Bingham [l],Heyer [6, Theorem 5.6.191 and Heyer and Pap [9]). Consequently, if (ps,t)oGsstis a Gaussian hemigroup in m ' ( G ) then ps,t E Bp(G), i.e., ps,t is a Gaussian measure in the sense of Parthasarathy for all 0 s t. Moreover, if (ps,t)oGsGt is a E Sp(G) for all 0 s t continuous convolution hemigroup in m'(G) such that then (ps,t)oGsGt is a Gaussian hemigroup (see Heyer and Pap [lo, 3.121).
< <
< <
8.3 Examples. If G = Rd then a family (ps,t)o+gt in m'(Rd) is a Gaussian hemigroup if and only if there exist a continuous function a : R+ + Rd with a(0) = 0 and a continuous function B : R+ + Rdxd with B ( 0 ) = 0 such that B ( t )- B ( s ) is symmetric positive semidefinite for all 0 s t , and such that
< <
(8.4)
i i s , t ( x v ) = exp
for all y E Rd 2 '')@I(
{
1
Z(Y,
44 - 4.))- 3 ( Y 7 (BW - B ( S i ) Y ) }
(see the notations in Example 5.1).
If G = Td then a family (ps,t)oGsct in %T'(p) is a Gaussian hemigroup if and only if there exist a continuous function a : R+ + [0, l)d with a(0) = 0 and a continuous
,
113 function B : R+ --f Rdxd with B(0) = 0 such that B(t) - B ( s ) is symmetric positive semidefinite for all 0 s t , and such that
< <
(8.5)
iis,t(xm)
= exp
{2 7 4 m , 4 t )
1 - 4.))- -2( m ,
(W- B ( s ) ) m ) }
for all m E Zd E (Td)" (see the notations in Example 5.2). If G = T" then a family (ps,t)oGs(t in ?lR'(T") is a Gaussian hemigroup if and only if there exist a coordinatewise continuous function a : R+ + [0, 1)" with a(0) = 0 and an entrywise continuous function B : R+ + RmX" with B(0) = 0 such that B ( t )- B ( s ) is symmetric positive semidefinite for all 0 s t , and such that
< <
iis,t(xm)= exp for all m E Z(") 2
{
d
1 "
2.lriX ( 4 t )- 4 s ) ) m- 5 k=l
1
(bj,dt) - bj,k(s))mjmk
j,k=l
(T")" (see the notations in Example 5.3)
If G = C, then a family (pLs,t)ogs
ii8,t(x7)= e w { i ( a ( t )
- - a ( s )+ j l ( t ) - j l ( s ) + . . . + C [ l , k - l ) ( j k - l ( t )
-jk-.1(s)))y
for all y E &)" (see the notations in Example 5.4). By continuity of j we obtain j ( t ) = 0 for all t 2 0, hence (8.6)
B s , t ( x 7 ) = exp
{
1
1
i ( 4 4 - 4 S ) ) Y - 5 ( b ( t )- W ) y 2
for all y E (C,)". Consequently, ps,t E B(C,), i.e., is a Gaussian measure for all 0 s t (not only in the wider sense of Parthasarathy). It is still an open question if this holds for all locally compact Abelian groups.
< <
It is not easy to check whether a given continuous convolution hemigroup is Gaussian. We would like to construct Gaussian hemigroups from simpler hemigroups by deterministic change of time. We recall the definition of diffusion hemigroups (see Heyer and Pap [lo]). 8.7 Definition. Let G be a locally compact group. A continuous convolution hemigroup (ps,t)ogsgt
in ?lR'(G) is called a dinusion hemigroup if lim
t-s+o
O<s
1
--ps,t(G
t-s
\ N)= 0
for all Bore1 neighbourhoods N of e and for all T > 0.
114 A diffusion hemigroup is always a Gaussian hemigroup, but there exist Gaussian hemigroups on R which are not diffusion hemigroups (see the example in Remark 4.5 of Bingham and Heyer [4]).
8.8 Definition. A function X : R+ + R+ is called a detewninistic change of time if it is strictly increasing and X(R+) = R+. The class of deterministic changes of time will be denoted by A. 8.9 Remark. If (ps,t)ogsGt is a convolution hemigroup in inz'(G) such that there exists X E A with 1 (8.10) lim ps,t(G \ N ) = 0 t-s-10 A(t) -A(.) O<s
for all Borel neighbourhoods N of e and for all T > 0 then (ps,t)oGs
for all Borel neighbourhoods N of e and for all T > 0, hence ( p x - ~ ( ~ ) , x - l ( ~ ) ) ios a diffusion hemigroup. In other words, if (8.10) holds with an appropriate X E A then (ps,t)o<sGt is a deterministically timechanged diffusion hemigroup. If G = Rd then any Gaussian hemigroup is a deterministically time-changed diffusion hemigroup (see Heyer and Pap [lo, 3.101). It is still an open question whether this holds for an arbitrary locally compact Abelian group.
8.11 Remark. Let G be a locally compact group with countable basis of its topology. If (vt)tao is a Gaussian semigroup in inz'(G) then ps,t := vt-,, 0 s t is a diffusion hemigroup. Moreover, if (vt)t)O is a convolution semigroup in tm'(G) such that pLs,t := vt-,, 0 s t is a Gaussian hemigroup then (vt)t>o is a Gaussian semigroup. (See Heyer and Pap [lo, 3.111.)
< <
< <
What hemigroups can be obtained from Gaussian semigroups by a deterministic change of time? First consider a symmetric Gaussian hemigroup (ps,t)oGsGt in DT'(R). By (8.4), there exists an increasing continuous function b : R+ + R+ with b ( 0 ) = 0 such that
{
1
Cs,t(Xy) = exp - Z ( ( b ( t )
- b(s))y')
,
for all y E R E (R)^
(see the notations in Example 5.1). Let (vt)t)O be the Gaussian semigroup in inz'(R) given by 1 for all y E R 2 (R)". P~(X,) = exp -Z(t - s)y2},
{
Then for all 0
< s
and y E R E
(a)^we have Ps,t(xy)= i;b(t)-b(s)(Xy),
hence
ps,t = v b ( t ) - b ( s ) .
In a similar way, if (ps,t)oGs
{-Z((b(t) - b ( s ) ) m z } , 1
for all m E Z
(T)"
115 (see the notations in Example 5.2), and ps,t = v b ( t ) - b ( s ) for all 0 denotes the Gaussian semigroup in 9Jt'(T) given by
< s < t,
where (vt)t>O
Moreover, if (ps,t)ogsgt is a symmetric Gaussian hemigroup in m'(C,) then by (8.6), there exists an increasing continuous function b : R+ + R+ with b(0) = 0 such that
(see the notations in Example 5.2), and ps,t = v b ( t ) - b ( s ) for all 0 denotes the Gaussian semigroup in 9Jt' (C,) given by
<s 6 t,
where
If b is strictly increasing with limt,, b(t) = 03 then b E A, and in each of the above cases G = R, G = T and G = C, we have pLg-l(s),b-l(t) = vt-, for all 0 s t , i.e., the Gaussian hemigroup (ps,&+.gt can be transformed into a Gaussian semigroup by a deterministic change of time.
< <
In connection with the above results we mention a structural theorem that one does not find explicitly in the literature.
8.12 Theorem. Let G be a locally compact Abelian group. Then the following statements are equivalent: (i) G is connected and one-dimensional; (ii) G is topologically isomorphic to R, with ck 2 2 for all k E N.
T or to
C,
for some c = (c1,c2,. . .) E N"
For the proof we need two propositions.
8.13 Proposition. Let H be a subgroup of Q, H # (0). Then H is isomorphic to C, for some c = ( c I , ~. ., .) E N" with ck 2 2 f o r all k E N.
Z or to
Proof. Let P denote the set of all primes and let a : P defined by a ( p ) := inf{vp(x) : z E H } ,
--f
ZU {-m} be the mapping
where v, signifies the multiplicity of p . Then (by algebraic arguments) (8.14)
H = {z E Q : up(%)2 a(p) for all p E P } .
We set
P+ := { p E P Po := { p E P P- := { p E P P-" := { p E P
: .(p)
> O},
: a(p)= O } , : -w
< a(p)< O},
: a(p)= -m}.
116 Since for
2
E H, x
# 0 and almost all
p E P we have vp(z) = 0,
n
e :=
P+ is finite. Let
PEP+
and
M := { p - 4 P )
:p E
P-} U { p k : p E P-m, k E N}.
If M is finite, then P-, = 0, and P- is finite. In this case we put
=
By (8.14) we obtain that H = [ . [' . Z Z. Otherwise, let c = (cl, c2,. . .) E generated by counting the elements of M . Then, again by (8.14)
No be
m c1 ' . . . . C" and the proof is complete.
0
8.15 Proposition. Every torsion-free Abelian group D of rank 1 is isomorphic t o some subgroup of Q. Proof. Let 5 denote the minimal divisible extension of D (Hewitt and Ross [ll,A.151). By Hewitt and Ross [ll,A.161, 6 is also torsion-free of rank 1. As a torsion-free divisible group D is isomorphic to Q' for some index set I , and from 1 = r a n k 6 = 111 the 0 assertion follows.
-
Proof of Theorem 8.12. (ii) ==+ (i). It is clear for R and T . Since C, is a subgroup of Q and # { 0 } , hence torsion-free of rank 1, its dual 2, is a compact connected (Abelian) group of dimension 1 (Hewitt and Ross [ l l , 25.25 and 24.281). (i) ==+ (ii). The group G is topologically isomorphic to Rd x K, where K denotes a connected compact (Abelian) group. Since 1 = dimG = dim(Rd x K) = dimRd
+ dimK = d +dimK
(see Hofmann and Morris [12, 8.25 and 8.261) we have either d = 1 and K 0-dimensional, hence connected and totally disconnected, hence G 2 R, or d = 0 and K connected (compact, Abelian), onedimensional. In the latter case G 2 K and is a torsion-free discrete group of rank 1 (Hewitt and Ross [ll, 24.25 and 24.28]), hence by Proposition 8.15 isomorphic to a subgroup # ( 0 ) of Q. If 2 is isomorphic to Z, then
G 2 K 2 KAAE T (where Pontryagin's duality theorem has been applied). Otherwise, Proposition 8.13 yields
K 2 KAAE C,
CX
C,
(Hewitt and Ross [ l l , 25.31) for an appropriate c = (c1, c2,. . .) E Mm with ck 2 2 for all k E M. 0
117 We have the conjecture that a locally compact Abelian group G is one-dimensional (i.e. G is isomorphic to R, T or to C, for some c = (c1, c 2 , . . .) E Nw with ck 2 2 for all k E N) if and only if all Gaussian hemigroups (ps,t)O+(t in %T'(G) can be represented in the form ps,t = V b ( t ) - b ( s ) for all 0 s t , where ( v ~ ) ~denotes )o some Gaussian semigroup in %T'(G) and b : R+ + R+ is an increasing continuous function with b(0) = 0.
< <
8.16 Remark. Let G be a locally compact group. Let (ps,t)o<sqt be a Gaussian hemigroup in %T'(G). What extra property would garantee that (ps,t)o<3
Acknowledgments. The authors are thankful t o Professors W. Hazod and W. Kaup for useful comments.
References [l] M. S. BINGHAM:Stochastic processes with independent increments taking values in
an Abelian group, Proc. London Math. SOC.22 (1971), 507-530. [2] A. BENDIKOV and L. SALOFF-COSTE:Brownian motion on compact groups in infinite dimension. To appear in: Contemporary Mathematics, Proceedings of the Heat Kernel semester, Centre Emile Bore1 2002, Springer-Verlag. [3] CH. BERG and G. FORST: Potential Theory on Locally Compact Abelian Groups. Springer-Verlag, Berlin, Heidelberg, New York, 1975.
[4] M. S. BINGHAMand H. HEYER: On diffusion hemigroups of probability measures on an Abelian locally compact group. Result. Math. 37 (ZOOO), 204-225. [5] J. DIXMIER: Quelques propriCtCs des groupes abCliens localement compacts. Bull. Sci. Math. 11. Ser. 81 (1957), 38-48. [6] H. HEYER: Probability Measures on Locally Compact Groups. Springer, Berlin Heidelberg New York, 1977. [7] H . HEYERand G. PAP: Convergence of convolution hemigroups on Moore groups, in: Analysis on infinite-dimensional Lie groups and algebras, Proceedings, Luminy 1997, pp. 122-144. World Scientific, Singapore, New Jersey, London, Hong Kong, 1998. [8] H. HEYERand G. PAP:Convergence of evolution operator families and its applications to limit theorems, Publ. Math. 58 (2001), 157-191.
118 [9] H. HEYERand G. Pap: Martingale characterizations of increment processes in a locally compact group. Infinite Dimensional Analysis, Quantum Probability and Related Topics 6(3) (2003), 563-595. [lo] H. HEYERand G. PAP: Gaussian hemigroups on a locally compact group. Acta Mathematica Hungarica 103(3) (2004), 193-224. [ll]E. HEWITTand K. A.
ROSS:Abstract Harmonic Analysis. Springer-Verlag, Berlin,
Gottingen, Heidelberg 1963. [12] K . H. HOFMANN and S. A. MORRIS:The Structure of Compact Groups. de Gruyter Studies in Mathematics 25, Berlin, 1998. Probability Measures on Metric Spaces. Academic Press, New [13] K. R. PARTHASARATHY: York and London, 1967. [14] E. SIEBERT:Einbettung unendlich teilbarer WahrscheinlichkeitsmaBeauf topologischen Gruppen. Z. Wahrscheinlichkeitstheorie und verw. Gebiete 28 (1974), 227-247. [15] E. SIEBERT:Jumps of stochastic processes with values in a topological group. Probab. Math. Stat. 5 (1985), 197-209.
Herbert Heyer Mathematisches Institut Universitat Tiibingen Auf der Morgenstelle 10 D-72076 Tiibingen Germany
Gyula Pap Institute of Informatics University of Debrecen Pf.12 H-4010 Debrecen Hungary
herbert. [email protected]
[email protected]
Infinite Dimensional Harmonic Analysis I11 (pp. 119-139) Eds. H. Heyer et 01. @ 2005 World Scientific Publishing Co.
CHARACTER FORMULA FOR WREATH PRODUCTS OF FINITE GROUPS WITH THE INFINITE SYMMETRIC GROUP
TAKESHI HIRAI 22-8 Nakazaichi- Cho, Iwakura, Sakyo, K y 0to, 606-0027, JAPAN E-mail: [email protected] ETSUKO HIRAI Department of Mathematics, Faculty of Sciences, Kyoto Sangyo University, Kita-Ku, Kyoto, 603-8555, JAPAN The purpose of this paper is t o give explicitly all the characters of factor representations of finite type of the wreath product groups 8,(T) = D m ( T ) )d 8, of any finite groups T with the infinite symmetric group Gm.
Introduction. Let G be a countable discrete group, K1(G) the set of all positive definite class functions f on G normalized as f(e) = 1 at the identity element e of G, and E ( G ) the set of all extremal elements in the convex set K1(G). In [14], a canonical bijective correspondence between E(G) and the set of characters of all factor representations of finite type is established. In this sense every element f E E(G) is called a character of G. The problem of determining all the characters, or the problem of giving a general character formula was worked out in [15] for G = emand then, for G = GL(m, F ) with a finite field F , in [13]. The result for infinite symmetric group attracted interests of many mathematicians and we cite here, among others, papers [16], [9] and [l]in which peoples worked principally from the point of view of approximation from finite symmetric groups 6, as n -+ 00. Recently in [5]-[6], we reexamined the case of em from the standpoint of taking limits of centralizations of positive definite functions obtained as matrix elements of simple unitary representations. Next to 6,, we treated the case of wreath products em(?')with afinite
120
abelian group T , and of their canonical subgroups. This case contains the cases of infinite Weyl groups WB- = 6,(22) of type B,/C, and WDof type D,, and limits 6,(2,.)of complex reflexion groups G(r,1,n) as n + 00 (for finite complex reflexion groups, cf. [8] and [12]). For this abelian case, a general explicit formula for characters f E E(G)is given in [7] with a sketch of proof, and so all the quasi-equivalence classes of factor representations of finite type are classified for G = e,(T), T abelian. In this paper we treat the case of 6,(T) with any finite group T . Our method is similar to those in the above cases and will be explained in $5. The main tool is centralizing positive definite functions on G with respect to subgroups GN /’ G, and taking limits of functions thus obtained as N + 00. 1. Structure of wreath product groups G,(T)
1.1. Definition of wreath product G , ( T )
For a set I, denote by 61 the group of all finite permutations on I. A permutation g on I is called finite if its support supp(a) := { i E I ; o(i)# i } is finite. The permutation group 6~ on the set of natural numbers N is called the infinite symmetric group and the index N is frequently replaced by 00. The symmetric group 6, is naturally imbedded in 6, as the permutation group of the set I, := { 1 , 2 , . . . ,n } c N . Let T be a finite group. We consider a wreath product B r ( T )of T with a permutation group 61as follows:
6;r(T)= D I ( T ) >a 6 1 , D I ( T )= n i , - I T i , Ti = T (i E I), where the symbol on D I ( T )as
n’ means the restricted direct product, and
g
(1)
E 61 acts
D r ( T ) 3 d = (ti)i,gA ~ ( d ) ( t i ) i c E~ D r ( T ) , ti = tc-i(i) (i E I). (2) 1
Identifying groups D I ( T )and 61with their images in the semidirect product 6 1 ( T ) ,we have a d r - l = a(d). The group 6.r,(T) is denoted as 6,(T), then G := 6,(T) is an inductive limit of G, := B,(T), and G = limn+, G, is countably infinite. In the case where T is abelian, we put
PI(^) = nti E T for d = ( t i ) i E ~E D I ( T ) ,
(3)
iEI
and define a subgroup of Br(T) as
6 ; ( T ) = Dq(T) x 61 with D ; ( T ) := { d = ( t i ) i e r ; Pr(d)= e T } , (4)
121
where eT denotes the identity element of T. Now take I = N . We identify frequently d and 6,(T) respectively, then udu-' = u(d) and
IJ
with their images in
(d,u)(d',u') = ( d ( ~ d ' u - l ) , u u ' ) (d,d' E D,(T),a,d
E
6,).
Notation 1.1. For d = (ti)iEI E D r ( T ) , I c N , put Suppr(d) := { i E I ; ti # e T } and we omit the suffix I if I = N or I is specified from the context.
1.2. Standad decomposition and conjugacy classes An element g = ( d , a ) E G = 6,(T) is called basic in the following two cases:
CASE1: u is cyclic and supp(d) c supp(u); CASE2: u = 1 and for d = (ti)iEN, t, # eT only for one q E N . Here 1 E 6 , denotes the trivial permutation, and the element (d,1) in Case 2 is denoted by tq,and put supp(t,) := supp(d) = { q } . For a cyclic permutation u = ( 2 1 iz . . . it) of C integers, we define its length as [(u) = e, and for the identity permutation 1, put [(l) = 1 for convenience. In this connection, is also denoted by (t,, ( q ) ) with a trivial cyclic permutation ( q ) of length 1. In Cases 1 and 2, put l ( g ) = e(u) for g = (d,u), and L(tq) = 1. An arbitrary element g = ( d , u) E G, is expressed as a product of basic elements as
<,
9 = tqltqz
* * *
tq,g1g2
* *
gm
(5)
with gj = (dj,uj) in Case 1, in such a way that the supports of these components, q1,q2, ..., qr, and supp(gj) = supp(uj) (1 6 j 6 m ) , are mutually disjoint. This expression of g is unique up to the orders of tqh's and g j ' s , and is called standard decomposition of g . Note that t(tqb) =1 for 1 6 Ic 6 T and C(gj) = !(a,) 3 2 for 1 6 j 6 m, and that, for 6,components, u = u11~2 .. . um gives a cycle decomposition of u. To write down conjugacy class of g = (d,u), there appear products of components ti of d = ( t i ) , where the orders of taking products are crucial when T is not abelian. So we should fix notations well. We denote by [t] the conjugacy class o f t E T, and by T/- the set of all conjugacy classes of T, and t t' denotes that t,t' E T are mutually conjugate in T. (If T is abelian, T/-= T.) For a basic component gj =
-
122 (dj,uj)of 9, let uj = (ij,l i j , ~... ij,ej) and put Kj := supp(uj) = { i j , ~ij,Z, , . . . , ij,ej } with l j = l ( ~ j ) For . d j = ( t i ) i , z K j ,we put
Puj( d j ) := [tijtij-l . . .tat:] E T / - with tk = t i j , k (1
(6)
Note that the product Puj( d j ) is well-defined, because, for tl, t z , . . . ,te E T , we have t l t z . . .ti t k t k + l - tetl. * t k - 1 for any k, that is, the conjugacy class does not depend on any cyclic permutation of (tl,t z , . . . ,tl).
-
-
Lemma 1.1. (i) Let u E 6, be a cycle, and p u t K = supp(u). Then, a n element g = ( d , u) E ~ K ( T=:) GK (put) is conjugate in it to 9' = (d', u) E GK with d' = ( t : ) i E K ,t: = eT (i # io), [ti,] = Po(d)f o r some io E K . (ii) I d e n t i h T E 6, with its image in G = e,(T). T h e n we have, f o r g = (d,o), ~ g 7 - l= ( T ( ~ ) , T U T - =: ~ ) (d',u') (put), and Put(d') = Po(d). Applying this lemma to each basic components gj = ( d j , uj),we get
Theorem 1.1. Let T be a finite group. Take a n element g E G = 6,(T) and let its standard decomposition into basic elements be g = EqlCqQz. . Cqv g1gZ"'grn in (5), with t q b = ( t q k , ( q k ) ) , and 9j = ( d j , C j ) , uj Cyclic, supp(dj) C supp(uj). T h e n the conjugacy class of g is determined by the set of
-
[tqblE T / - (1 6 k
and
P u j ( d j ) , l ( g j ) ) (1
6 j 6 m),
(7)
where Puj(dj)E T / - and l(uj)2 2. (Note that we put l(Eqk) = 1, l ( g j ) = l(uj)2 2.)
A finite-dimensional irreducible representations T of G = Note 1.1. 6,(T) gives a characters of type I,,n = dimr, as x r / d i m r , where x r ( g ) = tr(n(g)) ( 9 E G ) . Actually finite-dimensional irreducible representation r of 6,(T) are all one-dimensional, and are given in the form T = r<,€ with T<,E(9)
= C(P(4)( S g . 6 ) '
for 9 = (d,u) E % ( T ) = &(T)
6m,
where C is a one-dimensional character of T , P ( d ) is a product of components ti of d = (ti),and sgn6(u) denotes the usual sign of u and E = 0 , l . 2. Characters of wreath product group G,(T)
Here we give our general formula for characters of a wreath product group G = 6,(T) for any finite group T . Let T^ be the dual of T consisting of all equivalence classes of irreducible representations (= IRs). We identify every
123
equivalence class with one of its representative. Thus C E T^ is an IR and denote by x~ its character X C ( t ) = tr([(t)) (t E T ) ,then dim( = XC(eT). Denote by 1~ the identity representation of T , and put T^* := T^ \ { l ~ } , T* :=T\{eT}. Then
IT16,, = x ( d i m C ) x
~ as , a function on
C€?
T,
x
0 = x ( d i m C ) x~ and 1 = xlT = (dim[) C€4 C&*
(8)
x ~ on, T'.
(9)
Take an element g = (d,a)E G = e,(T) and let its standard decomposition into basic components be
tqk= (tqh,(qk)),tqr # eT, with l ( t q k=) 1 for 1 < Ic 6 T , and gj
= with a cycle uj of length [(a,)2 2 and supp(dj) c Kj = supp(aj) for 1 6 j m. For d j = ( t i ) i E K j E D K (~T ) ,put PCj( d j ) as in (6). Put where
( d j , aj)
<
x,(a) := sgn6(a)E ( a E 6, ; E = 0,1).
(11)
of decreasing sequences of non-negative real numbers q , E
= (q,E,i)iEN
, q,&J 3 q,,,2 2
(YC,E,3
3
* * *
20;
h
and a set of non-negative real p~ 2 0 (C E T ) , which altogether satisfy
c c Il%Ell
C€T^
with
+
IlPIl
< 1,
(13)
E E W )
II%Ell
=CiEN
QC,€,i
9
11P11 = CCE?PC
.
Theorem 2.1. Let G = 6,(T) be a wreath product group of a finite group T with 6,. Then, for a parameter
A := ( ( " C * E ) ( ( . , E ) € E + , l )
; P)
'
in (12)-(13), the following formula determines a character
(14) fA
of G: For
124
an element g E G , let (10) be its standard decomposition, then
where x E ( g j )= sgn,(aj)€ = (-l)E(t(ffj)-l). Conversely any character of G is given in the f o m of f ~ .
The parameter A of character is not necessarily unique because of the linear dependence (9). To establish uniqueness of parameter, we transfer from the parameter A , to another parameter B = + ( A ) given by
CcE?"KC
=
CCE?& - I T I
PIT)
and the uniqueness of parameter is established. Note that the factor for tqb= ( t q b (, q k ) ) in the formula is rewritten as
In this connection, in the case where T is non-trivial, we can propose another normalization of the parameter /I = ( p ~ ) ~p~ ~ ? , 0 as follows, whose merit is that the character formula (15) is valid even for t,, = eT (not necessarily t,, E T*), whereas this is not the case in other normalizations:
>
125
3. Wreath product of a finite group T with U, Consider a normal subgroup G’ = U,(T) := D,(T) >a U , of G = B,(T) with the infinite alternating group U , c 6,. We have the following result. Theorem 3.1. All the characters of the group G’ = U,(T)
restrictions of those of the group G = B,(T). o n G with parameters
are given as For two characters f A and
fAt
as in (12) respectively, their restrictions o n G’ coincide with each other if and only if f.4’ = ( ~ g n , ) ~(a f ~= 0 or l), or, under the condition ( M A X ) f o r both of A and A‘, if and only i f A‘ = A or A’ = t A , where t A = ( ( ~ ; , E ) ( ~ , & ) € S i x { o , l ) ; p”) i s defined as =
q , 1
,
a;I,1
= ac,o
(C E ?),
and p” = p .
Example 3.1. The restriction fAIQ has its unique parameter A if and (C E ?),and we have from (15) the only if t A = A. In this case, ac,1 = a c , ~ . g = ( d , a ) E G with standard decomposition following expression of f ~ For = LL * . * < q , g l g 2 . . . g m , 6 9 = ( & , ( d )g,j = ( d j , a j ) with e(aj) b 2, f A ( g ) = 0 if some of e(aj)iS even, otherwise f A ( g ) is expressed by
4. Limits of characters of G , ( T ) as n
-+ 00
In certain cases, we can calculate characters of G = B,(T) as limits of trace characters of representations of G, = B,(T) as n + 00. We take irreducible unitary representaions (= IURs) of a degenerate form (cf. [3]). Take a C E ?. Define tensor product representation of D N ( T )= D,(T) as @&&
with Ci = C for Ti = T (i E N)
with respect to a reference vector a = ( a i ) i € ~ , aEi V(Ci),llaill = 1, for which the representation space is V = @:ENV(&).For a E B,, put
I ( a ) (@TEN wi) := @ ; E N v u - ~ with ( i ) vi E V(Ci),vi = ai (i >> 1). (18)
126
Then, for a fixed (C,e) E G,d = ( t i ) i ~ ~ ,
T^ x (0, l},we get an IUR p of G: For g = (d, u) E
P(9) = P ( ( d 4 ) := ( @ ; E N Ci(ti))Wsgn(+
(19)
Now for I , = { 1 , 2 , . . . ,n }, we take a similar representation pn of G,. This is given on the space V, := @iE~,V((i)and, for g = ( d , o ) E G, =
(%(TI = DZ,(T)
% I ,
Pn(9) = P n ( ( d 7 4 ) =
( @&Z,
Ci(ti))wsgn(4E.
(20)
Then, we may consider as V, 7V, and then p, 7 p acccording to G, 7G. Take a g E G. Then, starting from a certain n, g belongs to G,, and so we can consider the limit of trace characters as limn-tm trace(p,(g)). As a result, we take the normalized one as trace(p,(g))/ dimp,.
Theorem 4.1. Let pn be an IUR of G, = B,(T) constructed from (C, E ) E as above. For g = ( d , a ) E G = 6m(T), let g = ...Iq, g192 . - - g m ,tq= (tq,( q ) ) , gj = (dj,uj), be a standard decomposition, i.e., a decomposition into mutually disjoint basic elements. Then the pointwise limit Fc,€(g)= limn+m trace(p,(g))/dimp, is given as
cql
T^ x {0,1}
where, f o r uj = (il i 2 . . . it,) with l j = l(uj),and dj = ( t i ) i E K j with Kj :=suPP(u~),P n j ( d j ):= [t;jt;j-l.*-t;t:] E T/- with ti = t i k . Note 4.1. (i) The positive definite class function FclSis a special case of f~ in (15) in Theorem 2.1, for which cy,€= (1,0,0,. . .) and other parameters and are all zero. (ii) For C = 1~ E T,the trivial representation of T,we have FIT,€(g)= s@(u)€ for 9 = (d,u) E G,and for any C E T^, Fc,~(9) = Fc,O(9)F l ~ , e ( g ) (YcI,~I
A
5. Method of proving Theorem 2.1 Let us explain our method of proving Theorem 2.1. Our proof consists of two parts. The first part is to prepare seemingly sufficiently big family of factorizable (hence extremal by the criterion in the second part) positive definite class functions on G = Gm(T). The second part is to guarantee that actually all extremal positive definite class functions or characters have been already obtained in the first part.
127
-
5.1. The first part of our proof has two important ingredients, which will be treated more in detail in $5 6 10. The first ingredient is a method of taking limits of centralizations of positive definite functions. For a subgroup GI of G a centralization of a function F on G with respect to GI is by definition
Taking an appropriate series of increasing subgroups GN 7 G as N + 00, we consider pointwise limit f ( 9 ) = lirnlv-,, FGN ( 9 ) . Here as a function F , we choose a positive definite matrix element of an induced representation p = I n d g r of a (not necessary irreducible) unitary representation (= UR) A of a certain subgroup H . The second one is inducing up positive definite functions from appropriate subgroups. After choosing subgroups H and their representations A appropriately, we use their matrix elements f n as positive definite functions on H to be induced up to G, and then to be centralized. We have constructed in [3] a huge family of IURs of a wreath product group G = G,(T) with any finite group T , by taking so-called wreath product type subgroups H in a ' saturated fashion ', and their IURs r of a certain form, getting IURs of G as induced representations p = Indgr. For our present purpose of getting (possibly) all extremal positive definite class functions on G as pointwise limits of centralizations of their matrix elements, we choose simpler subgroups of degenerate wreath product type and their IURs. In this case, we get URs p = I n d g r which are very far from being irreducible, but enough for our purpose to get a sufficiently big set of positive definite class functions, as such limits of centralizations. This method has been applied in [5]-[6] to the case of 6 , and reestablished the results in [15]from different point of view.
-
5.2. The second part of the proof contains also two important ingredients, which will be treated more in detail in $5 11 12. The first ingredient is a criterion for a positive definite class function f to be extremal or indecomposable, which says that f is extremal if and only i f it i s factorizable. The second one is a kind of partial Fourier transforms of class functions , with respect to the subgroup D,(T). We utilize it to on G = D,(T) >a 6 reduce the problem '' when is a factorizable class function f on G positive definite ? " to the level of the infinite symmetric group 6 ., As the results, such a factorizable class function f is positive definite if
128
and only if it has the same form as fA in Theorem 2.1 with a parameter A = ( ( a ~ , ~ ) ;( p~) satisfying , ~ ) ~the~condition ~ ~ ~(13) , ~in $2. ~ In [7], we applied this method to the case of the wreath product groups 6 , ( T ) with T abelian and obtained the character formula.
Note 5.1. For the case of non-discrete group G, we can refer [17] for G = U ( m )for example. The method of proof in [17] is quite different from ours. 6. Inducing up and centralization
Take a subgroup H of G and a unitary representation K of H on a Hilbert space V ( K ) and , consider an induced representation p = I n d z r . Taking a non-zero vector v E V ( n ) and consider a positive definite function on H associated to K as f*(h) = ( ~ ( h ) v , v )( h E H ) . Then we have a positive definite matrix element F on G associated to p which is denoted by F = Indg f n , and given as equals to fx on H and to zero outside of H . Let G N 7G be an increasing sequence of finite subgroups going up to G, and consider a series of centralizations FGN of F and thier pointwise limit as N + 00. Since F is zero outside of H , the value of centralization F G N ( g ) is # 0 only for elements g which are conjugate under G N to some h E H . Moreover, for h E H , we get
The condition g'hg'-' E H for g' E G N ,is translated into certain combinatorial conditions, and to get the limit as N -00, i we have to calculate asymptotic behavior of several ratios of combinatorial numbers.
7. Subgroups and their representations for G m ( T ) Our purpose is to get all the characters of G = B,(T) as limits of cenG where fir is a trdizations of matrix elements F = Indgfn of p = IndHK, positive definite matrix element of a UR K of H . For this purpose, we look for the best choice of a pair of H and K , following principally the case of [3], but simplifying the situation without paying attention on irredzlcibility of the induced representation. To give such subgroups H , we take first a partition of N as
129
where each Pc,€is an infinite index set, and the subsets I p ,Ic and I , are all infinite. Corresponding to this partition, we define a restricted direct product as
with
H p = Grp(T),He = Dr,(T), He = { e } .
Here e is the identity element of G, and we consider He as a trivial subgroup of 61, ( T ) .We call this kind of subgroups of degenerate wreath product type. For a representation 7r of H to be induced up to G, we take
, = 1 (p E Here be,€ = (bp)pEpC,,is a reference vector with b, E V ( n P )llb,ll Pc,€), and for p E Pc,&,7rp for H p = 6 r p ( T )is given for d = (ti)iErp E Drp(T),f7E 6 1 p , as XP
( ( 44)= (@;:rpCi(ti))
I ( d we(a)E ,
(26)
where up = (ui)iElP is a reference vector with ui E V(Ci),lluill = 1, and =Jas a representation of Ti = T (i E I p ) , and I ( a ) as in (18); and for C E T , 7rc for He = DI,(T) is given as
~ c ( d )= @&,Ci(ti) for d = (ti)iEr, E He,
(27)
where a< = ( u i ) i E ~ ,is a reference vector with ui E V(Ci),lluill = 1, and
<
8. Increasing sequences of subgroups G N
G
Depending on the choice of increasing series G N 7G = Gm(T)of subgroups, we get various positive definite class functions of G as limits of centralizations FGN for F = Indg f T , which turn out to be characters. We choose a series GN as GN = 6 J~ ( T ) ,JN 7N, and demand an asymptotic condition as
where P := U ( C , E ) E F x { O , l } PC ,& is the union of index sets. Note that even in this case, limN,, IIe n J"/l JNI may not exist. Anyhow we have CPEPXP
+ CC@C
<
1.
(29)
130
For each (C,e) E T^ x {O,l}, let reorder the numbers {A, ; p E P c , ~ } in the decreasing order and put it as a ~ ,:= € ( ~ c , , , ~ ) ~and € N ,also put II := (Pc)CEp Then,
-
c(~,€)€i;x(o,l) Ilac,4 + 11P11 < 1 , which is nothing but the condition (13). As a pointwise limit of the series of centralizations FGN, we obtain the character fA with A = ( ( a ~ , ~ ) ( o,ll; ~ , ~p ) ~inpTheorem ~ 2.1. The calculations are similar to those given in /5]-[6] in the case of em. Finally we remark that, to obtain all the characters of G, it is actually sufficient for us to use only one set of H and ?r above, and this means that the induced representation p = Indz.rr contains weakly all the factor representations of finite type of G. 9. Partial centralization with respect to D j , ( T )
As an increasing sequence GN / G = em(T) of subgroups, we have chosen GN = ~ J , ( T = ) D j N ( T )>a 6 5 , with JN 7 N . Put D N = D j N ( T )and SN = B J , for simplicity, then GN = DN >a SN,and we identify d' E D N and or E SN with their images in GN respectively. Our task is to calculate centralizations FGN of a positive definite matrix element F = Indgfn of p = I n d z r , and to determine their limits. From the formula (22) for FGN
and the explicit form of the subgroup H in (24), we see that for h E H ,
where as
z
is a partial centralization of
with respect to DN
fiC
f n ( h r )= 1 fiC(drh'd'-l)d p D N ( d ' )
(h' E H ) ,
S
T J N given
(31)
DN
with the normalized Haar measure dpDN on DN.(Hereafter we apply the notations in the case of compact groups by using the integration instead of the summation.) Note that for a finite number of h' E H , the partial centralization z ( h ' ) is stable as N is sufficiently large. To calculate it, we apply the explicit
131
form of representation n of H given in (25)-(27). Then we see that it is essentially enough to treat two cases of basic elements: (i) h' = tq= (t,, ( q ) ) with t, E T*,and (ii) h' = (d', a') with a' a cycle and supp(d') c supp(u'). For this, we prepare a lemma for a wreath product group 6,(T) of a compact group T with the symmetric group 6,. Put K = { 1,2,...,a } , and let a = (1 2 ... L) be a cycle with ) d = (ti)iEK. supp(a) = K and g = (d,a)a basic element in ~ K ( Twith Then, for d' = ( s ~ ) , € KE D K ( T ) ,we have
d'gd'-l = (d", a), d" = d'd * a(d'-l) = (sitisi_ll)iEK
(SO = st).
(32)
On the other hand, for a decomposable vector v = @ i E ~ vEi V ( @ i E ~ C i ) with vi E V(Ci),
n ( g ) v = @iEK (C(ti)vu-l(i))= @iEK
(C(ti)Vi-l)*
Lemma 9.1. Let @ i E ~ < be i a tensor product representation of D K ( T ) E T K of Ci = C of Ti = T (i E K ) , and take decomposable vectors v = @ i ~ ~and v i w = @ i E ~ wf ri o m V( @ i E &) ~ with vi, wi E V(&).Then, as a n integration with respect to the normalized Haar measure dpDK(T)(s)= D K ( T ) , we have d p ~ ( ~si = ) ,( s i ) i E E ~ TK
niEK
Take v = @ ~ E K v with ~ unit vectors vi E V(Ci),and put w = v in Lemma 9.1, then we get xc(P,(d))/(dim(.)' as the result of partial centralization above. Let H be a subgroup of G given by (23)-(24), and n its unitary representation given in (25)-(27). For a unit vector v E V ( n ) ,we put f T ( h )= (n(h)v,v)( h E H ) . Since we are now concerned with centralizations with respect to finite subgroups G N ,the role of reference vectors is not important, and we may take v as a tensor product of unit vectors from V(Ci)for i E I,,p E P C ,for ~ ,every (C,E), and similarly for 1"s. Then, by Lemma 9.1, we get the following result.
tqltqa
Proposition 9.1. Take a g = ( d , u) f r o m H and let g = ..-,$qp 9x92 * * - g m Q , = (tq,( q ) ) , gj = ( d j , a j ) , be a standard decomposition.
132
Then, the partial centralization z ( g ) of matrix element f x is given as follows. Let K(C) be the set of k, 1 6 k 6 r , such that Eqk E Hp with p E UEE(O,ll P c , or ~ EqL E He,and J(C,E)be the set of j , 1 6 j 6 m, such that gj = (dj, uj) E Hp with p E Pc,~. Then, Z ( g ) is expressed as
where, for uj = (il i 2 . . . i l j ) with k'j = e(Uj) and dj = ( t i ) i € K j with Kj = supp(uj), we put Puj(dj) = [t;jt;j-l...ta ti] E T / - with ti = t i , . 10. Limits of centralizations
We are on the way of calculating centralizations of F G N of a positive definite G matrix element F = Indgfirof p = IndH7r with respect to GN = DJ~ (T) >a 6J ~ and , to determine their limits. Recall the formula (30)as
-
where SN = 6J ~ and , the partial centralization fir with respect to D N = DJ~ ( T ) is defined by (31)and is calculated in Proposition 9.1. For any element in G, there exists an element in H conjugate to it. Therefore it is enough for us to determine the value FGN on H. Take = (d, E H and let = E q l t q z * * * Eqq,g1g2* * * gm, Eq = ( t q , ( q ) )9 gj = (dj,uj), be its standard decomposition. Put P = U(e,E)E~X{O,ll Pc,€,then,
H=
(n;,p
HP)
(n,,@k)
x He,
and the condition g E H means that each Eqk belongs to one of H p and Hc, and that each gj belongs to one of Hp. Furthermore, the latter condition can be expressed by means of supports as
Eq
For p E P, choose ( C , E ) such that p E Pc,€,and put for basic elements = ( t q , ( q ) ) and gj = (dj,oj) in Hp,
133
Then the formula in (34)for z ( g ) is expressed in the form
n( )n. . .
~ ~ ( t q k x)
c~?
< <
n ( JJ
PEP
k:qkEIC
xp(tq*) x
k:qhEI,
JJ
Xp(gj0
>
(38)
j:KjCI,
< <
where 1 k r,1 j rn. The term corresponding to C in the first product comes from tqkE H,-, and the term corresponding to p E P in the second product comes from tqQb E H p and gj E H p . Let Q(g,I,-) be the union of supports { q k } = supp(&,) c I,-,and QK(g,Ip) be the union of supports {qk} C Ipand Kj = supp(gj) c Ip. Since g E H , they give a partition of supp(g). Let their orders be n(C)and n ( p ) respectively, then
(CcE?Q(9,IC)) U (CpEpQK(9,Ip))
= supp(g),
C,-,p(C)+ C P E P W = IS~PP(9)l. E em, put 9' = T ~ T - ~ , ' ( = ~ T ~ ~ T and - ~ 'gj ,
(39)
Now, for T = 7gjT-l. Then, the standard decomposition of 9' into mutually disjoint basic elements is given as '9=
'
' *
.'~qp'gl 9' 2
* ' *
'gm,
'
= ( t q , ( T ( 4 ) ) ), '9j = (T(dj) ,T(7jT-l).
For Q,we have xP('tq) = xp(Jq)if ' Q is still in H p , or equivalently if ~ ( q E) Ip. For dj = ( t i ) i E K j ,recall that ~ ( d j = ) and P'(rj7-1(T(dj)) = Puj(dj) and SO xP('gj) = xp(gj) if 'gj is still in H p , or equivalently if 7 ( K j )c I p . Let us now consider a partial sum of (35), where T E SN = B J , runs over all such elements that it preserves Q,- := Q(g, I,-)and QKP := QK(g, Ip)inside of I,-and Iprespectively. Suppose that N is sufficiently large so that g is contained in H n G N ,then this condition on T E SN is written as T(Q,-) c 4 n J N , 7(QKp)c I p n J N .
(40)
Put & := { &,- (C E ?), QKP (p E P) }, and denote by 7(&, N) the set of T E SN = B J , satisfying the condition (40). Then, for T E 7 ( Q , N ) ,we see from the above consideration that X('g) = z ( g ) . Therefore the partial sum over 7 E 7(&, N) is calculated as
134
We can calculate the order 17(&,N)I,recalling that the union of Qc's and Q K p ' s is supp(g). Then we can evaluate under the asymptotic condition (28) as
Applying the formulas (41) and (38)' we obtain
where for p E Pc,,,
( 6 , ~ E) T^ x (0, l},
The above calculation for a partial sum over T E 7(&, N) C 6 ; can ~ be~ applied to other partial sums. These partial sums come from possible cases of Tg corresponding to which of I[ or Ip contains s ~ p p ( ~ = [ ~T ~ ( q)k ) , and . these cases give us limits of which of I p contains supp(Tgj) = T ( K ~ )All partial centralizations similarly as above, and they correspond altogether exactly to all the 'monomial' terms of the expansion of the right hand side of (15) in Theorem 2.1 into 'monomials'. Thus we get the following proposition, a half of Theorem 2.1. Proposition 10.1. Let T be a finite group. Let fA be the class function o n G = B,(T) given by the formula (15) in Theorem 2.1, with parameter A = ( ( a ~ , , ) ( ~ , ~; p)) ~in? (14). ~ ~ ~If ,the ~ )parameter A satisfies the conditions (12)-(13), then fA i s obtained as a limit of centralizations of a positive definite matrix element of Indgx with ( A , H) given above. T h e limit i s taken according to a n increasing sequence of subgroups GN = 6J~ (T obeying the asymptotic condition (28). All the class functions fA are positive definite under the conditions (12)(13). 11. Criterion for extremality We can prove the following criterion for extremality (cf. [15], Satz 1). This is the first ingredient of the second part of the proof of Theorem 2.1.
135 Theorem 11.1. Let T be a finite group, and f a positive definite class nomnalized as f ( e ) = 1. Then f is extremal if and function o n G = em(T) only if it has one of the following properties which are mutually equivalent: (FTP) [Factorizability Property] For any g = ( d , o ) E G, let g = Cqltq2***Cqv g 1 g 2 . - * g m , tq= ( t q , ( q ) ) , gj = ( d j , u j ) , be a standard decomposition. Then,
(FTP') For any two elements g , g' E G with disjoint supports, f (gg') =
f ( d f(9'). Let us rewrite these conditions in another form. As is proved in Theorem 1.1, conjugacy classes of basic elements in G is given by the set R of the following objects w: = 1)with [t]E T * / - , and w = ( [ t ] , lE) (T/-) x w = ([t],t
{t 2 2},
and the conjugacy class of g E G \ { e } with the above standard decomposition is determined by the collection of
([tq,l,t= 1) (1 6 k 6 T ) and ( p o j ( d j ) , t ( u j )1 (1 6 J' 6 m).
(43)
Denote by n, ( 9 ) the multiplicity of w E R for g = tqlcq2 * * * tqpg1g2 gm. Put 2 2 0 := { n E 2 ;n 2 0 ) and denote by ( Z > O ) (the ~ )set of all n, E 2 2 0 , with n, = 0 for almost all w. Then, n ( g ) := n= (n,(g)),En is an element of (Z>o)("),and the correspondence
gives a bijective map from the set of all conjugacy classes [g]of g E G, g # e, onto (Z>o)(*).For w = ( [ t ] , L ) E R, put w-l := ([t-'I,[).Then, i f w is the conjugacy class of tq= (tq,( q ) ) or of gj = ( d j ,uj),then w-l is that of Cq-l or of gjV1 respectively. Hence, n W ( g - l ) = n w - l ( g ) ,and the transformation [g] I+ [g-l] in the set G/- of conjugacy classes of elements in G induces an involutive transformation L on ( ~ > o ) ( ~given ) as L
:
(Zp)(*) 3 n = (n,),,n
Hn'
= (n;),€n with n; = n,-1 ( w E R).
For a positive definite class function f on G, put s ( f ) = (~,),~n with s, = f (g,), where gw denotes a basic element in the class w. Then, since w-l is represented by g i l , and since f ( g - ' ) = f ( g ) , g E G, we have s,-1 = % (complex conjugate). Define a positive definite class function f by f ( g ) = f o (9 E G), then 47) = s(f). Here, for s = (s,),E~, we = s, for w E Or, := { w E R ; w-l = w } . P u t put 3 := ( z ) w E nwith
136
R, := { w E R ; w-l # w } , then R = R,, U 0,. Let 1, := [-1,1] c R for w E R,,, and D, := ( z E C ; 1x1 < 1) C C for w E R,, and put S := S ,,
x S, with
Every s = (S,),~Q
E S defines a function Qs on (Z2,0)(”) 2 G/- by
where K = R or C according as R, = 0 or # 0. Then we get a class function fs := Qs o $! on G satisfying f S ( g - l ) = fs. Now the condition (FTP) above is rewritten in these notations as follows: ,, x S, such that (FTP”) There exists a n s = (~,),~n in S = S f = fs, that is, that f o r a g E G with standard decomposition g = Q1&a ..-[q,,g1g2 gm, let n w ( g ) be the multiplicity of w E R in these basic components, then 0
.
.
f ( 9 )=
nwEn s,~,(~),
where s,” := 1.
(45)
12. Final step of the proof of Theorem 2.1
By the “only if” part of the proof of Theorem 11.1, for each f E E ( G ) , there corresponds an element s E S such that f = fs = !4jS o +. As the final step of the proof of Theorem 2.1, we specify the parameter s = (sw)wEn and prove the following. Proposition 12.1. An extremal positive definite class function (or a charnormalized as f (e) = 1, i s given in the acter) f o n G = em(T), form of f A in the formula (15) in Theorem 2.1, with parameter A = ( ( C Y ~ , ~ ) ( ~ , .; )p~) ~in~ (14) ~ ~ ,satisfying ~ )
the condition (13).
To prove this, we proceed as follows. By the “only if” part already proved, we should examine a positive definite class function f of the form (45). We define a class function on T by putting X ( t ) = 1 for t = e T ;
X ( t ) = ~( [ q,l)for t E T * ,
137
where s ( [ ~ I=, ~s, ) for w = ( [ t ]1) , E 52. Then, since X is a class function on T, it is expressed as a linear combination of xc,C E ?, as
b c x c ( t ) ( t E T)
X ( t )=
with
(46)
CETI
For C 3 2, we define also a class function &(t) on T by putting &(t) = ( t E T), where ~ ( [ t l , = ~ ) s, for w = ([t],.!)E R. Then, similarly as for X, it is expressed as ~([t],!)
..tqq,g1g2- .. gm , we have from (42)and (45) Then, for g = tq1tqQZ ac,qo,)Xc (Po,( d j ) ) Fix a (CO,E) E T^ x (0, l}, and take a positive definite class function Fco+ in (21) in Theorem 4.1. Then the product f'(g) := ( f m ( g ) = f(g) G ( g ) is positive definite. Take a subgroup D, := DI,( T ) with n sufficiently large so that supp(g) C I , . The Fourier transform .FC~,~ o f f with respect to Fco,Eis by definition the integral o f f ' with respect to D,:
Then, after some calculations, we obtain the following formula for the Fourier transform .F~~,o;,(f)of f with respect to FcO,o: For (T E 6,,let (T = ( ~ 1 ~ . .2T(, be its decomposition into mutually disjoint cycles, then
-
The Fourier transform Fco,oin(f)is a positive definite class function on the symmetric group 6, for any n. Here we can apply Korollar 1 of Satz
138
2 in [15].After certain delicate calculations, we can complete the proof of Proposition 12.1.
Acknowledgements
We express our thanks to Professors N. Obata, N. Kawanaka, K. Nishiyama, H. Yamashita and K. Shinoda for valuable discussions and advices. References 1. P. Biane, Minimal factorization of a cycle and central multiplicative functions on the infinite symmetric groups, J. Combin. Theory Ser. A 76 197-212 (1996). 2. J. Dixmier, les d -algibres et leurs reprisentations, Gauthier-Villars, Paris, 1964. 3. T. Hirai, Some aspects in the theory of representations of discrete groups, Japan. J. Math. 16 197-268 (1990). 4. T. Hirai, Construction of irreducible unitary representations of the infinite symmetric group Em, J. Math. Kyoto Univ. 31 495-541 (1991). 5. T. Hirai, Centralization of positive definite functions, Thoma characters, weak containment topology for the infinite symmetric group, in RIMS K6kyOroku 1278,pp.48-74,2002. 6. T. Hirai, Centralization of positive definite functions, weak containment of
representations and Thoma characters for the infinite symmetric group, to appear in J. Math. Kyoto Univ. 7. T. Hirai and E. Hirai, Characters for the infinite Weyl groups of type B,/C, and Dm, and for analogous groups, in Non-Commutativity, InfiniteDimensionality and Probability at the Crossroad pp.296-317, World Scientific, 2002. 8. N. Kawanaka, A q-Cauchy identity for Schur functions and imprimitive complex reflexion groups, Osaka J. Math. 38 775-810 (2001). 9. S. Kerov and G. Olshanski, Polynomial functions on the set of Young diagrams, C. R. Acad. Sci. Paris, Ser. I, Math. 319 121-126 (1994). 10. N. Obata, Certain unitary representations of the infinite symmetric group, I, Nagoya Math. J. 105 109-119 (1987); 11, ibid. 106 143-162 (1987). 11. N. Obata, Integral expression of some indecomposable characters of the infinite symmetric group in terms of irreducible representations, Math. Ann. 287 369-375 (1990). 12. T. Shoji, A Frobenius formula for the characters of Ariki-Koike algebras, J . Algebra 226 818-856 (2000). 13. H.-L. Skudlarek, Die unzerlegbaren Charactere einiger diskreter Gruppen, Math. Ann. 223 213-231 (1976). 14. E. Thoma, Uber unitke Darstellungewn abzalbarer, diskreter Gruppen, Math. Ann. 153 111-138 (1964). 15. E. Thoma, Die unzerlegbaren positiv-definiten Klassenfunktionen der abzahlbar unendlichen, symmetrischen Gruppe, Math. 2.85 40-61 (1964).
139 16. A. Vershik and S. Kerov, Asymptotic theory of characters of the symmetric group, h n c t . Anal. Appl. 15 246-255 (1982). 17. D. Voiculescu, Representations factorielles de type 111 de U(m), J. Math. pure et appl., 55(1976), 1-20.
This page intentionally left blank
Infinite Dimensional Harmonic Analysis I11 (pp. 141-159) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
REMARK ON BIANE’S CHARACTER FORMULA AND CONCENTRATION PHENOMENON IN ASYMPTOTIC REPRESENTATION THEORY
AKIHITO HORA * Faculty of Environmental Science and Technology, Okayama University, Okayama 700-8530,Japan E-mail: horaOems.okayama-u.ac.jp
An asymptotic expression for moments of the Jucys-Murphy element in the group algebra of the symmetric group S(n)as n + 00 is presented. Applying this formula, we show that two remarkable results due to P.Biane in asymptotic representation theory on the symmetric groups are reconstructed. One is an asymptotic formula for irreducible characters. The other is a concentration phenomenon in irreducible decompositions.
1. Introduction
Theory of the limit shape of Young diagrams due to Vershik-Kerov [13] and Logan-Shepp 191 promoted developments in many topics in asymptotic combinatorics. Two books [7] and [12], for example, will serve as good surveys into this direction. If we consider random Young diagrams with n boxes distributed according to the Plancherel measure and take the limit of n + 00 under scaling by l/+, we observe that the distribution tends to concentrate quite near the special diagram called the limit shape. In other words, those components that correspond to the Young diagrams near the limit shape tend to occupy overwhelming parts in irreducible decomposition of the regular representations of the symmetric groups S(n) as n -+ 00. In the same scaling regime with this, Biane established a fundamental scheme of the concentration phenomenon in irreducible decomposition for growing families of more general representations of S ( n ) ([l], [2]). *Partially supported by Grant-in-Aid for Scientific Research of Japan Society for the Promotion of Science (13640175).
142
Let us assemble some notations. The set of Young diagrams is denoted by Y including the empty diagram 0.For X E Y,the numbers of boxes, rows, columns and j-rows are denoted by (XI, r.w(X), coZ(X) and mj(X) respectively. Set Y, = { A E YJIxI = n } , ~ 5 =, { A E ~ ( 1 5 ~ 1n } , Yo = {A E Ylml (A) == 0}, Y : = Y, nYo and Y(& = Y5, n Yo. The union p U D of p, D E Y is defined by mj(p U CT) = mj(p) m j ( a ) for all j . For p E Y&, C,,u(ln-l,l) denotes the conjugacy class of S ( n )having p u (In-IPI) as its cycle type. Set Z(p) = JpI - ~ m ( p (Figure ) 1). If x E Cpu(ln-l~l)for p E Yo<,,Z(p) is the minimal number of transpositions needed to produce x, or eqGvalently the distance between x and the unit element e on the Cayley graph of S(n) equipped with all transpositions as the generators. The irreducible character of S ( n )corresponding to X E Y, and its normalization = xx/dim X respectively. Here px is an are denoted by xx = trpx and associated irreducible representation with A. Let x(,) be a positive-definite central function on S(n) and = x ( , ) / x ( , ) ( e )its normalization. is expressed as a convex combination of the irreducible characters. Namely, we have probability P(,) on Y, satisfying
+
zx
XEYn
A probability on R is assigned to X E Y,which is denoted by mx and called the transition measure of X due to Kerov (see $52.1). The transition measure of x(,) is also determined through Eq.(l): XEY,
We are now considering the regime of l/fi-scaling. Roughly speaking, m x 6 is a finite quantity where X f i denotes the diagram obtained by rescaling X E Y, horizontally and vertically by l/+. Hence the kth moment of mx,denoted by Mk(mx), for X E Y, grows in the order of nkI2as n + 00. The result on the concentration phenomenon due to Biane ([l], [2]) can be formulated as follows. Let a growing family {x'")}satisfy Assumptions 1 and 2. Assumption 1 For any k E N,there exists mk E R such that
Mk(mx(,,))
-
mknkI2
(n + 00).
(3)
Moreover, the moment problem for the sequence { m k } is solved uniquely. Assumption 2 For any
~ , C T ,E T
Yo, x(,) satisfies
143
.
.
‘.. . . . ...
Figure 1. l ( p ) and uo
(a) (decay order)
(b) (asymptotic factorization)
By virtue of Assumption 1, we have probability Q on R satisfying
1, 00
mk =
skQ(W
(lc E
and continuous diagram A such that ma = Q (see 552.1). Then probability on Un determined by Eq.(l) tends to concentrate at A as n + 00. Here the topology on the set of continuous diagrams should be clarified as in 554.1. Taking the character of representation 7r of S ( n ) as a positive-definite central function in the above argument, we immediately see a concentration phenomenon in irreducible decomposition of 7r. In particular, we obtain asymptotic behavior of the Littlewood-Richardson rule, which describes irreducible decomposition of outer products of irreducible representations of S ( n ) ,as n + 00 under l/fi-scaling due to Biane [l]. Actually, our motivation of this note and also [4] was to reconstruct such a thermodynamical limit for the L-R rule without relying on representations of general linear groups. Let us consider the Jucys-Murphy element J n = ( 1 n + 1 ) + ( 2 n + 1 ) + . - . + ( n n + 1 ) E C[S(n+l)].
144
+
S ( n ) is embedded into S ( n 1) as the elements fixing letter n + 1. Map : C [ S ( n l)]-+ C[S(n)] is defined to fix any element of S(n) and to bring that of S ( n 1)\ S(n)to 0. Since J , commutes with S(n),IE,J: lies in 2 ( S ( n ) ) the , center of C[S(n)].In [4] we show the following asymptotic expression for IE,Jt (k E N). The adjacency operator corresponding to p E Yo -<,is denoted by ApU(ln-IpI) = ' x E C p U ( l n - I P I ) x. We define (TO E Yo by removing all 2-rows and the first column of (T E Yo (Figure 1). For (T E Vk, the set of noncrossing partitions of (1,2,. . . ,k) with block structure (T is denoted by N C ( a ) (see $53.1). Namely, a partition in NC(cr) has m j ( a ) blocks of length j.
IE,
+
+
Theorem 1.1. ([d])
For any k E N,
holds.
An advantage of Eq.(6) is that the range of sum does not depend on n. Our application of Theoreml.l include: (1) concrete computation of m k in Eq.(3) (2) Biane's asymptotic character formula (3) a fundamental scheme of the concentration phenomenon due to Biane. We prove Theoreml.1 and discuss (1) above in [4]. In what follows, we give alternative proofs for Biane's results (2) and (3) above in $3 and $4 respectively after preparing some materials in $2. 2. A Preparatory Trace Formula 2.1. Transition measure of a diagram
See [6] and [7] for details on transition measures, continuous diagrams and related notions. A Young diagram is displayed as in Figure 2. We deal only with centered diagrams. Set
D = { w : R -+~ I ~ w ( x-lw(x2>1 ) 5 1 x 1 - 521, ~ ( x=) 1x1 for large IxI} 2 DO = { w E I D ~ Wis piecewise linear, w'(x)= *I} II .
v
145
--Tozo(X)
0
coz (A)
Figure 2.
x1Yl x2
Y2
0
~ T - I
YT-IG
min-max coordinates of a Young diagram
The positions of local minima and maxima completely characterize X E DO as in Figure 2:
In the embedding U C DO,each box of a Young diagram is stretched by fi so that its min-max coordinates are integers. Probability xi='=, piszi on R with a finite support and mean 0 corresponds to X E Do with min-max coordinates (7) bijectively through the equality of partial fraction expansion (2
- 91) * * * ( z - YT-1) - 51) (2 - 2,)
(2
* * *
-- PI
z
- 21
+...+-
PT z - 2,
This probability is called the transition measure of X E Do after Kerov and denoted by mx. It follows from the definition that Mo(mx) = 1 and Ml(mx) = 0. Moreover, if X E Y,we have M2(mx) = 1x1. For X E Do with (7), setting &(A) as
(k E N),we have
Especially, { M n(mx)In=0,1 J , ... and {13k (X)}k=l J,... are expressed by polynomials of each other. Since Eq.(8) is extendable to w E D, Eq.(9) enables
146
us to introduce the transition measure m, of w E D. Actually, a further extension is possible to have probability m, with a noncompact support where w(x) - 1x1 tends to 0 appropriately as x -+ f w . The transition measure of the limit shape R
is the standard semi-circle distribution 1 mn(dx) = --J4_.z1[-~,~](z)cix. 2n If w E D is rescaled by ws(x) = s-'w(sx) for s m,(sdx) and hence
> 0, we have
(IC E N).
Mk(rnUs)= s-'Mk(m,)
m,.(dx)
=
(10)
2.2. A trace formula
For X E Yn with min-max coordinates x1 < y1 < . - .< yT-l < x,, putting a box a t the valley of min coordinate xj, we get R j E Yn+l ( j = 1, . . . ,T). By using the hook formula for dimensions of irreducible representations, we see r
mx = j=1
dim Rj ( n + 1) dimXbZj
(see [7],Chapter4, 1.4). Eq.(ll) connects transition measures with JucysMurphy elements as shown in [3]:
(A E Yn, IC E N).
gx(EnJ:) = Mk(rnx)
(12)
In this subsection, we supplement some explanation about Eq.( 12) since it plays a key role in subsequent discussions. In general, let us consider a finite group G, a subgroup H of G and IE : C[G] -+ @[HI which fixes the elements of H while maps those of G \ H to 0. Clearly (IEb)a = E(ba) holds for a E @[HI and b E @[GI. For a E fi (= the equivalence classes of irreducible representations of H ) , let e, denote the minimal central projection in @[HIassociated with a , which corresponds t o Iv, (identity) E End(V,) in
C [ H ]N
@ End(V,) a€H
N
@ V, C3 V: . ,€H
147
LH denoting the left regular representation of H,we have for a E C [ H ] LH(ae,)
-@
y(a)r(e,)
;.I
-
a(.) 8 1v2 CB o
,
TEA
hence taking traces, dimatra(a) = trLH(ae,) = IHI(aea)e
-
where denotes similarity of operators, ( . ) e indicates the coefficient of e, and a representation for y E H is expressed again as y for simplicity. This yields tra(Eb) =
[G : H ]dim a
dimptrp(be,)
( b E C[G]).
(13)
PEG
P(e,) E End(Vp) is the projection onto the a-component of representation (p,Vp) of H . Eq.(13) indicates also that z"(IE.) = t r a ( E . ) / d i m a is positive and hence is a state of C[G]. first corner
j , - i, Figure 3.
content of a corner
Now we take G = S ( n + l), H = S ( n ) , a = px (A E Yn),,4 = PA (A E Y,+1) and b = J," (k E N) in Eq.(13). The corners of A are denoted by
01,.
. . ,El, (Figure 3). We use the following facts (see e.g. [ll]).
148
VA = WI @
* @ W, (multiplicity free), W, q nV,\oP. ) An eigenspace of pn(J,) is an S(n)-invariant subspace of VA. 0 W, is the eigenspace of P A ( J , ) with respect to eigenvalue j p - i, (= content of 0,). Here j p - i, coincides with the new min coordinate of A \ 0,. Let X have min-max coordinates X I < y1 < ... < Yr-1 < and As E Yn+l denote the Young diagram made by putting a box at the sth local minimum of X (s = 1,. . . ,r ) . Then, Eq.(13) yields
0
0
from Eq. (11). This reconstructs Eq. (12). 3. Biane’s Formula for Irreducible Characters 3.1. Biane ’s asymptotic formula
We quickly recall definitions of noncrossing partitions and free cumulants. See [14] and [lo]. Let 7r = {Vl, . . . ,V d } be a patrition of (1,2,. . . ,n) with blocks K. If there are no quadruples pl < q1 < p2 < q2 such that pl and p2 lie in the same block of 7r while q1 and q2 in another one, 7r is said to be noncrossing. N C ( n ) denotes the set of noncrossing partitions of (1,2,. . . ,n). If the block structure of 7r E N C ( n ) agrees with p E Y,, 7r is said t o be of ptype. NC(p) denotes the set of these 7r’s of ptype. Restricting the partitions to the noncrossing ones in the ordinary moments-cumulants formula for probability p on R with all moments, we have the lcth free cumulant R k ( p ) of p determined by
Mn(p)=
C nENC(n)
RK(p)=
C #NC(p)Rl(p)m’(P)...R,(p)mn . (14) PEY,
for 7r = {Vl, ...,V d } E N C ( n ) . For A > Here RK(p)= Rtv1(p).-.Rtvd(p) 0, set Y,(A) = {A E Y,lrow(X),coZ(X) 5 Afi}. Young diagrams in Yn(A) are said t o be A-balanced after [l].Since row(X)coZ(X) 2 n always holds for X E Y,, an A-balanced diagram satisfies also row(X),coZ(X) 2 f i / A . Since supprnA6is uniformly bounded for X E Y,(A), Eq.(lO) and Eq.(14) assure that there exists C A , k satisfying
IMk(mA)l, IRk(mA)l 5 CA,knk” .
(15)
149
Biane showed the following asymptotic behavior of irreducible characters of S(n) as n + 00 in such a balanced regime.
Biane's formula ([I]) Let A
> 0 and k
E
w be given. For any p E Y o-< k ,
holds. The 0-term in Eq.(16) depends only on A and k, and is taken compact-uniformly in A. Note that the growth order of the first (main) term in the right-hand side of Eq.(16) is bounded by
from Eq.(15). In the next subsection, we state a proof of Biane's formula based on our Theorem 1.1.
3.2. Proof of Biane's formula We divide Biane's formula into parts as follows. Let k 2 2. d ( k ) : asymptotic formula for cycles with length up to k : for j -A X(j)u(ln-j) = n-jR.3+1(mx)
+ 0 (n-(j-')/'-'
)
5 1,
(A E Y n ( A ) ) .
B ( k ) : asymptotic factorization for Yo
= 0 (n-(l(c)+l(T))/2-' )
(A E Y n ( A ) ) .
C ( k ) : Biane's formula: Eq.(16) for Y&-. V ( k ) : decay order for -A
-
Xpu(1n-IpI)
for p E Yo
)
(A E Y n ( 4 ) .
Here every 0-term depends only on A and k , and is compact-uniform in A . Since Yo<3 = {@,(2),(3)}, it is trivial that 4 2 ) u C(2) and d(3) C ( 3 ) . It is also obvious that C ( k ) implies V ( k ) from Eq.(15) and Eq.(17).
*
Lemma 3.1. For k 2 4, d ( k ) , B ( k ) and D ( k - 2) imply C ( k ) .
150
Proof Let p € Y<j. be a non-cycle; for A(k) implies Eq.(16) if p is a cycle in Y°-fc. Then col(p) < k — 2. When p is divided into two diagrams in Y°, both belong to Y^ A _ 2 . Hence B(k) and T>(k - 2) yield
J>2
Applying ,4(fc) to the first term of the right-hand side, we get C(k). (QED) Lemma 3.2. For k>2, T>(k) implies B(k + 1). Proof We use Lemma 3.6 shown in the next subsection, taking y? = x\ If T>(k) holds, then we have for any a, r € Y° such that \a\ + \T\ < k + I
(AeY n (A)).
(18)
Since xx is multiplicative on Z(S(n)), the first term of the left-hand side of Eq.(18) is X^in-i,!)^!— iri)- Since H,|r| < fc - 1 < k for non-empty cr and r, £>(/;) assures
Then Eq.(18) implies X^ UrU(1 n-i,|-ki) = O(n-('^+1^/2) and hence
This completes the proof of B(fc + 1).
(QED)
Applying Theorem 1.1 together with Eq.(12) and Eq.(14), we have
(19) for A 6 Yn as n -*• oo. Note that 7?i(mA) = 0. The O-term in Eq.(19) depends only on k. Lemma 3.3. .4(2) holds. Proof Put k = 3 in Eq.(19). Since Y^ = {(3)}, NC(3) = {{1,2,3}}, /((3)) = 2 and (3)° = (2), we have
151
Let A € Y n (A) and use Eq.(15). Then
holds where the last O-term depends only on A and k.
(QED)
Lemma 3.4. For k > 3, C(k - 1) implies A(k). Proof Since A(k — 1) holds, we have only to show the asymptotic formula for a fc-cycle. Put k + 1 instead of k in Eq.(19). Then
>A) mjW
(A 6 ¥„). (20)
J>2
Fora € Y£+1 such that cr ^ (fc + 1), j<7°| < fc + 1-2 = fc-1, i.e. holds. C(k - 1) implies that for A 6 Yn(^4) the sum in the left-hand side of Eq.(20)
! JJ E j+1 (m A ) mi(
+O
J>2
where note /(CT)
i(q°) _ l(g) , '•<w(g) _ H _ fc + 1 ~ T~ ~ ~^~ + ~2~~ ~ T ~ ~2~
for the second equality and I (a) -\a°\= m 2 (<7) ,
mj(a°) = mj+i((r) (j > 2),
-R 2 (ntA) = n
for the third one. Hence, comparing the both sides of Eq.(20), we have
152
Again the 0-term depends only on A and k. This completes the proof of 4k). (QED) The argument so far in this subsection shows that the following proceeds inductively:
d ( 2 ) H C(2)
=$
d(3)H
U D(2)
D(3)
=$
13(4)
Consequently, d ( k ) , B ( k ) ,C ( k ) and D ( k ) are shown for all k 2 2. 3.3. Auxiliary estimates
We show Lemma 3.6 below already used in the proof of Lemma 3.2. Lemma 3.6 will be used also in 54.
Lemma 3.5. For 9 , h E S ( n ) such that #(suppg n supph) = length function 1 satisfies l(gh) 2 Z(g)
T
2 1, the
+ l(h)- 2r + 2 .
Proof Decompose g into nontrivial cycles: g = gt -..g291, where the cycles containing at least one letter in supph are 9 1 , . . . , g 8 (1 5 s 5 t ) . Decompose each cycle into transpositions as 91 = ( i p - . . i l )= ( i p i p - l )* * * ( i & ) ( i 2 i l ) . In the process that those transpositions are multiplied from the left of h, we look at “up” and “down” steps at which the length increases and decreases respectively. Since #(ups)
+ fl(downs)= Z(g)
and #(ups) - #(downs) = Z(gh) - l(h)
clearly hold, we have 2#(downs)= Z(g)
+ l(h)- l(gh) .
(21)
First let g be a cycle (i.e. t = s = 1). Fix the expression g = (ip - .il) = (ipip-l).. . (i3i2)(i2i1).Set S = suppg n supph. We can assume i l E S. When ( i j + l i j ) is multiplied from the left and ij+l $! S holds, this step cannot be a down. To each letter in { i l l . .. ,iP} \ S, assign the non-down step at which the letter lies in the left side. This assignment is feasible since the right end il is in S. Then, #(downs) 5 #(steps)- # ( { i l , . . , ip} Hence we have from Eq.(21), l ( g )
\ S) = (p - 1)- (p - T ) = T - 1 .
+ l ( h ) - l(gh) 5 2 ( -~ 1).
153
Next consider a general case: g — gt • • • gs • • • gi . The above estimate is valid for the process that gi,...,ga are multiplied from the left of h successively, while the steps at which gg+1 , . . . , gt are multiplied are not downs. Hence (j(downs) < r — s < r — 1. Again Eq.(21) completes the proof. (QED) Lemma 3.6. Fix k e N. Let a growing family of central functions (p on S(ri) satisfy as n -t oo. Then, for
)/2 1
- )
as n -> oo. ffere the O-terms depend only on k (and are independent off).
Proof
Let
uti™-!'!) x cVua— i'-i)ltt(suppffnsuppft)
=r
}-
Then
|
Each fiber of the map (g, h] € Cal)^n-\,\^ x C'rU(1n-|T|) •->• gh e 5(n) at 5/z € £$(()) has cardinality rii>2(m»(or) + mi(r))!/mi(CT)!mi(r)!. Hence the first sum in the right-hand side of Eq.(22) is (m i (g)+Tn i (r))! A .>2 mi(
n
Noting _
we have •"•ern ttC'rU(l"-KI) HA|T|
154
as n + 00 for bounded 1 ~ and 1 171. Up to here the argument is concerned just in C[n]. Applynig cp to the both sides, we have
+
171 5 lc. If (g,h) E S ~ ) ( Tfor ) T 2 1, the assumption implies Let 101 cp(gh) = O(n-1(gh)/2) since #supp(gh) 5 101 171 - 1. Combining this estimate with Lemma 3.5 and # S g j ( r )= O ( n ~ ' ~ + ~ T ~we- rsee ) , the second sum in the right-hand side of Eq.(23) is bounded as
+
4. Concentration Phenomenon
4.1. Scheme of concentration phenomenon By applynig Theorem 1.1,we show that the concentration phenomenon is observed if a growing family consisting of positive-definite central functions x(,) on S ( n ) satisfies Assumptions 1 and 2 as stated in Introduction. Transition measure rnx(-) of x(,) being defined by Eq.(2), Eq.(12) yields f(")(IE,J;) = Mk(rn,cn)). Hence Assumption 1 (Eq.(3)) implies for k E N
z(~)(IE,J;)N Mk(rna)n'/'
( n + 00).
(24)
Since IE,J,~ is in z ( S ( n ) )it, satisfies a factorization property
p((IE,J;IE,J;) for X E Y,. property
= p((IE,J;)p(E,J;)
Assumption 2 for
p'(lE,J;JE,J;)
(k,Z E N)
(25)
x(,) implies an asymptotic factorization
- ~'"'(IE,J;)p(lE*J;) = o(n('+')I2) (n + 00) (26)
155
as is shown in $84.2. Eq.(24) and Eq.(26) tell us that the "mean" and %tandard deviation" have the following asymptotics respectively for k E N:
~+")(E,J,") x nkl2, {~$")(IE,J,"IE~J:)- j $ n ) ( ~ ~ , ~ , k )=~ o(nk/') }l/~ . These indicate that concentration at the mean occurs (relatively) in the n -+ 00 limit. Through a standard discussion by using the Chebychev inequality, let us state the concentration more precisely. Combining Eq.(12) and Eq.(25) with Eq.(l), we have
- nk/2Mk(mA))2)=
p'((IE,J; =
c
c
P'"'(X)p((IE,J,"
- nkh4k(mA))2)
X€Y,
P'"'(X)(Mk(mx)- nk/2Mk(mA))2 ,
X€Y,
On the other hand, Eq.(26) yields f ( n ) ( ( I EJ," ~ -nkI2 Mk (ma)12) = ( ~ ( ( I~E1~ J ,-nk/2Mk ") (mall2+o(nk) = nkcn
where
Thus we have for any b
>0
Eq.(24) assures en + 0 as n + 00. Note Mk(mxfi) = n-k/2Mk(mx) by Eq.(lO). Taking b in Eq.(27) so that lim
n+co
nkcn =O b2
and
b =O n-+m nk/2 lim
e.g. b = nk12ck/4,we have P'"'({X E YnIIMk(m,fi) - Mk(mA)I 2 '&I4}) 5 CAI2
.
(28)
Now we should think of topology on the set of continuous diagrams. Let us refer to the topology on D determined by a family of semi-distances {dk}k=0,1,2,
...:
156
as the moment topology. We have from Eq.(8) fik(Wd-Pk(w2)
=
k ( k - 1) 2
1, O0
.k-2(w1(.)-w2(.))dx
(w1,wa E D;k 2 2).
{fik(w)}k=2,3,... and {Mk(mw)}k=2,3,...are expressed as polynomials of each other for w E D. Hence the moment topology on D is equivalent to the one determined by {IMk(m,,) - Mk(mwz)l}k=2,3,... . Eq.(28) indicates that both of 0 the error in shape with respect to the moment topology on D 0 the error in probability with respect to P(n) can be simultaneously arbitrarily small in the l/fi-scaling regime as n + 00. We thus observe concentration at the irreducible components corresponding to diagrams quite near the shape A as a weak law of large numbers with respect to the moment topology on ID.
Remark On the space of continuous functions on a compact interval [-A, A] of R, the uniform norm topology is equivalent to the one defined by the family of semi-norms { s_AA xk . dz k=O, ,2,, . If we are in a situation that we can restrict ourselves to dealing with A-balanced diagrams for some A > 0 and l/fi-scaling, Eq.(28) implies a weak law of large numbers also in the uniform topology on D.
I
I1
,,
In the case of the regular representation L, of S(n), we readily see Eq.(24) and Eq.(26) directly from Theorem 1.1. In fact, Eq.(6) yields
={
# ~ ~ ( ( 2 ~ / ~if )IC )isneven ~ / ~ 0 if k is odd
since gLn = 6,. The Lth moment of the standard semi-circle distribution (= ma) appears in the right-hand side. Again from Eq.(6), we have
x
g L n (Auou(1
n
- 100 I
"(in- 170 I
).
The value of g L n can survive only when 0' = ro and hence only when the parities of k and 1 coincide. Then, letting k 5 I , we continue as
157 Note that for u E Y:
1-k 1-k 21(u)+-- IuOI= -+Ial-TOW(u)+m2(a) 2 2
= -2
+
(TOW(u)-m2(0))
.
The top term in Eq.(29) corresponds t o u such that row(u) - m ~ ( uis) minimal. Hence Eq.(29) is equal to
On the other hand, in
the value of the character can survive only when no = ro = 0 and hence only when both k and 1 are even. Then, the right-hand side is #NC((2k/2))#NC((21/2))n("+1)/2(1 + O ( n - l ) ) . This verifies Eq.(26) for regular characters. We thus observe concentration at the limit shape R with respect to the moment topology on D in irreducible decomposition of regular representations.
Remark In [5] Ivanov-Olshanski showed this concentration at the limit shape by computing the transition rule for some generators in the polynomial functions (after [ S] ) on Young diagrams. Applying Hammersley's theorem after that, they derived concentration with respect to the uniform topology (as is noted in the previous Remark), namely reconstruction of the result on the limit shape due to Vershik-Kerov [13] and Logan-Shepp [91.
158
4.2. Asymptotic factorization f o r Jucys-Murphy operators
By using Theorem 1.1, we show that Eq.(26) follows from Assumption 2. Theorem 1.1, Lemma 3.6 and Assumption 2 yield
p) (EnJ,klEnJA)
+
+
where we used Z(a) Z ( T ) - (Z(a") Z(7"))/2 =1.1( hand,
+ 1~1)/2.On the other
g(n)(EnJ,k)ji("' (EnJA)
+q
n ( k + w - l
1.
Combining these two, we obtain Eq.(26). The 0-terms depend only on Ic, 1. In conclusion, this completes the proof of a fundamental scheme of the concentration phenomenon stated in Introduction. References 1. P. Biane, Adv. Math. 138, 126 (1998). 2. P. Biane, Int. Math. Res. Notices, 4, 179 (2001). 3. P. Biane, Characters of symmetric groups and free cumulants, See A.M.Vershik (ed.), [12]. 4. A. Hora, Jucys-Murphy element, walks on Young graph and application to asymptotic representation theory, Preprint, 2004.
159 5. V. Ivanov and G. Olshanski, Kerov’s central limit theorem for the Plancherel measure on Young diagrams, In: S.Fomin (ed.), Symmetric functions 2001, Kluwer Academic Publishers, 2002. 6. S. V. Kerov, Funct. Anal. Appl., 27, 104 (1993). 7. S. V. Kerov, Asymptotic representation theory of the symmetric group and its applications in analysis, MMONO 219, Amer. Math. SOC.,2003. 8. S. Kerov and G. Olshanski, C. R. Acad. Sci. Paris, 319,SBrie I, 121 (1994). 9. B. F. Logan and L. A. Shepp, Adv. Math., 26,206 (1977). 10. R. Speicher, Mem. Amer. Math. Soc., 627,1998. 11. I. Terada and K. Harada, Group theory (in Japanese), Iwanami-Shoten, 1997. 12. A. M. Vershik (ed.), Asymptotic combinatorics with applications to mathematical physics - St. Petersburg 2001, LNM 1815,Springer, 2003. 13. A. M. Vershik and S. V. Kerov, Soviet. Math. Dokl., 18,527 (1977). 14. D. V. Voiculescu, K. J. Dykema and A. Nica, Free random variables, CRM Monograph Ser. 1, Amer. Math. SOC.,1992.
This page intentionally left blank
Infinite Dimensional Harmonic Analysis I11 (pp. 161-176) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
REAL HARDY SPACES ON REAL RANK 1 SEMISIMPLE LIE GROUPS
TAKESHI KAWAZOE * Department of Mathematics, Keio University at Fujisawa 5322 Endo, Fujisawa, Kanagawa 252-8520,Japan E-mail: kawazoe Osfc.keio.ac.j p
Let G be a real rank one connected semisimple Lie group with finite center. We introduce a real Hardy space H 1 ( G / / K )on G as the space consisting of all Kbi-invariant functions f on G whose radial maximal functions M,+f are integrable on G. We shall obtain a relation between H 1 ( G / / K )and H1(R),the real Hardy space on the real line R, via the Abel transform on G and give a characterization of H 1 ( G / / K ) .
1. Introduction
The study of the classical Hardy spaces on the unit disk and the upper half plane was originated during the 1910’s by the complex variable method. In the 1970’s the Hardy spaces were completely characterized by various maximal functions of their boundary values and also by atomic decompositions, without using the complex variable method. This is a significant breakthrough in harmonic analysis. Nowadays, the spaces defined by the real variable method - maximal functions and atoms - called real Hardy spaces and a fruitful theory of real Hardy spaces has been extended to the spaces of homogeneous type: A topological space X with measure p and distance d is of homogeneous type if there exists a constant c > 0 such that for all z E X and T > 0 P ( B ( z ,2 T ) )
I CP(%
TI),
where B ( z , r ) is the ball defined by {y E X I d ( z , y ) < T } and p ( B ( z , ~ ) ) the volume of the ball (cf. [l,$11). However, when the space X is not of homogeneous type, little work on real Hardy spaces on X has been done. *supported by Grant-in-Aid for Scientific Research (C), no. 13640190,Japan Society for the Promotion of Science
162
Hence, looking at the example of a semisimple Lie group G as a space of non-homogeneous type, we shall introduce a real Hardy space H1(G//K) by using radial maximal functions on G. In this article we shall overview some results obtained in the previous papers [5], [6], [7] and announce a new characterization of H 1( G / / K ) ,which gives a relation between H 1(G//K) and the real Hardy space H1(R) on R via the Abel transform on G. 2. Notation
Let G be a real rank one connected semisimple Lie group with finite center, G = KAN = KAK Iwasawa and Cartan decompositions of G. Let dg = dkdadn = A(a)dkdadk' denote the corresponding decompositions of a Haar measure dg on G. In what follows we shall treat only K-bi-invariant functions on G. Since A is identified with R as A = { a x ;x E R}, all K-biinvariant functions can be identified with even functions on R denoted by the same letter as f ( g ) = f(ac(,,) = f ( 4 g ) ) = f ( - a ( g ) ) . We may regard the weight A(a,) as an even function given by
A(x) = c(
(sh21xl)2P+1,
(1)
+
where a = (ml m2 - 1)/2, ,8 = (mz- 1)/2 and ml,m2 the multiplicities of a simple root y of ( G , A ) and 27 respectively. We note that the one dimensional space R with normal distance and weighted measure A ( z ) d x is not of homogeneous type, because A(x) e2px with p = (Y ,8 1 > 0 as I + 00. Let LP(G//K) denote the space of all K-bi-invariant functions on G with finite Lp-norm and L:,,(G//K) the space of all locally integrable, K-bi-invariant functions on G. Let F be the dual space of the Lie algebra of A and for X E F, cpx the normalized zonal spherical function on G:
-
~ p x ( x= ) zFi ( ( p
+ +
+ iX)/2, ( p - iX)/2;a + 1; -sh2z) ,
where zF1 is the Gauss hypergeometric function. We recall that, if X $ Z, then cpx (x)has the so-called Harish-Chandra expansion:
+
cpx(x) = e-Px (+(A, z)C(X)eixx +(-A,
z)C(-X)e-"x) ,
(2)
where C(X) is Harish-Chandra's C-function. For some basic properties of cpx(x), +(X,x), and C(X) we refer to [2, $2, $31 and [12,9.1.4, 9.1.51.
163
For f E L1(G//K) the spherical Fourier transform f(X), X E F, of f is defined by
m J =
G
f(dcpx(S)dS=
lrn
f(x)vx(2)A(x)dx.
Since cpx(x) is even with respect to A, x and uniformly bounded on x if X is in the tube domain F ( p ) = {A E F,; lSXl 5 p } , it follows that f(X) is even, continuously extended on F ( p ) , holomorphic in the interior, and
Iml L Ilflll,
E
For f E CF(G//K) the Paley-Wiener theorem (cf. [2, Theorem 41) implies that f(X) is holomorphic on F, of exponential type. Furthermore, it satisfies the inversion formula f(X)
=
loo f(X)cpx(~)lC(W2dX
J-CO
and the Plancherel formula
Therefore, the spherical Fourier transform f I+ f of C r ( G / / K ) is uniquely extended to an isometry between L'(G//K) = L2(R+, A(x)dx) and L2(R+, lC(A)l-2dX) (cf. [2, Proposition 31, [12,Theorem 9.2.2.131). For f E CT(G//K) we define the Abel transform F;, s E R, of f as
q(") =
eP(l+sb
(3)
Here the Euclidean Fourier transform (FfS)-(X)is holomorphic on F, of exponential type, because Ffs(f) E CF(R), and it coincides with the spherical Fourier transform o f f :
f(X
+ isp) = (FfS)-(X), X E F,
(4)
(cf. [9, $31). Especially, F; is even on R. The integral over N in (3) can be explicitly rewritten by using a generalized Weyl type fractional integral operator W;: For a > 0, p E C and y > 0,
where n = 0 if Rp > 0 and -n < Rp 5 - n + 1 , n = 0 , 1 , 2 , . - . ,if Rp 5 0 (see [9, (3.11)]). Then Koornwinder obtains that for x > 0,
q") =KLa
O
W,2+,,,(f)(4
164
(see [9, (2.18), (2.19), (3.5)]). In the following, for simplicity, we denote W+(f)(E) = Ff(l.l), that is, 1
W+(f)(.)= ePXW,-p
O
~;+l/z(f)(l.l~,
2
ER
(6)
and for a smooth function F on R+,
w - ( F ) ( x ) = W!(p+,/,) 0 W]-(,+,(e-PxF), x E R+.
(7)
Clearly, W- o W+(f) = f and W+ o W- ( F ) = F . For f E L1(G//K), W+(f) belongs to L1(R), because the integral formula for the Iwasawa decomposition of G yields that
llflll
~ ~ ~ + ( f ) ~ ~ L 15( R . )
(8)
(cf. [9, (3.5), (2.20)]). Hence W+(f)"(X), X E F, is well-defined and it follows from (4) that f(X
+ ip) = W+(f)"(X),
X E F.
Let f , g E L1(G//K). Since f * g E L1(G//K) and (cf. [2,Theorem 51, [3,§5]), it follows that
(9)
(f * g)"(X) = f(X)@(X)
W+(f * 9) = W+(f) * W+(g).
(10)
We say that a function F on R is W+-smooth if W - ( F ) is well-defined and continuous. Then, for W+-smooth functions F, G on R with compact support such that e-P"F and e-PxG are even, it follows that
W - ( F * G) = W - ( F ) * W-(G). 3. Radial maximal functions
As in the Euclidean case, to define a radial maximal function we need to define a dilation 4t, t > 0, of a function on G. Let 4 be a positive compactly supported C", K-bi-invariant function on G such that
J,4(9)d9 = Jdm 4(x)Ab)dx = 1
(11)
and furthermore, there exists it4 E N such that
+(.) We define the dilation
$t
of
4 as
=O ( P ) .
(12)
165
Clearly, q5t has the same Ll-norm of 4: Il4tlll = 114111and, for 1 5 p 5 00, it gives an approximate identity in LP(G//K) (see [2,Lemma 161). We here introduce the radial maximal function M+f on G as follows.
Definition 3.1. For f E LtOc(G//K),
As shown in [5,Theorem 3.4 and Theorem 3.51, M4 satisfies the maximal theorem and, for 1 I p 5 00, llfllP 5 cllM+fllpif the both sides exist. By using W+(&),we shall define a maximal function on R as follows. Definition 3.2. For F E &(R), (M,RF)(z)= SUP O
I(F* W+(4t)(z)l,
z E R.
Since W+(f * 4t) = W+(f) * W+(q&) (see (10))and W+ is an integral operator with a positive kernel (see (5), ( 6 ) ) ,it follows that SUP IW+(f> * W+(4t)(x)lI W+
O
( SUP
O
If *)I.
(.I.
Therefore, from (8) we have a relation between M+ and M?: Proposition 3.3. For f E L t o c ( G / / K ) ,
( M ? W + ( f ) ) ( z ) I W + ( M + f ) ( z ) , 2 E RI n particular,
I IM,RW+ (f II
I CI IM+f II1
L1 (R)
if the both sides exist.
+
Now we note that W+(+t)-(X) = &(A i p ) (see (9)) has similar properties of the Euclidean Fourier transform of a Euclidean dilation:
(1) There exists c such that for all t
> 0, X E R and 0 I k I M ,
( 2 ) There exists c such that for all t
> 1 and X E R,
166
where M is the same as in (12). These properties guarantee that W+(q5) behaves like a Euclidean dilation on R. Hence, the maximal operator M p can characterize H1(R),that is, F E H1(R)if and only i f M F ( F )E L1(R):
Theorem 3.4. Let q5 be as above and suppose that M 2 2. Then F E H1(R) if and only if M T F E L1(R): llFllH1(R)
IlMrFIIL1(R)*
4. Real Hardy spaces Let q5 be the same as in the previous section (see ( l l ) , (12)) and M 4 , M p the corresponding radial maximal operators on G and R respectively (see Definitions 3.1 and 3.2). In this section we shall define two real Hardy spaces H i ( G / / K )and W-(H1(R))on G and give a relation between them.
Definition 4.1. W e define
H $ ( G I / K )= {f E L:oc(GIIK) i M4f E L1(GIIK))
IIf IIH; (G) = II'4fl
and
Clearly, since
11.
llflll
5 cllM@flll,it follows that ( G / / K )c L' ( G / / K ) .
Next we shall introduce a pull-back of the real Hardy space H1(R) on R to G via W+ (see (6)). Let M,, s 2 0, denote the Euclidean Fourier multiplier defined by
M,(F)"(A) = (A
+ ip)'F"(A).
Definition 4.2. For s 2 0, we define ~ - ( h ( H 1 ( R )= ) ){f E L:,,(G/IK)
7.
Ms OW+(f) E H1(R))
and give the norm by llMso W + ( f ) l l ~ ~ (W ~ e) .denote W_(Mo(H'(R))) by W-(H1(R)) for simplicity. Obviously, Proposition 3.3 and Theorem 3.4 yield the following. Corollary 4.3. Let M 2 2. There ezists a positive constant c such that ~ ~ ~ + ( f ) ~ ~5 H C1 l (l f R l l H) ; ( G ) for all f E H i ( G / / K ) and thus,
H i ( G / / K )c W-(H1(R)). Let s, = a
+ 1/2.
Then we see that
W-(M-sa(H1(R))) c W-(H1(R)).
167 Actually, let f E W-(M-Sa(H1(R)))and put F = W+(f). By the definition, Msp(F)belongs to H1(R).Since the Fourier multiplier M-sa satisfies the Hormander condition (cf. [ll,$5 in Chap.ll]), it is bounded on H1(R) (cf. [ll,Theorem 4.4 in Chap.141). Thereby, F E H1(R) and the desired inclusion (13) follows. Similarly, since the Fourier multiplier MYa' o WRT, 0 5 y 5 ,s, which corresponds to ( i X ) T / ( X + i p ) s p , satisfies the Hormander condition, it is bounded on H1(R). Hence, each WFT(F) also belongs to H1(R):For 0 5 y 5 ,s,
IIW R (~F, I IH1(R) = IIMT (W?~( F ) IIL1(€2) 5 cI I
II
sa ( F ) H1(R)*
Now we shall characterize the Hi-norm off E H i ( G / / K )and show that the real Hardy space H i ( G / / K )is located between W-(M-,_ (H1(R)))and W-(H1(R)) (see (13)). We recall that
(see (10)). Therefore, roughly speaking, the Hi-norm of f , that is, the L1-norm of M,#,f on G (see Definition 4.1) can be characterized in terms of the L1-norm of M T ( W - ( F ) ) (see Definition 3.2). We rewrite W-(F) by using the Weyl type fractional operator WF on R
Here WF(F)(chy) = Wi(f)(y) if f ( x ) = F(chx) and a = ,Ll = 1/2 in (5). Let 6 = (a - p) - [a - p] and 6' = (p - 1/2) - lo - 1/21, where [ ] is the Gauss symbol, and put 14 = [s,], = S 6' and D = (6,6',6 6') respectively. Then the local and global forms of W-(F) can be rewritten as follows:
a
+
(1) If F is W+-smooth and supported on 0 < x
where the sum is taken over 0 5 m 2 11 and
+
5 1, then
< E D and AL(z, s) satisfies
0 5 AL(x, s) 5 x-2sa+&++m-1 for all 0 < x
5 s.
(14)
168
( 2 ) If F is W+-smooth and supported on z 2 1, then lW-(F)(x)l
where A&(x,s) satisfies (14)and for j = 3,4, A L ( x ,s) 2 0 and there exists a positive constant c such that
s,'
A k ( x ,s)dx 5 c for all s
> 0.
By using these local and global forms of W - ( F ) , we can rewrite the L1-norm of M T ( W - ( F ) ) on R in terms of M?(W?7(F)), 0 5 y 5 s,, on R. Finally, we have the following. Theorem 4.4 Let M 2 2 and F = W+(f) for f E W - ( A L S a ( H 1 ( R ) ) ) . Then there exist c1,cz such that for all 0 5 y 5 s,, W_R,(F)(x)(thx)rlIL'(R) 5
clIIMF
IlfllHi(G)
n
5 c2
IIMF w?(,+€)(F)(x)(thx)m+611L'(R). m=O €ED
Especially, n
IlfIIH$(G) m=O
c IIMF
WTm+.t)(F)(x)(thx)m+'IIL'(R)
and thus,
W-(ALSa(H1(R))) c H $ ( G / / K )c W - ( H 1 ( R ) ) .
169
Remark 4.5. Let C(X) be Harish-Chandra’s C-function (see (2)) and Mc, the Euclidean Fourier multiplier corresponding to C,(X) = C(X + i p ) : Mc,(F)”(X)
= C(X
+ip)F“(X).
We define
~ - ( M c p ( H 1 ( R )= ) ){f E J%,(G//K) i ME; 0 W+(f) E H1(R)). Then it easily follows from Theorem 4.4 that w-(Mc,(H1(R))) c f q G / / K ) c w-(Hl(R)). This is one of the main results in [6]. However, the proof in [6]was a little bit complicated, because to obtain the first inclusion we used the HarishChandra expansion of the zonal spherical function cpx and also the Gangolli expansion of (see (2) and [2, $31). Thereby, to sum up the estimates of each expanded terms we required sharp ones. Here we can obtain the desired inclusion as an easy consequence of Theorem 4.4. 5. Atomic Hardy spaces
We introduce atomic Hardy spaces on G. In the Euclidean case the atomic Hardy space H&,o(R)coincides with H1(R) (cf. [4, Theorem 3.301, [lo, $2 in Chap.31). However, it may be not true in our setting, because the Lebesgue measure dx is replaced by the weighted measure A(x)dz (see (1)). We denote the interval [xo - r, xo + r] by R(Q, r ) and set the volume by lR(xo,r)l =
J
zo+r
A(x)dx.
zo -r
We say that a K-bi-invariant function a on G is a (1,cqO)-atom on G provided that there exist xo 2 0 and r > 0 such that
Here a is identified with a function on R+. Similarly, we shall define a (1,00,0, €)-atom a and a (1,00, +)-atom a by replacing (ii) with
5 IR(xo,r)I-l(l+
(ii)€
llallca
(iii)+
j0 a(x)A(z)dx = 0 i f r 5 1
and (iii) with P M
(16)
170
respectively. Then we shall define atomic Hardy spaces H L , o ( G / / K ) , HZo(G//K)H , k ; ( G / / K ) and h&,,(G//K) as follows. Definition 5.1. Let notations be as above. We define Xiai ;
H & , ( G / / K )= { f =
ai is (1, 00, 0)-atom on G and
xilXil < cm}
i
and denote the norm by
Ilf llH;,o(G)
=i n f CP i 1 7 i
xi
where the infimum is taken over all such representations f = Xiai. We also define H z o ( G / / K ) ( E 2 0) and H k $ ( G / / K ) by replacing ( l , c m , O ) atoms a on G in the above definition with ( 1 , q 0 , €)-atomsand ( l , c m , +)atoms respectively. Moreover, we define the small Hardy space h&,o( G / / K ) on G by restricting ( l , c m , 0)-atoms in the definition ofH&,,(G//K) to ones with radius 5 1. Clearly, for
E
2 0,
h:,O(G//w Let
x1
c H ~ o ( G / / Kc) HL,o(G//K)c Hk;(G/IK).
denote the characteristic function of B(l) = R(0,l) and set O(g) = lB(1)r1x1(9)> 9 E G.
Moreover, for each (not necessarily K-bi-invariant) function f on G, we define a K-bi-invariant function f:, x E G , as fL(g) =
11
f(x-'kgk')dkdk',
K
g E G.
K
Then the difference between hL,,(G//K) and H k s ( G / / K ) is given as follows.
Proposition 5.2. For f E H k ; ( G / / K ) there exist fo E h&,o(G//K)and xi E G, X i E R such that i
where
11 foIlaky(G) and
xiJ X i l are respectively bounded by 11f IIH~t,(G).
As in the Euclidean case, we shall introduce the truncated maximal operator M$" on G as (M:"f)(9) = SUP O
I(f * # 4 ( g ) l ,
9 E G.
171
Then M?" is bounded from H L $ ( G / / K ) to L 1 ( G / / K )(see [7]).As for M4, we see from [7, Theorem 5.31 that it is bounded from H k $ ( G / / K ) n W-(H1(R)) to L ' ( G / / K ) : Proposition 5.3. Let M 2 2 . M4 is bounded from H k $ ( G / / K ) n W-(H1(R)) to L ' ( G / / K ) , that is, there ezists a constant c > 0 such that
+ IlW+(f)lIH1(R))
IIMdll L c (llfIlHl.>(G) for all
f
E H;$(G//K) r l W-(H1(R)) and thus,
H ~ $ ( G / / Kn) w-(H~(R))c H ; ( G/ / K ) . Let a be a (l,oo,0, 1)-atom on G supported on R(zo,r ) . The conditions (15) and (16) imply that Ilalla,
5 lR(zo,r)l-l(l
+ r)-l
and
Then A = W+(a)is supported on R(z0,T ) and
L 00
A ( z ) d z = A"(0) = ii(ip) =
L
J, a(g)dg = 0.
a(g)dg = 0.
Moreover, it follows from ( 6 ) and [8, Lemma 3.41) that IA(z)I ce2Pzth(zo ~ ) ~ ' ~ I l aHence, ll~. ) Case I: zo - r 2 1. Since A is supported on R ( z o , T and
+
IR(zo,r)I
N
/"""
5
eZPzdx e2proshr, N
20-r
+
it follows that IA(z)I 5 Ce2P(zo+r)(e2pzoshr)-1(1 r)-' Case 11: zo - T < 1 and T 2 1. Since zo T 2 1,
+
Therefore, as in Case I, we have llAlloo 5 cr-l. Case 111: 20 - r < 1, r < 1 and zo > 2r. Since zo 20 T 2 3 and thus
+
IR(.o,r)I
-
5 cr-'.
> 2r, it follows that
zo+r
z Z s a d z5 c(z0 - T ) ' " T .
Since ( z ~ + r ) / ( z ~ -5r 3, ) we have IA(z)I 5 c t h ( x 0 + ~ ) ~ ~ ( ( 2 0 - r ) ' a r5 )-~ cr-1.
+
Case IV: zo - T < 1, T < 1 and zo 2r. Since zo T 5 3r < 3 and lR(zo,r)l 2 IB(r)I IB(3r)l, we may suppose that a is a centered atom supported on B(3r). Then IA(z)I 5 c(th3r)'-IB(3r)l-l 5 cr-l. N
172
These four cases imply that cA is a (1,00,0)-atom on R, where c is independent of a. Therefore, we obtain the following.
Theorem 5.4. Let M 2 2. Then
H~~~(Gc / /HK~)$ ( G / / Kn) w-(H~(R)). Especially, M,#, is bounded from H z o ( G / / K )to L 1 ( G / / K ) ,that is, there exists a constant c > 0 such that
for all
f E Hkto(G//K).
6. Characterization of
Hi(G//K)
We shall prove that the inclusion in Proposition 5.3 is the equality. Let s, = a 1/2 as above and put
+
da(X0,T) =
/
1x0 I+r
(thx)sudz .
max{O,l+ol--r)
We define a subspace Hk$(R), of H1(R)as the space of all F = such that lXil < 00 and each Ai satisfies
xi
xiXiAi
Definition 6.1. W e define
We can construct a (1,00,+)-atomic decomposition (see (17)) for f E W-(Hk$(R),). Let F = W+( f) and F = XiAi the decomposition of F given by the definition, that is, lXil < 00 and each Ai satisfies (i) to
xi
xi
(iii) in (18). Here we may suppose that ri I 1. Actually, when ri > 1, we decompose the support of Ai by using a smooth decomposition of 1, where each piece is supported in the interval with radius 5 1 and thus, we have Aij and each Aij satisfies (18) with radius 5 1. Moreover, we Ai = may suppose that zi = 0 with ri 6 1 or lxil > 2ri, because, if zi # 0 and
xj
173
lxil 5 2 ~ ithen , we may regard xi = 0 without loss of generality. Hence, we can rearrange the decomposition of F as
where each Ai satisfies (i), (ii) with xi = 0,
~i
5 1,
I
A i ( x ) d z = 0; each
Bj satisfies (i) to (iii) with Ixjl 2 2 ~ j ~j , < 1; each Ek satisfies (i), (ii) with 1xk1 2 2Tk, T k = 1, and moreover, Ci ( X i ( Cj lpjl+ Ck < 00. Since F is W+-smooth, finally, we have
+
i
k
j
where ai = W-(A,), b j = W-(Bj) and ek = W-(Ek). Here it is easy to see that each ai, bj, ek have the same supports of Ai, Bj, Ek respectively. Now we apply fractional calculus in [8] to estimate each ai, bj, ek. For simplicity, we abbreviate the suffices i, j,Ic and denote the supports of a, b, e by R(x0,T ) . Without loss of generality, we may suppose that 20 2 0. As for e, since e is supported on R(xo, 1) and 20 2 2, it follows that xo - 1 2 1 and thus, da(xO,1) 1. Thereby, (ii) and [8, Lemma 3.31 imply that on the support of e
-
< ce-2pz Ie(x)I 5 c(thx)-(a+1/2)e-2p" This means that c-le is a (l,cm, +)-atom on G. As for b, we recall that xo > 2r. Case I. zo - T 2 1: Since 20 - T 2 1, da(x0,r) Lemma 3.31 imply that on the support of b lb(x)I I c(thx)-(a+1/2)e-2p"r-1
5 clR(z, 1)I-l.
-
T.
Thereby, (ii) and [8,
I ce-'pZT-' I clR(x, r1l-l
This means that c-lb is a (l,oo,O)-atom on G. Case 11. 20 - T < 1: Since T < 1 and 20 > 2r, it follows that xo < T 1 < 2, 20 - T > x0/2, and xo + r < 3 ~ / < 2 3. Therefore, da(x0,r) I c(xo - T ) ~ = T and thus, on the support of b
+
-'
lb(x)I 5 c(thx)-(Q+1/2)e-2p"r-1(Z0 - T ) - ~ " < - c(xo - T ) - ( ~ ~ + T' ) Since
+
(20
T)/(ZO
.
- T ) 5 3, it follows that
IR(x0,r)l 5 c(x0
+
T)2a+lT
I c(z0 - T ) 2 a + l T .
Therefore, lb(x)I 5 clR(xo,~)I-' on the support. This means that c-lb is a (l,cm,O)-atom on G.
174
As for a, since 20 = 0 and
).(.I
< 1, it follows that d,(O, T )
T
I 5 c(thz)-(a+l/2)
-
rS-+l and
5 cA(z)-'r-'.
~-~WT-~T-(SU
(20)
Hence, if we put a+(.) = 44-'.-'x[o,l.](.), then la(.)[ L1-norm:
2
> 0,
5 a+(.) and a+ is a non-increasing function on R+ with finite Ila+llLl(A) =
I"
a+(z)A(z)dz = @. 00
Since a is supported on B ( T )and follows that IB(s)l-'
lw 00
a(z)A(z)dz =
A ( z ) d z = 0, it
a(z)A(z)dz is also supported on B ( r ) and
Here we used (20) and IB(s)l 1
-
A(s)s if s 5 r
5 1 (see (1)). Hence,
PW
This means that ca+ is an L' non-increasing denominator of a satisfying (21). Then [5, Theorem 4.51 yields that a has a centered (1,00,0)-atomic decomposition a = Cjrjaj on G such that Cj IrjI 5 ~ l l a + l l L ~ ( A5) cq,. 5 ccoEspecially, a E e , o ( G / / w and IlallH~,,o(c) These three cases imply that all ai, b j , ek in (19), and thus f belongs to H k $ ( G / / K ) : Proposition 6.2. All functions in W-(Hk$(R),) have (1,00,+)-atomic C Hk>(G//K). decompositions, that is, W-(Hk$(R),) Now we shall prove that H $ ( G / / K )C W-(HL$(R),). We shall give a sketch of the proof in the case of s, = a 1/2 is integer. Let f E H $ ( G / / K )and put F = W+(f). Then it follows from Theorem 4.4 that IIMT O WFSU(F)(z)(thz)SullLl(R)< 00. We recall that (thz)', is an Alweight. Therefore, WFSu( F ) has a (1,00, s,)-atomic decomposition with respect to this weight:
+
175
L 00
where Bi is supported on R ( z i , r i ) ,
llBillw 5 d , ( z i , r i ) - l and
xi
lXil
< 00.
B i ( z ) z k d z = 0, 0 5 k 5 s,
We set
Since s, is integer and each Bi satisfies the s,-th moment condition, it easily
L 00
follows that Ai is supported on R(z,,r i ) and
A i ( z ) d z = 0. Moreover,
~~W!!sJA~ =)llBilloo ~ ~ w 5 d , ( z i , r i ) - l . Therefore, each Ai satisfies (18) and thus, F E Hk$(R), and f has a (1,00,+)-atomic decomposition on G by Proposition 6.2. Furthermore, we can drop the assumption that s, is integer. Therefore, we have H i ( G / / K ) C H k $ ( G / / K ) in general. Finally, as a refinement of Proposition 5.3, we have the following main theorem. Theorem 6.3. Let notations be as above. Then
H i ( G / / K ) = H k $ ( G / / K ) n W-(H1(R)). As an easy consequence of the previous argument, we have Theorem 6.4. Let E 2 0 . Then H Z o ( G / / K )n W-(H1(R)) is dense in W-(H1(R)). Especially, H i ( G / / K ) is dense in W-(H1(R)).
References 1. Coifman, R.R. and Weiss, G., Extensions of Hardy spaces and their use in analysis. Bull. of Amer. Math. SOC.,Vol. 83, 1977, pp. 569-645. 2. Flensted-Jensen, M., Paley- Wiener type theorems for a diflerential operator connected with symmetric spaces. Ark. Mat., Vol. 10, 1972, pp. 143-162. 3. Flensted-Jensen, M. and Koornwonder, T., The conwolution structure and Jacobi transform expansions. Ark. Mat., Vol. 11, 1973, pp. 245-262. 4. Folland, G.B. and Stein, E.M., Hardy Spaces on Homogeneous Groups. Mathematical Notes 28, Princeton University Press, New Jersey, 1982. 5. Kawazoe, T., Atomic Hardy spaces on semisimple Lie groups. Japanese J. Math., Vol. 11, 1985, pp. 293-343. 6. Kawazoe, T., L1 estimates for maximal functions and Riesz transform on real rank 1 semisimple Lie groups. J. Funct. Analysis, Vol. 157, 1998, pp. 327-527. 7 . Kawazoe, T., Hardy spaces and maximal operators on real rank 1 semisimple Lie groups I. Tohoku Math. J., Vol. 52, 2000, pp. 1-18. 8. Kawazoe, T. and Liu, J., Fkactional calculus and analytic continuation of the complex Fourier-Jacobi transform. To appear in Tokyo J. Math.
176
Koornwinder, T., A new proof of a Paley- Wiener type theorem for the Jacobi transform. Ark. Mat., Vol. 13, 1975, pp. 145-159. 10. Stein, E.M., Harmonic Analysis. real-variable methods, orthogonality, and oscillatory integrals. Princeton Mathematical Series, 43, Princeton University Press, New Jersey, 1993. 11. Torchinsky, A., Real-variable Methods in Harmonic Analysis. Pure and A p plied Mathematics, 123, Academic Press, Orlando, Florida, 1986. 12. Warner, G . , Harmonic Analysis on Semi-Simple Lie Groups 11. SpringerVerlag, New York, 1972.
9.
Infinite Dimensional Harmonic Analysis I11 (pp. 177-186) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
WHITE NOISE ANALYSIS, FILTERING EQUATION AND THE INDEX THEOREM FOR FAMILIES RQmi LQandre Institut Elie Cartan. UniversitQ Henri Poincar6 54000. Vandoeuvre-les-Nancy. FRANCE emai1:leandreQiecn.u-nancy.fr New adress after first Febmary ZOO4 Institut de MathQrnatiques.FacultQdes Sciences. UniversitQ , 21000. Dijon. FRANCE email:[email protected]
deBourgogne
Abstract: We give a meaning to the heuristic formulas of Bismnt relating some integrals of forms on the loop space to the Index theorem for families as a Hida distribution with values in the space of Lz forms on the loop space of the parameter space. I. Introduction Let M be a compact spin manifold. L,(M) is the smooth free loop space. It inherites a canonical circle action, and the generator of this circle action is called the canonical Killing vector field X, on L,(M). X, can be seen as a 1-form over L,(M) by considering the Lz metric on the free loop space. Atiyah, following a suggestion by E. Witten, established the following heuristic fomula ([At]):
where D+ is the Dirac operator on M and IndD+ its Index. Atiyah was motivated by an extension in infinite dimension of the localisation formulas of Duistermaat-Heckman ([D.H]). Let us introduce a complex bundle t on M . Bismnt ([Biz], [Bis]) deduced from t a complex bundle t, on L,(M) and introduced, associated to it, the Bismut-Chern character Ch(<,) which is a series of even forms over L,(M) such that: (d + ix,)Ch(t,)
(1.2)
Bismut considered the Dirac operator tensorised by
=0
< D+,c and established the heuristic formula:
motivated by the works of Berline-Vergne ([B.V1]). Getzler ([Gel) considered the map (1.4)
u
+ 2-1
l_,,,
expl-(d
+ ix,)Xml
Au
as a current (called Witten current) operating on Chen forms. A convenient framework to define this current is provided by Hida Calculus (See [H.K.P.S], [O]), where Chen forms replace the traditional role of Wiener chaos in white noise analysis. By using Potthoff-Streit theorem ([PSI),[Llo] showed that the Witten current is an Hida current.
178 We are interested in an extension of this theory to the case of a family of Dirac operators. Let us consider a fibration ?r : M + B of compact manifolds with hasis B a compact manifold. y is the generic element of B and z is the generic element of ?r-'(y) = V,. L,(M) and L,(B) are the smooth free loop space of M and B . s + z, denotes a loop in M ans s + y. denotes a loop in B. nm is the map which to {s + z,} associates the loop { s + r z s }and i, is the injection map from B into L,(B). We introduce a complex bundle [ on M , and a family of Dirac operators D,,< parametrized by y E V, and tensorized by the restriction of to V,. We consider the complex bundle , [ over L,(M) of sections of [ over each loop in M . We consider the Bismut-Chern character Ch(&,) which is a series of forms over L,(M) which is equivariantly closed: (1.5)
(d
+iX,)Ch(L) =0
Let y. be the generic element of L,(B). Bismut ([Biz], [Bi3], [Bib]) has introduced the formal expression
and said that, formally, this gives a series of smooth form over L,(B). Moreover, iLCh,(Ch([,)) is equal to a representative of the Chern character of the index bundle associated to the family of Dirac operators. We are interested in this paper in giving an interpretation of this heuristic formula in term of white noise analysis, generalizing to the case of a family of Dirac operators [Llo]. The main remark is that we have to work on the continuous loop space instead of the smooth loop space. Let L ( B ) be endowed with the free Brownian bridge measure. Over it we can define the space of Lz forms. (See [J.L]) 5in fact we choose an equivalent measure). We consider the algebraic model of Chen forms over M (and not over B ) as in [L.Ou]. By using the methods of the filtering equation as in [Bib] and the tools of white noise analysis, Wiener chaos being replaced by Chen forms, we can define a current operating over Chen forms on L ( M ) , called Bismut-Witten current, with values in the space of Lz forms on L ( B ) . The reader interested in a probabilistic proof of the Index theorem for families can see the original work of Bismut ([Bi4]) and the pedagogical work of LBandre ([Ll]), with a simplified proof. The reader interested in probabilistic tools used in Index theory can see the surveys of Kusuoka ([Ku]), LBandre ([Lz]) and Watanabe [Wa]. 11. Partial Malliavin Calculus
Let us consider the compact Riemannian spin manifold M of dimension m. We consider the fibration ?r from M into the compact manifold B. Let V, = ?r-'(y). We consider the bundle ?r*T(B)on M , where T(B) is the tangent bundle of B. We write **T(B) @ T ( V )= T ( M )where T ( V )is the bundle on M obtained by buit from the tangent bundle T(V,) of V,. On T(V,) we choose a Riemannian metric. From the previous decomposition, we deduce a Riemannian metric g M on T ( M ) . We can consider the Laplace-Beltrami operator A L on M associated to the Riemannian metric on M . Let AB,A' be the Levi-Civita connections on B and M for the considered metrics. On T ( V ) ,we choose the connection: (2.1)
v" = PT(V)VL
is the projection on T ( V ) . We can consider the pullback connection of VB on ?r*T(B)still where called VB. We consider Bismut connection on T ( M ) : (2.2)
v = V B @ v"
The decomposition of T ( M )into ?r*T(B)and T ( V )is parallel for this connection. Following Bismut ([Bib]), we put: (2.3)
s=vL-v
179 We denote by T the torsion of V. Let us remark that V is a metric connection. Let us recall (See [Bi5] p 29): -)Vv does not depend on the metric on B. -)T takes its values in T ( V ) . -)The tensor < S(.).,. > does not depend on the metric on B. -)For U in T ( M ) ,S ( U ) maps T ( V )into n*T(B). -)For U, V,W E n*T(B),, < S(U)V,W >= 0. Let us consider the operator AB +A, = A where A, is the lift of the Laplace-Beltrami operator on B and Av is the Laplace-Beltrami operator on the fiber V . A - A, is a first order operator. It is a vector field which is a section of n*T(B). Let us construct the heat semi-group associated to A1 as it was done in [B.FI. We can consider the orthogonal frame bundle S O ( M ) of M with projection T I . We consider the canonical horizontal vector fields X;on S O ( M ) for the connection V. We consider m - d Brownian motions wi and d Brownian motions Biwhere d is the dimension of B. We consider the stochastic differential equation in Stratonovitch sense of Eells-Elworthy-Malliavin:
where
ug
is an isometry which preserves the decomposition of R"' into Rm-d @ Rd and of T ( M ) , , l ~ u into o~
fBT(v)rt(uo).
~ * ~ ( ~ ) n - , ( U O )
The diffusion semi-group associated to A is represented by nl(ut)= z'(t). Let us consider the Brownian motion on M denoted by z ( t ) .The law of z ( t )is equivalent to the law of z'(t). It is given through Girsanov's formula by
for some smooth vector field a and some smooth function j3 (See [El], [I.W]). Let us remark that n(z'(t))follows the law of a Brownian motion on B. Let us follow [B.F] in order to do an intrinsic differential Calculus of the Bismut-Malliavin type for the process t + z'(t). We write X ; for the vector fields Xj in the direction of TV and Y , for the vector fields in the direction of s * T ( B ) .Let us consider some deterministic element V iof L2 and some deterministic elements ui of L z . We consider the perturbed equation:
The main interest of the connection V is that it preserves the decomposition of T ( M ) into r * T ( B )@ T ( V ) . Let 7,R be the equivariant representation of T and R, the curvature tensor of A. Let 8 be the canonical 1-form on S O ( M ) and w be the connection 1-form for the connection V. We get (See [K.N]):
d8
= -wA8
w = -w A w
We put, following the lines of [B.F]: (2.8)
and (2.9)
+r +R
180 We deduce as in [Bill, [B.F] and [Dr] that: (2.10)
do, = (w, 8 a,)ds
+ ~ ( d u( u~s,e ) * )+ w(dwi fBdBi)
where (use)* is the horizontal lift or U,O,.By working as in [B.F] and [Dr], we put: (2.11)
do'
= ((w,
+
@a,) - 1/2RC6")ds T ( ~ u (#u,s a ) * )
where R, is the Ricci tensor of 0. We get (See [B.F]): (2.12)
s'
E [ < df(z:),u& >] = E[f(zj)
< (us@ a s ,dwb 8 dB5 >]
In order to define our Sobolev Calculus, we take as Hilbert space the direct sum of the space of Lz functions w , on Rm-d and of the space of L2 functions u. in Rd. Since we have the integration by parts (2.12), we can define Sobolev spaces Wr,p(wi,B') in the sense of these Hilhert spaces with respects of the gradient defined by (2.12), as it was done in [Lx] and [Ls]. Let x c t be the Brownian motion on B starting from yo. Let FA be the u-algebra spanned by it for t 5 1. By doing as in [B.F] (see [B.M1], [B.Mz], [KuS] too), we can state the following lemma: Lemma 1I.l:Let G belonging to all W,,p(w',b'). Let us consider a local trivialization of M into B x V on g(y) where g is a bump function on B with small support. Let us suppose that zt start from zo. Let us consider the functional:
f
(2.13)
--t
E[Gg(.(.:))f(.:)lFAl
where where f is a smooth function on V . It is almost surely equal to Svp(G,s(z').,zo,z)f(z)dmv(z) dmv(z) is the Riemannian measure on V and where z --t p ( G , r ( z f ) , z o , z ) is ?r(z1) measurable and is smooth in I. Moreover, E[p'(g,n(zf),so,z)]'/' is finite and can be estimated by the Sobolev norms of G. Proof: If G is a cylindrical function in zf, it follows from the arguments of [B.F]. Since V preserves the splitting of T ( M )int n'T(B) @ T ( V )it, suffices to do the stochastic Calculus of variation in the w directions and to see if we consider the Ricci tensor R , y of V restricted to V that the solution of the stochastic differential equation: (2.14)
dA,
=
-1/2R,vA,ds
+ ~ ( d ~ l(u,A)') ,,
belongs to all the LP. We deduce the results by approximating G by cylindrical functionals.
0 Instead of considering the previous random measure on V , we can consider the measure on B x V given by (f,f') + E[Gg(~(zl.)f(zt)f'(z:)] in order to deduce an corresponding lemma, but for the bridge of ~ ( z ibetween ) yo and yo. Let FB be the u-algebra spanned by the Brownian bridge t + ~ ( z i between ) yo and yo. We get the analoguous lemma: Lemma 11.2: Let be the random measure on V,, f + E[Gf(z:)IFs]. It has almost surely in the trajectories of the Brownian bridge t --f ~ ( z : ) a smooth density with respect of the Riemannian measure on V,, p (G ,n( zf ) ,zo,z) . Moreover, E[lp(G,*(z1),zo,z)I'Ii/' can be estimated in terms of the Sobolev norms of G. Proof:: This follows from Lemma 11.1 and from Malliavin Calculus (instead of decomposing the time interval in 2 pieces as it was done in [B.F]).
0 111. Bismut-Witten current
181 Let us consider the Hilbert space He" of L2-even forms over M and the Hilbert space of L2-odd forms on M . Let H = He' €3 H o d d . We consider the exterior derivative d , its adjoint d' and the operator dd' + d'd = A,. We consider the complexified Sobolev space H p of forms such that
lM
< (A; + l ) u , u > dmM = \\u\\; < 00
(3.1)
Let H,= nH,. By Sobolev's imbedding theorem, this space is nothing but the space of smooth forms on M . We consider the symmetric Fock space constituted of series 5 = C dn where i?,, belongs to HF1"'"' such that (3.2) = nSp,q. A Hida We get a second quantized Fock space Sp,9, The space of Hida functionals is S-, distribution is an element of the topological dual S-, of S,-, Let L,(M) be the smooth free loop space of M of smooth loops s + y(s) from S' into M . If dn = (u,?, @u>') @ . . @ ( u : @ u $ " (we ) consider a symmetric tensor product), we introduce the Chen iterated integral:
(3.3)
F(8,) =
( u A ~ ~ +I L<.,<..<s"
d ' ' ( d y ( s 1 ) )A
.. A (U,"ds, + a?(dy(s,))
It is a Chen form over the free loop space ([Ch], [G.J.P]). If 5 E S--, we can extend the definition (3.3) to F ( d ) = EFT(*).FT((r)belongs to the set of ?-measurable forms on L,(M)(see [Llo]). Let us recall the formalism of Jones-Lbandre ([J.L]). Let us consider ?r(zi), the free Brownian bridge on B. On the the free continuous loop space L ( B ) ,we put the measure: (3.4) where p f ( y , y ' ) is the heat kernel associated to ?r(z:) and d P f v the law of the Brownian bridge on B starting from y . We consider the stochastic parallel transport ~t associated to the Levi-Civita connection for the process t + ?r(z:). We consider finite energy paths t + Ht in T+;)(B) such that: (3.5) and such that ~ 1 H 1= Ho. t + X , = T ~ constitutes H ~ a generic element of the fiber of the stochastic tangent over the generic Brownian path t -i.(xi). (3.5) allows to introduce an Hilbert tangent space over the Brownian loop. Since we have a measure on L ( B ) ,we can speak of L2 r-forms on the free Brownian bridge. We call the space A>'(L(B)). But we are motivated in fact by the Brownian motion on M . We replace A;'(L(B)) by A;(L(B))where we take the law o f t + n ( z t )on L ( B ) ,where t + ztis the free Brownian bridge on M endowed with the probability law: (3.6)
#(z,z)dm~(z@ ) dP$
= dPM
py(z,z') is the heat kernel for the Brownian motion on M and d P y z is the law of the Brownian bridge starting from z and arriving a t time z in M . Let us suppose that M is spin, B is spin and that T ( V )is spin. We get the spin bundle S ( M ) ,the spin bundle S ( T ( V ) )and the spin bundle ?r*S(B).Moreover, S ( M ) = S ( T ( V ) )8 r*S(B). We get the corresponding complexified bundles C I ( M ) ,CL(T(V))and ?r*CL(B).Moreover, C1(M) = CI(T(V))@ ?r*CI(B).
182 We get the complexified exterior bundles A ( M ) , A ( T ( V ) )and r * A ( B ) .A ( M ) = A(T(V))&r*A(B). Moreover, A ( M ) is isomorphic to C l ( M ) , and this isomorphism is compatible with the various decompositions done before. If u E A ( M ) , , we denote by u:Pt the corresponding element of C l ( M ) , and by sigmaEr the correspond) r * A ( B ) . Let us recall the definition of the Witten current W(M) over L,(M). ing element of C l ( T ( V ) @ To 5, = (u,?,f 3u2') @ .. @ (u: fBu;"), where we consider symmetric tensor products, we associate:
In (3.7), s + 5 , is the free Brownian bridge on M , i , , t the stochastic parallel transport for the Levi-Civita connection on the various considered bundles along the path u + zu runned in the opposite sense for s < t. K is the scalar curvature. Tr. is the supertrace of an operator acting of the spinor vector space. Let us recall that W(M) is an Hida distribution. The first step in the construction of the partial Witten current WP"'(M) is to replace in (3.7) u$Yt by ? : u and by uk'rp"', is,$ by the stochastic parallel transport on r * T ( B )@ S ( T ( V ) )for the connection Vtot on r * T ( B )8 S ( T ( V ) )naturally inherited from the Levi-Civita connection on S ( M ) . Let us explain what we mean by that:if we consider the connection form on S ( M ) , consider some product by Clifford element e, where we consider a local orthonormal basis of T ( V )and by some en, where we consider We replace the Clifford multiplication by e, by a wedge product by the a local orthonormal basis of T(B). where corresponding element in the exterior algebra called fn. Moreover, we replace K(z,) by Kr(zal(zs) K, is the scalar curvature of V,. We get an element of L ' ( ~ MQ) A i ( L ( B ) )(Ww consider only its part of degree r). It is given as follows:
uk:y
The term ai(si)is an element of L 2 ( p ~Q)ef,A1(M)Q e;,r*A(B)where e, is the evaluation map s + zs. Let us remark that the vector field s + X , on L ( B ) are r(z.)measurable. We take the conditional expectation of the kernels of (3.8), we select the term of degree r on L ( B ) and we get the definition of the Bismut- Witten current WparJ(M). Theorem III.l:WPa'~'(M) defines an element of S-, with values in A ; ( L ( B ) ) . Proof: Let us recall that the law of z. is equivalent to the law of z1. Moreover, the Radon-Nikodym derivative G belong to all the Lp. Let F be a functional depending on x,and let F' be a functional depending on ~ ( z . )We . have: (3.9)
E [F (z . )F ' (r(z . ))]= EIGF(zl)F'(r(z!))]
E[E[GF(~~)IFB]((~(ZI))F'(~(~'))] = E[E[GF(~~)IFBI(~(~.))G:F'(~(..))~
where FB is the sigma-algebra spanned by r(z1) and where of r ( d )with respect to the law of a(..). So
G is the Radon-Nikodym derivative of the law
E [ G F ( ~ ~ ) I F B I ( ~ (= ~ .E)[)FG( x )IFB]
(3.10)
Moreover, we can write: (3.11)
H,
=
A'
< a,(.), H, > du +
I'
< a:(.), HL > du.
where r ~ a . ( l )= a,(O) (see [L6]). If 5n = (u: @ u?') Q .. 8 (u: @ u?"), by using (3.11) we can write the kernels of (3.8) for the Hilbert structure (3.5). By Lemma 11.2 and (3.9), we get an estimate of the conditional expectation of these kernels
183
+
with respect of a(..) in ~ ( ~ ~ ~for some ~u convenient ~u ~ p~ independent ~ ~ ~off,,.~ )Moreover, ~ the L2 norm of WPaP3'(M)(d,) can be estimated in n(llwLl1, ~ ~ uWe ~ conclude ~ ~ as in~[Llo] ) Lemma . 2.1 that for some ( p , q )
5
+
I I W P n ' " ( ~ ) l I ~ : ( ~I( Cll5llp,q ~))
(3.12)
0
Let E be a complex bundle on M . Since M is compact, ( is given by a smooth projector. We can suppose (see [N.R]) that the connection is the projection connection with see [Lo]). Let us consider the Bismut-Chern character: (3.13)
Ch(Fm) =
c/
T r ( A ( d y ( s 1 ) )- R d s i ) A O<S,<..<S*
.. A ( A ( d y ( s , ) )
- Rds,)
Ch([,) = F(5c) where 5~belongs to S,-. Let ,i be the injection from B into L ( B ) . By doing as in [B.M1], [B.M2] or [Bis], we can see that the conditional law of z. conditionally to a(..) = y is given by the law of the free Brownian loop on V, given by the measure dpv,: (3.14)
d p v , = P?(", z ) d m v , (z) @ d P 2
where p p ( z , z ' ) is the heat kernel associated to the Brownian motion on V, and d P 2 is the law of the Brownian bridge on V, starting from z. We get the following theorem: Theorem 111.2:
1i ~ W p a r 3 ' ( M ) ( 5 ~Ch) I n d p D p
(3.15)
=
where Ch I n d p D p is the Chern character of the index bundle of the family of Dirac operators on V twisted by F. Proof: The connection Vtot on a*T(B)@ S ( T ( V ) )reduces to a connection Vy on a*T(B)@ S(V,) (see [Bi4], [L1]). Let us consider the horizontal Laplacian A v u wich operates on the sections on V, of r * T ( B )@ S(V,). Let us Cy be the tensorial operator given in local coordinates by: (3.16)
Cy = 1 / 2 c e ; , e j @ R ( e i , e j )+ c l / 2 f a A f ' ~ R ( f u , f ' ) + ~ f a ~ e ; @ R R ( f u , e ; ) i,i
a,;
where e, is a local orthonormal basis of T(V,) and f" is a local orthonormal basis of T ( B ) (Let us recall that we have taken for the Clifford algebra the convention e: = -2 ([LI] (1.22)). Let us put (3.17)
RY,
= -Av.
+ 1/4Ky + C,
Let us recall Bismut's theorem (See [Bis], [Ll]): (3.18)
Ch I n d p D . p = Tr, exp[-RTv]
Let z: be the free Brownian bridge on V,. Let T$ be the stochastic parallel transport for the bundle s * T ( B )@ S(Vy)@ E relative to the product connection of s * T ( B )@ S(V,) and (, the time being runned in the opposite sense between s < t. Let us consider the differential equation: (3.19)
dU:
=
U:T~,tCy(?~,t)-ldt
184 We get the following probabilistic representation of Tr, exp[-1/2R~J (see [Bill, [LI]): (3.20)
But we can expand the solution of (3.20) in series as well the term which appears of the parallel tansport on E. This gives nothing else that Ci~WPar,‘(M)(3.().
0 References [A] Albeverio S.: Wiener and Feynman-path integrals and their applications. In ”Norbert Wiener Centenary congress” Mandrekar V. Masani P.R. edit. Proc. Symp. App. Math. 52 (1996), 163-194. [A.H-K] Albeverio S. Hoegh-Krohn R.: Mathematical theory of Feynman path integral. Springer. Lect. Notes. Math. 523 (1976). [A.S] Albeverio A. Sengupta A: Construction of the non-Abelian Chern-Simons functional integral. C.M.P. 186 (1997), 563-579. [A.D] Andersson L. Driver B.: Finite dimensional approximation to Wiener measure and path integral formulas on manifolds. J.F.A. 165 (1999), 430-498. [At] Atiyah M.: Circular symmetry and stationary phase approximation. In ”Colloque en l’honneur de L. Schwartz”. AstQrisque 131 (1985), 43-59. [B.V,] Berline N. Vergne M.: ZQrosd’un champ de vecteurs et classes caracteristiques Qquivariantes. Duke Math. Journal 50 (1983), 539-548. [B.Vz] Berline N. Vergne M.: A proof of Bismut local index for a family of Dirac operators. Topology. 26 (1987), 435-463. [Bill Bismut J.M.: Large deviations and the Malliavin Calculus. Prog. Math. 45 Birkhauser (1984). [Biz] Bismut J.M.: Index theorem and equivariant cohomology on the loop space. C.M.P. 98 (1985), 213-237. [BiS Bismut J.M.: Localization formula, superconnections and the Index theorem for families. C.M.P. 103 (1986), 127-166. [Bi4] Bismut J.M.: The Atiyah-Singer theorem for families of Dirac operator: Two heat equation proofs. Invent. Math. 63 (1986), 91-151. [Bis] Bismut J.M.: Filtering equation, equivariant cohomology and the Chern character. in ”VIII t h international congress on mathematical physics”. World Scientific (1987), 17-56. [B.F] Bismut J.M. Freed D.: The analysis of elliptic families. 11. Dirac operators, eta invariants and the holonomy theorem. C.M.P. 107 (1986), 103-163. [B.M1] Bismut J.M. Michel D.: Diffusions conditionnelles I. J.F.A. 44 (1981), 174-211. [B.Mz] Bismut J.M. Michel D.: Diffusions conditionnelles 11. J.F.A. 45 (1982), 274-292. [C.D-M.F] Cartier P. Dewitt-Morette C. Folacci A. edit: Functional integration. Basics and applications. NATO AS1 series. Physics: 361 (1997). [Ch] Chen K.T.: Iterated path integrals of differential forms and loop space homology. Ann. Maths. 97 (1973), 213-237. [Co] Connes A.: Entire cyclic cohomology of Banach algebras and character of 0- summable Fredholm modules. K-theory 1 (1988), 519-538. [D-MI Dewitt-Morette C. edit: Special, issue of the Journal of Mathematical Physics 36.5 (1995) about functional integration. [Do] Donnelly H.: Local index for families. Mich. J. Maths. 35 (1988), 11-20. [Dr] Driver B.: A Cameron-Martin type quai-invariance formula for Brownian motion on compact manifolds. J.F.A. 110(1992), 272-376. [D.H] Duistermaat J.J. Heckman G.J.: On the variation in the cohomology of the symplectic reduced phasespace. Invent. Math. 69 (1982), 259-269. [El] Elworthy K.D.: Stochastic differential equations on manifolds. L.M.S. Lectures Notes Series 20. Cambridge University Press (1982).
185 [Em] Emery M.: Stochastic Calculus in manifolds. Springer (1989). [E.P.F] Estrade A. Pontier M. Florchinger P.: Filtrage avec observation discontinue sur une variQt6. Existence d’une densite rQguliere. Stochastics Stochastic Rep. 56 (1996), 33-51. [Gel Getzler E.: Cyclic homology and the path integral of Dirac operator. Preprint (1988). [Ge.J.P] Getzler E. Jones J.D.S Petrack S.: Differential forms on a loop space and the cyclic bar complex. Topology 30 (1991), 333-373. [Ge.S] Getzler E. Szenes A,: On the Chern character of a theta-summable Fredholm module. J.F.A. (1989), 343-357. [Gi] Gilkey P.: Invariance theory, the heat equation and the Atiyah-Singer theorem. C.R.C. Press (1995). [G.S.V] Grothaus M. Streit L. Volovich I.: Knots, Feynman diagrams and Matrix models. Inf. Dim. Ana. Quant. Probab., Rel. Top. 2 (1999), 359-380. [HI Hida T.: Analysis of Brownian functionals. C.R.C. Press (1995). [H.K.P.S] Hida T. Kuo H.H. Potthoff J. Streit L.: White noise: an infinite dimensional Calculus. Kluwer (1993). [LW] Ikeda N. Watanabe S.: Stochastic differential equations and diffusion processes. North Holland (1981) [J] Jaffe A,: Quantum harmonic ana:lysis and geometric invariants. Adv. Math. 113 (1999), 1-110. [J.L.O] Jaffe A. Lesniewski A. Osterwalder K.: Quantum K-theory. The Chern character. C.M.P. 118 (1988), 1-14. [J.L.W] Jaffe A. Lesniewski A. Wisniowski M.: Deformation of super-KMS functionals. C.M.P. 121 (1989), 527-540. [J.La] Johnson G. Lapidus M.: The Feynman integral and Feynman operational Calculus. Oxford Univ. Press (2000). [J.L] Jones J.D.S. LQandreR .: Lp Chen forms on loop spaces. In ”Stochastic Analysis” Barlow M. Bingham N. edit. Cambridge University Press (1991), 104162. [Ka] Kato T.: Perturbation theory for linear operators. Springer (1966). [Kh.S] Khandrekar D.L. Streit L.: Constructing the Feynman integral. Ann. Phys. 1 (1992), 49-55. [K.N] Kobayashi S. Nomizu S.: Foundations of differential geomtry. Tome 11. Interscience (1969). [K.L.P.S.W] Kondratiev Y.G. Leukert P. Potthoff J . Streit L. Westerkamp W.: Generalized Functionals in Gaussian space: the characterization theorem revisited. J.F.A. 141 (1996), 301-316. [Ku] Kusuoka S.: More recent theory of Malliavin Calculus. Sugaku 5 (1992), 155- 173. [KuS] Kusuoka S. Stroock D.W.: The partial Malliavin Calculus and its application to nonlinear filtering. Stochastics 12 (1984), 83.162. [LI] LQandreR.: Sur le theoreme de l’indice des familles. SQminairede probabilitQsXXII. AzQma J. Meyer P.A., Yor M. edt. L.N.M. 1322 (1988), 348-413. [Lz]LQandreR.: Applications quantitatives et qualitatives du Calcul de Malliavin. In ”Col. Franc-Japonais” MQtivier M. Watanabe S. edit. Springer. Lect. Notes. Maths. 1322 (1988), 109-133. English translation: In ”Geometry of Random motion”. Durrett R. Pinsky M. edit. Contemp. Maths. 73 (1988), 173-197. [L3] LQandreR.: Integration by parts formulas and rotationally invariant Sobolev Calculus on the free loop space. In ”XXVII Scool of theoretical physics of Carpacz”. Borowiec A. Gielerak R. edt. J. Geo. Phys. 11 (1993), 517-528. [L4] LQandreR.: Cohomologie de Bismut-Nualart-Pardoux et cohomologie de Hochschild entiere. SQminaire de ProbabilitQs XXX in honour of P.A. Meyer and J . Neveu. AzQma J. Emery M. Yor M. edit. Springer. Lect. Notes. Math. 1626 (1996), 68-100. [Ls] LQandreR.: Brownian cohomology of an homogeneous manifold. In ”New trends in Stochastic Analysis”. Elworthy K.D. Kusuoka S. Shigekawa I. edit. World Scientific (1997), 305-347. [L6] LQandreR.: Invariant Sobolev calculus on the free loop space. Acta. Appl. Math. 46 (1997), 267-350. [LT]LQandreR.: Stochastic Adams theorem for a general compact manifold. Rev. Math. Phys. 13 (ZOOl), 1095-1133. [Ls] IKandre R.: Stochastic equivariant cohomology and cyclic cohomology. Preprint. [Lg] LQandre R.: Hypoelliptic diffusion and cyclic cohomology. To appear in ”Stochastic analysis”. R. Dalang. M. Dozzi. F. Russo edt. [Lw] LQandre R.: Theory of distribution in the sense of Connes-Hida and Feynman path integral on a manifold. Inf. Dim. Ana. Quant. Probab. Rel. Topics. 6 (2003), 505-517.
186 [L.Ou] L6andre R. Ouerdiane H.: Connes-Hida Calculus and Bismut-Quillen superconnections. Preprint. [LS] Leukert P. Schaefer J.: A rigorous construction of Abelian Chern-Simons integrals using white noise analysis. Rev. Maths. Phys. 8 (1996), 445-456. [Lo] Loday J.L.: Cyclic homology. Springer (1998). [Mal] Malliavin P.: Stochastic Calculus of variations and hypoelliptic operators. In ”Stochastic analysis” K. It6 edt. Kinokuniya (1978), 155-263. [Maz] Malliavin P.: Stochastic analysis. Springer (1998). [N.R] Narasimhan M.S. Ramanan S.: The existence of universal connection. Amer. J. Math. 83 (1961), 535-581. [N.T] Nest R. Tsygan B.: Algebraic index for families. Ad. Math. 113 (1995), 151-205. [Nil Nistor V.: Super-Connections and Non-Commutative geometry. In ”Cyclic cohomology and noncommutative geometry” Cuntz J.R. Khalkhali Edt. Fields Institute Comm. 17 (1997), 155-166. [Nu] Nualart D.: The Malliavin Calculus and related topics. Springer (1995). (01Obata N. White noise analysis and Fock space. L.N.M. 1577 (1994). [Pa] P a r d o n E.: Filtrage non lin6aire et Qquations aux dQriv6es partielles stochastiques. Ecole d’Qt6 de Probabilitbs de Saint-Flour XIX. P. Hennequin edt. L.N.M. 1464 (1991), 67-163. [PI Perrot D.: A bivariant Chern character for families of spectral triples. Com. Math. Phys. 231 (2002), 45-95. [PoS] Pontier M. Szpirglas J. Filtering on manifolds. In ”Stochastic modelling and filtering”. L.N. Control. Infor. Sci. 91 (1987), 147-160. [PSI Potthoff J. Streit L. A characterization of Hida distribution. J.F.A. 101 (1991), 212-229. [Q] Quillen D.: Superconnection and the Chern character. Topology 24 (1985), 89-95. [S.W] Smolyanov 0. Von Weizsaecker H.: In preparation. [St] Streit L.: An introduction to white noise analysis. In ”Stochastic analysis and applications in physics”. Cardoso A. Sbnbor R. Streit L. edt. Nato Asie Series 449 (1994); 415-441. [Y] Yor M.: Sur la th6orie du filtrage. Ecole d’btQde Saint-Flour IX. P. Hennequin edt. L.N.M. 876 (1981), 239-280. [Sz] Szabo R.: Equivariant cohomology and localization of paths integrals in physics. Springer. Lect. Notes. Physics. 63 (2000). [Wa] Watanabe S.: Stochastic analysis and its application. Sugaku 5 (1992), 51-71.
Infinite Dimensional Harmonic Analysis I11 (pp. 187-195) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
Laplace approximations for Diffusion Processes Song LIANG
1
The problems
Let us first state our problem for diffusion processes on tori Td = Rd/Zd. Let LO= +A boa V, where A and V mean the Laplacian and the gradient and operators, respectively, and bo E C"(Td;Rd). Let R = C([O,oo),Td) Xt(w) = w ( t ) for w E R, t 2 0. Let Ft = a { w ( s ) ; s 5 t } and F = V t 3 t . Also, let {P,},ETd be the solution of the Lo-martingale problem, ie.,
+
2. P,(wo = x) = 1.
Jot
Let Lt = f 6x,ds, where 6. means the delta measure. We are interested in the asymptotic behavior of Lt under P,(.IXt = y) as t -, 00. By ergodic theorem, { Pz}sETd has a unique invariant probability measure T , and for any x,y E Td,we have that Lt converges to 7r in law as t --+ 00 under P,(.lX, = y). Hence for any closed set A c p(Td) that does not contain T , (where p(Rd) means the set of all probabilities on Td),we have P,(L, E A ) -+ 0 as t -+ 00. Large deviation principle (LDP) studies the order of this convergence in terms of the so-called entropy function determined by the generator Lo. Under some conditions, we have that {Lt}t>O under p,(.Ixt= y) satisfies LDP, ie., -
1 inf ~ ( v 5 ) lim inf - log P, T+w T
uEAo
(LT
E A ~ =Xy) ~
188
for any A E B(p(Td)),where I : p(Td)+ R U {m} is the entropy function given by
I ( v ) = sup
{
-
Ld
Lou Lou y d v ;u E C"(Td),u > 0, - is bounded 21
(2) \- /
Here a(.) means the Bore1 0-field, and A' and 3 mean the interior and the closure of A , respectively. Let M ( T d )denote the set of signed measures on Td with finite total variations. Let Q, be a function on M ( T d )that is "good" enough, and let Zx9y T = EP=[eT@(LT)lX~ = y] . Then by the LDP given above, 1 lim -10gZ;~ = A, T-+m T
where X = maX,,,(p){@(v) - I ( v ) } .However, in many problems, this is not enough and we need the 1 o(1) order estimate of 23' as T + 00 (see, e.g., [7] and [lo] for some examples). We study this problem in this paper. To be more precise, we show that under some conditions, e-ATZ;y converges to a constant as T + cm.See Section 2 and [5] for the details.
+
Similiarly, we can also consider this precise estimating problem for diffusions on Euclidean spaces Rd. We show that the similar result holds under some conditions. See Section 3 and [lo] for the details.
+
The 1 o(1)-order precise estimating problem for the sums of Banach space-valued i.2.d. random variables has been discussed by many authors, e.g., Bolthausen [l],Kusuoka-Liang [4] and Liang [6], etc. Also, BolthausenDeuschel-Tamura [3] considered the same problem for Markov processes on compact state spaces under some conditions that derive the so-called "Central Limit Theorem Assumption" as a result. (Note that we do not need the "Central Limit Theorem Assumption" in our setting). The problem for diffusion processes on non-compact state spaces is very less studied. One of the obvious and most vital difficulity is that, when the state space is not compact, many properties such as bounded property, which is trivial for continuous functions in the compact case, become very difficult. For example, let G denote Green operator, then we can show that VGf is continuous. We need some estimate with respect to VGf (X,) in our proof. In the case of diffusions on Td,there is no problem since VG f is automaticly bounded. However, since Rd is not compact, this estimate is not so easy.
189
In [lo], we succeeded in dealing with diffusion processes on Rd, a noncompact state space. The main idea is as follows. First, by using [9], we have that lVGf(z)l, although not necessarily bounded, can be dominated by (1 1 ~ 1 ~ ) $ 1 1 f lwhere l~, is some constant depending on the drift term of the generator that is not too big. The main point is that, although the state space is not compact, when the drift term is "good" enough, the probability that the process goes to infinity converges to 0 exponentially, uniformly with respect to the starting point. This absorbs the extra factor (1 Izlz)g.
<
+
+
2
The case for diffusions on tori
Use the same notations as in Section 1. For later use, write E = Td. Let K = {v E p(E) : @(v) - I(v) = A}. Then K is not empty and is compact in p(E). We assume the following as in [5]:
A1 K contains only one element vo. A2 @ : M ( E ) -, R is three times continuously fie'chet differentiable and satisfies the following condition: there exist @@) E C(p(E) x ( E ) k R) ; (k = 1,2,3) such that Dk@(v)(R1,. . . ,R k ) =
k...
@@)(v;
XI,
for any v E p(E) and R1, Rz,R3 E M ( E ) ,k
*.
a
,
zk)Rl(dz1).. . R k ( d z k )
= 1,2,3.
We construct, as in [2] and [3], a family of diffusion probabilities { Q Z } z E ~ whose invariant measure is vo. In general, for any 'p E Cb(E),let
P:(z,A) = Epz[eSaf'p(X")dS,Xt E A],
A E B(E),
and let A'@ = sup(JE'pdu - I(v);v E p(E)}. Then there exists a unique (up to constant multiplications) h p i P E C ( E )such that P:hp@ = eAPPvthe'p for any t 2 0. Let Q(P)Pf = e-A"'t(hP~~)-'P,'(hp~~f), and let {Q(P)g}ZEE be the corresponding diffusion measures. Let 9" E Cb(E) be given by @'O(Z)
= D @ ( ~ o ) ( b-, vo)
+ @(vo),
2
E E.
190
Then X = Ap,4”O. Also, let { Q S } I E ~ be given by { Q S } s=~{ Q~( I ‘ ) $ ” O } s E ~ . then { Q z } r E ~ has vo as its invariant measure. Note that he4”O E Cb(E)r l VhPduo C 1 ( E )and the generator of {QI}IE~ is L = LO a hP,duO . V, where a = Id. Let G denote the Green operator corresponding to { Q S } S E ~ , let G* denote the dual operator of G in L2(dvo),and let G = G G*. (See [5] for the details). As in [3], we also assume the following:
+
+
A3 I
- D2@(v0) is
non-degenerate.
It is not difficult to show that IH - DZ@(v0)lHxH is non-negetive definite. strictly positive-definite. Here H is A3 assumes that IH - D z @ ( v ~ ) l Hisx H the Hilbert space whose norm is essentially the second Frkchet differential of the entropy function I . H can be regarded as a dense subset of M o ( E ) .As I is not smooth, this description is not mathematically precise. See [5] for the precise definition of H and statement of (A3). Finally, we assume the following:
A4 For any 6 > 0, there exist an EJ > 0 and a symmetric KJ E Cb(Ex E ; R) such that thefunctionx given b y x ( R 1 ,R z ) := JE JE KJ(z,y)Rl(dx)Rz(dy), & , R Z E Ma@), satisfies l l ~ ~ H x H I I ~ . ~L. 6 and
D 3 @ ( R ) (-v VO,v
- vO,
v
- VO) 5
KJ(V- VO,v - vO)
f o r any R , V E g ( E ) with dist(R,V O ) < 6 and dist(v,v0) < 6. Theorem 1 Assume A1
-
A4. Then
xdetz(IH - D’@(V~))-’/~,
3
for any a , y E Rd
The case for diffusions on Euclidean spaces
In this section, we discuss the l+ o ( 1)-precise estimating problem for diffusion processes on Euclidean spaces Rd.
191
+
&,
Let LO := C$=laijViVj b . V, where Vi = i = 1,... ,d, and V = (VI,. . . , Vd). (The assumptions with respect to (aij)tjz1 and b will be given later). The path space is R = C([O,oo),Rd). Let {Pz}zERd be the solution of the Lo-martingale problem. Also, let X t ( w ) = w ( t ) , w E 0,t 2 0, 6xsds, where 6. means the delta measure, as before. and let Lt = Before stating our assumptions, let us prepare some notations. Let { Rt}t?o denote any diffusion semi-group with associated generator aGViVj b R . V. We define several conditions. LR = E CP(Rd;Rd2) and is uniformly elliptic, ie., there (Cl): aR := exist cl, c2 > 0 such that
c$=,
+
For Q , 6 2 > 0, we define the following conditions: (C2,,): bR E C(Rd;Rd)and there exist c3, c4 > 0 such that
(C3,,): bR E C"(Rd; Rd)and there exist c5, c6 > 0 such that
We use these conditions to define two sets:
HI ( K , I ) = { diffusion semi-group{Rt}t?o satisfying (Cl) and (C2K1)}, H ~ ( I n2)= E ~ , { diffusion semi-group{Rt}t20 satisfying (CI),(C2,,) and (C3,2)}. Now, we are ready to give our first assumption:
E1 There exist
and
such that
192
By [8],we have the following under El: {Pz}&Rd has an invariant probability 71, P,(z,dy) = P,(Xt E dy) has smooth density pt(z,y) E Cm(Rdx Rd;R+) with respect to T , and SuP,,=Rd,lyl$,.pt(~, y) < 00 for any r, t > 0. Therefore, we can define the pinned probability P'( * IXt = y) for all 2,y E Rd and t > 0. We have by [8] that Pz(Lt E .IXt = y) satisfies LDP, ie., (1) holds for any A E B(p(Rd)),where I is given by (2) with Td substituted by Rd. Let is continuous with @ : M(Rd)+ R be a bounded function such that
@I
p(W
respect to Prohorov metric dist(., .) and let
Z$Y =
[exp ( T @ ( L T ) )
Ix.
= y]
.
Then as a corollary of the large deviation principle, log 2;' + X for every 2,y E Rd, where X = sup{@(v)- I(v);v E p(Rd)}. As before, we want to give a precise estimate of 2;' as T + m, up to the order 1 o(1). Before presenting the next assumption, we notice that for any T > 0, {XT-t(w)}tEpTl under P,(dw) is a diffusion associated with the semi-group {P'"}t,O, where P,*" denotes the dual operator of Pt in Lz(d.lr). Let $(x) = z E Rd,and for any a E R, define €3: as
+
dm,
€3: = {f E C(Rd;C);I l f l l ~ : := SUP$(.)-"lf(.)l
< m}.
xERd
For any E Hl(y) with y > 1 and any cp E B; with 13 E [O,y - l), we can define a new semi-group of operators {Q(R)r)t>o in the followingway (see [lo] for the details): Let
%f(z) = ERz[ exP(
0
c p ( X s ) d s ) f ( X t ),]
2 E
Rd,
(5)
where {Rz}sERd is the family of diffusion measures associated with {Rt}tzo. is a continuous linear operator on l3: and AR@ := For any a > 0, limt--rm$ log lB:-,B: is well-defined, finite and not depending on a > 0, (we can choose a = 0 if B = 0), and there exists a unique (up to constant for any t > 0. Let multiplications) hR@E B: such that &hRiv = eAR'vthR@ {Q(R)r}t>o - be the Markovian semi-group given by Q (R): f := e-ARlvt (hRgv)-1
%(hR"f),
(6)
193
and let {Q(R)Z}ZERddenote the corresponding diffusion measures. Now, we are ready to give our second assumption of this section.
+
E2 There exist a {St}tro E H~(y{,y;)with 7: > 1 and y; E [y{,yi i(y{ l)), and a cpo E C"(Rd) n Bjo with 1 9 ~E [O,? - (7; - 7;)) such that { P , * T } t i= ~ {Q(s)Y'}tro.
Theorem 2 Assume E l , E2 and A1 lim e-T'Z?Y
T-i"
4
=
-. h(z) exp h(Y)
-
A4 (with E
= Rd).
Then
{i/ ( c @
~ ) ~ ) ( .,v.)I o ; vo(du)}
(4 xdetz(IH - D 2 G ( v ~ ) ) - 1 / 2 , for any 2,y E Rd.
Examples
In this section, let us give some examples of {Pt}t?o that satisfy our assumptions E l and E2 of Section 3. Let U and b be any pair of functions satisfying the following:
< 00, and b E C"(Rd; Rd), H1 There exist constants y1 > 1 and 7 2 E [y1,y1 i(yl - 1))such that HO U E C"(Rd; R) with 2
+ (b-
*
JRd e-u(Z)da:
+
v E Hz(y1,yz),
H2 There exist constants y{ > 1 and y; E [yi,yi ;A - b . v E Hz(yi, y;),
[O,v
H3 There exists a 60 E - (7; iVU) - div(b - i V U ) E Bi0.
- 7;))
+ i(yi - 1))such that
such that
'po
:= VU
. (b -
Let {Pt}t>Obe the semi-group of continuous linear operators on Cb(Rd) corresponding to 1 Lo= 2 A - V U . V t b . V. Then {Pt}t>o satisfies E l and E2 of Section 3. Actually, let P,*" denote the dual operator of Pt in L2(dp),where p means the invariant measure of {Pt}t>o, and let {St}t>o be the semi-group of continuous linear operators on Cb(Rd) corresponding to generator ;A - b . V, then we can show that P,*" = Q(R)Tfor any t > 0. (See [lo] for the proof).
194
A concrete example that satisfy HO, H1, H2 and H3 is given by the following: for example, let d = 1, let A > 0, 6 > 0, v > 0, > V (7- 6 - 2) be any constants, let y ( x ) E CF(R) such that y ( x ) = IxJ-E-2zfor any 1x1 2 2 and y ( z ) = x for any 1x1 5 1, and let
4
VU(z) = 1xI6z- Ay(z)lzlv, 1 b ( x ) = 2 (Iz16z- Ay(z)Ixlv) + !I(.).
+
Then b and U satisfy HO, H1, H2 and H3 with y1 = 7 2 = 7; = 7; = 1 6 6 - 7;-1 and 80 = 6 - 5 < ; i- (74 - 7;).
References [l]E. Bolthausen, Laplace approximations for sums of independent random vectors, Probab. Theory Relat. Fields 72 (1986), no. 2, 305-318
[2] E. Bolthausen, J.-D. Deuschel and U. Schmock, Convergence of path measures arising from a mean field and polaron type interaction. Probab. theory Related Fields, No. 95, pp. 283 - 310 (1993) [3] E. Bolthausen, J.-D. Deuschel and Y . Tamura, Laplace approzimations for large deviations of nonreversible Markov processes. The nondegenerate case. Annal. Prob., Vol. 23, No. 1, pp. 236 - 267 (1995) and ~ ~S. Liang, Laplace approximations for sums of indepen[4] S. K U S U O dent random vectors, Probab. Theory Related Fields 116 (ZOOO), no. 2, 221-238 [5] S. Kusuoka and S. Liang, Laplace Approximations for Diffusion Processes on Torus, nondegenerate case, J. Math. Sci. Univ. Tokyo, 8, 43-70(2001). [6] S. Liang, Laplace approximations for sums of independent random vectors -the degenerate case, J. Math. Sci. Univ. Tokyo 7 (2000), no. 2, 195-220 [7] S. Liang, Precise Estimations related to Large Deviations, to appear in Proceedings of the First Sino-German Conference on Stochastic Analysis.
195
[8] S. Liang, Large Deviation Principles for a type of Diffusion Processes on Euclidean Space, J. Math. Sci. Univ. Tokyo, 10(2003), 555-578. [9] S. Liang, A Bounded Property for Gradients of Diffusion Semigroups on Euclidean Spaces, to appear in J. F’unct. Anal.
[lo] S. Liang, Laplace Approximations for Large Deviations of Diffusion Processes on Euclidean Spaces, Preprint.
This page intentionally left blank
Infinite Dimensional Harmonic Analysis I11 (pp. 197-212) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
A N O T E O N A F F I N E QUOTIENTS AND EQUIVARIANT DOUBLE FIBRATIONS
KYO NISHIYAMA Department of Mathematics Graduate School of Science Kyoto University Sakyo, Kyoto 606-8502, Japan E-mail: kyoQmath.kyoto-u. ac. j p We consider two linear algebraic groups G and G' over the field of complex numbers,both of which are reductive. Take a finite dimensional rational representation W of G x G'. Let Y = W//G := SpecC[WIG and X = W//G' := SpecC[WIG' be the affine quotients. Since the action of G and G' commutes on W, the quotient space X (respectively Y ) naturally inherits the action of G (respectively GI). In this note, we study the interrelation between the orbit structures of X / G and Y/G'. In a good situation, we can embed Y/G' into X I G , and the embedding map 0 preserves important properties such as the closure relation and nilpotency. We give a sufficient condition for the existence of such embedding, and provide many examples arising from the natural representations of classical groups. As a n application we consider the geometric problem of unimodular congruence classes of bilinear forms proposed by DokoviC-Sekiguchi-Zhao.
Introduction Let us consider two linear algebraic groups G and G' over the field of complex numbers, both of which are reductive. Take a finite dimensional rational representation W of G x GI. The afEne quotient of W by the action of G is denoted by Y = W//G := Spec(C[WIG,and similarly, X = W//G' := Spec(C[WIG',where C[W] is the ring of regular functions on W , and the superscript of G denotes the subring of G-invariants. Since the action of G and GI commutes on W, the quotient space X (respectively Y ) naturally inherits the action of G (respectively GI). In this note, we study the interrelation between the orbit structures of X / G and Y/G'. In a good situation, we can embed Y/G' into XIG, and the embedding map 8 preserves important properties such as the closure relation and nilpotency. We give a sufficient condition for the existence of such embedding. Let us briefly explain the condition.
198
Let cp : W + X and $ : W + Y be the quotient maps. Then % = $ - l ( $ ( O ) ) is called the null cone (for the action of G). We assume the pair (G,GI) and the representation W satisfy the following. Assumption A. (a) The quotient map $ : W + Y is flat. This m e a m the regular function ring C[W]is flat over (c[w]~. (b) There exists a n open dense G-orbit 00in %. (c) The null cone % is isomorphic to the scheme theoretic fiber W x y ( 0 ) ) i.e., the fiber product W x y ( 0 ) is reduced. (d) A generic fiber of the quotient map cp : W + X is a single (hence closed) GI-orbit. (e) Let W" be the union of closed GI-orbits 0' in W such that cp-'(cp(O')) = 0'.Then, for any y E Y , the fiber $-'(y) intersects W" non-trivially.
Our main theorem, which is proved in $2, is the following. Theorem B. Let us assume Assumption A holds. For any GI-orbit 0'in Y , there exists a G-orbit 0 in X such that = 0 holds. Thus we have a map 0 : YJG' + X J G which sends 0'to 0.The lifting map 8 is injective and preserves the closure relation. If 0'c Y is a nilpotent orbit, then 0= O ( 0 l ) is also nilpotent. Moreover, we have 0N (W x y @)//GI.
cp($-'(m))
Let us summarize the brief history of the above theorem. The statement of the theorem is primitively noticed by early works of Roger Howe, and then clearly stated by Przebinda et al. for complex nilpotent orbits of reductive dual pairs (see e.g., [4], [2]). Recently, Daszkiewicz-KraSkiewiczPrzebinda [3], Ohta [12] and Nishiyama-Ochiai-Zhu [lo] extend the lifting map to the case of symmetric pairs arising from dual pairs of real reductive groups, but still only for the nilpotent orbits. By private communication (cf. [13]), T. Ohta made me aware the fact that the orbits other than nilpotent ones are also in correspondence. In fact our proof of the lifting map in [lo] is applicable to all orbits without restriction, because of the geometric nature of the proof (see also [9]). In this paper, we extend the correspondence to the general framework which is independent of the notion of dual pairs. In $3, we provide many examples which satisfy Assumption A. The examples cover the cases in which W is the tensor product of the natural representations of classical groups, and also it contains several cases of
199
contractive actions of the general linear groups. Although most of the cases are already obtained from the theory of dual pairs, two types of the liiting maps are newly found (see Theorems 3.1 and 3.2). One of the new examples is strongly related to the SL(m,@) action on the space of m x m-matrices M,((c);
A
I+
gAtg
( g E S U m , @), A E Mm(@)),
(0.1)
which is studied by D. Z. DokoviC, J. Sekiguchi and K. Zhao [l]; also the structure of the orbit space is being studied by H. Ochiai recently. Our theory tells that the orbit space M,,,(@)/SL(m,@)can be embedded into the orbit space of the f i n e cone of the Grassmann variety G$(V @ V*) with GL(V)-action.
Theorem C. There is an embedding map
8 : M,(@)/SL(m, @) + G$(V
@ V*)/GL(V),
(0.2)
which preserves the closure relation, and carries nilpotent orbits to nilpotent ones. The image of the trivial orbit O1= 8({0}) is a spherical variety, and v v*) is normal. its closure i n G ~ ( CB
In 54, we have another embedding arising from the decomposition of the full matrix space into symmetric ones and skew-symmetric ones. Finally, we would like to propose some natural problems. Problem D. (1) Find a pair (G, G‘) and a representation W satisfying Assumption A, for which one of the pair is an exceptional group. ( 2 ) Consider irreducible representations V of G and U of G’. Classify all the pairs (V, U )for which W = V 8 U satisfies Assumption A. (3) Consider irreducible representations V of G = GL(n,@)and U of G’. Classify all the pairs (V, U)for which W = (V @ V*) 8 U satisfies Assumption A. (4) Give a complete description of the lifting map 8 in a combinatorial way. ( 5 ) Find a representation theoretic interpretation of the lifting map 8. In the case of the liftings arising from dual pairs, it is provided by the theta correspondence (or Howe correspondence). See [4] and [ll]. (6) Find the relation between the singularity of @ and that of 0, where 0= 8 ( 0 ‘ ) is the l i e d orbit.
The author thanks Ralph Bremigan for useful discussion and for pointing out the reference [14].
200
1. Preliminaries In this section, we summarize definitions and well known facts on afEne quotient maps. Let X be an afEne variety on which a reductive algebraic group G acts rationally. We denote the atline coordinate ring (or ring of regular functions) on X by C[X]. Then G naturally acts on C[X] via the formula 9 * f ).(
-
= f (9-1 ).
(f (I. E WI,9 E GI.
We denote the ring of G-invariants in @[XIby C[XIG. The a f i n e quotient X//G of X by the action of G is defined to be X//G = Spec C[XIG. The d n e variety X//G is often called the categorical quotient in the literature. The inclusion map C[XIG cs @.[XIinduces a quotient morphism cp : X + X//G, which has the following properties.
Lemma 1.1. Let cp : X
+ X//G
be a n a f i n e quotient map.
(1) For any y E X//G, the fiber cp-'(y) is a G-stable closed subvariety of X, and it contains a unique closed G-orbit. (2) Let 2 C X be a G-stable closed subvariety. T h e n the restriction cpIz : 2 + cp(2) c X//G is a n a f i n e quotient map, and consequently cp(2) = Z//G.
If we make a closed point y E X//G correspond to the unique closed G-orbit in the fiber cp-'(y) C X, we have a bijection between X/JG and the set of all closed G-orbits in X. In this sense, X / / G only classifies closed G-orbits in X. In the following, we give three basic examples of afEne quotient maps which will play important roles in the subsequent sections. 1.1. Special linear group Let V = C" be a vector space on which G = SL(n,C) acts naturally as the matrix multiplication. Take an another vector space U = Cm and put W = V @ U. G acts on W in the first component. If we identify W with then the the space of n x m-matrices over C, which we denote by Mn,m, action is given by the matrix multiplication on the left. Let us assume that n 5 rn and identify W = Mn,m.Then the ring of Ginvariants C[WIG is generated by all n x n-minors, which have the Plucker
201
relations. It is well known that there is no other relation among them (see, e.g., Theorem 3.1.6 in [8]),and we can identify the quotient W//G with the affine cone of the Grassmannian variety of n-dimensional subspaces in U. We denote it by G:ff(U). The f i n e quotient map cp : W = M,,,,,, + G:ff(U) is interpreted as follows. By the Plucker embedding, we consider G f ( U ) as the closed subvariety of A" U.Under this identification, the quotient map cp sends A E Adn,,,, to the exterior product of its rows. Thus, if rank A < n, then cp(A) = 0 E A" U. If n > m,then the only G-invariants in C[W] are scalars. So we have W//G = {*} (one point).
1.2. General linear group Let V = Cn be a natural (or defining) representation of GL(n,C). Take another vector spaces U+ = CP and U- = CQ, and put W = V@U+@V* @ U-,where V* denotes the contragredient representation of V. G acts on W naturally in the first components. We identify W with Hom (U+*, V) @ Hom (V, U-) . If n 2 p , q, then it is easy to see that W//G = U+ @ U-.The quotient map cp is given by
cp:W!xHom(U+*,V)@Hom(V,U-)3 ( f , g ) ~ + g o f E Hom (U+*, U-)21 U+ @ U-.
(1.1)
Let us consider the case where n < max{p, 9 ) . In this case, we have W//G = Det,(U+ @ U-), where Det,(U+ @ U-)denotes the determinantal variety of rank n, which is isomorphic to
{f E Hom (U+*, U-)I rankf 5 n) under the identification Hom (U+*, U-)N U+ @ U-. The quotient map is essentially the same as in (1.1). 1.3. Quadratic space Let V = C" be a vector space with a non-degenerate bilinear form, which we assume symmetric or skew-symmetric. We denote by G the group of isometries on V so that G is an orthogonal group O(n, C ) or a symplectic group S p ( n ,C ) according as the form is symmetric or skew-symmetric. Note that n is necessarily even in the skew-symmetric case since the bilinear
202
form is non-degenerate. Take an another vector space U = C" and put W=V@U. Let us first consider the symmetric case, hence G = O ( n ,C ) . I f n 2 m, then the quotient W / / G is isomorphic t o the symmetric tensor product Sym(U) in U @ U . Let us identify Sym(U) with { h E Hom (U*,U ) I t h = h}, where t h denotes the transpose of h. For f E Hom (V*,V ) N W , the image of the quotient map 'p is given by
V ( f ) : u*
'f b
v 2: v*
+u,
where the isomorphism V N V* is induced by the symmetric bilinear form. It is easy to see that 'p(f) belongs to Sym(U). If n < m, the above image 'p(f) belongs to Sym,(U) := Sym(U) n Det,(U 18 V), and we have W / / G = Sym,(U). The skew-symmetric case is similar. I f n 2 m, we have W / / G 2: Alt(V), where Alt(U) denotes the set of skew-symmetric tensor products in U @ U . Note that it is canonically isomorphic to { h E Hom (U*,U ) I t h = -h}. If n < m, we have
W / / G = Alt,(U) := Alt(U) n Det,(U c3 U ) . 2. Equivariant double fibration
Let G and G' be connected linear algebraic groups over C. which are reductive. Suppose that there exists a finite dimensional complex vector space W , on which G x G' acts linearly. We put
X = W / / G ' , with quotient map cp : W + X , Y = W / / G , with quotient map $ : W + Y. Then G naturally acts on X , and similarly, Y inherits an action of G'. By abuse of notation, we'denote the image 'p(0) (respectively $ ( O ) ) of 0 E W simply by 0 E X (respectively 0 E Y ) .
Definition 2.1. A G-orbit 0 c X is said to be nilpotent if The same definition applies to a GI-orbit 0'c Y .
a contains 0.
Let Z = W / / ( G x G') be the affine quotient of W by G x GI, which is naturally identified with X / / G and Y//G' respectively. We denote the induced quotient maps by $0 : X + Z N X / / G and 'po : Y + Z = Y//G'.
203
L e w 2.1. For a nilpotent G'-orbit 0' C Y , the subset p($-'(@)) C X is a union of nilpotent G-orbits. Proof. Let 0 c X be a G-orbit. Then 0 is nilpotent if and only if = (0}, where 0 E 2 is the image of 0 E W. Thus, it is enough to show that the image of p($-'(@)) under the map $0 is (0). Since $0 o cp = 90o $, we have
$o(a)
$0 0
p($- l (@
1) = 9 0 0
$($-'(w = Po(@) = (0).
0
Let '32 = $-'(O) C W be a null cone (or null fiber). Throughout this article, we assume the following.
Assumption 2.1. (a) The quotient map $ : W + Y is fiat. This means the regular function ring C[W]is fiat over C[wIG. (b) There exists a n open dense G-orbit 00in %. (c) The null cone '32 is isomorphic to the scheme theoretic fiber W x y {0}, i.e., the fiber product W X Y ( 0 ) is reduced. Few remarks are in order. If the action of G on W is cofree (i.e., C [ W ] is a graded free module over C[WIG), then $ is flat. The cofree actions are classified by J. Schwarz [14]. The assumptions (b) and (c) imply that the null cone '32 N W x y (0) is reduced and irreducible. The irreducibility follows from the assumption (b). Moreover, if G is semisimple, the assumption (b) implies that W x y (0) is reduced (see Korollar 2 in [i']),hence (c) holds automatically. Under these assumptions, we have
Theorem 2.1. Take a G'-orbit 0' in Y .
$-'(m)
(1) The scheme theoretic inverse image =W xy is reduced and irreducible. (2) The inverse image $-'(@) contains a n open dense G x G'-orbit D, hence there is a G-orbit 0 in X such that p($-'(@)) = W e say the G-orbit 0 is lifted from Or, and denote it by 0 = e(0'). (3) The lifting map 8 preserves the closure relation. If 0'is a nilpotent G'-orbit, then 0= 8 ( 0 ' ) is also nilpotent.
a.
Proof. This theorem is a generalization of Theorems 2.5 and 2.10 in [lo]. Note that the results in [lo] are stated for nilpotent orbits, but actually
204
they are valid for all kind of orbits. Hence, the proof is almost the same as in [lo], but for convenience of the reader, we indicate the outline of the proof. First, we prove that the scheme theoretic fiber W x y {y} is reduced for any y E Y. Then this will imply that W x y 2 is reduced for an arbitrary closed subvariety 2 c Y. In the terminology of commutative algebra, the claim that W x y {y} is reduced is equivalent to that C[W]@C[Y] C, does not contain any non-zero nilpotent element, where C, denotes the function ring on the one point set {y}. Note that we assume that C[W] & [ y ] Co N C[%] is an integral domain. Since C[W] @ ~ [ C, y ] is a deformation of the homogeneous integral domain C[%],it is also an integral domain. Thus we have proven W x y {y} is reduced and irreducible. Next, we shall prove that the fiber 11,-'(y) contains a dense open Gorbit. Put M = 11,-l(y) and denote by M^ the asymptotic cone of M (see 5.2 of [15] for the definition of asymptotic cones). Then, by the flatness of 11,, the asymptotic cone M^ coincides with the null cone %. Let 0, be a generic G-orbit in M . Consider the cone C M generated by M in W , then it is clear that the dimension of a generic orbit in CM is equal to dimO,, which in turn coincides with the dimension of the generic orbit in CM c W . Since % = M^ c (CM, the dimension of a generic orbit in % cannot exceed that of 0,. Note that % has an open dense orbit by Assumption 2.1 (b). This means that dim 0, 2 dim %. On the other hand, we have the equality dim M = dim M^ = dim 'JI of dimensions, hence dim 0, >_ dim M . Since 0, c M , we conclude that dim 0, = dim M , and that 0, is an open dense orbit in M , by the irreducibility of M just proved above. Since 11, is GI-equivariant, we get 11,-'(0')= G' -11,-'(g) for any y E 0'. Note that 11,-'(g) contains an open dense G-orbit 0,. Choose an arbitrary point w E U,,and we see the G x GI-orbit D = G'Gw is open dense in
11,-1 (0'). Since we assume that 11, is flat, it is an open map by Ex. (111.9.1) in [5]. Thus the equality 11,-l(0')= holds. Now we conclude that 0 = Gp(w) c p(11,-'(@)) is the open dense orbit which we want. The claim that the lifting map preserves the closure relation is obvious from the definition of 6. Lemma 2.1 tells us that 6 preserves nilpotency. 0
$-'(a)
Corollary 2.1. Let 0' be a GI-orbit in Y ,and 0 = 6 ( 0 ' ) its lip. Then we have C[O] N (C[W]@'@[Y]C[@])G'.
205
If@[W]is free over @[WIG',we can write @[W] = 31@C[WIG',where 31 is the space of G'-harmonic polynomials in @[W]. Then, the above corollary tells us that
@[Dl2L (31 €3 @[@])G'. Note that, as a G'-module, 31 is isomorphic to the regular function ring @[%I of %. Let us denote by X/G the set of all G-orbits. Note that X/G may not be an algebraic variety, but only a topological space. In general, the lifting map 0 : Y/Gr + X/G is not necessarily injective. Let us give a sufficient condition for the injectivity of 0. We denote
X " = {z E X
I cp-'(z)
W" = cp-'(x") =
consists of a single G'-orbit), and cp-'(z).
XEXO
Theorem 2.2. If, for any y E Y, the fiber $-'(y) intersects W" nontrivially, then the lifting map 0 : Y/Gr + X/G is injective.
Proof. Let 0;# 0;be two different G'-orbits in Y , and denote 0 i= 0 ( 0 ! , ) c X (i = 1,2). Without loss of generality, we can assume that 0;n @ = 0. Then $-'(@) n $-'(@) = 0 and, by the assumption, $-'(0; contains ) a closed Gr-orbit which is of the form cpd1(z) for some z E X. This means that z $! cp($-'(@)) = &, while z E cp($-'(@)) = 0 1 . Thus 0 1 # 0 2 which proves the theorem. 0 3. Double fibration related to the natural representations
Here we give several examples which satisfy Assumption 2.1. These examples arise from the natural representations of various classical groups. To exclude trivial cases, we further assume the following.
Assumption 3.1. (d) A generic fiber of the quotient map cp : W + X is a single (hence closed) G'-orbit. (e) Put X " and W" as in (2.1). For any y E Y, the fiber $-'(y) intersects W" non-trivially. Assumptions 2.1 and 3.1 assure that the liiting map 8 : Y/Gr + X / G is injective, and preserves the closure ordering. Also 0 lifts nilpotent orbits to nilpotent orbits.
206
3.1. Tensor products We first investigate examples satisfying Assumption 2.1. Let V be a finite dimensional representation of G, and U a finite dimensional vector space. Put W = V 8 U on which G acts naturally.
Lemma 3.1. The quotient map 1c, : W = V 8 U -+ Y := W//G satisfies Assumption 2.1 i f the representation (G,V) and a vector space U are in Table 1. Here we denote by Sym(U) (respectively Alt(U)) the set of symmetric (respectively alternating) tensor products in U 8 U. In these cases, the action of G o n W is cofree. Table 1. G
o(n,C) Sp(2n,C)
1
I
I
V
I
u
I
2dimU < n dimU 5 n
Cn (natural) CZn(natural)
Y ~
Sym(U) Alt(U)
Next, let us consider examples satisfying Assumption 3.1 (d), i.e., we need t o check that a generic fiber of the quotient map is a single orbit. Let U be a finite dimensional representation of GI. Take an arbitrary finite dimensional vector space V and put W = V 8 U as above.
Lemma 3.2. The quotient map cp : W = V 8 U + X = W//G' satisfies Assumption 3.1 (d) if the representation ( G ' , U ) and a vector space V are in Table 2. For the notation of Sym,(V),Alt,(V) and G$(V), see $1.1 and $1.3. Table 2. G'
U
l
v
l
x
Combining these lemmas, we have the following
Theorem 3.1. Let W = V 8 U be a representation of G x G' in Table 3. Here, f o r ezample, O(n,C ) @ O(m,C ) means the tensor product of the natural representations of O(n,C) and O(m,C ) .
207 Table 3.
w
condition O(n,C) ® O(m,C) 2m < n O(n,C)Sp(2m,C) 4m < n O(n,C)®SL(m,C) 2m < n Sp(2n,C) ®O(m,C) m
X Symm(C") Alt2m(C") Gjf(C») Symm(C2") Alt2m(C2") G aff( C 2n)
Y dual pair Sym(C-) (GL(n,R),GL(m,K)) Sym(C2m) (0(n,C),Sp(2m,C)) none Sym(C-) Alt(C-) (Sp(2n,C),0(m,Q) Alt(C 2m ) (GL(n,H),GL(m,H)) Alt(C m )
none
Then the double fibration by the affine quotient maps X = W//G' <—£— W —*—> W//G = Y satisfies Assumptions 2.1 and 3.1. In particular, a G'-orbit <0' CY is lifted to a G-orbit O C X, and the lifting map 0 is injective.
3.2. Contraction by the action of a general linear group In this subsection, we consider the action of general linear groups. Lemma 3.3. Let V = Cn be the natural (or defining) representation of G = GL(n, C), and U± finite dimensional vector spaces. Put W = V®?7+® V* ® U~ on which G acts naturally. Then the quotient map ip : W -> Y — W//G satisfies Assumption 2.1 if and only if dimU+ + dim J7~ < dim V. In this case, the action of G onW is cofree and Y is naturally identified withU+ ®U~. Lemma 3.4. Let U = Cm be the natural representation ofG' = GL(m, C), and V± finite dimensional vector spaces. Put W — V+ ® U ® V~ ® U* on which G' acts naturally. Then the quotient map X = W//G' satisfies Assumption 3.1 (d) if and only i/dim V+ = dim V~ or dim V ± > dim U. The quotient space X is naturally identified with the determinantal variety Det m (V + ® V~) of rank m = dimU (see §1.2). Theorem 3.2. If W is one of the representations of G x G' which are listed in (l)-(6) below, then the quotient maps X = W//G' ^^- W —*—>• W//G = Y satisfy Assumptions 2.1 and 3.1. In particular, a G'-orbit O' C Y is lifted to a G-orbit O C X, and the lifting map 6 is injective. (1) Let G = GL(n,C) and V = C" the natural representation ofG. We put W = (V ® V*) ® U for the natural representation U ofG' in Table 4. The quotient space X is given in Table 4 and Y = U ® U.
208
Table 4. G' 0(m,C) Sp(2m, C) SL(m,C)
condition 2m < n 4m < n 2m < n
X Sym^VeV)
Ait m (v®v*) c»f(vev)
(2) Let G - GL(n, C) and V = C" the natural representation ofG. We put G' = G'+xG'_, and W - V®U+®V*®U- for the natural representations U± of G'± in Table 5. The quotient space X is given in Table 5 and Y= Table 5.
G'+ 0(p,C) Sp(2p,C) 0(p,C) SL(f,C)
condition G'_ X 0(9) C) p + ? < n Sym J ,(V)®Sym,(V) Sp(2g,C) 2p + 2g < n Alt2p(V)®Alt 2 ,(V*) 5p(2g,C) p + 2g < n SL(q,C) p+g
dual pair (0(p,«),Sp(2n,R)) (Sp(2p,2«),0'(2n)) none none
(3) Let G' = GL(m,C) and U = Cm the natural representation of G'. We put W = V ® (U @ U*) for the natural representation V of G in Table 6. The quotient space Y is given in Table 6 and X = Detm(V ® V"). Table 6.
G 0(n,C) 5p(2n,C)
condition 4m < n 2m < n
Y Sym(U-®tT) Alt(t/©t/*)
(4) Let G' = GL(m,C) and U = Cm the natural representation of G'. We putG = G+xG-, and W = V+ ® U ® V~ ® 17* /or the natural representations V± of G± in Table 7. The quotient space Y is given in Table 7 and X = Detm(F+ ® V~). Table 7.
G+ 0(r,C) Sp(2r,C) 0(r,C)
condition y G_ Sym(J7)®Sym(I7*) 2m < r, s 0(s,C) Alt(J7)®Alt(f/*) m < r, a 5p(2s,C) Sp(2s,C) 2m < r- l,s Sym(f/)® Alt([/*)
dual pair (Sp(2m,R),0(r,s)) (0*(2m),5p(2r,2*)) none
209
( 5 ) Let V = C" (respectively U = C") be the natural representation of G = GL(n,C) (respectivelyG' = GL(m,C ) ) and assume 2m 5 n. We put W = V €9 U*@ V* 8 U. The quotient spaces are given by
X = Det,(V* 8 V) and Y = U*63 U. The corresponding dual pair is (GL(m,C),GL(n,C ) ) . ( 6 ) Let W = CP 8 Cr @ (CP)* €9 (C")" @ (CQ)* €9 (Cr)*@ CQ8 C" be a representation of G x G' with p + q 5 r, s, where G = GL(r,C)x GL(s,C) and G' = GL(p,C) x GL(q,C ) . The quotient spaces are given by X = Detp(C' €9 (Cs)*) x Detq((Cr)*63 C") Y = CP 8 (CQ)* CB (CP)* 63 CQ.
and
The corresponding dual pair is ( U ( p ,q), U(r,s)). 4. Application to Dokovit-Sekiguchi-Zhao problem
Recently, from the view point of the invariant theory, D. Z. Dokovit, J. Sekiguchi and K. Zhao [l]are studying the SL(m, C)-action on Mm(C) = C m BC" (see (0.1)). The orbit structure of the action is also being studied by H. Ochiai. Here we resolve this action into two different ways, which fit our theory of the double fibration. 4.1. Resolution via the contmction by the action of G L ( n ,C ) .
Let G = GL(n,C) and V = C" the natural representation of G. We put W = (V@V*)@Ufor the natural representation U = C" of G' = S L ( m ,C ) . In the following, we assume that 2m 5 n. Theorem 3.2 (1) tells us that the double fibration by the affine quotient maps
X=W//G'i
'
W
*
}W//G=Y
satisfies Assumptions 2.1 and 3.1. Here we only check Assumption 3.1 (e). Lemma 4.1. For any y E Y , the fiber $ - l ( y ) intersects a closed GI-orbit, which is precisely a fiber (p-'(x) for some x E X .
Proof. We identify W = V@U@V* @U with the space of 2n x m-matrices MZ,,,". Then a non-zero w E Mz,,," generates a closed GI-orbit if and only if rank w = m. In fact, for any non-zero x E G",(U), we can prove that the
210
fiber cp-l(z) consists of a single G’-orbit, hence it is closed. Thus all nonclosed orbits are contained in the null fiber cp-’(O) which is characterized by rankw < m. If we write w = ( A ,B ) E Mn,m x Mn,m, the quotient map TJ is given by
T J ( ~=)
t~~
E M,(c)
= Y. contains a full rank
Now it is elementary to verify that any fiber +-‘(y) matrix w under the condition m 5 n.
0
It is easy to see that Y = U @ U = Mm(C) inherits the SL(m,C)action in Eq. (0.1) considered by DokoviC-Sekiguchi-Zhao [l],while X is isomorphic to G z ( V @ V * ) .Thus, an SL(m,C)-orbit 0’ c Mn(C!)is lifted to a GL(n,C)-orbit 0 = O ( 0 ‘ ) C G$(V @ V * ) .Moreover, the lifting map O : Mm(C!)/SL(m,C ) + Gf(V @ V*)/GL(n,C)
(4.1)
is injective, and preserves the closure relation and nilpotency. As an example, we examine the simplest case, i.e., the lifting of the trivial orbit. Example 4.1. Let O1 = O({O}) be the lift of the trivial nilpotent orbit C Y. Then we have @ = %//G‘. Let us show that O1 is a GL(n, C!)-spherical variety with the normal closure in X = G$(V @ V*). To prove it, let us prepare some notations. We denote the set of paxtitions of length at most m by
0’= (0)
P, = {a = (a1,...,a,) E Zrn I a1 2
2 a , 2 O},
which we identify with the subset of dominant weights of GL(m,C) as usual. The irreducible finite dimensional representation of GL(m, C!) with highest weight a is denoted by 7Lm),and its contragredient 7Lm)*has the * holds. highest weight a* = (-a,, -am-l,. . . , -a1),hence T ( ~ ~=) T:;) For a,/3 E Pm,we put a O P = ( a , O ,..., O , p * ) E Z n ,
which is a dominant weight of GL(n, C). Now let us return to the proof. We note that as a GL(n,C) x (GL(m, C!) x GL(m, C))-module,
@.[%I
21 31 N
x@
TZo
a,BEP,
(T:?’
@ TAY’).
21 1
For this, see Theorem 2.5.4 in for example (note that the action of GL(rn,C) in the second factor is the dual to that in 6 ) . Then the action of SL(rn,C ) is obtained by the restriction of the action of GL(rn,C ) x GL(rn,C ) to the diagonal subgroup ASL(m,C). Note that
C if a - /3* = kl, for some k E Z, 0 otherwise, N C if where 1, = (1,1,. . . ,1) E Z". Thus we have (T$) @ and only if ai +/3,-i+l = k (1 I: i 5 rn) for a fixed non-negative integer k. From this, we obtain the decomposition of the regular function ring @[@I of the closure of the lifted orbit O1 as a representation of GL(n,C).
k20 c r - f i * = k l ,
(44 This shows that @ is a spherical variety, hence so is O1. As for the normality, T is known to be normal, and as a quotient of the normal variety, 0 1 is also normal. 4.2. Resolution via the action of the orthogonal and
symplectic groups. We denote the natural representation of O(p,C) by V+ and that of Sp(2q, C ) by V-.We put G = G+ x G- = O(p,C ) x Sp(2q, C ) and W = (V+@ V-)@ U for the natural representation U = C" of G' = SL(rn,C ) . We assume that 2m < p,2q + 1. By Theorem 3.1, the double fibration by the affine quotient maps
X = W//G' <
q w
*
bW//G=Y
satisfies Assumptions 2.1 and 3.1. Moreover, we have
Y = Sym(U) @ Alt(U) N M,(@) with the SL(m,C)-action in Eq. (O.l), while X is isomorphic to Gz(V+@ V-). Thus, an SL(rn,C)-orbit 0' C Mn(C) is lifted to an O(p,(c) x Sp(2q, C)-orbit 0 C Gz(V+@ V-).Moreover, the lifting map
e : iw,(c)/sL(rn,c) + sg(v+@ v-)/o(v+) x sp(v-)
(4.3) is injective and it preserves the closure relation and maps nilpotent orbits to nilpotent orbits.
212
Thus, if we put p = 2q, then the orbit space M,,,(C)/SL(m, C ) has two different embeddings to the same afEne Grassmaanian cone GE(V +@ V - ) . One is treated in this subsection, and the other is explained in $4.1. References 1. Dragomir Z. DokoviC, Jiro Sekiguchi, and Kaiming Zhao. On the geometry of unimodular congruence classes of bilinear forms. Preprint, 2003. 2. Andrzej Daszkiewicz, Witold Krdkiewicz, and Tomasz Przebinda. Nilpotent orbits and complex dual pairs. J. Algebra, 190(2):518-539, 1997. 3. Andrzej Daszkiewicz, Witold Krdkiewicz, and Tomasz Przebinda. Dual pairs and Kostant-Sekiguchi correspondence. I. J. Algebra, 250(2):408-426, 2002. 4. Andrzej Daszkiewicz and Tomasz Przebinda. The oscillator correspondence of orbital integrals, for pairs of type one in the stable range. Duke Math. J., 82(1):1-20, 1996. 5. Robin Hartshorne. Algebraic geomet y . Springer-Verlag, New York, 1977. Graduate Texts in Mathematics, No. 52. 6. Roger Howe. Perspectives on invariant theory: Schur duality, multiplicity-free actions and beyond. In The Schur lectures (1992) (Tel Aviv), pages 1-182. Bar-Ilan Univ., h a t Gan, 1995. 7. fiiedrich Knop. Uber die Glattheit von Quotientenabbildungen. Manuscripta Math., 56(4):419-427, 1986. 8. Laurent Manivel. Symmetric functions, Schubert polynomials and degeneracy loci, volume 6 of SMF/AMS Tezts and Monographs. American Mathematical Society, Providence, RI, 2001. Translated from the 1998 French original by John R. Swallow, Cours SpBcialis6s [Specialized Courses], 3. 9. Kyo Nishiyama. Multiplicity-free actions and the geometry of nilpotent orbits. Math. Ann., 318(4):777-793, 2000. 10. Kyo Nishiyama, Hiroyuki Ochiai, and Chen-bo Zhu. Theta lifting of nilpotent orbits for symmetric pairs. math.RT/Q312453,2003. 11. Kyo Nishiyama and Chen-bo Zhu. Theta lifting of unitary lowest weight modules and their associated cycles. To appear in Duke Math. J., 2003. 12. Takuya Ohta. Nilpotent orbits of &-graded lie algebra and geometry of the moment maps associated to the dual pair (U@,p), U ( T , S ) )Preprint, . 2002. 13. Takuya Ohta. On the geometric quotient which appear as the restriction of the moment map related t o dual pairs. In Proceedings of Symposium on Representation Theory 2003 (Ohnuma), pages 52-61, 2003. (Japanese). 14. Gerald W. Schwarz. Representations of simple Lie groups with a free module of covariants. Invent. Math., 50(1):1-12, 1978/79. 15. B. Vinberg and V.L. Popov. Invariant theory. In Algebraic geometry. IV, pages 123 - 278. Springer-Verlag, Berlin, 1994.
Infinite Dimensional Harmonic Analysis I11 (pp. 213-232) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
ADMISSIBLE WHITE NOISE OPERATORS AND THEIR QUANTUM WHITE NOISE DERIVATIVES
UN CIG JI* Department of Mathematics Chungbuk National University Cheongju 361 -763 Korea E-Mail: uncigji @cbucc.chungbuk.ac.kr
NOBUAKI OBATA+ Graduate School of Information Sciences Tohoku University Sendai 980-8579 Japan E-Mail: [email protected]. ac.j p
An operator on a Fock space is considered as a non-linear and non-commutative function of annihilation and creation operators at points { a t , a ; ; t E T}. The derivatives with respect to at and a ; , called respectively the annihilation- and creation-derivatives, are formulated within the framework of quantum white noise theory. We prove the differentiability of an admissible white noise operator and give explicit formulae for the derivatives in terms of integral kernel operators. The qwn-derivative is a non-commutative counterpart of the Gross derivative in stochastic analysis. Mathematics Subject Classifications: 46325 60H40 81825 Key words: Fock space, admissible white noise operator, Gross derivative, annihilation-derivative, creation-derivative, quantum white noise
1. Introduction
In this paper we focus on an operator on the (Boson) Fock space r ( H ) with H = L 2(T ,u),where T is a topological space equipped with a a-finite Bore1 measure v. A fundamental role is played by the annihilation and *Work supported in part by grant (No. R05-2002-000-00142-0) from the Basic Research Program of the Korea Science & Engineering Foundation. tWork supported in part by the Grant-in-Aid for Scientific Research (No. 15340039) from Japan Society for Promotion of Sciences.
214
creation operators at a point, denoted by at and a;, in many questions arising from quantum stochastic analysis, infinite dimensional harmonic analysis, quantum field theory and so on. The pair {at, a; ; t E T} is sometimes called a quantum white noise field on T . However, they are not well defined only within the framework of the Fock space. In literatures we find two formulations. One is to consider at and a; as unboundedoperator-valued distribution by smearing t; the other is to formulate them as continuous operators on a certain Gelfand triple ( E ) c r ( H ) c ( E ) * without smearing t. The second approach is along with the classical Hida calculus and has been considerably studied under the name of quantum white noise theory [4, 8, 171. We recall that every white noise operator admits a Fock expansion, i.e., E E ,C((E),( E ) * )is decomposed into an infinite sum of integral kernel operators:
where
x at, - . . a t m v ( d s l ) .. . v ( d s l ) v ( d t l ) ..v ( d t , ) ,
(1.2)
and 6 1 ,is~a kernel distribution. One may accept (1.2) as a “polynomial” in at and a;, hence (1.1) as a function of them: E = E(at,at ; t E T). Thus we are naturally led to a kind of functional derivatives:
These motivated us to study the annihilation- and creation-derivatives. The main purpose of this paper is to formulate the annihilation- and creation-derivatives (together called qwn-derivatives), and to study gwndifferentiability of a white noise operator. In fact, a general white noise operator is not qwn-differentiableand there is difficulty in formulating (1.3) in a direct manner. Our strategy is to establish an operator version of Gross derivative (the derivative along H ) of an admissible white noise function (see e.g., [l]). The paper is organized as follows. In Section 2 we prepare some notation in white noise theory. We introduce in Section 3 the spaces of admissible (test and generalized) functions as
( E ) c B c r(H) c B* c (E)*
215
and discuss an admissible white noise operator, i.e., a continuous operator from B into G*. In Section 4 we give the definition of the qwnderivatives OF2 for a white noise operator E. Then we prove the qwndifferentiability of an admissible white noise operator and obtain formulae for the annihilation- and creation derivatives (Theorem 4.4). This result follows from the qwn-differentiability of an admissible integral kernel operator (Theorem 4.1) and the Fock expansion of an admissible white noise operator (Theorem 4.3). The notion equivalent to an admissible white noise function was first introduced by Lindsay-Maassen [13] and has been studied in many literatures for different purposes, e.g., Awe-Bksendal-Privault-Ub~e [l], Belavkin [2], Benth-Potthoff [3], Grothaus-Kondratiev-Streit [ 5 ] , Ji [7], LindsayParthasarathy [14]. We hope that our new idea of qwn-derivative opens a new direction in the study of Fock space operators. Further study and application will be discussed in [lo]. 2. White Noise Theory
2.1. Underlying Gelfand rriple
Let T be a topological space equipped with a a-finite Bore1 measure v. Let H = L2(T,v) be the (complex) Hilbert space of L2-functions and the norm is denoted by 1. l o . Let A be a selfadjoint operator (densely defined) in H satisfying the conditions (Al)-(A4) below. ( A l ) inf Spec (A)
> 1 and A-'
is of Hilbert-Schmidt type.
Then there exist a sequence
and an orthonormal basis { e j ) g o of H such that Aej = Xjej. For p E R we define M
C~yl
t E H. A ~= ~ I ~(t,ej>12, j=O Now let p 2 0. We put E, = {t E H ; I I p < m} and define E-, to be the completion of H with respect to I - I-p. Thus we obtain a chain of Hilbert spaces { E p; p E R} and consider their limit spaces: It:I
=I
<
S A ( T )= E = proj limE,, P+M
Sfi(T) = E* = indlim E - p . ,+a
216
These are mutually dual spaces. Note also that SA(T)becomes a countably Hilbert nuclear space. Identifying H with its dual space, we obtain a complex Gelfand triple:
E = SA(T)C H = L2(T,v) c E* = Si(T). As usual, we understand that SA(T) and Si(T)are spaces of test functions and generalized functions (or distributions) on T,respectively. For white noise theory Si(T)must contain delta functions. But this is not automatic and we need further assumptions: (A2) For each function ( E SA(T) there exists a unique continuous function c o n T such that ( ( t ) = r(t) for v-a.e. t E T. Thus SA(T)is regarded as a space of continuous functions on T and we do not use the exclusive symbol The uniqueness in (A2) is equivalent to that any continuous function on T which is zero v-a.e. is identically zero.
c.
(A3) For each t E T the evaluation map bt : { t)( ( t ) ,( E SA(T), is a continuous linear functional, i.e., St E Si(T). (A4) The map t t)St E Si(T), t E T,is continuous with respect to the strong dual topology of Si(T). See [17] for more discussion on these assumptions. The canonical C-bilinear form on E* x E is denoted by (., .). In other words, we set
We also write .) for the canonical C-bilinear form on H. Let J be the conjugate operator defined by (a,
M
00
It then follows that
The real parts of E, H, E* are subspaces invariant under the action of J and are denoted by ER, H R and E&, respectively. Then we obtain a real Gelfand triple:
ER C HR C E&.
(2.1)
217
(These are real vector spaces but not necessarily spaces of R-valued functions.) R e m a r k 2.1. A prototype of our consideration is the case where T = R with Lebesgue measure v ( d t ) = dt and
A = l + t - -dt2 d= 2
(t + -i ) * ( t + - $ ) + 2 .
In this case SA(T)coincides with the space of rapidly decreasing functions, which is commonly denoted by S(R). Recall also that
ej(t)= (t/;;2jj!)-"'~j(t)e-t'/2,
j = 0,1,2,-.-
,
where H j is the Hermite polynomial of degree j , constitute an orthonorma1 basis of L2(R) and Aej = ( 2 j 2)ej. This prototype is suitable for stochastic processes, where R plays a role of the time axis. Our general framework allows to take T to be a manifold (space-time), a discrete space or even a finite set.
+
2.2. Hida-Kubo-Takenaka Space
Let Ep be the Hilbert space defined in $2.1, where p E R. We consider the (Boson) Fock space:
c 00
I'(EP) =
{
A
9 = ( f n ) Z o; fn E E p , II 9 1; =
n! I
fn
;1 <
n=O
I
7
which is essentially a direct sum of symmetric tensor powers of Ep and the weight factor n! is for convention. Having obtained a chain of Fock spaces {F(Ep) ; p E R}, we set
(E) = projlimI'(E,), P+W
( E ) *= indlimI'(E-,). P-+W
Then we obtain a complex Gelfand triple:
( E ) c W)c (El*, which is referred to as the Hida-Kubo-Takenaka space [ll]. By definition the topology of (E) is defined by the norms M
n=O
218
On the other hand, for each 9 E (E)* there exists p 2 0 such that 9 E I?(EPp).In this case, we have M
9 = (Fn). n=O
The canonical C-bilinear form on (E)*x (E) takes the form: 03
((9, 4)) =
C n!
( ~ n fn) , 3
9= (
~ n E)
( E ) * , 4 = (fn) E
(~1-
n=O
Here we recall two important elements of ( E ) * . (a) White noise. By assumption (A3),
wt =(o,&,o, ...),
t E T,
belongs to (E)*and is called a white noise. According as T represents time or space, the family {Wt ; t E 2') c ( E ) *is called the white noise process or white noise field on T . (b) Exponential vector: For x E E* an ezponential vector (or a coherent vector) is defined by n!
Obviously, q ! ~E~ ( E ) * . Moreover, ~$tbelongs to ( E ) (resp. I?(Ep))if and only if [ belongs to E (resp. E p ) . In particular, 4 0 is called the vacuum vector. 2.3. White Noise Operators
In general, a continuous operator from (E) into (E)* is called a white noise operator. The space of all white noise operators is denoted by C ( ( E ) ,( E ) * ) and is equipped with the bounded convergence topology. It is noted that C ( ( E ) (, E ) )is a subspace of L ( ( E ) (, E ) * ) . For each t E R the annihilation operator at is uniquely specified by the action on exponential vectors as follows: at45 = W 4 t l
t E E.
It is well known that at E C ( ( E ) (, E ) ) .The creation operator is by definition the adjoint a; E L ( ( E ) *(, E ) * ) We . see from (2.2) that the composition . . .a:,at, . . .atmis well defined and belongs to C ( ( E ) (, E ) * ) .
219
Let I , m 2 0 be integers and
Kl,m E
KP,m = sn+l
0
C(Eam,(E@')*). We define
(In 8 Kl,m),
(2.3)
where In : E@" + E@nis the identity and Sn+l : E@(n+')+ EQ(n+l)the symmetrizing operator. An integral kernel operator El,m(Kl,m) is defined by the action 9 = (fn) H ( g n ) given by Sn+l
It is known that
El,m(Kl,m)
=
(n
+ m ) ! Kzmfn+mr n!
n
2 0.
E L ( ( E ) ,( E ) * ) .
Remark 2.2. We have C(E@",(E@')*) E (E@('+"))*by the kernel theorem. For Kl,m € L(E@m, (E@')*) let I E ~ be , ~the corresponding element, i.e., (Kl,m,
q@' 8 Pm) = ( ~ l , r n t @ P1) ~ , >
t , q E E.
We then easily understand that (1.2) in Introduction is a descriptive expression for El,m(Kl,m).In many literatures the notation E l , r n ( ~ l , r nis) used for El,m(Kl,m),see e.g., [17].
For a white noise operator symbol are defined by
-
A
%d
= ((+,
4,))
E E C ( ( E ) (, E ) * )the symbol and the Wick
>
E(t977) = W
E , 9,))
e-(t?
t,v E E ,
respectively. A white noise operator is uniquely specified by the symbol or by the Wick symbol. For an integral kernel operator we have El,m(Kl,rnY(t,q)= ( K l , m P m , ~
~ ~ ) e ( E * q ) ,
E',m(Kl,mY(t, 7)) = (Kl,mtBm, 7"). 2.4. Gaussian Realization
Based on the real Gelfand triple (2.1) we define a Gaussian measure p by its characteristic function:
The celebrated Wiener-It6 decomposition theorem says that L2(E;, p ) is unitarily isomorphic to r ( H ) through the correspondence:
220
Taking (2.2) into account, we regard (E) as a subspace of L 2 ( E * , p ) .In this sense an element of ( E )is called a test white noise function and, accordingly, an element of (E)*is called a generalized white noise function.
3. Admissible White Noise Operators 3.1. Admissible White Noise Functions For p E R we set o(I
ni 4 1;
=
C n!ezpnifni%
4 = (fn)E
w).
n=O
For p 2 0 we define 4, = {d = (fn) E r ( H ); 1 1 d 1, < co} and 4-* to be Then {G, ; p E R} form a the completion of r ( H ) with respect to 1 1 . I-,. chain of Hilbert spaces satisfying
6 = projlimGp c 4, c 4 0 = r ( H ) c 4-, c G* = indlimG-,. P-+W
,--too
Note that 4 is a countable Hilbert space but not necessarily a nuclear space (4 is nuclear if and only if H is finite dimensional), and that 9 and B* are mutually dual spaces.
Lemma 3.1. For any pairp, q satisfying 0
1 1 d Illp L I1 d llq
and
p
5 -qlog 11 A-' llop we have
*
I1 lLq 5 Ill 9 II-,
where 4 E r ( H ) and @ E ( E ) * .(The norms can be in a usual way.)
00,
9
which is understood
The proof is immediate from the definition of the norms. Then, we have
(E)c B c r(H) c G* c p)*.
(3.1)
An element in 4 (resp. G*)is called an admissible test (resp. generalized) function. The canonical C-bilinear form on G* x 4 is denoted by ((., .)) too. 3.2. Admissible White Noise Operators
We note from the inclusion relations (3.1) that C(G,P*) is regarded as a subspace of L ( ( E ) ,( E ) * ) .A white noise operator belonging to the former space is called admissible. For an admissible operator we can find a pair of real numbers p 2 q such that Z E L(GP,Gq).
22 1
Proposition 3.1. Let Kl,m E ,C(Emm,( E @ ' ) * ) .Then the integral kernel operator El,m(Kl,m) is admissible if and only if K1,m E ,C(H@'", H@'). In that case, for an arbitrary q E R and r > 0 we have
Ill %m(Kl,mM Il, 5 C II K1,m llop Ill 9 Illq+r
>
where
-and 11
Kl,m
llop
stands f o r the Hilbert space operator norm. In particular,
=l,m(Kl,m) E L(Gq+r, Gq).
Proof. If Zl,m(Kl,m) is admissible, there exist a pair of real numbers p 2 q and a constant C 2 0 such that
1 1 %,m(Kl,m)$Il, 5 C Ill 4 Illp > 4 E Gp* Taking a particular q5 = (0,.. . ,0, f m , 0,.. . ), one obtains easily
which shows that K1,m E L ( H B m ,H@'). Conversely, suppose K1,m E L(H@", H@'). For q E R and q5 = ( f n ) E G we have by definition
Note from (2.3) that bounded by
11 KZm [lop 5 11 Kl,m Ilop. Then for any r > 0, (3.2) is 00
n=O
By an elementary calculus (see e.g., 117 Section 4.11) we have sup n2o
+
~
+
( n I ) ! ( n m)! -2rn n! n! e
er/2
'+m
5 e'llmm (7) .
Combining (3.3) and (3.4), we obtain the desired estimate.
(3.4) 0
222
3.3. Admissible White Noise Opemtors with Supporter
Let U C T be a Borel set with v ( U ) > 0. Then, starting from the Hilbert space L2(U,v)we obtain the spaces of admissible functions which are denoted by
G ( U ) c G p ( U ) c I V 2 ( U , v ) )= GO(U) c G-,(V)
c G*(U).
We identify G p ( U )with a closed subspace of Gp = Gp(T)through the natural inclusion L2(U,v ) L) L2(T,v). An element in Gp(V)is called an admissible white noise function supported by U. A description of the inclusion Gp(U)c)Gp(T)is given in terms of tensor product decomposition. We first recall the following fact whose proof is standard and is omitted.
Lemma 3.2. Let T = UIU U, U U Urn be a partition into a disjoint union of Borel subsets u p t o null sets. Then, the correspondence
4t
tEL2(W,
++4EtUl @ * * * . 4 E t U m ,
gives rise t o a unitary isomorphism
G,(T) for
Gp(V1)
€3
*
- - @4 Gp(Um)
all p E R.
Now let T = U U V be a partition, where v ( U ) > 0 and v(V) > 0 without loss of generality. It follows from Lemma 3.2 that
G P W (3.5) and 9 I-+ 9 €3 4 0 tv gives the canonical inclusion Gp(U)L) Gp( T ). With each continuous operator E E L(Gp(U),Gq(U)), where we assume GAT)
G P ( W @4
2 q without loss of generality, we associate an admissible white noise operator E €3 I according to the factorization (3.5), where I is the identity operator on Gp(V).Summing up, for a Borel set U C T and a pair of real numbers p 2 q we have inclusions:
p
L(GP(U),G,(W
c L(GP(T),G,(T)) c L ( ( E ) (El*). ,
An operator in L(Gp(U),G,(U))is called an admissible white noise operator supported by U. Whenever no confusion occurs we use the same symbol E for E @4 I. The concept of an admissible white noise operator with support is useful in the study of conditional expectation and quantum martingale, see [7].
223
4. Quantum White Noise Derivatives 4.1. l h n s l a t i o n Opemtor
Since each 4 E ( E ) is a continuous function on E;L, for any translation Ted defined naturally by
” E E;t.
TC4(”) = 4(” + C),
C E E;L the (4.1)
It is known [17]that Tcq5 E ( E ) and Tc E ,C((E),(E)). However, (4.1) is not applicable to a generalized function. By the Wiener-Its-Segal isomorphism, for 4 = ( f n ) E ( E ) we have
where Gm is the right m-contraction of symmetric tensor products. It is then natural to define the translation operator by extending the right hand side of (4.2). Namely, given C E E* (hereafter we allow a complex C) and 9 = (F,) E ( E ) * ,we define Tc@by the right hand side of (4.2) with replacing fn+m by Fn+,, whenever well defined as an element in ( E ) * .
Proposition 4.1. Let C E H and 9 E Gp with s o m e p E R. Then,for any it holds that Tc9 E Gq and Tc E ,C(Gp, Gq). q < p - log Proof. Let p , q E R. By definition,
Applying the Schwartz inequality, we have
00
0
0
.
224
Given p E R, we choose q E R such that 2e-2(P-q) Then we come to
which means that Tc9 E
< 1, i.e., q < p-log&
8, and Tc E L(Gp,Gq).
4.2. Gross Derivative
+
Modelled after abstract Wiener space theory, we say that E ( E ) * is Gross differentiable if for any C E H the translation Tee@is defined for small I E ~ < €0 and if
converges in (E)*with respect to the weak topology. Dc9 is called the Gross derivative of 9 in the direction C. Proposition 4.2. Every 9 = (F,) E G* is Gross differentiableand Dg9 = ( ( n+ l)C&F,+l)F=o. Moreover, Dc is a continuous linear operator on G* equipped with the strong dual topology.
+
Proof. Let C E H , 9 = (8’‘) E Gp and set \E = ( ( T I l)CGIFn+l):=o. We first note that \E E Bq for any q < p. In fact, by direct computation we obtain
Next we show that
P = Dc9.It follows from (4.2) that
Applying a similar estimate as in the proof of Proposition 4.1, we obtain
where q < p - log&. Thus we have shown that (4.3) converges in norm and the desired assertion follows. The last assertion follows from (4.4) and 0 general theory of locally convex spaces.
A Gross differentiable function 9 E ( E ) * is called pointwisely Gross differentiable if there exists a weakly measurable function t I-$ \Et E ( E ) *
225
such that the function t and
( m a ,9)) =
/
T
I+
11 9 t [ I p
belongs to H = L2(T,v) for some p E R
c ( t ) W t , 9))
c E H,
4%
9 E (E).
(4.5)
In that case we write 9 t = D t 9 . Note that D t 9 is determined for almost all t E T . The pointwise Gross derivative plays a basic role in stochastic analysis and similar derivatives have been introduced by many authors in different contexts, see e.g., [6,12,15,16]. Proposition 4.3. [l: Lemma 3.101 Every 9 E S* is pointwisely Gross differentiable and D t 9 = ( ( n+ l)Fn+l(t, for 9 = (F,). Moreover, if 9 = (F,) E S p and q < p - log then D t 9 E S, for v-a.e. t E T.
a,
Proof. Our proof is different from the one in [l]. Consider a function t I+ 9 t = ( ( n l)Fn+l(t, which is defined for almost all t E T by Fubini theorem. We note that
.))rz0,
+
S,11
qt
1;
v(dt)=
S,2
n!e2,”(n
+ 1)21Fn+l(t,*)I:
v(dt)
n=O
00
=
C(n+ l)e-2Pe-2(P-,)n(n + ~ ) ! e ~ p ( ~I F,+~I; +l) n=O
5 e-2pCp”-p1 1 9 Il l2p < 00. Then 9t E S, for almost all t E T. Since II *t
(4.6) IIqA0
I II q t I I,Ao I II *t II,
by Lemma 3.1, we see from (4.6) that the function t I+ 11 9t ll,Ao belongs to H = L 2 ( T , v ) . Finally, (4.5) follows from Proposition 4.2 with direct computation. Thus D t 9 = 9ffor almost all t E T. 0
Corollary 4.1. For 9 E ( E ) we have Dtq5 = atqi It is shown by norm estimates that
)
n
90 9=
(C k=O
Fk&,,-k
p* is closed under the Wick product:
03 7
@=
n=O
(Fn),
* = (Gn).
Then the next result is straightforward. Proposition 4.4. For 9, 9 E
G* we have
D c ( 9 0 9)= ( D c 9 ) o 9
+9
D t ( 9 o *) = ( D t 9 )o 9
+ 9 o (DtQ),
0
(Dcq),
c E H, for almost all t E
T.
226
4.3. Annihilation- and Creation-Derivatives
Let Z E L ( ( E ) ,( E ) * ) It . is proved that for any Q E E there exists 9, E (E)* uniquely specified by
40)
((*?I,=
((5*4q, 4d) ((4-7, 4d)
.t E E.
9
By using the Wick product we may write
9, = (E*+,)o&,. Now assume that 9, is Gross differentiable for all Q E E. This assumption is equivalent to that so is 3*&, since 9, o 4, = E*&. If, in addition, ( ( D c ~ ,$0) , is the Wick symbol of some operator in C ( ( E ) (, E ) * )for any C E H , denoted by DYE, i.e.,
(((D;Z)4b 4,))e-(c9q) = ((DC+,, 4th & Q E E, (4.7) then E is said to be differentiable in annihilation parts and DFZ is called the annihilation-derivative of Z with C E H . Similarly, the creation-derivative D l Z E C ( ( E ) (, E ) * )is defined by
WpMC, 4v))e-(Et9)= ((DC% 4,)L
t,QE E ,
where
*c = ( q t )
04-C.
We say that E E C ( ( E ) ,( E ) " )is qwn-differentiable if DFZ E L ( ( E ) ,( E ) * ) exists for all C E H . The derivatives DFE are regarded as non-commutative extension of the Gross derivative. Let us study the qwn-derivatives of an integral kernel operator. We need notation. For Kl+1,, E C(HGm,H6(l+l))and E H we define C*Ki+l,,,, E C(HGm,H6') by
{ (C * K l + l , m ) E @ m ,Pi}= (Kl+l,mPrn,Pi€3 c} > E , rl, c E H. Similarly, for C ( H G m ,H6") by
E L(H&("+l), H G r )and ( E H we define Kl,m+l* I E
((Kl,m+l*
C)E@'",
7 1 y = {Kl,m+1Prn €3
c, Q@).
Theorem 4.1. A n admissible integral kernel operator is qwn-differentiable. Moreover, for any Kl,m E C ( H G m ,H c z ) and C E H we have
D,Si,m(Ki,m) = mZ,m-i(Ki,m DfSl,,(Kl,,)
= G-l,m(C
*%?a).
* C),
(4.8) (4.9)
227
Proof. For simplicity we set Z = Zl,m(Kl,m).It follows from (4.7) and Proposition 4.4 that D;Z E L ( ( E ) ,( E ) * )is characterized by
(((qws,
47))e-(c97)
= (((Q=*47>0 4-77 4s))
+ ((=*4, 0 ( W - o ) , 4s))
*
(4.10)
The right hand side being equal to
(4.10) is equivalent to
As is verified by direct computation, DcZ*4,, = (h,) is given by
where n = 0,1,2,. .. (the second term vanishes for n = 0). Then the first term of the right hand side of (4.11) becomes
Therefore (4.11) becomes
from which we see that Z admits the annihilation derivative and (4.8) holds. A similar argument can be applied to (4.9). 0
228
4.4. Fock Expansion of an Admissible Opemtor
We assemble some general results on an admissible white noise operator to discuss its qwn-differentiability in the next subsection. As a special case of [9] we obtain
Lemma 4.1. Let p , q E R. For each L[,m E C ( H S m , H g 1 )there exists a unique operator Il,m(Ll,m) E L ( q ,Gq) such that
Il,m(Ll,mjiJ,7 ) = (Ll,mPm,
In this case, II Il,m(Ll,m)]lop I
rn 11 L1,m
5 , E~E -
9
110p-
Theorem 4.2. Let p , q E R. For any E E L($,,Gq) there ezists a unique family of operators Ll,m E C ( H G m , H g ' I ) ,Z,m 2 0 , such that 00
5=
16,m(LI,m),
(4.12)
l,m=O
where the series converges weakly in the sense that m
( ( ~ 4q)) , =
C
((1l,m(~l,rn)4, +))
3
4E~
p ,
1c1 E 9-q.
l,m=O
The expression (4.12) is called the chaotic expansion of 5. In fact, Ll,,,, is obtained by the formula: 1
Ll,m = -I?EIm, l!m!
(4.13)
where I, E L ( H Q m ,4)defined by ImFm = (0,. . . ,0,F,, 0,. . .). On the other hand, the Fock expansion of Il,m(Ll,m)is easily computed:
Inserting (4.14) into the chaotic expansion (4.12), we obtain the Fock expansion: 1Am
n!
l,m=O
(4.15)
Theorem 4.3. Let Z be an admissible white noise operator and let
c M
5=
l,m=O
Zl,m(Kl,rn)
(4.16)
229
be the Fock expansion. Then for all 1, m 2 0 we have Kl,mE L(HNm,l?@‘). Moreover, if E E L(Gp,Gq) for some p,q E R, then (4.16) converges in L ( G q - s + r , Gq-s) for any r > 0 and s > 0 satisfying (4.17) Proof. Given Z E L(G,G*), we define Ll,m as in Theorem 4.2. Comparing (4.15) and (4.16), we obtain
(4.18) from which the first assertion is obvious. We shall prove the convergence. Suppose that E E L(Gp,Gq) with some p , q E R and denote by 1 1 E 1 1 the operator norm. It follows easily from (4.13) that epm-d 1, m 2 0. II Llm llop 5 Ill 2 Ill 7
m
Then (4.18) becomes
Applying the Schwartz inequality, we see that the last quantity is bounded by
Thus, (4.19) becomes (4.20) Now let r , s > 0. Applying Proposition 3.1 and a simple inequality nn enn!, we obtain
5
Ill %n(Kl,m)+ lllq-s Hence for any r, s > 0 satisfying (4.17), the Fock expansion (4.16) converges 0 in L ( G q - s + r , G q - s ) .
230
4.5. Q WN-Derivatives of an Admissible Operator
Theorem 4.4. Eve y admissible white noise operator is qwn-differentiable. More precisely, i f the Fock e q a n s i o n of Z E L(B,G*) is given as in (4.16), then for any E H we have
<
l,m=O
l,m=O
where the right hand sides converges in the same manner as mentioned in Theorem 4.3. Moreover, DF is a continuous linear operator o n L(B,G*). Proof. Each El,m(Kl,m)is admissible by Theorem 4.3 and hence qwndifferentiable by Theorem 4.1. Then, it is sufficient to show the onvergence of the right hand sides of D t E . Suppose that Z E L(Bp,Gq). Applying Proposition 3.1 and (4.20),we obtain easily
1 1 m S , m - l ( K , m * C M lllq-s
1 1 11 I C 10 1 1 4
'
I!q-S+T
This estimate is almost the same as (4.21) and the series 00
C
1 1 mS,m-l
(Kl,m
* 04 I I I ~ - ~
l,m=O
converges whenever (4.17) is satisfied. In this case we have
IH (DT'I4
1IIq-S
'
I C IH HII C IOHI 4 !lq-S+T
7
with some constant C = C(p,q, r , s), which proves that DC is a continuous linear operator on L(B,G*). The argument for D;' is similar. 4.6. Pointwise Q WN-Derivatives
A qwn-differentiable operator E
E L ( ( E ) (E)*) , is called pointwisely qwndifferentiable if there exists a measurable map t I+ D:E E L ( ( E ) (, E ) * ) such that
((P;%,4 7 ) ) =
/T
( ( ( D ? W b hJ)C(t)v(dt),
c E H , 6 77 E E.
The following examples support the intuitive idea (1.3) in Introduction.
231
Example 4.1. For f E E* define KOJE L ( E , C ) and K1,o E L ( C , E * )by
K0,l :
< * (f,<) ,
K1,O : c
* cf,
respectively. The integral kernel operators
axe respectively called annihilation and creation operators associated with f . If f E H , both A f and A; are pointwisely qwn-differentiable and
D r A f = f ( t ) l , D t A f = 0;
DFAj = 0 , D:Aj = f ( t ) I .
Example 4.2. The number operator and the Gross Laplacian axe defined by
N = %,1(1)=
/
a f a t v(dt), AG = Z0,2(1)=
T
/
a:
v(dt),
T
respectively. Then we have
DFN = a;,
D:N = at,
DFAG = 2at,
D,+AG= 0.
References 1. K. Aase, B. Bksendal, N. Privault and J. Ub0e: White noise generalizations of the Clark-Haussmann-Ocone theorem with application to mathematical finance, Finance Stochast. 4 (2000), 465496. 2. V. P. Belavkm: A quantum nonadapted It0 formula and stochastic analysis in Fock scale, J. Funct. Anal. 102 (1991), 414-447. 3. F. E. Benth and J. PotthoE O n the martingale property for generalized stochastic processes, Stoch. Stoch. Rep. 58 (1996), 349-367. 4. D. M. Chung, U. C. Ji and N. Obata: Quantum stochastic analysis via white noise operators in weighted Fock space, Rev. Math. Phys. 14 (2002), 241-272. 5. M. Grothaus, Yu. G. Kondratiev and L. Streit: Complex Gaussian analysis and the Bargmann-Segal space, Methods of Funct. Anal. and Topology 3 (1997), 46-64. 6. T. Hida: “Analysis of Brownian Functionals,” Carleton Math. Lect. Notes, no. 13, Carleton University, Ottawa, 1975. 7. U. C. Ji: Stochastic integral representation theorem for quantum semimartingales, J. Func. Anal. 201 (2003), 1-29. 8. U. C. Ji and N. Obata: Quantum white noise calculus, in “Non-Commutativity, Infinite-Dimensionality and Probability at the Crossroads (N. Obata, T. Matsui and A. Hora, Eds.),” pp. 143-191, World Scientific, 2002. 9. U. C. Ji and N. Obata: A role of Bargmann-Segal spaces in characterization and expansion of operators o n Fock space, J. Math. SOC.Japan 56 (2004), in press.
232 10. U. C. Ji and N. Obata: Annihilation-derivative, creation-derivative and representation of quantum martingales, preprint, 2003. 11. I. Kubo and S. Taken&: Calculus on Gaussian white noise I, Proc. Japan Acad. 56A (1980), 376-380. 12. H.-H. Kuo: “White Noise Distribution Theory,” CRC Press, 1996. 13. J. M. Lindsay and H. Maassen: An integral kernel approach to noise, in “Quantum Probability and Applications I11 (L. Accardi and W. von Waldenfels Eds.).” Lecture Notes in Math. Vol. 1303, pp. 192-208, Springer-Verlag, 1988. 14. J. M. Lindsay and K. R. Parthasarathy: Cohomology of power sets with applications in quantum probability, Commun. Math. Phys. 124 (1989), 337-364. 15. P. Malliavin: “Stochastic Analysis,” Springer-Verlag, 1997. 16. D. Nualart: “The Malliavin Calculus and Related Topics,” Springer-Verlag, New York, 1995. 17. N. Obata: “White Noise Calculus and Fock Space,” Lect. Notes in Math. Vol. 1577, Springer-Verlag, 1994.
Infinite Dimensional Harmonic Analysis I11 (pp. 233-247) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
PDE APPROACH TO INVARIANT AND GIBBS MEASURES WITH APPLICATIONS
MICHAEL ROCKNER Fakultat fur Mathematik Universitat Bielefeld 0-3361 5 Bielefeld (Germany) Email: [email protected]. de
In this paper we give a pedagogical account of the PDE approach to invariant and Gibbs measures in finite and infinite dimensions. As an application we describe some recent new results on the classical problem whether “invariance implies Gibbsian” and illustrate how they apply to a wellstudied lattice model from statistical mechanics with non-compact single spin spaces.
AMS 1991 Subject classification: Primary: 58G32,58B99, 82B21. Secondary: 53C21, 53C80,58G03,60560,82B05. Keywords and phrases: invariant measures, diffusions on manifolds, elliptic equations for measures, Lyapunov functions, Gibbs distributions, logarithmic gradients To appear in conference proceedings.
1. Introduction This paper is an extended version of my talk given at the symposium in Tiibingen. The main purpose is to give a concise pedagogical account on the PDE (= partial differential equation) approach to invariant and Gibbs measures, developed in a series of papers during the last few years by S. Albeverio, V.I. Bogachev, Y.G. Kondratiev, T. Pasurek, F.Y. Wang and the author (cf. e.g. AKR97a AKR976, AKRTOO BRWOl BROl AKPROI), and de-
,
,
7
7
scribe a recent new application to the classical problem whether any invariant measure is Gibbsian (cf. BRWo2). The latter will be done in more detail, but the main result will be only precisely formulated, without recalling the proof from BRWo2. Instead, we shall present an application to an intensively studied model from statistical mechanics (see e.g. B H K 8 z ) .
234
2. The PDE-approach to invariant and Gibbs measures 2.1. The finite dimensional case
Consider the d-dimensional Euclidean space Rd (or more general a ddimensional manifold M d ) and a Bore1 measurable function E : Rd + (-m, +00] such that e-E E L1(Rd,dx) where dx denotes Lebesgue measure. For a better understanding of the following one should think of E ( z ) as the "energy" of the "configuration7' x = (xi)l
(2) p satisfies the following first order PDE:
aip = -&E (where
ai :=
(3) Setting
*
p
V 15 i 5 d
& and & p denotes the distributional derivative of p).
:= ( & ) l < i < d ,
&
d
:= &E and
LZ := c(8: - &&), then p i=l
symmetrizes LZ with domain C,"(Rd), i.e.
s
LZUv dp =
s
u LZVdp V U , v
E
C,"(Rd).
(4) p is Lz-infinitesimally invariant, i.e.
/
LZU dp = 0 V u E C,"(Rd),
or shortly,
L>p = 0 (so p satisfies a second order PDE).
Then (1) H (2) @ (3) + (4). Under our present smoothness assumptions, i.e. E E C 1 ( R d )the , im(3) are pretty much obvious. For (2) + (1) a plications (1) + (2) regularity result to ensure the existence of a sufficiently regular RadonNikodym derivative with respect to dx for any p satisfying (2) is necessary. For details on the latter in a much more general case, namely where merely e- 3 E E Hk:(IRd, dx) (= local Sobolev space of order 1 in L~o,(Rd, dx)) and e - i E > 0 dx-a.e. is assumed, we refer e.g. to Proposition 1.5 and its
235
proof in A R Z g 3 (see also Remark 2.1 below). For (3) =+ (2) we refer e.g. to Lemma 1 in BR03. The equivalence of (1) and (2) is of great importance since as we shall see in the next section it generalizes to infinite dimensions, so gives the possibility to study Gibbs measures by PDE-methods. Concerning the relation between (3) and (4) by choosing v, E Ci(Rd) with v, = 1 on a ball of radius n and letting n 4 00 we obviously deduce (4). So, we have “Gibbsian =+ infinitesimally invariance”. The converse in infinite dimension is a famous conjecture of Gibbs (originally formulated for so-called Hamiltonian dynamics). Under our present regularity assumption on E , i.e. E E C’(Rd),in this finite dimensional case the converse is also true. This follows from two highly non-trivial general results from Corollary 2.3 (see in particular also Remark 2.4.(i)) in Sta99 and Theorem 3.1 in BRSoo (see also Theorem 4.1 in B R S 0 2 ) which imply that the PDE in (4) has a unique solution, which since p in (1) is a solution, is therefore Gibbsian. If one relaxes the assumptions on E the situation becomes much more complicated. We summarize this in the following remark.
Remark 2.1. For E : Rd -+ mentioned condition
(-00,00]
e - i E E Hk:(Rd,dx)
and
let us only assume the already
e - i E > 0 dx-a.e.
(2.1)
Setting p := e-E (E H k t ( R d ,dx)) we can reformulate (2) as follows:
(2)
p satisfies the following first order PDE
for some (Borel) dx-version
(9)of 9.
9
Here we set as usual := 0 on { p = 0). In this case as mentioned above always p << dz, so the superfix “J’ can be dropped a posteriori. Note that in general q! L:,,(Rd,dx), so even if E E L:,,(Rd,dx), in general # & E , since the right hand side is always a Schwartz distribution. Then we still have
9
%
(1) =+ (2)
@
(3) =+ (4).
(9)and (3), (a) are just conditions (3), (4) respectively, augmented by the condition (9)-L t c ( R d , p ) 1, 5 i 5 d. Since Here we take
Zi :=
E
236
by Theorem 1 in BKR97 any solution of L*p = 0 is absolutely continuous with respect to dx, we can again drop a posteriori. Under condition (2.1) the PDE in (2) can have infinitely many solutions (cf. Example 6.1 in BR95 and also Remark 3.6(ii) in AKR97b) even for d = 1. So, (2) i+ (1). If, however, in addition, to (2.1) also E L:,,(Rd,dx), then p from (1) is the unique solution of the PDE in (27 (cf. Theorem 6.2 in BR95 and Theorem l.S(ii) in A B R 9 9 ) . In addition, = &E, 15 i 5 d , “N”
9
(cf. Lemma 6.4 in B R 9 5 ) . On the other hand, even if E L:,,(Rd,dx), it appears to be unknown whether + (3). But also no counterexample seems to be known. For conditions so that (4) + (3) in the present situaton we refer to Theorem l.S(i) in A B R g g for the most general result we are aware of that does not require higher integrability of If one
(a)
y.
assumes that
E Lt:E(Rd,dx),e
> 0 , then by the same results from
(a)
mentioned above the PDE in has a unique solution (among probability measures), so + ( l ) , hence (1) @ (2) H (3) @ in this case. Stagg, BRSoo, BRSo2
(a)
(a)
If we replace Rd by a suitable manifold, then there are easy examples even for smooth 2 where (4) + (3).
Example 2.1. (cf. Remark 2.5(ii) BRWo2) Let M be a connected complete Riemannian manifold with infinite volume measure A M , such that there exists a positive harmonic function h, integrable with respect to the volume measure AM (cf. GhzL83,Ls84 for existence). Choose the vector field 2 to be identically equal to zero, so LZ = A (= Laplacian on M ) . Define
Then for u E C i ( M )
237
since A h = 0. But for u , v E C i ( M )
s
L Z U Vd p =
/
u Av h dXM
+
/
uv A h d X M
+2
s
u ( V v ,V h ) d X M
= / u L ~ v d p + 2 / u ( V v , V h )d X M
since h cannot be constant because X M ( M )= 03. Therefore, p is infinitesimally invariant for L z , but not symmetrizing, hence not Gibbsian. In infinite dimensions much less is known about when or not (4) implies (3). We have included the quite detailed discussion in Remark 2.1 since it displays a typical characteristics in comparing finite and infinite dimensional analysis. Difficulties in the analysis of PDE or differential operators in infinite dimensions are reflected in part in finite dimensions if the coefficients become singular. The last remark of this subsection concerns the relation between “infinitesimal invariance” and “invariance”. We refer to Subsect. 2.5 in BRSoo and Sections 3 and 4 in BRSo2 for more details.
(a)
Remark 2.2. Assume again that (2.1) holds and that p is as in (cf. Remark 2.1). Set L := L z . Suppose there exists a closed extension ( ,b ’ , D ( i p ) ) of the operator ( L , C i ( R W don ) ) L 1 ( R d , p )which generates a Co-semigroup (T:)t>o = (et e w )t20 on L1(Rd,p).If p is the special measure in ( l ) ,then it follows by Corollary 2.3 and Remark 2.4.(ii) in that ( i p , D ( i p ) )must be the closure of ( L ,Ci(Wd)) and this closure really generates a Co-semigroup on L1(Rd,p ) . In this case a simple consideration implies that
T f p=
Jf
dp
for all t > 0, f E L 1 ( R d , p ) ,
(a),
(2.2)
For a general p satisfying however, the i.e. p is (T:)t>o-invariant. mere existence of (T,”)t?o= ( e e w ) t l o for a suitable closed extension 2 . of L is unknown and, if it exists, it might not be the closure of L. So, it is unclear whether p will be its invariant measure in the sense of (2.2). On the other hand, if (T,”)t>oexists and satisfies (2.2), by differentiating at t = 0, we deduce from (2.2) that p satisfies L*p = 0. So, even in finite
238
dimensions “infinitesimally invariance” seems to be a more general notion than “invariance”. 2.2. The infinite dimensional case
Consider now the d-dimensional lattice RZm,m E N,instead of Rd. (Again we could also consider a product Mi of finite dimensional Rieman-
n
iEZd
nian manifolds, cf. BRW02 for details). We are going to restrict the class of “energy”-functionals E : R”” -+ (--oo,oo] a bit, with applications to statistical mechanics in mind. So, let ~~
~ ( x:= )
C
uA(xA),
x
= (xi)icWZm
(2.3)
ACZ”
I4<m
where /A1 denotes the cardinality of A, X A := ( x i ) i E A , and UA : RA + (-00, 001 is Borel-measurable. Of course, (2.3) is purely informal, because this sum almost never converges. U A ( X is ) called the “potential of the configuration x in A” and as before E ( x ) is the LLenergy of the configuration z = ( x i ) i E p m ” . Similarly as in finite dimensions one then defines corresponding Gibbs measures p on R”” (equipped with the product of the Bore1 a-algebras on R) as
(1)’ p is a Gibbs measure (with energy E ) , i.e.
with
dxi
:= Lebesgue measure on
R1.
Of course, also (1)’is purely informal, since an infinite product of Lebesgue measures does not exist in a suitable sense and, as said before, E ( x ) is not well-defined. But it turns out (and has been well-known for many years) that the expression in the right hand side of the equality in (1)’ can be given sense. In general, however, the correspondingly defined measure is not unique. One reason is e.g. that if one defines the right hand side as a limit along a “localizing sequence” A, /” Z”, n + 00, with A, finite, the limit might depend on the sequence. Furthermore, “localizing” implies that one has to fix ‘Lboundaryconditions” outside every A,, and different choices of these might also lead t o different limits. A precise definition of a Gibbs measure taking all these issues into account requires the notion of a “local specification” and then a Gibbs measure can be defined by
239
determining its conditional probabilities on R" for each finite A C Z" by this specification fixing the configuration outside A, i.e. the measure satisfies the DLR (= Dobrushin-Lanford-Ruelle) equations. We refer e.g. to the exposition in Geo88 for details, since we do not need this below, since this rigorous version of (1)'is as in finite dimensions equivalent to the following, infinite dimensional analogue (2)' of (2) which is rigorous. This equivalence has been proved in in various frameworks: AKR97a,AKR97b1AKRT00
(2)' p satisfies the following first order PDE (in infinitely many variables):
Precise assumptions for "DLR-version of (1)'
(2)"' to hold are e.g.
c Z" and for some R > 0 and diamh > R
for all finite A
UA = 0 if
("finite range interaction"). To define the distributional derivative 13ip in (2)' we need a test function space. As usual in infinite dimensions we take for e E N U {co}
FC; := { u : RZn
I
R 3 finite A c Z" and g
E C;(RA)
such that ~ ( I c = ) g ( X A ) for all
IC E
(2.6)
ItZm}.
Then a probability measure p on R"" satisfies the PDE in (2)' if
where
za :=
&uA,
iEZm.
(2.8)
A:iEA ACZm, A finite
Note that by the finite range condition the sum in (2.8) has only finitely many non-zero summands, so Zi is well-definied. As in the preceding subsection we now consider two further assertions about a probability measure p on R"":
240
(3)’ Setting Z := (Zi)iEzrn,Zi as in (2.8)’ and L z := then p symmetrizes LZ with domain 3C;, i.e.
cz“=,Ca? - Zit$),
(4)’ p is Lz-infinitesimally invariant, i.e.
/Lzudp=O
VUE~C?,
or shortly,
L>p = 0 (so p satisfies a second order PDE in infinitely many variables). We emphasize that since u (and v) are in 3Cz the sum in the definition of LZU has again only finitely many non-zero summands. So, all is welldefined. In A K R g 7 a , AKR97b, AKRToo also the equivalence (2)’ ($ (3)’ has been proved, and obviously (3)’ + (4)’by taking v 3 0. So, altogether as in finite dimensions we have: DLR-version of (1)’($ (2)’ @ (3)’ +-(4)’. However, as mentioned in the previous subsection the implication “(4)’+ (3)”’ even under stronger smoothness assumptions on the UA is a major problem in this infinite dimensional case. In the next section we shall present a result giving a sufficient condition for this to hold.
Remark 2.3. (i) Let p satisfy (2)’ (H (3)’ @ DLR-version of (1)’). In this infinite dimensional case ( as in finite dimensions, cf. Remark 2.2) again even the mere existence of (T[)t20= (ett”)tgo as a Co-semigroup on L1(RZrn,p)for a suitable closed extension LP of L is not clear in general and only known under quite stringent assumptions. However, if ( T [ )-t >=~ (etep)t>o - exists and the analogue of (2.2) holds for p, then as in finite dimensions always L*p = 0 by taking $ It=o. So, “invariance” implies “infinitesimally invariance”, but the converse is unlike in finite dimensions in fact known to be wrong in general (cf. Chap. 5b in Ebe99 for counter examples). Concerning the question whether (4)’ + (3)’’ our result in the next section is therefore more general than just stating it for invariant measures, since we prove it for a larger class. In particular, it generalizes the classical well-known
241
results in e.g. H S 8 1 , Fri82 since it holds also for infinite products of manifolds (cf. B R W o 2 ) . (ii) To be precise we mention that in (2)', (3)', (4)' above one has to assume, in addition, each time that Z i E L2(Rzm,p) V i E Z", in order to have that (Lz,FC;) is an operator on L2(Rz",p) and that the Gibbs measure in (1)'defined through the DLR-quations is ternpered in a suitable sense. We suppressed this point above since in applications, the square integrability of Zi is automatic (cf. Section 4 below on applications). 3. Infinitesimally invariance implies Gibbsian
Consider the situation described in Subsection 2.2, so UA, Z = ( Z i ) i E z m , L z are as defined there. Assumptions on the potentials UA: Let UA, A c Z", IAl < 00, satisfy assumption (2.5) for some (fixed) R > 0, and in addition:
where
Remark 3.1. We note that obviously for Ic E N Ek(z)
=Ek(xAk),
E IWZm 9
and for i E Z"
SO,
zi(x) = z i ( x i + A l ) ,
x E Rzm,
zi(x)= Z i ( x A k + l ) ,
2 E
if i E h k ,
Rzm.
Now we can formulate one of the main results from
BRW02.
Theorem 3.1. Let p be a probability measure on R"" such that
242 (2)
zi E L2(R"",p)
(ii) L > p (iii) & E k Set f o r k
v i E Z", i.e. p is Lz-infinitesimally invariant, E L2(R"", p ) 'd i E A k , k E N.
= 0,
E
N
-
where Ep[. I a ( A k ) ] denotes conditional expectation of p with respect t o the X A ~ x , E RZm. If there exist a-algebra u(hk) generated by the map x c k E [ ~ ; , c of l) (O,CO) such that (3.3)
then p is Lz-symmetrizing, i.e. Gibbsian. Instead of giving an account of the proof of Theorem 3.1 we refer to BRW02 and shall rather discuss an application in the next section. We only mention here that the crucial quantities LIE, Ic E N,in (3.2) exactly capture how strongly the PDE's in (3)' and (4)' are coupled with respect to the one dimensional coordinates of x = ( x i ) i E Z m or "how much" p differs from a product measure on R"". This will become, particularly, clear in the applications below. (3.3) just says that DE should not grow too fast with
k. 4. Application
In this section we shall apply the above, in particular, Theorem 3.1, t o a well-studied model from statistical mechanics (cf. e.g. B H K 8 2 and the references therein). This is a lattice system over Z" with a two-body interaction of finite range R > 0, i.e. in the frame from the previous section we have for finite A c Z":
UA = 0, If A
=
unless d i a m h 5 R and IAl 5 2.
{i} we set
v, := U,,}, and if A = { z , j } , i
# j , we set wi,j
:= U { i , j ) .
(4.1)
243
So, in particular, Wi,j = Wj,i. Assumptions on V,, Wi,j:
V, E C'(R),
(4.2)
W2,j E Cl(rW2).
There exist K , K o E (0, co),p E (m,co) such that K > 12Ko(1+R)2+p, and C E (0, m), a E [ 2 , co),such that for all i, j E Z m and all s, t E R (wi,j(s,t)J I Ko(1
+ ISIQ f ItI"),
+
J & W i , j ( S , t )5 JK
o(~ sv,'(s) 2 KISI"
-
+ Itla-'), c. (4.3)
It can be easily shown that (4.2), (4.3) imply conditions (2.5) and (3.1). Let us first calculate the corresponding Zi and &Ek. We have for i E Z",k E N,in this situation
C
z ~ ( Z )= - ~ l ( Z i )-
a1Wi,j(Zi7Zj)
(4.4)
j€Z"
li-jls R &Ek(Z) = -K'(Zi)-
& W i , j ( Z i , Zj),
if i E h k .
(4.5)
jE,Ak
li-3llR
By Examples 6.12 and 4.6 BRol there exist probability measures p on R"" such that the following properties hold:
(A) L>p
= 0.
(B) ("temperedness of p7')
Furthermore, for all such p and every such that
T
E
(0,co) there exists M , E (0,co)
in particular (by (4.2), (4.3))
zi, &Ek for i E
Z" and k
E
W
E L2(R"",p)
is such that a E Ak.
244
Now let us calculate the crucial quantities D:, k E W, from the conditions of Theorem 3.1. Fix k E N and i E A k , then since aiEk is a ( A k ) measurable we can use Jensen's inequality to obtain
J
'
jCZrn\Ak
li-jlgz
where we used (4.4), (4.5) in the last step. By ( 4 . 3 ) and Holder's inequality the latter is bounded by
5 3(2R + 1 ) 2 m ~( i1 + 2 ~ 2 , - 2 ) =: C ( R ,m,KO, a ) where we used (4.7) in the second step. So, for any such p as above and kEN
D: I C(R,m,Ko,Q) I h k \Ak-ll
=: C k .
Note that the latter is of order k m - l , so
where "M" means "equal up to a constant". So, Theorem 3.1 applies if and only if m 5 2 . Hence we have
Corollary 4.1. For the one- and two-dimensional lattices Z1 and Z2 for the above model we have: infinitesimal invariance implies Gibbsian. Concluding Remark 4.2. (i) Let ZLX denote the set of all probability measures on R"" satisfying conditions (A), (B) above. By similar techniques one can prove (cf. B R W 0 2 ) : if one p E ZLX satisfies the logarithmic Sobolev inequality, then #ZLX = 1 , i.e. we have uniqueness in this case. This extends results known for compact spin spaces (i.e. lattices of type Mi,
n
iEZm
245
compact manifolds, cf. SZ92a 1 SZ92c S.Z92b Sz95 Zeg92 7 Zeg96 ) with M~ for a particular case with to t h e non-compact case (see also Mi = R as above). (ii) There are examples on R”” where “infinitesimal invariance + Gibbsian” for all lattice dimensions m. We refer to BRW02 for details.
Acknowledgement
It is a pleasure for t h e author to t h a n k Professor H. Heyer for a very pleasa n t and stimulating conference in Tubingen and for t h e financial support through t h e DFG. Financial support of t h e BiBoS-Research Centre a n d t h e German Science Foundation (DFG) through t h e DFG-Research G r o u p “Spectral Analysis, Asymptotic Distributions and Stochastic Dynamics” is also gratefully acknowledged.
References ABR99. AKPR04.
AKR97a. AKR97b.
AKRTOO.
ARZ93.
BHK82.
BKR97.
BR95.
S. Albeverio, V. I. Bogachev, and M. Rockner, On uniqueness of invariant measures for finite and infinite dimensional diffusions, Comm. Pure Appl. Math. 52 (1999), 325-362. S. Albeverio, Y . G. Kondratiev, T. Pasurek, and M. Rockner, Euclidean Gibbs measures on loop lattices: existence and a priori estimates, Ann. Prob. 32 (2004), no. l A , 153-190. -, Ergodicity of L2-semigroups and extremality of Gibbs states, J. F’unct. Anal. 144 (1997), 394-423. S. Albeverio, Y . G. Kondratiev, and M. Rockner, Ergodicity for the stochastic dynamics of quasi-invariant measures with applications to Gibbs states, J. Funct. Anal. 149 (1997), 415-469. S. Albeverio, Y . G. Kondratiev, M. Rockner, and T. V. Tsikalenko, A-priori estimates on symmetrizing measures and their applications to Gibbs states, J. F‘unct. Anal. 171 (2000), 366-400. S. Albeverio, M. Rockner, and T.-S. Zhang, Markov uniqueness for a class of infinite dimensional Dirichlet operators, Stochastic Processes and Optimal Control (H. J. Engelbert et al., eds.), Stochastic Monographs, vol. 7, Gordon & Breach, 1993, pp. 1-26, J. Bellissard and R. Hpregh-Krohn, Compactness and the mazimal Gibbs states for random Gibbs fields on a lattice, Comm. Math. Phys. 84 (1982), 297-327. V. I. Bogachev, N. V. Krylov, and M. Rockner, Elliptic regularity and essential self-adjointness of Dirichlet operators on Rd, Ann. Scuola Norm. Sup. Pisa C1. Sci., Serie IV XXIV (1997), no. 3, 451-461. V. I. Bogachev and M. Rockner, Regularity of invariant measures on finite and infinite dimensional spaces and applications, J. Funct. Anal. 133 (1995), 168-223.
246 BRO1. BR03. BRSOO.
BRSO2.
BRWOl.
BRWO2. Chu83. Ebe99.
Fri82. Geo88. HS81. LS84.
Sta99.
SZ92a.
SZ92b. sz92c. sz95. YOSO 1.
Zeg92.
-,
Elliptic equations for measures o n infinite dimensional spaces and applications, Prob. Th. Rel. Fields 120 (2001), 445-496. -, O n LP-uniqueness of symmetric diffusion operators o n Riemannian manifolds, Matem. Sbornik. 194(7) (2003), 969-978. V. I. Bogachev, M. Rockner, and W. Stannat, Uniqueness of invariant measures and essential m-dissipativity for diffusion operators o n L1, Infinite dimensional Stochastic Analysis (P. Clement et al., eds.), Royal Netherlands Academy of Arts and Sciences, Amsterdam, 2000, pp. 39-54. ___ , Uniqueness of solutions of elliptic equations and uniqueness of invariant measures of diffusions, Matem. Sbornik. 193:7 (2002), 3-36, (Russian version), 945-976 (English version). V. I. Bogachev, M. Rockner, and F.-Y. Wang, Elliptic equations f o r invariant measures o n finite and infinite dimensional manifolds, J. Math. Pures Appl. 80 (2001), 177-221. ___ , Invariance implies Gibbsian: some new results, BiBoSPreprint 02-12-106 (2002), to appear in Comm. Math. Phys., 18 pp. L. 0. Chung, Existence of harmonic L1-functions in complete Riemannian manifolds, Proc. Amer. Math. SOC.88 (1983), 531-532. A. Eberle, Uniqueness and non-uniqueness of semigroups generated by singular diffusion operators, Lecture Notes in Math., vol. 1718, Springer, Berlin, 1999. J. Fritz, Stationary measures of stochastic gradient systems, infinite lattice models, Z. Wahr. verw. Geb. 59 (1982), 479-490. H.-0. Georgii, Gibbs measures and phase transitions, de Gruyter, 1988. R. Holley and D. W. Stroock, Diffusions o n a n infinite dimensional torus, J. Funct. Anal. 42 (1981), 29-63. P.Li and R. Schoen, Lp and mean value properties of subharmonic functions o n Riemannian manifolds, Acta Mathematica 153 (1984), 279-301. W. Stannat, (Nonsymmetric) Dirichlet operators o n L1: Existence, uniqueness and associated Markov processes, Ann, Scuola Norm. Sup. Pisa C1. Sci., Serie IV 28 (1999), 99-140. D. W. Stroock and B. Zegarlinski, T h e equivalence of the logarithmic Sobolev inequality and the Dobrushin-Shlosman mixing condition, Comm. Math. Phys. 144 (1992), no. 2, 303-323. -, T h e logarithmic Sobolev inequality for continuous spin systems on a lattice, J. Funct. Anal. 104 (1992), no. 2, 299-326. ~, The logarithmic Sobolev inequality for discrete spin systems o n a lattice, Comm. Math. Phys. 149 (1992), no. 1, 175-193. -, O n the ergodic properties of Glauber dynamics, J. Stat. Phys. 81 (1995), no. 5-6, 1007-1019. N. Yoshida, T h e equivalence of the log-Sobolev inequality and a maxing contition for unbounded spin systems o n the lattice, Ann. Inst. H. Poincare Probab. Statist. 37 (2001), no. 2, 223-243. B. Zegarlinski, Dobrushin uniqueness theorem and logarithmic Sobolev
247 inequalities, J. Funct. Anal. 105 (1992), no. 1, 77-111.
Zeg96.
-, T h e strong decay t o equilibrium for the stochastic dynamics of unbounded spin systems o n a lattice, Comrn. Math. Phys. 175 (1996), no. 2, 401-432.
This page intentionally left blank
Infinite Dimensional Harmonic Analysis I11 (pp. 249-264) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
DEFORMATIONS OF CONVOLUTION SEMIGROUPS ON COMMUTATIVE HYPERGROUPS
MARGIT ROSLER Mathematisches Institut, Universitat Gottingen Bunsenstr. 3-5, 0-37073Gottingen, Germany E-mail: roeslerOuni-math.gwdg.de MICHAEL VOIT Fachbereich Mathematit, Universitat Dortmund D-44221 Dortmund, Germany E-mail: michael.uoitOmathematik.uni-dortmund.de It was recently shown by the authors that deformations of hypergroup convolutions w.r.t. positive semicharacters can be used to explain probabilistic connections between the Gelfand pairs (SL(d,C),S U ( d ) )and Hermitian matrices. We here study connections between general convolution semigroups on commutative hypergroups and their deformations. We are able to develop a satisfying theory, if the underlying positive semicharacter has some growth property. We present several examples which indicate that this growth condition holds in many interesting cases.
1. Introduction
Klyachko recently derived a connection between SU(d)-biinvariant random walks on SL(d,C) and random walks on the additive group H ~ , of o all hermitian d x d-matrices with trace 0, whose transition probabilities are invariant under conjugation by S U ( d ) . He used this connection t o transfer the recent solution of the spectral problem for sums of hermitian matrices (7, lo) t o the possible singular spectrum of products of random matrices from SL(d,C) with given singular spectra. The singular spectrum of a matrix A g SL(d,C) here means the spectrum of the positive definite matrix Klyachko’s connection between SL(d,C) and H ~ , w oas explained in a different way and extended by the authors in 18; it is shown in l8 that the commutative Banach algebra of all SU(d)-biinvariant bounded measures on SL(d,C) may be embedded into the Banach algebra of all bounded measures on the Euclidean space H ~ , in o an isometric, probabil-
m.
250
ity preserving way. The proof of this fact, which has some applications in probability theory (see 18), depends on so-called deformations of hypergroup convolutions with respect t o positive semicharacters as introduced in 20. These deformations lead to connections between random walks and convolution semigroups on different, but closely related hypergroups. This forms the motivation to investigate systematically when and how convolution semigroups of probability measures on a commutative hypergroup ( X ,*) can be transformed canonically into convolution semigroups on a deformation ( X ,0 ) of ( X ,*). In particular we show that the generators and Levy measures of the original and the deformed convolution semigroup are closely related whenever this transformation is possible. We mention that the deformation of convolution semigroups is closely related t o Doob’s h-transform, and that L6vy processes associated with a convolution semigroup and its deformation are related by a Girsanov transformation on the path space; see 21, The paper is organized as follows: In Section 2 we collect some facts on deformations and present examples. In particular we indicate how for a maximal compact subgroup H of a complex, non-compact, connected semisimple Lie group G, the double coset hypergroup G / / H may be r e garded as deformation of an orbit hypergroup. This includes the examples above. Section 3 is devoted t o deformations of convolution semigroups w.r.t. positive semicharacters ao. We show that this concept works in a satisfying way under a canonical growth condition on the convolution semigroup together with some growth condition concerning (YO. Section 4 finally contains examples where this condition on a0 is satisfied. In fact, we have no example for which this condition would not hold. 2. Deformations of commutative hypergroups We give a quick introduction. First, let us fix notations. For a locally compact Hausdorff space X , M + ( X ) denotes the space of all positive Radon measures on X , and M b ( x ) the Banach space of all bounded regular complex Bore1 measures with the total variation norm. Moreover, M 1 ( X ) c &(x) is the set of all probability measures, M , ( X ) C &(x) the set of all measures with compact support, and 8, the point measure in z E X . The spaces c(x)3 cb(x)3 C o ( x ) 3 cc(x)of continuous functions are given as usual.
Definition 2.1. A hypergroup (X, *) consists of a locally compact Hausdorff space X and a weakly continuous, probability preserving convolution
251
* on Mb(X)such that ( M b ( x )*), is a Banach algebra and * preserves compact supports. Moreover, there exists an identity e E X (such that 6, is the identity of ( M b ( x ) *)) , as well as a continuous involution z I+ Z on x that replaces the group inverse. For details we refer to and 6 . We here only deal with commutative hypergroups ( X ,*), i.e., * is commutative. In this case there exists an (up t o normalization) unique Haar measure w E M + ( X ) which is characterized by w ( f ) = w(f,) for all f E C,(X) and z E X , where we use the notation
Similar t o the dual of a locally compact abelian group, one defines
x ( X ) :={aE C ( X ): a # 0 , a(z * y) = a(z)a(y)for all z,y E X } , X * :={a E X ( X ):
a(z)=a(.> for z E X I ; 2 := X * n cb(x).
Elements of X * and 2 are called semicharacters and characters respectively. All spaces are locally compact w.r.t. the compact-uniform topology. Example 2.1. (1) Let K be a compact subgroup of a locally compact group G. Then
Mb(GIIK) := { p E Mb(G) : 6,
* p * 6,
= p for all z,y E K }
is a Banach-*-subalgebra of Mb(G) with the normalized Haar measure dlc E M 1(G)of K as identity. The double coset space G / / K := { K z K : z E G } is locally compact w.r.t. the quotient topology, and the canonical projection p : G -+ G / / K induces a probability preserving, isometric isomorphism p : ikfb(GIIK)-+ Mb(G//K) of Banach spaces by taking images of measures. The transport of the convolution on Mb(GIIK) to ikfb(G//K) via p leads t o a hypergroup structure ( G / / K ,*) with identity K and involution ( K z K ) - := K z - l K , and p even becomes a Banach-*-algebra isomorphism. If G / / K is commutative, i.e., (G,K ) is a Gelfand pair, then a K-biinvariant function cp E C(G)with cp(e) = 1 is spherical if cp(z)cp(y) = 1, f(zlcy) dlc for z, y E G. The functions a E x ( G / / K ) are in one-to-one correspondence with the spherical functions on G via a I+ a o p for the canonical projection p : G + G / / K . (2) Let (V,( . ,.)) be a finite-dimensional Euclidean vector space and K c O ( V ) a compact subgroup of the orthogonal group of V. For
252 p E Mb(V), denote the image measure of p under
k
E
K by k(p).
Then the space of K-invariant measures
M F ( v ) := { p E M b ( v ) : k ( p ) = p for all k E K } is a Banach-*-subalgebra of M b ( v ) (with the group convolution) with identity 60. The space V K := {K.z : z E V } of all K-orbits in V is again locally compact, and the canonical projection p : V + V K induces a probability preserving, isometric isomorphism p : M F ( V ) 3 M b ( V K ) of B a n d spaces and an associated orbit hypergroup structure ( V K *) , such that p becomes an isomorphism of Banach-*-algebras. The involution on V K is given by 5 = -K.z. Moreover, the continuous functions
are multiplicative on ( V K *) , for X E V I ,the complexification of V , a p if and only if K.X = K.p. It is known (see 6, that and ax V K = {ax : X E V } . h
BY
20,
positive semicharacters lead to deformed convolutions:
Proposition 2.1. Let a0 E X * be a positive semicharacter o n the commutative hypergroup ( X ,*), i.e., ao(z)> 0 for z E X . Then p
v = ao((a,'p)
* (a,lv))
( Av E
M
C
W
)
(2.2)
extends uniquely to a bilinear, associative, probability preserving, weakly continuous convolution 0 o n M b ( X ) , and (x,0 ) becomes a commutative hypergroup with the identity and involution of ( X ,*). ( X ,0 ) will be called deformation of ( X , * ) w.r.t. ao. Eq.(2.2) shows that p I+ aop is an algebra isomorphism between ( M ,( X ) ,*) and ( M c( X ), a ) which for unbounded a0 cannot be extended to M b ( x ) ; 6.Section 3. Many data of ( X ,0 ) can be expressed in terms of a0 and corresponding data of ( X ,*). For instance, if w is a Haar measure of ( X ,*), then a i w is a Haar measure of (X,.). Moreover, the mapping Ma0 : a I+ a/ao is a homeomorphism between ( X ,*)* and ( X ,.)*, and also between x ( X , *) and x ( X , 0 ) ; see 2o and 1 8 .
Remark 2.1. Deformation is transitive as follows: Let (X, 0 ) be the deformation of ( K ,*) w.r.t. ao, and let PO be a positive semicharacter on ( K ,0 ) .
253
Consider further the deformation ( K ,0 ) of ( K ,0 ) w.r.t. PO. The function a& is a positive semicharacter on ( K ,*), and ( K ,0 ) is the deformation of (K,*)w.r.t. a&,. For PO = l/ao, one obtains o = *. We next present some examples; for further examples see Section 4.
Example 2.2. Let (V,( . , . )) be an n-dimensional Euclidean vector space, K a compact subgroup of the orthogonal group O ( V ) , and ( V K , * )the associated orbit hypergroup. Fix p E V with -p E K.p, and consider the function e p ( z ) := e ( p j z ) on V and
M ! ~ ~ ( v:=) {e,p
:p E
M,(v) K-invariant}.
The multiplicativity of ep on V yields that w.r.t. the group convolution on M,(V), we have e,p * epu = e,(p * u ) . Hence, M,",K(V)is a subalgebra of Mb(V),and its norm-closure
Ma"'K(V):= M p ( V ) a Banach subalgebra. On the other hand, ao(K.z) :=
s,
e,(k.z)dk (z E V )
is a positive semicharacter on ( V K *); , see Example 2.1(2) above as well as Proposition 2.8 of 18. Proposition 2.8 of l8 also states that for the deformation ( V K , o ) of ( V K , * )w.r.t. ao, the canonical projection p : V -+ V K induces a probability preserving isometric isomorphism of Banach algebras from M l Z K ( V )onto M b ( V K , o ) . In other words, the deformed hypergroup algebra may be regarded as Banach algebra of (not longer Kbiinvariant) measures on V .
Example 2.3. It is well-known that the double coset hypergroup SL(2,C ) / / S V ( 2 ) and the orbit hypergroup (IR3)s0(3) may be identified with [0,XI[, and that the associated hypergroup structures on [0, m[ are deformations of each other; see I , 18, or 20. Here is the higher rank extension of this example:
Example 2.4. Let G be a complex, noncompact, connected semisimple Lie group with finite center and K a maximal compact subgroup. Consider the Cartan decomposition g = e p of the Lie algebra of G, and choose a maximal abelian subalgebra a c p. K acts on p via the adjoint representation as a group of orthogonal transformations w.r.t. the Killing form ( . , .) as scalar product. Let W be the Weyl group of K , which acts on a as finite reflection group; here and further on we identify a with its dual a* via the
+
254
Killing form. Fix some Weyl chamber a+ C a and the associated set C+ of positive roots. Then the closed chamber C := Q+ is a fundamental domain for the action of W on a, and C can be identified with the orbit hypergroup ( p K , *), where a K-orbit in p corresponds to its representative in C. C can also be identified with the commutative double coset hypergroup G / / K where x € C corresponds t o the double coset K(e")K. Denote the corresponding convolution by 0 . Using the known formulas for the spherical functions on G / / K and pK (see Helgason 4), we proved in l 8 that ( G / / K ,0 ) = (C,0 ) is the deformation of the orbit hypergroup ( p K , * ) = (C,*) w.r.t. the positive semicharacter a-ip (in the sense of Example 2.1(2)) with P := Cae.E+ a E a+ *
As - p E K.p, the construction in Example 2.2 shows that Mb(G1I.K) may be embedded into Mb(p) in an isometric, probability preserving way. Here are the most prominent examples (c.f. Appendix C of 9 ) . (1) T h e Ad-1-case. K = SU(d) is a maximal compact subgroup of G = SL(d,C). In the Cartan decomposition g = t p we obtain p as the additive group H: of all Hermitian d x d-matrices with trace 0, on which SU(d) acts by conjugation. Moreover, a consists of all real diagonal matrices with trace 0 and will be identified with
+
{2= ( x 1 , . . . , x d )
E
xd : cxi =o} i
on which the Weyl group acts as the symmetric group take
sd.
We thus
C : = { x = ( Z 1 , . . . , x d ) E R d x: l 2 X z > . . . > % d r c x i = o } . i
+
+
Then in particular, p = (d - 1 , d - 3 , . . .,-d 3, -d 1). (2) T h e Bd-case. For d 1 2 consider G = S 0 ( 2 d + 1,C) with maximal compact subgroup K = S0(2d 1).Here a may be identified with Rd, and we may choose
+
c={ x E Rd : Z 1 > 2'2 > > x d 2 0 ) * * *
with the Weyl group W N s d K Z,! and p = (2d - 1,2d - 3 , . . . ,1). ( 3 ) T h e Cd-case. For d 2 3 let G = Sp(d,C) with the maximal compact subgroup K = Sp(2d 1). Here, a = Rd with C and W as in the Bd-case. We have p = (2d, 2d - l , , .. ,2). The preceding results on hypergroup deformations imply that the hypergroups
+
255
+
+
+
S p ( d , C ) / / S p ( 2 d 1 ) and S 0 ( 2 d l , C ) / / S 0 ( 2 d 1) are (UP t o isomorphism) deformations of each other; see also . (4) The Dd-case. For d 2 4 let G = S 0 ( 2 d , C) with maximal compact subgroup K = S O ( 2 d ) . In this case a = Rd and we may take
c = {z E Rd : 2 1 2 2 2 2 '
'
2 Zd-1 1 I.dl}
with p = (2d - 2,2d - 4,.. . ,2,0).
3. Deformation of convolution semigroups We now always assume that a0 is a positive semicharacter on a a-compact, second countable commutative hypergroup ( X ,*) and that ( X ,0 ) is the associated deformation. We show how under a natural growth condition, convolution semigroups on ( X ,*) can be deformed into convolution semigroups on ( X ,.). To describe this condition, we introduce the spaces
M2;(x) M:o(x)
Mbl+(X) : aop E Mb,+(X)}, := M ' ( X ) n M2b+(X) := { p E
as well as the transformation
Lemma 3.1. Let p, v E M b ! + ( X ) .'Then p * v E M::(X) if and only if p, u E ( X ) . Moreover, i f one of these conditions holds then
Ra0(P * v ) = Rcq ( P ) RaO ( v ) Proof. If p, u have compact support, then the lemma is clear by Eq. (2.2). In the general case, choose compacta (K,),>l - in X with X = U,K, and K,+1 3 K, for n E N. Put pn := p l ~ and , v, := ~ I K , . As the p, * u, have compact support, we have
Monotone convergence implies that
where one term is finite if and only if so is the other one. This proves the first part of the lemma. Moreover, if these terms are finite, then the same monotone convergence argument shows that for all f E cb(x)with f 2 0,
256
This implies R,, ( p * Y) = R,, ( p ) 0 R,,
(Y).
Remark 3.1. Notice that the mapping R,, : M A o ( X ) + M 1 ( X ) is not (weakly or vaguely) continuous whenever QO is unbounded. In fact, choose (zn)n21C X with ao(z,) + 00 and a0(zn)2 1. Then the measures pn := (1- ( Y O ( Z ~ ) - ~ ) ~~, o ( z ~ ) - ' 6 ~ ,tend , to 6, while
+
does not tend t o 6, = ROO(6,). We now investigate convolution semigroups.
Definition 3.1. A family (pt)t>o - C M1(X) is called a convolution semigroup on ( X , * ) , if po = 6,, if p8+t = p8 * pt for s , t 2 0, and if the mapping [0, 00[+ M1( X ) , t I+ pt is weakly continuous. It is well-known (see Rentzsch 12) that each convolution semigroup (pt)t>O - admits a LCvy measure r] E M + ( X \ {e}) which is characterized by 1
1
f dpt
for
f E C c ( X ) with e
# supp f.
(pt)tlo is called Gaussian, if r] = 0 which is equivalent to saying that for all neighborhoods U of e E X , 1imt-o i p t ( X \ U )= 0.
We next study under which conditions convolution semigroups on (X, *) can be deformed w.r.t. QO. We here need the following condition on (YO.
Definition 3.2. A positive semicharacter a0 on ( X ,*) is called exponential if there exists a neighborhood U of e E X and a constant C > 0 such that for all z, y E X with y E z * U ,ao(y)/ao(z) 5 C. We conjecture that positive semicharacters are always exponential. Unfortunately we are not able t o prove this. However, we present at least some criteria and examples in Section 4 below. The following theorem is motivated by 5 , 19, where a variant for the group case is studied.
Theorem 3.1. Let a0 be an exponential positive semicharacter on (X, *) with a0 2 1. Then the following statements are equivalent for a convolution semigroup (&>o - on ( X ,*) with Lkwy measure r]. (1) pt E M A o ( X )holds for some t > 0. (2) pt E MA,(X) holds for all t 2 0, the mapping cp : [O,oo[+]O,oo[ given by q(t) = J a 0 dpt is continuous and multiplicative, and (R,,(pt))tLo is a convolution semigroup on ( X ,0 ) .
257
(3) For any neighborhood U of e E X , s x , r r adq ~ < 00.
If one and hence all of these statements hold, then aoq is the L&vy measure of the convolution semigroup (Rao(pt))t>O o n ( X ,0 ) . In particular, Gaussian semigroups on ( X , * ) always lead t o Gaussian semigroups on ( X ,0 ) .
+
Proof. (1) (2): Lemma 3.1 implies that cp 2 1 is well-defined and multiplicative. To check continuity, we observe that the multiplicativity implies that for N E N and 0 5 s 5 1/N, c p ( ~ > ~ c p( l s N ) = cp(1) and hence cp(s) 5 cp(l)llN+ 1 for N + 00. Therefore, cp is continuous at t = 0 and hence, as a multiplicative function, on [0,00[. Using Lemma 3.1 and the fact that the mapping [0,00[+ M 1 ( X ) ,t C ) Ra,(pt) = p(t)-'aOp is vaguely and hence weakly continuous, we conclude that (Rao(pt))t>o is a convolution semigroup on ( X ,0 ) . (3): The measure p := l{ao>z}qE M b l + ( X )is the Levy mea(2) sure of the Poisson semigroup (vt := e-IIPllt exp(tp))t>O,e x p denoting the exponential function on the Banach algebra ( M b ( X ) ,*,.' Moreover, it is easy t o see that q - p is the Ldvy measure of a further convolution semigroup (ijt)t>o with pt = ut * fit for t 2 0. Lemma 3.1 shows that vt E M k o ( X )for t 2 0. As obviously p 5 (ellPllt/t)vt for t > 0, we obtain p E M:;(X) and thus (3). Furthermore, for f E C c ( X )with e $! supp f ,
-
Hence, aoq is the Levy measure of the semigroup (Rao(pt))t>O on ( X ,0)n The proof of (3) =+ (1) is more involved. Recapitulate that for a convolution semigroup (pt)t>o - on ( X ,*), the translation operators Tt(f ) := p; * f (t 2 0 ) form a strongly continuous, positive contraction semigroup on L 1 ( X , w ) ,w being the Haar measure of ( X , * ) ; see [BH]. Let A be its infinitesimal generator with the dense domain DA C L 1 ( X , w ) . We have:
Lemma 3.2. Let a0 be a positive semicharacter and (pt)t>o - a convolution ~ q satisfies Stao2z}a0 dq < 00. semigroup o n ( X ,*) whose L & Jmeasure Then f o r each neighborhood U of e E X there exists f E C c ( X )n DA with f 2 0 , f = f * # O , s u p p f c U ,a n d J l A f l a o d w < o o . Proof. Let U be a compact neighborhood of e E X with U - = U . Then by 12, there exists f E D A with f dw = 1, f 2 0, f = f*, and supp f c U.
sx
258
Let x # U*U and y E U.Then f (x*y) = 0, which means that the translate fz given by fi(y) := f (x * y) satisfies fi = 0 on U ,and hence 1
A f (4= t+O lim -t( P t
s
* f d e ) - f z ( e ) >= f (x * Y) drl(Y).
Consequently, by Fubini’s theorem,
lAfl-aodw
+
/
x\u*u
lAfl.aodw
Now
(1) in the theorem now follows from Lemma 3.2 and the follow(3) ing result.
Lemma 3.3. Let a0 be an exponential positive semicharacter with a. 2 1, and (pt)t>o a convolution semigroup on ( X , *) with generator A. Assume that for each neighborhood U of e E X there exists f E C c ( X )n D A with f 2 0 , f = f*#O,suppf c U a n d J l A f l a o d w < o o . T h e n f o r a l l t 2 0 , JaodClt < 00. Proof. Let U be a neighborhood of e E X and C1 > 0 a constant with clao(x) 5 ao(z) for x E X and z E U*x. Let f E c c ( X ) n D A with f >_ 0, f = f * # 0 , SUM, f c U and [Aflaodw < 00. Then for all m E N,the functions am:= a0 A rn E C I ( X )also satisfy clam(%) I a,(z) for x E X , z E U * x. Hence, there is a constant C2 > 0 depending on f such that for all m E N a n d x E X ,
s
am(x) I c 2
/
am(x * Y)f (Y)
WY)
= c2 . a m
* f (.I.
(3.1)
Moreover, as a. 2 1, we have for all m E N and x, y E X,
a m ( x * y) I m A ao(x * Y) = m A (ao(x)ao(y)) I am(x)am(y). (3.2)
259
Define hm(t):= J(pt * f ) .am dw = J a m * f d p t . As f E DA and Af E L1 ( X ,w ) holds, we obtain $pt * f = pt * Af and hence
Therefore, by (3.2) and (3.1),
This yields h m ( t ) I hm(0) etc for t 2 0 and some constant C 2 0 independent of rn. Hence, again by (3.1),
for all m E N. This yields the claim
J(YO
dpt
< 00 for t 2 0.
Notice that the growth condition on a0 was needed above only for the preceding lemma. Theorem 3.1 therefore admits the following variant. Theorem 3.2. Let (YO be a positive semicharacter and (pt)t>o c M 1 ( X ) a Poisson semigroup o n ( X ,*), which means that pt = e-tllPl&xp(tp) f o r all t 2 0 and some p E M b * + ( X ) .Then p is the L6vy measure of (pt)t>o, - and the statements (1)-(3) of Theorem 3.1 are equivalent.
Proof. It suffices t o check (3) =+- (1). However, if R := Jcto d p < 00, then for all n 2 0 , J (YO dp(") = R" and hence J (YO dpt < 00 for all t 2 0.n Remark 3.2. Let a0 be an exponential positive semicharacter and (pt)t>o C MAo( X ) a convolution semigroup on ( X ,*). Then the convolution operators (Tt)t>Oon C o ( X )with Tt f := pt * f form a Feller semigroup. Its generator A with 1 (2 E x,f E D ( 4 ) A m ) = t+O lim -(p; * f(.) - f (.)I t admits a II.Il,-dense domain D ( A ) in C o ( X ) ; see 13. Now consider the generator Ano of the Feller semigroup on C o ( X ) which is associated with the renormalized convolution semigroup (Rno(pt))tro on (X, 0 ) . Using the notation above, we have
260
Theorem 3.1(2) shows that cp(t)= ect for some c E E.,and 1 lim -(l/cp(t)- 1) = -c. t+O
t
Hence
Therefore
A"" = M l/ao O A O Mao - c
(3.3)
at least on D(AQo)n C c ( X ) ,where M , denotes the multiplication operator with g E C(K). The same holds for other function spaces like LP(X,w). 4. Exponential positive semicharacters
It seems reasonable t o conjecture that positive semicharacters are always exponential. Unfortunately we are not able to prove this. Here are, at least, some criteria and several examples: Lemma 4.1. (1) If ( X ,*) is discrete, then (YO is always exponential. (2) Let ( Y O , ( ~ 1be exponential positive semicharacters o n ( X ,*), and let (X,.) be the deformation of ( X , * ) w.r.t. ( Y O . Then (YI/(YO is an
exponential positive semicharacter o n ( X ,0 ) . Proof. Part (1) is clear by taking U = { e } . For the proof of (2) choose U1 of e and constants CO, C1 associated with (YO, a1 neighborhoods UO, respectively. For U := UO n U1 n U; n U; and C := CoC1, we obtain that for x , y E X with y E x * U , we have x E y * U and thus ( Y O ( X ) W ( y ) / ( a o ( y ) a l(.) I C as claimed. Example 4.1. In 2 3 , Zeuner presented quite general, but technical conditions on a function A E C([O,co[)n C'(]O,co[) with A ( x ) > 0 for x 2 0 which ensures that there exists a unique commutative hypergroup ([0,oo[,*)
261
whose semicharacters are precisely the eigenfunctionsof the Sturm-Liouville operator
LAf := -f”
- (A’/A)f’
with initial conditions f(0) = 1 and f‘(0) = 0; see also Section 3.5 of This hypergroup is called the Sturm-Liouville hypergroup associated with A. Moreover, to the knowledge of the authors, all known hypergroup structures on [O,m[ appear in this way (up to isomorphism); see also for details. We here mention that Zeuner’s approach in particular includes all Chebli-”kimeche hypergroups and thus all double coset hypergroups associated with noncompact symmetric spaces of rank one. We claim that all positive semicharacters on a Sturm-Liouville hypergroup on [0,m[ with A satisfying Zeuner’s conditions are exponential. To prove this, recall from Section 3.5 in that Zeuner’s conditions imply that 1 p := - lim A’(x)/A(x)2 0 2 x--fw exists, and that the positive semicharacters are precisely the unique solutions cpix of
with X 2 0. Moreover, the renormalization ([O,m[,o) of ([0, co[,*) w.r.t. pix is again a Sturm-Liouville hypergroup associated with the renormalized function Ax := ‘p% A where Ax again satisfies Zeuner’s conditions; see Section 3.5.51 of Applying (4.1) to A as well as to Ax, we see that lim,,,&(x)/~pix(x) exists. As supp(b,*b,) c [ ) x - y l , x + y ] for x , y 2 0, it follows from the mean-value theorem that cpix is exponential.
’.
Example 4.2. Let V be a finite-dimensional Euclidean vector space, K C O ( V ) a compact subgroup, and V K the associated orbit hypergroup as in Example 2.1(2). Then, for each p E V , the positive semicharacter ajPwith CX~,(K.Z) = ] K e - ( ~ + x ) d k( x E V) is exponential. In fact, we may take U := {K.x : x E V, 11x112 6 1) c V K as a neighborhood of the identity. For orbits K.x, K.y E V K with K.x E U * K.y we then have representatives X , y E V with llx - y112 6 1 which implies that e - - ( P l k . x ) 5 e-(PrL.u)ellPllz for k E K and thus ai,(K.x) 5 aiP(K.y)ellPllzas claimed. Example 4.3. Let G be a (not necessarily complex) noncompact, connected semisimple Lie group with finite center and K a maximal compact subgroup. Let G = NAK and g = n + a t be the corresponding Iwasawa decompositions. For g E G let A(g) E a be the unique element with
+
262
g E N e z p ( A ( g ) ) K . Let C+ be the set of positive roots (for the order m,a the half sum of positive roots corresponding t o n), and p = CaEC+ with m, as multiplicity of a. Then, by a formula of Harish-Chandra (see Theorem IV.4.3 of 4), the spherical functions on G, i.e., the multiplicative functions on G I I K , are given by
where X runs through w, the complexification of a. Clearly, the cpx for X E i . a are positive multiplicative functions. These functions are also exponential. To prove this, we conclude from Lemma IV.4.4 of that
Hence, for each comact neighborhood U of e there is a constant C that cpx(g-'h) 5 Ccpx(h)for all g E U ,h E G and X E i . a.
> 0 such
Example 4.4. Let R be a (reduced, not necessaryly crystallographic) root system in R" with the standard inner product (. , .), i.e. R c Rn \ (0) is finite with R n lwru = {&a}and ua(R)= R for all a E R, where u, is the reflection in the hyperplane perpendicular to a. Assume also without loss of generality for our considerations that (a,a ) = 2 for all cr E R. Let W be the finite reflection group generated by the u, and let k : R + [0, CG[ be a fixed multiplicity function on R, i.e. a function which is constant on the orbits under the action of W . The (so-called rational) Dunkl operators attached t o G and k are defined by
Here at denotes the derivative in direction 5 and R+ is some fixed positive subsystem of R. The definition is independent of the special choice of R+, due t o the G-invariance of k. As first shown in ', the Tc(k), 5 E Rn generate a commutative algebra of differential-reflection operators. This is the foundation for rich analytic structures related with them. In particular, there exists a counterpart of the exponential function, the Dunkl kernel, and an analogue of the Euclidean Fourier transform with respect to this kernel. The Dunkl kernel Ek is holomorphic on a?' x a?' and symmetric in its arguments. Similar to spherical functions on symmetric spaces, the function Ek ( . ,y) with fixed y E cc" may be characterized as unique analytic solution of the joint eigenvalue problem
T . ( k ) f= ( 5 1 Y ) f
for all
5 E c,f(0) = 1;
(4.3)
263
c.f. ll. Apart from the trivial case k = 0 with Ek(z,y) = e(zJ),EI, is explicitly known in a few cases only like n = 1; see l7 for a survey. The G-invariant counterpart of Ek is the generalized Bessel function
which is G-invariant in 2,y and naturally considered on the closed positive Weyl chamber C associated with R+. For n = 1, J k is a usual Bessel function. Moreover, in the cristallographic case and for certains half-integer multiplicities, the Jk are the multiplicative functions of certain Euclidean orbit hypergroups as in Example 2.1. Here, and for n = 1, the Jk(2,y) (y E P) therefore form the multiplicative functions of some commutative hypergroup on C. It is conjectured that there exist such commutative hypergroups on C for all root systems and multiplicities k 2 0. Only part of this conjecture has been verified up to now in 16. Now fix a root system R and k 1 0 such that the Jk(.,y) (y E P) are the multiplicative functions of a commutative hypergroup (C, *). To find positive semicharacters, we employ the following psoitive integral representation for Ek (and thus Jk): For given R,k 2 0, and s E Rn there exists a unique pribability measure pa on Rn such that
~ k ( z , y= ) / e + > v ) dpx(z)
for y E C .
(4.4)
Moreover, supppz c { z E Rn : llzll2 5 11x112). Thus, for each y E Rn, Jk(., y) is a positive semicharacter on (C, *). We claim that these semicharacters are exponential. To show this, let U := { z E C : llzll2 5 1) and ~ , Z ZE C with 2 1 E U * x 2 , We conclude from Theorem 4.1 of l6 that then x1 E C n n w E W { zE Rn : 1 - 20.221 5 1) holds. As llz - 'wII llz - 20.211 for all 2,z E C and w E W by Ch. 3 of 3, we even have 1\21- 2211 1. In the same way as in Example 4.3 we now obtain from Eq. (4.4) that Jk(z1,y) 5 ellvllJk(z2,y) which proves that Jk(.,y) is exponential for each y E Rn.
<
<
References 1. W.Bloom, H.Heyer, Harmonic Analysis of Probability Measures on Hypergroups. De Gruyter-Verlag, Berlin, 1994. 2. C.F. Dunk], Differential-difference operators associated to reflection groups, Trans. Amer. Math. SOC. 311 (1989), 167-183. 3. L.C. Grove, C.T.Benson, Finite Reflection Groups. Springer-Verlag, 2nd ed., 1985.
264 4. S. Helgason, Groups and Geometric Analysis. American Mathematical Society, 2000. 5. A. Hulanicki, A class of convolution semi-groups of measures o n a Lie group. In: Probab. Theory on Vector Spaces 11, Proceedings. Lecture Notes in Mathematics Vol. 828, Springer 1980, pp. 82-101. 6. R.I. Jewett, Spaces with a n abstract convolution of measures. Adv. Math. 18 (1975), 1-101. 7. A. Klyachko, Stable bundles, representation theory, and H e m i t i a n operators, Selecta Math. (new Series) 4 (1998), 419445. 8. A. Klyachko, Random walks o n symmetric spaces and inequalities f o r matrix spectra, Linear Algebra Appl. 319 (ZOOO), 37-59. 9. A.W. Knapp, Lie Groups beyond an Introduction. Birkhauser, Boston, 1996. 10. A. Knutson, T. Tao, T h e honeycomb model of GLn(@)tensor product I. Proof of the saturation conjecture, J. Amer. Math. SOC.12 (1999), 10551090. 11. E.M. Opdam, Dunkl operators, Bessel functions and the discriminant of a finite Coxeter group, Compos. Math. 85 (1993), 333-373. 12. C. Rentzsch, A Ldvy-Khintchine type representation of convolution semigroups o n commutative hypergroups. Probab. Math. Stat. 18 (1998), 185198. 13. C. Rentzsch, M. Voit, Ldvy processes o n commutative hypergroups. Contemp. Math. 261 (ZOOO), 83-105. 14. L.C.G. Rogers, D.Williams, Difisions, Markov Processes, and Martingales, Vol. II. Wiley: Chichester - New York, 1987. 15. M. Rosler, Positivity of Dunkl’s intertwining operator. Duke Math. J. 98 (1999), 445-463. 16. M. Rosler, A positive radial product formula f o r the Dunkl kernel. Trans. Amer. Math. SOC.355 (2003), 2413-2438. 17. M. Rosler, Dunkl operators: Theory and Applications. Lecture Notes for the SIAM Euro Summer School: Orthogonal Polynomials and Special F’unctions. (Leuven, 2002). In: Springer Lecture Notes in Mathematics, Vol. 1817, 2003, pp. 93 - 136. 18. M. Rosler, M. Voit, SU(d)-biinvariant random walks o n SL(d,C) and their Euclidean counterparts, preprint 2003. 19. E. Siebert, Continuous convolution semigroups integrating a submultiplicntive function, Man. Math. 37 (1982), 383-391. 20. M. Voit, Positive characters on commutative hypergroups and some applications, Math. Z. 198 (1988), 405-421. 21. M. Voit, A Girsanov-type formula for Ldvy processes o n commutative hypergroups. In: H. Heyer et al. (eds.), Infinite Dimensional Harmonic Analysis, Proc. Conf., Grabner 2000, pp. 346-359. 22. H.Zeuner, T h e central limit theorem f o r Chdbli-lkimeche hypergroups. J. Theor. Probab. 2 (1989), 51-63. 23. H. Zeuner, M o m e n t functions and laws of large numbers o n hypergroups. Math. Z . 211 (1992), 369-407.
Infinite Dimensional Harmonic Analysis I11 (pp. 265-276) Eds. H. Heyer et ol. 0 2005 World Scientific Publishing Co.
AN INFINITE DIMENSIONAL LAPLACIAN ACTING ON MULTIPLE WIENER INTEGRALS BY SOME LEVY PROCESSES KIMIAKI SAITO Department of Mathematics Meijo University Nagoya 468-8508, Japan E-mail: ksaito @ccmfs.meijo-u.ac.jp
In this paper we shall discuss the Ldvy Laplacian as an operator acting on multiple Wiener integrals by some LBvy processes. This space includes regular functionals in terms of Gaussian white noise and it is large enough to discuss the stochastic process. From Cauchy processes an infinite dimensional stochastic process is constructed, of which the generator is the Ldvy Laplacian.
Introduction An infinite dimensional Laplacian was introduced by P. LCvy in his famous book [17]. Since then this exotic Laplacian has been studied by many authors from various aspects see [1-6,18,20,23]and references cited therein. In this paper, generalizing the methods developed in the former works [16,19,24,28,29], we construct a new domain of the LCvy Laplacian acting on some class of multiple Wiener integrals by some LBvy processes and associated infinite dimensional stochastic processes. This paper is organized as follows. In Section 1 we summarize basic elements of white noise theory based on a stochastic process given as a difference of two independent LCvy processes. In Section 2, following the recent works Kuo-Obata-Saitb [16], Obata-Saitb [24],Saitb [30] and SaitbTsoi [31], we formulate the LBvy Laplacian acting on a Hilbert space consisting of some LCvy white noise functionals and give an equi-continuous semigroup of .class (CO) generated by the Laplacian. This situation is further generalized in Section 3 by means of a direct integral of Hilbert spaces. The space is enough to discuss the stochastic process generated by the LCvy Laplacian. It also includes regular functionals (in the Gaussian sense) as a harmonic functions in terms of the LCvy Laplacian. In Section 4, based on infinitely many Cauchy processes, we give an infinite dimensional stochastic process generated by the LCvy Laplacian.
266
1
Functionals of Gaussian noise and Poisson noise
Let E = <S(R) be the Schwartz space of rapidly decreasing R-valued functions on R. There exists an orthonormal basis {ev}v>o of I<2(R) contained in E such that Aev = 1(v + l)ev,
v = 0,1, 2 , . . . ,
A=
For p 6 R define a norm | • \p by \f\p = \Apf\L2(R) tor f € E and let Ep be the completion of E with respect to the norm | • \p. Then Ep becomes a real separable Hilbert space with the norm | • \p and the dual space E' is identified with E-p by extending the inner product (-, •) of I<2(R) to a bilinear form on -p x Ep. It is known that E = proj lim Ep, p-KX>
E* = indlim E-p. P-»°°
The canonical bilinear form on E* x E is also denoted by {•, •}. We denote the complexifications of L2(R), E and Ep by L^R), EC and EC,P, respectively. Let {LpX(t)}t>0 and {L^x(t)}t>o be independent Levy processes of which the characteristic functions are given by E[eizLi-*w] = ethW , t>Q, j = 1,2, h(z) = imz - y z2 + (eixz - 1), where m 6 R, 0 and A e R. Set A ff) x(t) = L\^(t] - L2^x(t) for all t > 0. Then we have = exp {-ta2z2 + t(eaz + e~ixz - 2)} , t > 0. Set (7(0 = exp {^(fctfiOO) + h(-^(u)))du} , f = (^1,6) £ E x E. Then by the Bochner-Minlos Theorem, there exists a probability measure fifft\ on E* x E* such that
where (z,£) = (xi,£i) + (xi,£2), x = (zi,z 2 ) £E*xE*,£= (6,6) £ExE. Let (L2)ff>x = L2(E* x E*,^^) be the Hilbert space of C-valued squareintegrable functions on E* x E* with L2-norm || • H^ with respect to /ZO-,A-
267
The Wiener-It6 decomposition theorem says that:
c 00
(L2)u,x =
(1.1)
@&,
n=O
where Hn is the space of multiple Wiener integrals of order n E N and C. According to (1.1) each cp E (L2),,xis represented as
c
HO=
00
cp =
In(fn),
fn
f2L”cR)%
n=O
where L&(R)&)” denotes the n-fold symmetric tensor power of L&(R)(in the sense of a Hilbert space). An element of (L2),,x is called a white noise functional. We denote by ((-,-)),,A the inner product on (L2),,x. Then, for cp and $ E (L2),,0 we have
n=O
n=O
n=O
where the canonical bilinear form on J ~ & ( R x) @ L&(R)@n ~ is denoted also by (-7
9).
For cp and $ E (L2)o,xwe also have 00
00
is defined by
The U-transform of
Theorem 1.1 [26] (see also [9,14,22]) Let F be a complex-valued function defined o n E x E . Then F is a U-transform of some white noise functional in (L2),,x if and only if there exists a complex-valued function G defined o n E c x E c such that 1) for any ( and q in E c x E c , the function G(z6 + q ) is a n entire function O f Z E C ,
2) there exist nonnegative constants K and a such that IG(J)I I: Kexp
+
3) F(<)= G(ia2& ia2&
[alrl;] , K E E c
x Ec,
+ X(eixcl - e-ixc2)) for all < = ( ( I , & )
E E x E.
268
2
The LQvy Laplacian acting on multiple Wiener integrals
Consider F = U'p with 'p E (L2),,x. By Theorem 1.1, for any <,T,I E E x E the functions z H F ( < zq) admits the Taylor series expansions:
+
where F ( n ) ( [ ): E x x E + C is a continuous n-linear functional. Fixing a finite interval T of R,we take an orthonormal basis {(n},"=o c E x E for L2(T)which is equally dense and uniformly bounded (see e.g. [14,15]). Let DL denote the set of all 'p E (L2),,x such that the limit
c
N-1
L L ( U V ) ( t )=
(U'p)%)(Cn, G),
n=O
exists for any E E x E and A L ( U ' ~is) in U[(L2),,x].The LCvy Laplacian AL is defined by
ALP = U - ~ K L U V ,
'p
E DL.
Given u 2 0,X E R, n E N and f E LE(R)6n, we consider 'p E (L2),,x of the form: V=/,n
The U-transform U p of
UP(<)=
J
f(u1,. 'p
--
(2.1)
r.Iln)dA,,x(ul)...dA,,x(un).
is given by n
f(u1,. '. ,un)
T"
JJ %x(s)(uj)dul-
*
.dun,
s E E x E.
j=1
+
where E , , ~ ( ( ) ( u j )= iu2<1(uj) iu2&(uj)+ X(eixcl(uj)- e - i X c a ( u j ) ) . For any u 2 0, X E R and n E N let E , , A , denote ~ the space of cp which admits an and suppf C Tn.Set expression as in (2.1), where f belongs to L&(R)6n Eo,x,O = C for any u 2 0,X E R. Then E , , A ,is ~ a closed linear subspace of (L2),,x.Using a similar method as in [31], we get the following
Theorem 2.1 [31] (see also [16,29]) For each u 2 0, n E N and X E R the L6vy Laplacian AL becomes a scalar operator o n E,,o,,, U E o , ~such , ~ that AL'p = 0,
'p
E Eu,o,n;
(2.2)
269
Proposition 2.2 [31] Let X E R be fixed. Consider two white noise functions of the form: m
m
n=O
n=O
If cp = 1c, in (L2)0,x,then cpn = lc,n for all n E N U (0). Taking (2.2) and (2.3)into account, we put
~ admits an For N E N and X E R let D Y be the space of cp E ( L 2 ) o ,which expression
n=l
such that
c 00
IIIPl112N,o,x =
amllcpnll:,x <
(2.4)
n=l
By the Schwartz inequality we see that D$X is a subspace of (L2)0,x and becomes a Hilbert space equipped with the new norm 111 IIIN, o , x defined in (2.4). Moreover, in view of the inclusion relations:
-
( L 2 ) o , x3 DY*x3 . . - 3 D$X 3 D$:l we define
DZX= projlimDi OX = N-tw
3
... ,
n DY. 00
N=l
Note that for any X E R we have 00
(JEo,x,n c DZXc (L2)o,x.
n=l
Then A L becomes a continuous linear operator defined on D$l: satisfying
I~~AL(PIIIN,o,A I ~ ~ ~ ~ ~ ~ ~ N +'pl ,EODZx, ,X~ N E N O Summing up, we have the following
into D$X (2.5)
270
Theorem 2.3 [16,31] The operator AL is a self-adjoint operator densely defined in D Y f o r each N E N and X E R. It follows from (2.5)that AL is a continuous linear operator on DSx. In view of the action of (2.3), for each t 2 0 and X E R we consider an operator G; on DZx defined by
n=l
n=l
We also define G: on (L2)u,o.
(L2),,0
as an identity operator I by Icp = cp, cp E
Theorem 2.4 [16,30] Let X E R. T h e n the family of operators {G?; t 2 0) o n Dzx i s a n equi-continuous semigroup of class (CO) of which the infinitesimal generator i s A,. 3
The LCvy Laplacian acting on WNF-valued functions
Let &(A)
be a finite Bore1 measure on R satisfying
JRdYo<m. X4
Fix N E N. Let 9&be the space of (equivalent classes of) measurable vector functions cp = (cpx) with cpx = Cr=lp: E D>x for all X E R \ {0}, and cpo E ( L 2 ) , , 0 , such that
Then 9 ; becomes a Hilbert space with the norm given in (3.1). Let 9; be the space of (equivalent classes of) measurable vector functions cp = (cpx) with cpx = CrZlcp: E (L2)0,xfor all X E R\{O}, and cpo E (L2),,0, such that
Then 9: also becomes a Hilbert space with the norm given in (3.2).
271
Proposition 3.1 T h e m a p
i s a continuous linear m a p and a bijection f r o m DG into 9:. In view of the natural inclusion: D$,+l c 23% for N E N, which is obvious from construction, we define
The Levy Laplacian A, is defined on the space D& by ~ L c = p (AL‘P’),
cp = (9’)
E 9:.
Then AL is a continuous linear operator from D& into itself. Similarly, for t 2 0 we define
Gtcp = (G,x‘PX),
cp = (‘PA>E
Then by Theorem 2.4 we have the following:
Theorem 3.2 T h e family (Gt;t 2 0) is a n equi-continuous semigroup of class (Co)o n D& whose generator i s given by AL. 4
An infinite dimensional stochastic process associated with the LBvy Laplacian
For p E R let EF be the linear space of all functions X c) (A E Ep x Ep, X E R, which are strongly measurable. An element of EF is denoted by = ( ~ x ) x ~ R . Equipped with the metric given by
<
the space EF becomes a complete metric space. Similarly, let CR denote the linear space of all measurable function X !+ q,E C equipped with the metric defined by
272
Then CR is also a complete metric space. In view of dp 5 dq for p 1 q, we introduce the projective limit space ER = proj limp+m EF. The U-transform can be extended to a continuous linear operator on 9: by
up(<)= (U(pX(b))XER,
< = (6X)XER E ER,
for any p = ( ( p X ) x E ~E D&. The space U[DO,] is endowed with the topology induced from D$ by the U-transform. Then the U-transform becomes a homeomorphism from a&onto U[D&]. The transform U p of p E DL is a continuous operator from ER into CR. We denote the operator by the same notation Up. Let
et be an operator defined on U[D$] by 6t = UGtU-',
t 2 0.
Then by Theorem 3.3, {Gt; t 2 0) is an equi-continuous semigroup of class ((20)generated by the operator A;L. Let { X i } ,j = 1,2,3,4, be independent Cauchy processes with t running over [0, oo),of which the characteristic functions are given by
E[eiZx:]= e-'l2l, Take a smooth function r]T E E with
y;" =
{
z E R, r]T
j = 1,2,3,4.
= 1/ITI on T. Set
(xitqT, - x i t q T ) (x!,tr]T,-X?,tr]T),
if 2 0, otherwise.
Define an infinite dimensional stochastic process { Y t ;t 2 0) starting at ( = (<X)XER ER by
Yt = (b + KX)XER,
t 1 0.
Then this is an ER-valued stochastic process and we have the following
Theorem 4.1 If F is the U-transform of an element in D ,; G t W ) = E [ F ( Y t ) l Y o=
(1,
t 2 0.
we have (4.1)
273
PROOF.We first consider the case when F E U[Iz&]is given by
F(t) = (FX(<X))XER,
F o E U[(L2)u,0],
with f E L&(R)Qn.Then we have
Next let and for any
Then foralmost all is expressed in the following form:
Since Fo E U[(L2),,o]and F," E U[DzA],there exist cpo E (L2)u,oand c p i E DzXsuch that Fo = U[cpO]and F," = U[cpk]for v-almost all X and each n. By the Schwarz inequality, we see that
274
-
where ‘ptx = C(&,)-lei(’itx) for v-almost all X E R and each N E N. Therefore by the continuity of G;, X E R, we get that
Thus we obtain the assertion.
I
Acknowledgments This work was written based on a talk at the 3rd Japanese-German Symposium on “Infinite Dimensional Harminic Analysis” held from September 15th to September 20th 2003. This work was partially supported by JSPS grant 17540136. The author is grateful for the support. References
1. L. Accardi and V. Bogachev: The Ornstein-Uhlenbeck process associated with the L i v y Laplacian and its Dirichlet form, Frob. Math. Stat. 17 (1997), 95-114. 2. L. Accardi, P. Gibilisco and I. V. Volovich: Yang-Mills gauge fields as harmonic functions for the L i v y Laplacian, Russ. J. Math. Phys. 2 (1994), 235-250. 3. L. Accardi and 0. G. Smolyanov: Trace formulae for Levy-Gaussian measures and their application, in “Mathematical Approach to Fluctuations Vol. I1 (T. Hida, Ed.),” pp. 3 1 4 7 , World Scientific, 1995. 4. D. M. Chung, U. C. Ji and K. Sait8: Cauchy problems associated with the L i v y Laplacian in white noise analysis, Infin. Dimen. Anal. Quantum Probab. Rel. Top. 2 (1999), 131-153.
275
5. M. N. Feller: Infinite-dimensional elliptic equations and operators of Ldvy type, Russ. Math. Surveys 41 (1986), 119-170. 6. K. Hasegawa: Ldvy’s functional analysis in terms of an infinite dimensional Brownian motion I, Osaka J. Math. 19 (1982), 405-428. 7. T. Hida: “Analysis of Brownian Functionals,” Carleton Math. Lect. Notes, No. 13, Carleton University, Ottawa, 1975. 8. T. Hida: A role of the L i v y Laplacian in the causal calculus of generalized white noise functionals, in “Stochastic Processes,” pp. 131-139, SpringerVerlag, 1993. 9. T. Hida, H.-H. Kuo, J. Potthoff and L. Streit: “White Noise: An Infinite Dimensional Calculus,” Kluwer Academic, 1993. 10. T. Hida and K. SaitB: White noise analysis and the Ldvy Laplacian, in “Stochastic Processes in Physics and Engineering (S. Albeverio et al. Eds.),” pp. 177-184, 1988. 11. E. Hille and R. S. Phillips: “Functional Analysis and Semi-Groups,” AMS Colloq. Publ. Vol. 31, Amer. Math. SOC,1957. 12. I. Kubo and S. Taken& Calculus o n Gaussian white noise I-IV, Proc. Japan Acad. 5 6 A (1980) 376-380; 5 6 A (1980) 411416; 5 7 A (1981) 433436; 5 8 A (1982) 186-189. 13. H.-H. Kuo: O n Laplacian operators of generalized Brownian functionals, Lect. Notes in Math. Vol. 1203, pp. 119-128, Springer-Verlag, 1986. 14. H.-H. Kuo: “White Noise Distribution Theory,” CRC Press, 1996. 15. H.-H. Kuo, N. Obata and K. Sait6: Ldvy Laplacian of generalized functions o n a nuclear space, J. Funct. Anal. 94 (1990), 74-92. 16. H.-H. Kuo, N. Obata and K. SaitB: Diagonalization of the Ldvy Laplacian and related stable processes, Infin. Dimen. Anal. Quantum Probab. Rel. TOP.5 (2002), 317-331. 17. P. LBvy: “Legons d’Analyse Fonctionnelle,” Gauthier-Villars, Paris, 1922. 18. R. LBandre and I. A. Volovich: The stochastic Ldvy Laplacian and YangMills equation o n manifolds, Infin. Dimen. Anal. Quantum Probab. Rel. TOP.4 (2001) 161-172. 19. K. Nishi and K. Sait6: An infinite dimensional stochastic process and the Lkuy Laplacian acting o n WND-valued functions, t o appear in “Quantum Inofrmation and Complexity” World Scientific, 2004. 20. K. Nishi, K. SaitB and A. H. Tsoi: A stochastic eqression of a semigroup generated by the Ldvy Laplacian, in “Quantum Information I11 (T. Hida and K. SaitB, Eds.),” pp. 105-117, World Scientific, 2000. 21. N. Obata: A characterization of the Ldvy Laplacian in terms of infinite dimensional rotation groups, Nagoya Math. J. 118 (1990), 111-132. 22. N. Obata: “White Noise Calculus and Fock Space,” Lect. Notes in Math.
276
Vol. 1577, Springer-Verlag, 1994. 23. N. Obata: Quadratic quantum white noises and L i v y Laplacian, Nonlinear Analysis 47 (2001), 2437-2448. 24. N. Obata and K. Sait6: Cauchy processes and the Le‘vy Laplacian, Quantum Probability and White Noise Analysis 16 (2002), 360-373. 25. E. M. Polishchuk: “Continual Means and Boundary Value Problems in Function Spaces,” Birkhauser, Basel/Boston/Berlin, 1988. 26. J. Potthoff and L. Streit: A characterization of Hida distributions, J. F’unct. Anal. 101 (1991), 212-229. 27. K. Sait6: It6’s formula and Ldvy’s Laplacian I, Nagoya Math. J. 108 (1987), 67-76; 11, ibid. 123 (1991), 153-169. 28. K. Sait6: A (Co)-group generated by the Ldvy Laplacian 11, Infin. Dimen. Anal. Quantum Probab. Rel. Top. 1 (1998) 425-437. 29. K. Sait6: A stochastic process generated by the L i v y Laplacian, Acta Appl. Math. 63 (2000), 363-373. 30. K. Sait6: T h e Ldvy Laplacian and stable processes, Chaos, Solitons and Fractals 12 (2001), 2865-2872. 31. K. Sait6 and A. H. Tsoi: The L i v y Laplacian as a self-adjoint operator, in “Quantum Information (T. Hida and K. Sait6, Eds.),” pp. 159-171, World Scientific, 1999. 32. K. Sait6 and A. H. Tsoi: The Le‘vy Laplacian acting o n Poisson noise functionals, Infin. Dimen. Anal. Quantum Probab. Rel. Top. 2 (1999), 503-510. 33. K. Sait6 and A. H. Tsoi: Stochastic processes generated by functions of the Ldvy Laplacian, in “Quantum Information I1 (T. Hida and K. Sait6, Eds.),” pp. 183-194, World Scientific, 2000. 34. K. Yosida: “hnctional Analysis (3rd Edition) ,” Springer-Verlag, 1971.
Infinite Dimensional Harmonic Analysis I11 (pp. 277-287) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
LEVY PROCESSES ON DEFORMATIONS OF HOPF ALGEBRAS
MICHAEL SCHURMANN* Instatut fir Mathematik und Infomatik Universitat Greifswald Friedrich-Ludwig-Jahn-Strape15a 17469 Greifswald, Germany Ernail: schurman0uni-greifswald.de
Using the theory of additively deformed *-algebras developed by J. Wirth, we introduce the notion of LBvy processes on such objects and prove that, as in the non-deformed case, these LBvy processses can be realized as the solution of quantum stochastic differential equations in the sense of Hudson and Parthasarathy. Deformations of polynomial algebras, group algebras and of non-commutative coefficient algebras and their Levy processes are considered as examples.
A *-algebra means an associative unital complex algebra A with an involution a I+ a*, i.e. an antilinear self-inverse mapping satisfying (ab)*= b*a*. By @ we denote the tensor product of vector spaces. Notice that the multiplication of an algebra A is given by a linear mapping p : d B A + A. A coalgebra is a triplet (C, A, 6) consisting of a complex vector space C and linear mappings A : C + C @ C, 6 : C + C satisfying the coassociativity and counit conditions respectively. For *-algebras A1 and A2 the tensor product dl @ A2 is considered as a *-algebra with (a1 @ a2)* = at @ a; and (a1 @u2)(bl @ b 2 ) = albl @ u 2 b 2 . A *-bialgebra ( D , * , p , 1,A,6) is a *algebra (f?,*, p, 1) and a coalgebra (I?,A, 6) such that A and 6 are *-algebra homomorphisms; see [4]and also [2]. A quantum probability space is a pair (A,9)consisting of a *-algebra A and a state 9 on A (i.e. a positive normalised linear functional on A). A quantum random variable on the *-algebra 23 over the quantum probability space (A,9)is a *-algebra homomorphism j : f? + A; 6. [4]. *work supported in part by the european community's human potential programme under contract hprn-ct-2002-00279 qp-applications
278
We summarize some of the results of [5]. An additive deformation of a - of mappings *-bialgebra (B, *, p, 1 , A, 6) is a family (pt)t>o pt : B €3 B
+B
such that 0
(a,* , p t , 1 ) becomes a *- algebra Po=P
0 0
(p8~pt)o(A~A)=Aop,+tforalls,t~0 limt,o+ 6 o pt = 6 63 6 pointwise.
T €3 id) o (A €3 A) with 7 the flip. Notice that the Here A 6 A means (id 18 third condition repeats the bialgebra property of t3 if we put s = t = 0. We call a n additive deformation a Hopfdeformation if there exists a family (St)t>o - of linear mappings on t3 sucht that
pt o (St 63 id) o A = pt o (id 63 St) o A = e o 6
where e : C + B,e(X) = X1. Such a family ( S t )-t > ~if, it exists, is uniquely determined.
It is shown in [5] that the limit
exists pointwise for all elements in
B @ B,and that
Pt = P * =P*(tL)
L*p=p*L
L(b El c) = L(c* €3 b*) L ( l €3 1 ) = 0 and
G(a)L(b€3 c)
- L(ab €3 c) + L ( a €3 bc) - L(a €3 b)6(c) = 0
(5)
for all a, b, c E B. The latter equation means that L is a 2-cocycle in the Hochschild cohomology given by the B-bimodule structure a.X.b = S(a)XS(b)
on C. Moreover, for a given linear mapping L : 13 63 B + C satisfying (2), (3), (4) and ( 5 ) we have that (1) defines an additive deformation of the involutive bialgebra B. In our context a hermitian cocycle will always mean a bilinear form on B satisfying (3) and ( 5 ) . A 2-cocycle will be called
279
commuting if also ( 2 ) holds, and normalized if (4) holds. Of course, dl 2-cocycles are commuting if
B is co-commutative.
. . ,d, of the non-commutative We say that the involutive sub-algebras d1,. probability space (d, @)are tensor-independent (cf. [4] and also [l]) if 0
[ak,al]= 0
0
@(a1
. . .a,)
dk and ar E 4, k#I = @ ( a l ) . .@(a,) for ak E dk,k = 1,...,n.
for
ak E
The random variables j k : Bk + d, k = 1,...,n, are called tensorindependent if jl (&), . . . ,jn(Bn) are tensor-independent.
A family ( j s t ) o ~ s of ~ t mappings jst : B + d is called a Lkvy process on the deformed involutive bialgebra B (over the quantum probability space (.A,@)) if 0
0
is a random variable on (D, 1) j t l t a , . . . ,jt,t,+l are tensor-independent for all tl jrs*jst = j r t for all 0 i r 5 s 5 t 9 0 jst = @ 0 j O , t e s for all 0 5 s 5 t limt-+o+ 9 0 jo,t = b pointwise j8t
If we have a Hopf deformation of Levy process on B if j8t
5 ... 5 t,+l
B,then a family (jt),>, - on B is called a
= ( j , 0 S8)* j t
forms a Levy process in the above sense.
For a linear mapping L : B 8 B + C!, B an involutive bialgebra, a linear functional ?,!J on B is called L-conditionally positive if W*b)
+ L(b* 8 b) >_ 0
for all b E B(O) = kerns.
Proposition 1 Let B be an involutive bialgebra. Let L : B 63 13 + C be linear and such that L satisfies (5) (i.e. L is a 2-cocycle). Then for an Lconditionally positive linear functional ?,!J on B
+
(a) (a,b ) = $(u*b) L(a* 8 b) defines a positive sesquilinear form on B(0)
(b) p(a)r)(b)= r)(ab) defines a representation of the involutive algebra B ( O ) on D
280
where D = B(O)/(, denotes the pre-Hilbert space obtained from B(O) by dividing by the null space of ( , ), and where r] : B(O) + D denotes the canonical mapping.
Proof: To prove (b), let b E B(O) with ( b , b ) = 0. Then for a E B(O),using (51,
+ L(b*a* €3 ab) = $(b*(a*ab))+ L(b* €3 a*ab)
(ab,ab) = $(b*a*ab) = (b,a*ab)
wnich shows (ab,ab) = 0 by Cauch-Schwartz inequality.
0
Now extend r] and p to B by putting r ] ( l ) = 0 and p(1) = 1. Denote by H the completion of D and denote by r ( H ) the Bose Fock space over the Hilbert space L2(&.) @ H. For u E D ,T E L(D), the annihilation, preservation and creation processes At (u), &(r) and A! (u) respectively are defined on the dense sub-space
of r ( H ) , where the big union is over all finitedimensional linear sub-spaces of D, by r ( E ) we denote the Bose Fock spaces over La(&) B E considered as a sub-Hi1bel.t-space of r ( H ) , and N is the number operator on r ( H ) ;cf. [4]. The quantum stochastic differential equations (QSDE)
with
WJ) = At(r](b*>) + At(P(b) - W)l)+ A;(r](w have, in the sense of [4], a unique solution on L(&(D)). Theorem 1 Let 23 be an additively deformed involutive bialgebra with pt = p*exp, (tL), and let $ be an L-conditionally positive, hermitian normalized linear functional on 23. Then, in the vaccuum state 9 on A = L(&(D)), the solutions jst of the quantum stochastic differential equations (6, 7) form a LBvy process on B. Moreover,
281
for all b E f?.
Proof The increment property and the independence and stationarity of increments follow from the general theory of [4]. We prove that jst is a *-algebra homomorphism with respect to the *-algebra structure on (B,*) given by the mutliplication pt-s i.e. that jst(ab) = j l r t ( a ( l ) ) j 8 t ( ~ ( l ) ) e ! t - s ) L ( a@( 2b(2)) )
(8)
The differential of the right hand side equals
1) e!t-s)L
d ( j s t (a(i)) j s t
+jsta(l)jStb(l)L(a(2)
(a(2)QD b(2)
@ b(2))dt-s)L(a(3)@ b(3))dt
(b(1))+ djStb(1)) j s t (b(1))
= (jStb(1)
(t-s)L
+djst(a(l))djst(b(l)))e*
b(2) @
fjSt(a(l))jSt(b(l))L(a(2)
QD
b(2))
b(2))ekt-s)L(a(3)@ b(3)Idt
= jSt(~(l))jSt(~(l))
(dAt (“(2) 4 2 ) > *1 + dAt (Pb(2)4 2 ) 1 - 6b(2)42)11) +dAZ(r)(a(2)b(2)))+ ?G3(2)b(2))dt) (t-s)L
e*
b ( 3 ) QD
43))
(t-s)L
= jst(a(l))jst(b(l))%
(a(2) @ b(2))d1t(a(3)b(3))
where we used quantum Ito’s formula and that L is commuting. Now equation (8) follows from the fact that both sides satisfy the same QSDE.
0 The following result of [5]establishes a 1-1-correspondencebetween (equivalence classes) of Levy processes on an additively deformed *-bialgebra f? with generator L and Lconditionally positive hermitian linear functionals on 8. Theorem 2 (J. Wirth [5]; Schoenberg correspondence) For a linear functional 1c, on B the following are equivalent:
(i) J?+! is L-conditionally positive, hermitian and normalized (ii) exp,(t$) is a state on (B, *, pt, 1) for all t 2 0 0 In order to control the possible additive deformations of a given bialgebra we need a description of the class of commuting 2-cocycles. The following is sometimes helpful.
282
Proposition 2 Let B and C be bialgebras. Let 19 : B homomorphism such that
+ C be a bialgebra
(19 @ id) o A = (19 Bid) o T o A.
(9)
Then for each 2-cocycle 1 on C the mapping
L = 1 0 (19 €3 19) is a commuting 2-cocycle on B.
Proof Since 19 is a bialgebra homomorphism L is a 2-cocycle on 13 if 1 is a 2-cocycle on C. It remains to show that L is commuting. However,
( L€3 p ) 0 (A6A) = (1 @ p ) 0 (id @ id @ p ) 0 (19 @ 1 9 B i d @ i d ) 0 T 0 (A €3 A) = ( I 8 p ) 0 (id @ id @ p ) 0 (tY@tY@idBiid)0 7 0 (..A) ~ = ( p a L) 0 (A6A) where we used (9).
( T O A )
0
It is easy to see that under the hypothesis of the preceding proposition 19(C) must be cocommutative. Condition (9) is related t o the center of a semi-group in the case of bialgebras of funtions. For if G is a semi-group, H a sub-semi-group of G, and f?, C denote the algebras of ‘regular’ functions on G, H respectively (see [3]), then with 19 the restriction t o H condition (9) says that H lies in the center of G. Examples 1. Universal enveloping algebras. Let 4 be a complex *-Lie algebra, i.e. a complex Lie algebra equipped with an antilinear self-inverse mapping z + z* such that [z*,y*] = -[z,y]*. The universal enveloping *-Hopf algebra U(6) of 9 is the tensor-*-algebra over the involutive vector space L d i v i d e d b y t h e r e l a t i o n s z @ g - y @ z = [z,y] w i t h A z = z @ l + l @ z and 6z = 0. Since U(G) is cocommutative each hermitian normalized 2cocycle on U(6) gives rise t o a deformation of U(4). For example, if 9 is the abelian *-Lie algebra structure on a *-vector space 6 then U(6) equals the symmetric tensor-*-algebra S(4) over 4 and a hermitian normalized 2-cocycle on S(6) is given by
c n
L(f, 9) =
i= 1
ddf)
m,f,9 E S ( 6 )
(10)
283
where di, i = 1,.. .n are derivations. Here a derivation on S(S) is a h e a r functional d : S(G) + C satisfying d u g ) = G(f)d(g)
+d ( f ) W -
In general, L will not be a coboundary, but it is positive. Suppose now that 0 is finite-dimensional which means that S(G) equals a polynomial algebra C[zl,. . .,zd]with hermitian elements 21,.. . , z d . Denote by 8i(f), a = 1,. . . , d , the ith partial derivative of the polynomial f evaluated at 0. Then (10) becomes d
W,g)=
C Qijai(fPj(g)
i,j=l
with Q a positive complex d x &matrix. Thus we can form the L6vy process with generator 0 on the deformation of C : [ q , . . ,zd]as the solution of the QSDE of Theorem 1 which in the special case of primitive elements simply becomes
xt(4- At(%) + A;(77i) where X i s = - h ( Z j ) and qi = ~(zi) E H for H = C7Q the Hilbert space obtained from the positive form Q. We have
[xii),x,(j)l=min(s, t
) ( -~Q ~~ ~~ )
which are the canonical commutation relations. For more examples of additive deformations of universal enveloping algebras of Lie algebras see [5]. 2. *-Semigroup algebras. Let G be a *-semigroup, i.e. a semigroup (with unit 1) which has also an involution g I+ g*, that is (g*)* = g and (gh)*= h*g*. The *-semigroup algebra CG of G is the *-algebra obtained
from G by passing t o formal linear combinations of elements of G. If we Put Ag=g@g sg = 1
this extends t o a cocommutative *-bialgebra structure on (CG. In the case of a group G we put g* = 9-l. Then (CG becomes a Hopf *-algebra with antipode S(g) = 9-l. Each L-deformation is a Hopf deformation with
284
For G = Z all 2-cocycles are coboundaries and for a hermitian normalized 2-cocycle L on Z there is a function X : Z + C such that
+
Z(k,m) = X(k m) - X(k) - X(m), k,m E Z X(-m) = X(m) X(0) = 0
The deformations are given by pt(k, m) = (k + m)et(X(h+rrs)--X(k)-X(m)).
For G = Z2 the second cohomology group is not trivial. The hermitian normalized 2-cocycle
i
L ((k,m ),(n,r))= -(mn 2
is not a coboundary. The deformation of
- kr)
(11)
UZ2given by this L is
pt((k,m),(n,r))= (k +n,m+r)eit("n-kr)
which yields the Heisenberg group. The linear functional
k2 + m2
+ ( k , n ) = --
4
is L-conditionally positive and we can construct the LBvy process with generator on the deformation of UZ2as the solution of the QSDE
+
dj8t(b) = jst dj,,(b) = G(b)id
(b)
where 1 It(1,O) = -dA;
fi i
It(0,l) = -dA;
a
1 - -dAt d2
i + -dA
fi
t -4
t
z
We put Ut = jot(1,O) and V, = jot(0,l). Then
u,ut
= utu,
V,& = V,v, V,V, = ei(min(8it))) and we obtain the discrete version of the canonical commutation relations in its Weyl form. The continuous version of this is created by starting with
285
R2 instead of Z2 and by taking the same hermitian normalized 2-cocycle given by (11).
3. Free matrix *-bialgebras. Consider the free *-algebra C(xk1,xi1) = c(xkl,xi1;k , l = 1, generated by 8 indeterminates *-bialgebra if we put
xkl,
d E
N.
*
9
d>
This space is turned into a
d
We set
In the same manner the free commutative *-algebra c [ x k l , x z l ] can be turned into a commutative *-bialgebra. Theorem 3 The commuting hermitian 2-cocycles L on (C(Zk1, xgl) are precisely given by
L ( M €3 N ) = 1 ( 6 ( M )€3 6 ( N ) )
(12)
for monomials M , N where 1 is a hermitian 2-cocycle on C ( x , x * ) and where d ( M ) for a monomial M = x Z l 1 . . .x&, means the monomial d(M)X" .. .X'" of q x , x * ) .
Pmof: The mapping x k l I+ bklx extends to a *-bialgebra homomorphism .9 from c ( Z k l , xZl) to C ( x ,x * ) . Moreover,
c d
(6 €3 id) 0 A X k l =
d ( x k n ) 8 Xnl
n=l
= x €3 x k l d
n=l
= (6 €3 id) 0 7 o A X k l
which shows that 6 satisfies (9). By Proposition 2 equation (12) defines a commuting hermitian 2-cocycle. - Now let L be a commuting hermitian
286
,...,ni
nl
=
C ,...
nl
L(x:l,l
x€i * * *
nili,x$181*-*x$j8j)xZnl
xEi
61
6.
kinixr,ml * * * x G m j
,ni
which forces L(X21,1
.-.X&i,X281
=dklll ..*xr:8j) 6.
* * . ~ k i l i 6 r i 8- *1 * ' & j 8 j
l(XC1
. . . xE',x61 . ..x6j)
for some function 1 of pairs of monomials in C(x,x*). Since L is a hermitian 2-cocycle this must also be true for the extension of I to a linear functional x*)EI C(x,x*). 0 on C(x, Non-commutative analogue of generated by the elements
4.
The *-ideal I in cC(xki,x;,)
ud.
d
n=l d
n=l We denote the *-bialgebra C ( X ~ ~ , X by ~ ~K(d). ) / I If we is a *-bi-ideal. start form the commutative free matrix *-bialgebra, division by the corresponding ideal will give the coefficient algebra of the group u d of unitary d x &matrices. Thus K(d)can be regarded as the non-commuative analogue of this coefficient algebra or of u d itself. For a monomial M = x x l l .. .zZln in K(d) we put
c n
[MI = S ( M ) ( n- 2
Q)
i
Theorem 4 The commuting hermitian 2-cocycles L on K(d) are exactly given by
L(M
N ) = ~(IMI, INI)
(15)
for monomials M , N where 1 is a hermitian 2-cocycle on Z.
Proof: The mapping 13 of Theorem 3 maps the relations (13) and (14)in K ( d ) to the relations xx* = 1 = x*x in K(1) = a%. Thus 19 gives rise
287
t o a mapping n : K ( d ) + CZ and n(M) = [MI. By Proposition 2 the mappings L of the form (15) are commuting hermitian 2-cocycles on K(d). - Conversely, if L is a commuting hermitian 2-cocycle on K ( d ) it can be lifted t o a commuting hermitian 2-cocycle i on (C(Z/EI,zil;k,I = 1,. . .,d). By Theorem 3 i is of the form i = l"o (I9 €319)with hermitian 2-cocycle ion C(z, 2;). Since respects the relations for K(d) it follows that I" respects the relations for CZ. 0
It follows that the deformed multiplications are given by p t (8 ~N ) = ~ ~ e ~ ( ~ ( I ~ I + I ~ I ) - ~ ( I ~ I ) - X ( J N I ) ) which implies that the solutions
ut = ( j t ( z k l ) k , l = l , ...,d of the QSDE (6, 7), in general, are not unitary. References
[l]A. Ben Ghorbal, M. Schiirmann: Non-commutative notions of stochastic independence. Math. Proc. Camb. Phil. SOC.133 (2002), 531-561 [2]U. Franz, R. Schott, M. Schiirmann: LCvy processes and Brownian motion on braided spaces. Preprint Greifswald 2003 [3]U. Franz, M. Schiirmann: LCvy processes on quantum hypergroups. Transactions of a JapaneseGerman Symposium, Kyoto 1999 (eds. H. Heyer, T. Hirai, N. Obata). Grabner, Altendorf 2000 [4]M. Schiirmann: White noise on bialgebras. Lect. Notes in Math. vol 1544. Springer, Berlin Heidelberg New York 1993 [5]J. Wirth: Formule de LBvy-Khintchine et dgformation d'algkbre. Thesis Paris 2002
This page intentionally left blank
Infinite Dimensional Harmonic Analysis I11 (pp. 289-311) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
UNITARY REPRESENTATIONS OF THE GROUP OF DIFFEOMORPHISMS VIA RESTRICTED PRODUCT MEASURES WITH INFINITE MASS*
HIROAKI SHIMOMURA ( JOINT WORK WITH TAKESHI HIRAI) Department of Mathematics, Facuty of Education Kochi University Kochi 780-8520 Japan E-mail: shimomu0cc.kochi-u.ac.jp
Let M be a connected, non compact but a-compact, smooth manifold with d := dim(M) > 2, Diffo(M) be the group of all smooth diffeomorphisms on M with compact support, and p be a smooth locally Euclidean measure with infinite mass. Take a restricted product measure V E of the countably infinite copies of p assigning a family E = { E n } n of disjoint Bore1 sets in M which satisfies; 0 < p ( E n ) < +m W
11 - p ( E n ) l < +XI as in the context of Moore
and
[lo].
VE
is quasi-invariant
n=l
under the diagonal actions of Diffo(M). Our main aim in this paper concerns the irreducibility and equivalence of some unitary representations of Diffo(M), proposed by Hirai [4] which are considered to be irreducible components of the natural representations (Diffo(M), LEE (MOO)). The results obtained here are an extension of, and variation on, the work described by [20] on finite direct product space and on configuration space.
1. Introduction
Let M , Diffo(M), p and v~ be as defined above. Many unitary representations of Diffo(M) or its subgroups have been studied and constructed by various authors (cf.[1,2,4,5,6,7,8,11,20]). In particular an extension of the natural representation (Diffo(M), L:(M)) to the one on finite product measure spaces was first described in [20].Let us recall it briefly. For any n E N take a product measure v, of the copies p on M", *This research was partially supported by Grand-in-Aid for Scientific Research (No.14540167), the Ministry of Education, Culture, Sports, Science and Technology, Japan.
290
and take an irreducible unitary representation ( p , W ) of the permutation group 6, of { 1. . . ,n } . Then, the restriction of natural representation (Diffo(M), Lz, ( M " ) ) to a subspace gives an irreducible representation of Diffo(M), where H,,,, consists of all the functions expressed in the following form:
f(m) = p-l(a)f(z)
for
'U
E
B,,
under the notation X U := ($,(I), . ,~ ~ ( ~ 1 ) . In addition, in the same literature an infinite-dimensional version of this result is described via measures on the configuration space. Besides the corresponding group there to 6, is 6,, the infinite permutation group on N, but 6, works more implicitly than 6,. On the one hand, another infinite-dimensional version of the above result is considered by Hirai [4],which is based on the theory of infinite tensor product, with reference vectors attributable to von Neumann (cf.[12]). He constructed an irreducible unitary representation by means of E = {E,}, which are mentioned above, and which he calls p-unitals; by means of the corresponding product measures of PIE,; and finally, by glueing the L2spaces generated by the product measures together into an acting group 6, of all finite permutations on N. His reslults are important and interesting. However it seems that some corrections or more exact arguments to his work are necessary and that direct and clear discussions will be given through restricted product measures in place of the ones above. Therefore, in this paper, we concern ourselves with the unitary representations of Diffo ( M ) on the following representation spaces: Given p, E = {E,}, and an irreducible unitary representation (II,H ) of B,, we consider a Hilbert space %(C), C := ( p , E , l l ) , of all Borel measurable functions f on M" such that
f ( z a )= I I ( u ) - ' ~ ( z ) for and
Ilfll??,:=
'U
E
B,,
LE
Ilf(z)II$VE(dZ) < +O0,
where DE is a Borel set that satisfies: DEU n DE = 8, if u # id, and the complementary set of U,,EG~ DEU is YE-null under the notation DEU := {XUIX E DE}. As a suppliment of our discussions, we recall the following fact: Consider the natural representation T ( g ) of Diffo(M) on L z E ( M w )
291
acting from the left together with the unitary operators R ( u ) , u E 6,; R(a)f(x) := f(m)acting from the right, which is a unitary representation of 6,, as is easily seen. Then the system { T ( g ) , R ( a ) forms } a dual pair (cf [ 5 ] ) . Hence we claim that our representation discussed in this paper is the irreducible component of the natural representation ( T ,L;E ( M " ) ) along the dual pair. However this compelling conjecture has never been proved. In any case we will justify irreducibility and equivalence in section 3 and 4, respectively, as each requires a substantial argument. In particular, we will need a useful lemma (listed as Lemma 4.1), that has many applications to an ordinal check of the irreducibility of natural representations of the normal type. In addition, to complete the proof, we need another lemma (listed as Lemma 4.3) that states the possibility of incompressive transportation of a mass from one part to another part of an open, connected set. However, by the techinical reasons, we have to impose a condition (mcc) on our manifold M , though it is fairly mild (perhaps we will be able to show our results without (mcc), but it seems that they will require a lot of arguments, and in addition, we do not know whether (mcc) generally holds or not for non compact, connected manifolds). Finally, as a rule we will omit the easy proofs of the propositions made in this paper, but t o those which are found hard we will prove them completely. 2. Basic Discussions 2.1. Restricted product measure with infinite mass. We begin by introducing the notion of restricted product measure. Suppose that we are Bn,Pn) and a set E n E Bn given, for each n, a a-finite measure space (Xn, with 0 < Pn(En) < +GO. Denoting the restriction of pn to En by PnlEn, we form a product measure:
~ lim Since fin is increasing for n, we have a a-finite measure f i := n+m the product measurable space ( X " , Bm)of (Xn, B,,).
-
fin
on
Proposition 2.1. Take a sequence {Fn}nwith the same poroperty as that of {En}n.T h e n f i is ~ equivalent t o f i ~ (& f i ~ ) ,if and only
292
where 8 denotes the usual symmetric dif-
4 +m,
follows from the above conclusion, with
ference. Moreover,
In what follows, all the measure spaces (Xn,!Bn,pn) are identical: ( M ,% ( M ) ,p ) , where M is non compact but u-compact, manifold of class C", B ( M ) is the Borel field and p is a smooth, locally Euclidean measure whose total mass is infinite.
Definition 2.1 (Unital sequence) A sequence E = { E n } n of Borel sets in M is said to be p-unital, if it satisfies the following two conditions: (1) ' n , O < p ( E n ) < +m, w
C 11- p(En)I < +m.
(2)
n=l
In addition, i f the En are mutually disjoint, we call it a disjoint p-unital sequence.
n 00
Let us agree that the product
En, which is called a unital product
n=l
set, is denoted by the same letter E , so that no confusion occurs.
Definition 2.2 (Cofinality) Two p-unital sequences, E = { E n } n and 03
F =
{Fn}n
are said t o be cofinal if they satisfy
C p ( E n 8Fn) < +w, n=l
and this is symbolicly denoted by E F. In addition, i f En = Fn f o r suficiently large n, they are said t o be strongly cofinal, and this expression is denoted by E sz F. N
Take a p-unital sequence E = {En}n and form the restricted product measure CE. Because the infinite product of { p ( E n ) } , absolutely converges, 00
p ( E n ) C ~ has a meaning as a measure in M w ,and it depends
YE := n=l
only on the equivalence class of the sequences cofinal to E. It is easy but important to observe that UE(M&) = 0, where we
n 00
put
M E := U;=,(Mn
r(u); x tion.
-+
x
&) and that the action u E 6, on M"
:
n+ 1 X(T
leaves ME invariant. We now introduce one more defini-
293
Definition 2.3 (o-ring) Let E be a disjoint p-unital sequence. (1) The cr-ring generated by all unital product sets F with F N E ( with F M E ) is denoted by M ( p ,E ) ( M " ( p ,E ) ), respectively. (2) Similarly, the cr-ring generated b y all unital product sets F that come from the disjoint p-unital sequences with F N E is denoted by M d ( p ,E ) .
Lemma 2.1. Assume E = {En}nbe a disjoint p-unital sequence. (1) ME E M ( P ,E ) . (2) M ( p , E ) = M " ( p ,E ) mod VE. (3) M ( p ,E ) = M d ( p ,E ) mod VE. Proposition 2.2. Given a disjoint p-unital sequence E = {En}n and a B,-invariant set B E M d ( p , E ) , that is Ba = B for all a E 6,, there exists a set 13 E M d ( p ,E ) such that
B=
Do
: disjoint
union.
UEB,
Corollary 2.1. Given a disjoint p-unital sequence E , there exists a set DE E M d ( p ,E ) such that 'cr#id,
D E U ~ D E = $and
DEU
ME=
modv~.
UE6-
We end this subsection with the equivalence-singular dichotomy between the measures V E . Proposition 2.3.
W e have YE
= VF
or
VE
IVF
for any p-unital sequences E = {En},, and F = {Fn}n. 2.2 Action of Diff,(M) from the left and of 6, from the right. Let g E Diffo(M) and cr E 6 , act on M m in the following manner: g(Z1,.,xn. * .
,
(21,... z n
*.
) + (9x1,... , g z n , ... 1,
*)a + ( x u ( ~ ) .,* . 7 xu(n), *).
Clearly the actions of g and of o are mutually commutative, and they lead to transformations of V E , which will be denoted by gvE and U U E ,respectively.
294 Theorem 2.1. Given a disjoint p-unital sequence E = { E n } n , (1) U E is 6,-invariant, and
is Diff 0 ( M )-quasi-invariant. More precisely, the Radon-Nikodim derivative is expressed in the following form;
(2)
UE
where the infinite product converges in the L:,-sense
n
on each set
00
Bx
Ek ( B € '13(Mn0), p ( B ) < +ca, and no is arbitrary).
no+l
(Of course the above convergence is equivalent to Lz, -convergence for the square roots of the corresponding functions.) By the above theorem, we have two unitary representations, T ( g ) , g E Diffo(M) and R(a),a E emon L;,(Mw), such that
The representations T and R are mutually commutative, and moreover they form a dual pair (cf. [ S] ) . 2.3 Representation space R(C). Using the same notation for p and E as before and taking an irreducible unitary representation ( I I , H ) of B,, where H is the separable representation Hilbert space, we put C := ( p , E,II). Next, take a Borel measurable H-valued function f on M a which possesses the following property:
f ( x a ) = II(O>-' f (x) for t/a E em. Put
llf112 :=
/
Ilf(x)ll%'E(dx>*
(2.1)
(2.2)
DE
Then, the space R(C) consisting of those functions f that fulfill llfll < 00 forms a Hilbert space by the above norm. It is useful to note that, for each f E R(C), the integral domain D E may be replaced by an arbitrary Borel set D that possesses the following two properties:
295
(2) f = O
(
on
c
Dn)'
mod YE.
0€6-
It follows that the following action T (denoted by the same letter, but no confusion ocurrs) of Diffo(M) on X ( C ) is well-defined and ( T , X ( C ) )is a unitary representation;
T ( g ) :f ( x )+
Jdg""cx) f (9-lx). dYE
The continuity will be addressed in the later part of this section. Now take any disjoint p-unital sequence F = {Fn}nthat is cofinal to
QF : L:,(
E. We introduce a map
u 00
F n , H ) -+ R(C) such that
n=l
n
co
00
We recall that vEl Q F ( f ) = 0 on (
c
Fn = ( p x x p x ...)I Fn and agree that n=l n=l Fo)'. In addition, it follows easily from the remark -
1
.
U€6,
above that
is an isometry.
Lemma 2.2. The space spanned by QF(LE,(F,H)), F runs through the family C ( E ) of the disjoint p-unital sequences that are cofinal to E , is dense in X(E). Lemma 2.3. Assume that d := dim(M) ing condition.
> 2 and let M satisfy the follow-
(CC) There exist a sequence {U,}, of relatively compact, open sets, U, t M such that E' is connected for every n.
Given a disjoint unital sequence E = {En}n7 there exists G = (Gn}nE C(E) that posesses the following properties:
(F)'
(1) V n , G, is a relatively compact, open, connected set and p ( C \ G n ) = 0 , (2)
'n,
is connected,
= 0, if n # m ,and (3) G,nG, (4) given a compact set K , there exists N K E N such that K
n G,
=
0
296
for all n 2 N K . 03
Moreover given any E
> 0, we can take G such that
p(En8 Gn)
< E.
n=l
Remark 2.1. Let B be an ope set in Rd (d 2 2) which is surrounded by an outer large sphere and by an inner small sphere. The manifold B does not satisfy (cc). Let Co(E)denote the family of G = {Gn}nE C(E)possessing the properties from (1) to (4), which are listed in the preceding lemma.
Lemma 2.4. Under the same assumptions in L e m m a 2.3, the space spanned by { & ~ ( L ~ , ( FH,) ) } F ~ c o ( E is) dense in X ( C ) . Now we discuss the continuity of (T,'H(C)).Let us recall that Diffo(M) K } , where K runs is a union of Diff(K) := {f E Diffo(M)I suppf through the compact subsets in M. Diff(K) is equipped with the group toplogy 7 K derived from uniform convergence of the maps together with their all derivatives. We introduce the inductive limit topology r M of { r K } K to Diffo(M), though it is not a group toplogy (cf. [19]).
Theorem 2.2. Assume that d > 2. The unitary representation ( T , X ( C ) ) of Diffo(M) is continuous with respect t o r M . We end this section with concerning the following condition that will be imposed on M from now on, which is clearly weaker than (cc). (mcc) There exists a closed set S in M such that p(S) = 0, M \ S is connected and satisfies the condition (cc). (mcc) is fairly mild, and usual non compact, connected manifolds (of course the manifold B described in Remark 2.1) satisfy this condition. The next theorem is easily seen, but basic for our discussions.
Theorem 2.3. Assume that d > 2, and let M satisfy (mcc). Put M' := M\S, and denote X ( C ) for M and M' by X M ( C )and X M I(C), respectively. T h e n the unitary representations ( T ( g ) ,'Htln/r(C)) and ( T ( g ) ,X M (C')), ~ where g E Diffo(M'), C := ( p , E , l l ) and C' := (pIM',E n M',II), are unitary equivalent by a n intertwinig operator A,
n 03
A :
f(2)
E ' H M ( C )-+
k=l
X M ~ ( Z ~ ) ~E( XX M ) ~(C').
297
3. Irreducibility The main goal of this section is to prove the following theorem.
Theorem 3.1. Let E be a disjoint p-unital sequence, (II,H ) a n irreducible unitary representation of em.Form a triplet C = (Elp,II), as before. If dim(M) 2 3 and M satisfies the condition (mcc), the unitary representation ( T ,X ( C ) ) is irreducible. The proof consists of the succesive steps as below. First we need the following lemma.
Lemma 3.1. Let G = {Gn}n be a disjoint p-unital sequence, and further assume that G, is connected and open for each n. Then, the natural * representation T of the restricted product group Diffo(Gn) on L:G (G),
n n
defined below, is irreducible.
Next we go to the second stage that describes the successive steps. As before, we can assume that E = {En}npossesses the properties in Lemma 2.3. Put Em,n
.-
(q
(which is a connected open set in M),
k=n+l
En := UoEB,
n
((Em,nIn
Ek)g,
n+l (which is a symmetric set in M") and, pn :f E x ( C ) + XB,
*
f E x(C),
(which is an increasing projection tending to id). In addition, take any non empty disjoint connected open sets GI, . . . G, in Em,n,and form a unital sequence
G E := {GI, * .
*
,Gn, En+' ,
* * *
,Ek,. } * *
and a map Q g E . Now let A be an intertwinig operator of ( T , X ( C ) ) . It is not hard to see the following results.
298
(1) Image(P,AQzE) C Image(QzE) (because of the irreducibility described in the above lemma). (2) P,A&EE = &gE((id)qG,, 63 3 A n , ~ where ), A,,G is a bounded operator on H (due t o the same reason and to the irreducible assumption on
(ITH I ) . (3) A,,G does not depend on a particular choice of ( GI,. . . ,G,), so we simply write A, := A,,G (due to the connectedness of E,,,). (4)
' 0
E 6,, II(n)A, = A,II(c).
(5) vn, A, = An+l(=: A,), some calculations). (6) 3c E C , A,
(Both (4) and (5) are easily shown by
= c id (merely because of the irreducibility of (II,H ) ) .
(7) PnAQgE = cQzE (due to (2)). (8) P,AP, = cP, (because {Image(Q",))G, G := (GI,... ,G,) generates the space P,(X(C)), whenever G runs through all possible pairs of sets in (Em,,),). This is a rough sketch of the proof of Theorem 3.1. 4. Equivalence
This section is devoted to a study of the mutual equivalence of ( T , X ( C ) ) and lets E run chiefly through p-unital sets. Theorem 4.1. Assume that dim(M) >_ 3 and let M satisfy the condition (mcc). Given C1 = ( p , E,IIl) and C2 = ( p , F , I I z ) , ( T , X ( C 1 ) )and ( T , X ( & ) ) are unitary equivalent, if and only if (1) there exists a permutatation a on N (maybe an infinite permutation) such that E Fa-' and is equivalent to "II2 defined by (2)
-
"rII,(a):= rIz(a-laa)
for
v0
E6 .,
First we prove sufficiency. A space transformation defined via a; x + xu is a measure-preserving map between vFa-l and v ~ thus ;
Proof.
299
it induces a unitary operator from L:Fo-l(Mw) to L:,(Mw); f(x) + f ( x a - l ) . Furthermore, it also induces an intertwining operator A from (T,?l(Ei)) to ( T , X ( & ) ) , where Ci := ( P , F ~ - ~ , ~=I (IP~, )E , ~ I I ~ ) . By the assumption there exists an intertwinig unitary operator U from (aI12,H2) to (HI, H I ) , so ((id )p E €3 U ) o A-l gives a desired operator 0 To prove necessity, as before, we may assume that M possesses (cc). We need many lemmas. Lemma 4.1. Let M satisfy (cc), and suppose that ( T , X ( C 1 ) ) and (T,?t(&))are unitary equivalent and let A be a n intertwining unitary operator; A : X ( C 1 ) --+ X ( C 2 ) . Given a Bore1 set B E B ( M ) , we introduce a projection PB o n X(&) (i = 1,2) by
(PBfI(2):= l - K l X B ( 4 m . Then, we have
APE = PEA. Proof. The proof of this lemma is somewhat long and complicated owing to the parts concerning infinite-dimensional arguments. Thus it would be better to observe its finite dimensional version and to show how to prove the main parts. So suppose that we have as given a relatively compact, open set D. Given ~ 1 take , a compact subset K of D such that p(D\K) < 71 and cover K with a finite collection of relatively compact, open sets {Wt}r=l that are diffeomorphic to disks in Rdim M:
K
c urz1Wt
D.
Without loss of generality, we may assume that the image of p(Wt by the coordinate map q5t : Wt --+ Rdim is the restriction of the Lebesgue measure to q5t(Wt)and that p(wt\Wt) = 0. Put
v1 :=W1,
V,:=Wt\(W1U...UWt-1)
(t=2,... ,T).
Then V, (t = 1 , e . e , T ) are mutually disjoint, open sets,
p(UT=lWt\ UT=, V,) = 0 , and hence p(K\ UT=, I$) = 0. Given
r]2
> 0, take an open set Ut such that Ut C V, and p(V,\Ut) < 59 (t = 1 , . . . ,T ) .
300
Moreover, for a given 0 such that
< a < 1, and for each t , take
(j:>q)-l(x) = ax on
j,"Iv
E Diffo(4t(Wt))
+t(Vt).
Finally, put
n T
g:lq := 4;'
o j:lv o q5t,
and ga?l?:=
g:'q.
t=l
Then, we have ga>vE Diffo(D) and
Hence, letting first, a -+ +O and second,
771, r/2
-++0, we get
Thus formally we have
(%771,772
--++O).
Hence, it follows from a trivial equality
< T(ga'v)A$,A4 >=< T(ga>"+, 4 > that we get
< P D ~ A ~ , >=< A + P D C ~>, ~ as the limit when a, r/1,7~tend to 0, and PO,A = APD.. Now we can replace D" with a general Bore1 set B in Rd by standard arguments in measure theory. The proof of Lemma 4.1 will be carried out in a similar way, but by using infinite-dimensional techniques.
Lemma 4.2. Under the same assumptions and the same notation as in Theorem 4.1 and the assumption (cc) o n M , the following holds: f o r suficiently large n E N, there exists a(n) E N such that lim p ( E n 8Fc(n))= 0.
n+m
301
Sketch of the Proof. As before, we may assume that E = {En}nand F = {Fn}npossess the properties expressed in Lemma 2.3. Take vh E H1 with ilhllHl = 1 and put 03
d x ) := AQ",'(II
XE,
h)(x).
@I
n=l
Applying the above lemma to B := Ei for each k , we have PE$
= 0.
Next, approximating g with a sum of QF-image of tame functions such that M
QFM z i ,
* * *
,x i )
I-J
. h'),
X F , (2,)
1+1
where p is a square summble function and h' E Hz, we find that
n 00
'E
> 0, ' N ,
EN
s.t., ' k 2 N,,
p ( E i n F,) < 6 .
n=l
It follows that
3 4 N E N,
P(Ei n F o ( k ) )
< CKU,
with a universal constant Ku. By proceeding in a similar fashon, but changing E to F , we can show that p(Ek n F:ck,) is so small, if Ic is so large. 0
For the proof of Theorem 4.1 we need more analysis on M . The following lemma is quite useful for our discussions, and it shows the possibility of incompressive transportation of a mass from one part to another part with slim tubes in a connected open sets in M . Lemma 4.3. Assume that d := dim(M) 2 2. Let F be a connected, open subset of M and Ui (i = 1,2) be open subsets of F such that U1 n Uz = 0 and p(U1) < p(U2) < +oo. Then, given E > 0, we have a p-preserving diffeomorphism ge E Diffo(F) and a Bore2 subset B, c U1 such that
P(Ul\&)
< E and
!A(&)
c u2.
The Proof of this lemma is based on the local considerations that are guaranteed by the next lemma.
Lemma 4.4. There exists a Lebesgue measure preserving daffeomorphism with a compact support that realizes local displacement in Rd (d 2 2).
302
Proof.
Put n := d - 2 and take compact intervals
[Q',P'l c [a,PI,
[^/',6'1 c [Y,61, and
We take a vector field
defined by
wi E 0 (i = l , . - . n ) , where fi, f2 and gi (i = 1,. - . n )are C" functions on R1with a compact support such that
fl(z)= 1 on [a',p'], = O
on [a,/3]',
It is clear that
n
v =(i,~ 0 , ,. . . ,o) on T := [a',p'] x [TI,6'1 x n [ a ; , & l . i= 1
Therefore, exp(tv)(zo) = zo provided that zo morphisms.
+ t(1,0,0,..- ,0)
+ t ( l , O , 0,.. . ,0) E T.
for all
20
E T,
This is one of the desired diffeo0
Lemma 4.5. Under the same assumptions and the same notations in Lemma 4.2, we have
where the summation is taken over the indices except for a finite number of the n 's.
303
Proof. We continue to assume that E = {En}n and F = {Fn}n possess the properties in Lemma 2.3. For each n E N put 00:
a n := maxp(Ek fl Fc(n)), k#n
and GZ :=
Diffo,,(FO(k)) k=n
(: restricted direct product), where Diffo+(Fk) is a group of all p-preserving diffeomorphisms in Diffo(Fk). It is easy t o see that T(g) ---+ id, g E Gg strongly in 3t(&), as n --+m. It follows from the assumption that for a fixed unit vector h E Hi
Take sufficiently large n and m > n. As below, we choose E{ and Fj, and then arrange them in the following manner. First, put Ei := En+l,F; := Fc(n+l), and Ei is one of the Ek's that attains the maximum value of p(Ek n Fc(n)), k # n. Let Fi be the corresponding set to E i . Namely, if Ei = E k for some k , we put Fi := F,,(k). Going these procedures, we have the following sets of class E and F:
E: , F; ,E;, F211.. * ,EAl,FA1,EL, where EL(= Ep) is the first appearing set that posseses the properties: 5 n or p > m, or the property that Ep coincides with a set Et already appeared in this line. Next, we begin the second line with Ef(= E k ) that does not appear in the first line and that has the minimum index k , n < k 5 m. In a similar manner we end the second line with EL(= E p ) , where Ep is the first appearing set in the second line that possesses the properties: p 5 n or p > m, or the property that Ep coincides with a set E{ ( j = 1,2) or EL already appeared in these lines. Continuing these processes, we have the following sets of class E and F:
p
E:, F;,. . . 7 EA1,FA1,EL
.....................
304
E: ,F,",. . * ,EL, ,FL,,EL, where ELl is the last set choosen from {En+1,.. ,E m }by this procedure. Now we apply Lemma 4.3 to the sets F j ( = F ) E: n F j ( = U2)and F{ n E!+l(= U1) (i 1 = 2 , . . . , n j , or oo), and get a diffeomorphism 91 E GE with an arbitrary small E { ( = E ) (If p(F{ n E!+l) = 0 , we take the identity map as 91). Pu t
+
EL).
a{ := p ( ~nj E!+~)(i < nj), afj := p ( ~ ni ~
By the proof of Lemma 4.3 (but, here it is omitted), we see that 91 leaves invariant the other parts, namely F j n (E! u E!+l)c,except possibly small subsets with p-measure. From now on till the end of this proof, we will carry the proof through somewhat rough arguments, because the exact estimations are highly complicated and they disturb to see the essential parts. step 1. First we take the sets Fii-lwith odd indices in each line and work only these gii-l. In addition, if for the final set FAi in each line, ni is an odd number, we omit g&. It is evident that a map g composed of these gii-l is in GE. Let us estimate the following value: pg
:=
1fi
reem
n 03
XEk ( z k )
k=l
X g - l E , ( & ) (zk)vE(dz)-
k=l
It follows from the choice of gii-l that we may neglect the terms in pg except for the terms corresponding to a product of transpositions I- = ( p ,q), where
n 03
Ep = Eii-l and E, = Eii. Thus, p,/ than a product of the following terms:
Since
1
pg > -
n
2 k=l
p(Ek) is approximately smaller
k=l
p ( E k ) for large n,m, it follows from an elementary in-
equality: 1- t
5 - logt for all t E
(0,1],
305
that we have
This demonstrates that
for large n and m. step 2. In a similar fashon, taking only g& correponoding to even numbers, we have
c
a;i
< 2 log2.
id
+
step 3. Consider E L that is expressed as Eq (n 1 5 3q 5 m) and that does not coincide with any Ehk ( k = 1 , . . . ,I ) . Suppose that
Ej = EG = . . . = Ejk 00. We work only gel, . . . ,g k k and set gp,i := g g l these gp,i and form a map g .
n
0.’ ’ o
g k k . Let us compose
0
This time, the term pg/
P(Ek)is approximately smaller than a prod-
k=l
uct of the following terms:
It follows from the same arguments as before that for large n and m we have
which leads t o
cak,,.<
2log2 for large n and m.
P>S
+
step 4. Consider E L that is expressed as Eq ( n 1 5 j q 5 m ) , and that coincides with some E t k ( k = 1,.. . , I ) , and form a class & of all such EL. Let E L E E . Then E L = EAj for some j < Ic. Again, if E L E E , we have E L = EAi for some i < j. Continuing these processes, we finally
306
reach t o a set EEp such that EI$ 4 E . Thus, we are led to a tree diagram: The first(top) stage consists of the above EKp's. The second stage consists of all EAj such that E L = EEp,and then we connect Egp and ELj by a segment, and so on. Taking only gLj corresponding to the set ELj (e Fij) in the even stage, and composing a map g of these g A j , we see that the summation over the terms p(Fij nEE,) is smaller than 2 log 2. A similar result is obtained from the odd stage. step 5. Finally, consider E L which is expressed as Ep (p 5 n or p > m). Let HI, . . . ,H , be a family of the maximal (disjoint) sets of the above ones: H a. -- E2.i = . . . = EGi.' 7ai.i We take g::i,4,. . . ,gnYsi,; (i = 1 , . . . , q), and compose a map g of these maps. It follows from the same discussions as before that
a,k
for large n and m. Since n and m are arbitrary, provided that they are so large, we have < +co. 0
Can n
Lemma 4.6. Under the same assumptions and the same notations in Lemma 4.2, we have
where the summation is taken over the indices except f o r a finite number of the n 's. Proof. We continue to assume that E = {E,}, and F = {F,}, possess the properties in Lemma 2.3. As before, put a, := maxkf, p(Ek r l Fu(,l). 1 Take any positive number 6 such that S < min(-,inf p(Ej)), and take an 4 3 0
integer JO such that
aj j=Jo
< -. 2
By the assumption on E and F , there exist J1, J2,J3 E N such that JO < J1 < 5 2 < J3 which possess the following properties:
307
Take integers p and q such that J 3 5 p < q, and take all of Fi which intersects at least one of Ep +l,. . . ,Ep. Evidently we have i 2 J z . Further, we add all of other El that intersects at least one of these Fi ( I 2 J1). We may assume that Fi = F r ( ~(i<= ) l , . . ., r ) , and hence Ki 2 J1. Now let Es1, *
* *
,ESm
(s1
< . . . < sm)
be a list of all Ei which intersect to F r ( ~n , )Ek, with positive p-measures. If necessary, cutting off small parts of E,, (k = 1,.. . ,m ) we may assume that there exists a unit u (depending on K1) such that P(Es, n F r ( K l ) ) / U =: Nk E N. We divide each ESkfl F,,(K,) into Nk pieces with p-measure u and then number them as E k , l , . ,E k , N k . Now applying Lemma 4.3 to these pieces, we have maps
n
M *
gl E GJ, :=
1 = 1;-. ,rn ax (N l,-.. ,Nm) such that
Diffo,p(Fj),
j = Jo
gl(Ek,l) C Ek',z approximately holds (except a subset of E k , l with an arbitrary small p-measure), where E k , l satisfies 1 5 Nk, and k' is to be the smallest integer such that k' > k and 1 5 Np . If such E k t .1 does not exist we take EK,n F r ( ~ in , ) place of E k l , 1 in the above inclusion reration. On the other parts gE acts as the identity map. Pu t g1 := n g l for F r ( ~ (, If we have no such E,; with the above properties for F r ( ~n, )Ek, , we put g1 := id). For each 1 = 1,. . . ,max(N1,. . ,Nm), let Ml be the largest integer of M such that 1 5 NM. We remove these EM,,[and gl(EMl,l) (C E K n ~F r ( ~ from E K and ~ E,, , . ,E,,,, . Going on the same procedures for F r ( ~ ;(i) = 1 , 2 , . . . ,T ) , we have the maps g l , . . . ,gp E Gj,, and a new unital sequence {Ek}n (Ek C En). Put +
g := g1 0 . . . og', Clearly we have
/3j
and
= 0 for all j
/3j
:= p (Ej\Ei)
for each j .
< 51, and T
T
M
308
In addition, on the estimation of the term
/n M
DL := rEem
n 00
X E ; (zk)
k=l
Xg-'E:(,) (zk)vE(d%)
k=l
we can neglect the terms corresponding t o all nation from E = {En}n. By elementary calculations, we have
k=l
T
k=l
# id in virtue of the elimi-
k=l
1 where we assume that p ( E j ) > - for all j 2 Jo, and K1 is a universal 2 constant such that 1 -1ogt 5 K ~ ( I t ) for - < ' t 5 1. 2 Let €g
:= IIT(g)QE(XE @ h) - QE(xE @ h)ll,
for an arbitrary but fixed unit vector h E HI. Of course S whenever JO --+ 00, and set
U
cg
+~ 0,
Then, M
k=l
Y
and hence we have
It leads to
Since E is so small, if necessary, by taking large JO and small 6, we can assume € 1 < -. 2 2 P(Ek)(l - 2K16)
nE1
~
309
It follows from some calculations that
On the other hand, we have n(gEk n Eck) < n(gEk\gE'k) + »(gE'k n Ekc) = »(Ek\E'k) + »(gE'k n E'k). Thus, n£
Since
we get
El) <
» + .infj, ^(Ej) .1 . - dy ,} - 9y 2/iid)
=:?? In particular,
k=p+l v
For p + 1 < fc < q we have Ek = Ekn Fa(Kl} + --- + Ekn Fa(Kr) + Ek n Fc, oo
and A = Ki0 for 1 < 3z0 < r (Recall that F = V Fk). In addition, for any k=l
i 7^ i0, we can assume that
(.Efc n Fff(X;)) C ^. Hence,
and
^(ff£;fc n Eg) > ]£ /z(^ n F ff(lfj) ) = MEfc n ^ (jt) ) - ^(£fc n FC). i^io
Therefore, we have
k=p+l
k=p+l
310
c 0
Since
p(Ek
n F " ) < +oo and p , q is arbitrary, provided that
J3
5 p < q,
k=l
we conclude that n
Finally, we have n
n
< +oo.
n
n
0
Proof of Theorem 4.1 continued. The rests of the proof are that g is extended to a permutation on N and to show the equivalence relation of the irreducible unitary representations (II,H ) . However, since the space is lacking for a full explanation and these proofs are rather easier than those what we have seen, we will omit them. We end this issue with the following remark, which is seen by the above discussions. Our results obtained on the irreducibility and the equivalence also valid for any disjoint unital sequences whose cofinal equivalence classes including other sequences possessing the properties described in Lemma 2.3 in any non compact, connected manifold M with dim(M) 2 2, even if M does not satisfy the condition (mcc). References 1. G.A.Goldin, Non relativistic current algebra as unitary representations of
groups, J.Math.Phys. 12 (1971) 462-488. and D.H.Sharp, Non relativisitic current 2. G.A.Goldin,J.Grodrick,R.T.Powers, algebra in the N / V limit, J.Math.Phys., 15 (1974), 217-228. 3. T.Hirai, Construction of irreducible unitary representations of the infinite symmstric group Boo,J.Math.Kyoto Univ., 31 (1991), 495-541. 4. T.Hirai, Irreducible unitary representations of the group of diffeomorphisms of a non-compact manifold, ibid., 33 (1993), 827-864. 5. T.Hirai and HShimomura, Relations between unitary representations of diffeomorphism groups and those of the infinite symmetric group or of related permutation groups, ibid., 37 (1997), 261-316. 6. R.S.Ismagilov, Unitary representations of the group of diffeomorphisms of a circle, Funct.Anal.Appl., 5 (1971), 45-53 (= Funct.Anal., 5 (1971), 209-216 (English Translation)).
311 7. RSIsmagilov, On unitary representations of the group of diffeomorphisms of a compact manifold, Math.USSR Izvestija, 6 (1972) 181-209. 8. R.S.Ismagilov, Unitary representations of the group of diffeomorphisms of the space Rn, n 2 2, Funct.Anal.Appl., 9 (1975), 71-72 (= Funct.Anal.,S (1975), 154-155 (English Translation)). 9. R.S.Ismagilov, Representations of infinite-dimensional groups, Transl. Math. Mono, 152, Amer.Math.Soc., Providence, R I (1996) 10. C.C.Moore, Invariant measures on product spaces, 5th Berkeley sympo. Univ.of California Press, 2 part 2 (1967) 447-459. 11. Yu.A.Neretin, The complementary series of representations of the group of diffeomorphisms of the circle, Russ.Math.Surv., 37 (1982), 229-230. 12. J.V.Neumann, On infinite direct products, Compositio Math., 6 (1938) 1-77. 13. H.Omori, Infinite-dimensional Lie groups, Transl. Math. Mono, 158, Amer.Math.Soc., Providence, RI (1997). 14. E. Shavgulidze, Mesures quasi-invariantes sur les groupes de diff6omorphismes des variht6s riemaniennes, C.R.Acad.Sci., 321 (1995) 229-232. 15. H.Shimomura, Poisson measures on the configuration space and unitary representations of the group of diffeomorphisms, J.Math.Kyoto Univ., 34 (1994) 599-614. 16. H.Shimomura, Ergodic decomposition of probability measures on the configuration space, ibid., 35 (1995) 611-630. 17. H.Shimomura, 1-cocycles on the group of diffeomorphisms, ibid., 38 (1998) 695-725. 18. H.Shimomura, Quasi-invariant measures on the group of diffeomorphisms and smooth vectors of unitary representations, J.Funct.Anal., 187 (2001) 406-441. 19. N.Tatsuuma,H.Shimomura and T.Hirai, On group topologies and unitary representations of inductive limits of topological groups and the case of the group of diffeomorphisms, J.Math.Kyoto Univ., 38 (1998) 551-578. 20. A.M.Vershik,I.M.Gel’fandand M.I.Graev, Representations of the group of diffeomorphism, Usp.Mat.Nauk., 30 (1975), 3-50 (= Russ. MathSurv., 30 (1975), 3-50). 21. Y.Yamasaki, Measures on infinite dimensional spaces, World Scientific (1985).
This page intentionally left blank
Infinite Dimensional Harmonic Analysis I11 (pp. 313-324) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
AN APPLICATION OF THE METHOD OF MOMENTS IN RANDOM MATRIX THEORY
MICHAEL STOLZ Ruhr- Universitat Bochum, Fakultat fur Mathernatik N A 4/30 D-44780Bochum Germany E-mail: [email protected] The Method of Moments says that under certain circumstances it is possible to prove the weak convergence of probability measures by establishing the convergence of their moments. The present paper reviews the fundamentals of this method insofar as they are relevant for the proof of multivariate Central Limit Theorems. As an application, Central Limit Theorems for the joint distributions of the traces of different powers of random matrices from the compact classical groups are proven.
What is sometimes called ”Method of Moments” (cf. Billingsley’ , Diaconis3) is the systematic use of the fact that under certain circumstances it is possible to prove the weak convergence of probability measures by establishing the convergence of their moments. Although this method is largely folklore, it is surprisingly difficult to find a systematic presentation of it which also covers the multidimensional case. Therefore it seems worthwhile t o provide a rigorous introduction to the method with applications to multivariate Central Limit Theorems in mind. This is the content of Section 1. Those who are interested in the moment problem per se may wish t o compare the deep and far-reaching results obtained by de Jeu6. In Section 2 the method is applied to derive Central Limit Theorems for random vectors which are constructed from random elements of the compact classical groups as follows. For n E N let K , be one of On, Sp, or U,, i.e., the compact group of orthogonal, symplectic or unitary n x n-matrices, respectively. (In the symplectic case we have to assume that n = 2m be even.) Let r n be a random variable with values in K n , and suppose that the distribution of r, is given by Haar mesure. In the first two cases, for
314
any
T
E
N,we consider the R'-valued random vector
(Tw),l-w2),.. . ,Wr")), whereas in the last case we consider the C2"-valued random vector
(n(r), . . . ,n(r"),W,. . . ,q r q ) . It will be shown that these random vectors tend in distribution to a normal limit as n + M (resp. m --+m in the symplectic case). 1. Moments and Weak Convergence
In this section we will introduce moments of probability measures on the euclidean spaces and see in which way they are related to weak convergence. Let M1(R', D') (T E N) denote the set of all probability measures on the measurable space (R", a"). Let a = (al, . . . ,a,) E NL,and set
and
n
T
and
exist and are finite. aa(p) is called the a-moment of p, Pa(p) the absolute a-moment of p.
If X = (XI,.. . ,X,): (0,-4, P ) + (R',B") is a random vector with distribution Px E M:(RT,Dr), then we set
a a ( X ):= cy,(Px).
(5)
315
Note that
The absolute moments of a random vector are defined in an analogous manner. cp :
Ng + R is said to be a moment function if there exists p E such that for all a E N; : a a ( p ) = cp(a). In this case we say
M*(R',B')
that cp is represented by p. A moment function cp is said t o be determinate if cp is represented by one, and only one, p E M*(R', B'). We will then also say that p is determinate. There seems no informative condition available which is both necessary and sufficient for a moment function to be determinate, not even in the case T = 1. On the other hand, there exists a useful link, discovered by L. C. Petersen, between the one-dimensional and multidimensional problems. Let ri (i = 1,.. . , T ) denote the projection R' + R : z = (X I,.. . ,x,) ++ zi. Theorem 1.1. p E M*(R', B') is determinate if for all i = 1,.. . , r the image measure 7ri(p) E M*(R,B) is determinate.
Proof. Petersenlo, Thm. 3 The cited paper Petersen" contains an example which shows that the converse of the theorem is false. We now consider the case T = 1, fix p E M*(R,B), and write a k = a k ( p ) and P k = P k ( / l ) . Then a certain analyticity condition on the Fourier transform fi of p guarantees that p is determinate. This criterion, an explicit statement of which is contained in Theorems 1.2 and 1.3 below, appears under several guises in a number of monographs and textbooks. Let us remark that Lukacs8, p. 198, gives an example of a probability measure which does not meet the criteria below, but is nevertheless determinate in view of Carleman's theorem (see Shohat and Tamarkin13, Thm. 1.11). Let X be an algebraic indeterminate. The formal power series
316
will be called the moment generating function of p.
Theorem 1.2. For p E M*(R,B), the following conditions are equivalent:
(4
(ii) There exists 6 > 0 such that M, converges on the open disc B6(O), (iii) There exists 6 > 0 such that Mp is a holomorphic continuation of Res]-6,6[(fi) to Ba(0). (iv) There exist 6 > 0 and a holomorphic function f o n the strip { z E C : IIm zI < S} such that Resw(f) = fi. Theorem 1.3. If p E M* (R, 23) satisfies one of the conditions in Theorem 1.21then it is determinate. Corollary 1.4. For all m E R, u
2 0, N(m,a2)is determinate.
Corollary 1.5. For all r E R, m E R', C E R'" definite, N(m, C ) is determinate.
symmetric and positive
Proof. For any linear form X : R' + R the image measure X(N(m,C)) is a one dimensional normal distribution. Now Corollary 1.4 and Theorem 1.1imply what was claimed. 0 Now we wish to show that weak convergence to a determinate distribution can be established using moment arguments. This will follow from Prohorov's theorem. Since it is not very easy to find an explicit treatment of the multidimensional case in the literature, we have chosen to include full proofs.
Definition 1.6. A sequence ( p n ) n in ~ ~M1(R',Br) is said to be uniformly tight, if for every 6 > 0 there exists a compact set K , C R' such that for all n E N: pn(K,) > 1 - 6. Theorem 1.7. If ( p n ) n E is ~ a uniformly tight sequence in M1(R',23'), then it has a weakly convergent subsequence.
Proof. P a r t h a ~ a r a t h y Thm. ~, I1 6.7
0
Lemma 1.8. Let ( p n ) n E be ~ a sequence in M*(R',f?') which converges weakly to p E M1(RT,23'). Suppose that for all a E N; there exists 0 <_
317
K , < 00 such that supnENIa,(pn)l 5 K,. Then p E M*(R',B'), and for all a E Nh we have the convergence a a ( p n )--+ a,(p) as n -+ m. Proof. By Skorohod's theorem (see Kallenberg7, Thm. 4.30, for a version in sufficient generality), we can choose a probability space (Qd,P ) and random variables X,,X : (fl,d,P) -+ (R',f?') with Px, = p, ( n E W), Px = p such that X , + X almost surely as n -+ 00. For all a E Nb we set
n T
fa : R'
+ R,
,
(Zl,. . . z,)
H
z4'.
j=1
Then for all a E Ng we have that a,(pn) = E(f a o X n ) and f a o x , a.s. . Now for all p > 0
- 1 - - E ( f ( 2 a 1,...,2 0 , )
P
1 1 O X , ) = - a ( 2 a 1 ,...,2 a , ) ( ~ n )5 P
P
-+
K ( 2 a 1 ,...,2a,)
faoX
< m,
Hence (fa o X n ) n E is ~ uniformly integrable. A standard theorem on uniformly integrable sequences implies that ( f a o x ) is integrable, and therefore E(If a l o X ) = ,&(p) < 00, i.e., a,(p) exists. In addition, this theorem yields the convergence aQ(pTI) = E ( f c 2
n-icc --+
x71)
x , = aa(p).
E(fa
Theorem 1.9. Let ( P ~ ) ~ E N p, be in M*(R',B'), determinate. If for all a E Nb n+cc
aQ(pn)
0
and suppose that p is
aa(p)>
then ( p n ) n Econverges ~ weakly to p . Proof. By assumption, the sequences (q2,o,o, ...,O)(Pn))nEN,
(Q(0,2,0,
...,o)(Pn))nEN,
are convergent, hence bounded. Let 0 5 R Then for all n E N, p > 0, we have that
...
7
( y o , o , o , ...,2
)bn))nm
< cc be a bound for their sum.
318
This implies that (P,),~N is uniformly tight, because for
[
fl]
E
> 0 we can
take K , = I1 . 11 5 in Definition 1.6. In view of Theorem 1.7 it remains only to show that every weakly convergent subsequence of ( p , ) , € ~ converges t o p. Suppose that the subsequence (p,,) converges weakly to v E M1 ( E X T , B T ) ,as k tends to infinity. Since for all a E NE (a,(p,,))k€~ is bounded, we obtain from Lemma 1.8 that v E M*(RT,B T )and that cya(pn,) + aa(v).On the other hand we know that a a ( p n k )+ aa(p).This implies cy,(v) = aa(p). As a E NE was arbitrary, determinacy implies p = v. 0 2. Application to Random Matrices Fix n E N,and let K , be one of the compact matrix groups U,, O,, Sp, (n even in the last case). A random element of K,, hence a random matrix, is a random variable r, with values in K,. In what follows we only consider the case that the distribution of I?, is the normalized Haar measure WK, on K,. If I?, is a random matrix in this sense, then its eigenvalues, trace, and related functionals are complex valued random variables, and it makes sense t o study their behavior as n -+ co. Depending on the functional considered, different methods of investigation apply. See, e.g., D’Aristotile, Diaconis and Newmanl for a problem which can be treated using elementary Fourier analysis. We will concentrate on a case which is particularly well suited for applying the Method of Moments. To be specific, we consider random vectors of the form (Tr(T),Tr(r2),. * * (T
E
, n-(rr)>
(7)
N) in the cases K , = 0, and K , = Sp,, and of the form (Tr(r),. . . , Tr(rT),m,. . . , Tr(r7))
(8)
in the case K , = U,. For these vectors, the method yields multivariate Central Limit Theorems. Of course, one has to compute the joint moments of the random vectors (7) and (8) in the first place. This is in fact possible and can be done using elementary combinatorics. The ultimate reason for the friendly character of this situation is the fact that the joint moments of (7) and (8) have natural interpretations in terms of the tensor invariants of the classical groups O(n, C), Sp(n, C) and GL(n, C), respectively (see Stolz14 for details). The moment formulae were first obtained by Diaconis
319
and Shahshahani4, who chose an approach via special functions and used the character theory of Brauer algebras (Raml1>l2)in the orthogonal and symplectic cases. Let us explain these moment formulae. Fix r, q E N , a = (a1 , .. . ,a,) E NE, b = @ I , . . . , b q ) E Ni. Set 7.
k , := C
j U j
(9)
j=1
and define kb analogously. In the orthogonal and symplectic cases we write k := k,. Define f a by setting for all j E N
f a ( 8 :=
1;
if aj = 0, if j a j is odd, aj 2 1, if j is odd and aj is even, aj 2 2,
- I)!! 1+ C:zJ j d ( 3 ) ( 2 d- l)!! if j is even, aj 2 1.
jqaj
0 .
Here for m E N (2m - l)!! := (2m - 1)(2m - 3 ) . . . 3 . 1 .
(10)
If K , = On, we then have the following Theorem 2.1. If k = k , is odd, then
If k = 21 is euen and 2 n 2 k, then this expectation equals r j=1
By elementary computations, one gets from this the following more intuitive result:
Theorem 2.2. Consider a family 21,.. . ,Z, of iid standard normal random variables. If 2n 2 k ,
then
(11)
320
where r)j
:=
1, if j is even, 0 , i f j is odd.
Note that if k is odd, then (12) holds regardless of whether condition (11) is met or not, both sides being equal to zero in this case.
If K , = Sp, (n even), then we have Theorem 2.3. If k = k , is odd, then
If k = 21 is even and n 2 k , then this expectation equals T
j=1
Again elementary computations teach us that this is tantamount to Theorem 2.4. If
n 2 k, then
If K , = U,, we have the following Theorem 2.5.
If k, # kb, then
and if k , = kb and n 2 k,, then
321
To provide an interpretation in terms of the moments of a suitable normal distribution, note that if Z is an R2-valued random vector with distribution N(0, $I2), then it has the same distribution as a product RU of independent random variables R and U , where U is uniformly distributed on the unit circle and R has the same distribution as the euclidean norm llZll2 of 2 , R2 thus being distributed according to what might be called a "squeezed" X2-distribution with 2 degrees of freedom. In any case, this distribution belongs to the family of gamma distributions, and an appropriate specialization of parameters shows that it is in fact an exponential distribution with parameter 1. Partial integration and a trivial induction show that E((R2)k)= Ic! for all Ic E NO. Now regard Z as a complex random variable and call it a standard complex normal random variable. Write U = eiT, where T is distributed according to &.X1~0,2~[.We are now going to compute the joint moments of Z and Let a , b E NO. Then E ( Z " 2 ) = E(R"eiaTRbe-ZbT) = E(R("+b))E(ei("-b)T)by independence. Now
z.
Hence E(Z"zb) = dab E(R2") = dab E((R2)") = dab a! ~ iid standard complex normal random variables. Now let ( Z j ) j E be Then
Comparison with (16) yields
After this preparatory work, we can apply the Method of Moments to yield Central Limit Theorems. Theorem 2.6. For n E N let r, be a random variable with values in 0, whose distribution is the normalized Haar measure won on 0,. Fix r E N.
322
Then, as n
+ 00, yn := (Wrn),wr3,..
. 7
n(r3)
converges in distribution to
+ 17 d i ~ 3 , ... ,
z := (21, where
2 1 ,. . .
, Z, is an aid family Vj
:=
&zT
+
vT),
of standard normal random variables and
1, if j is even, j is odd.
{ 0 , if
This means that the distributions of the Yn converge weakly to 1
0 1 0 1
N
0
2
3
4 0
.. r
VT
Proof. Given a E N;, (11) will be fulfilled if n is la (rge enough. Hence by (12) the a-moment
will eventually be equal to aa(Z), so a fortiori we have the convergence aa(Yn)
n+oo ---f
aa(Z).
a being arbitrary, the claim follows by Theorem 1.9 and Corollary 1.5. 0 Using (14) instead of (12) the same argument yields the analogous result for the symplectic group. Theorem 2.7. For m E N let I'2, be a random variable with values in Sp,, whose distribution is the normalized Haar measure wsP,_ on Sp,,.
Fix r E N. Then, as m
+ 00,
n(r;,), . . . ,n ( G m ) )
Ym := (n(r2,),
converges in distribution to
2 := (Z17fizz
- 1,&z37..* , &ZT - V T ) ,
323
that is, the distributions of these random vectors converge weakly to 0 -1
N
0
-1
What makes the unitary case slightly more difficult is that a straightforward application of Theorem 1.9 and Corollary 1.5 using the identification of C with R2 requires information about the joint moments of the real and imaginary parts of the traces. Therefore we must extract this information from equation (17) in the first place. To this end consider endomorphisms cp and I)of the polynomial ring C[S, TI, where cp is given as the evaluation homomorphism determined by S w S + iT,T w S - iT,and I)as the Tw A routine evaluation homomorphism defined by S i--+ computation verifies at once that cpI)= I)cp = id. So cp and I)are automorphisms of C [ S , T ] . Now fix T E N, and for j = 1,.. . , r let &, 0, be real random variables and <j := <j + iej. Then for any a , b E N; the monomial
y,
9.
T
j=1
can be expressed as a linear combination of monomials of the form
j=1
with c , d E
Ni.
Now consider sequences ( @ n ) ) n E ~ ,(B("))nEN of RT-valued random vec:= 5'") iB(") for all n E N and suppose that for all c, d E N; tors. Set <("I
+
Interchanging the limit with the finite linear combination from the last
324
paragraph one sees that for all a, b E
NE
Specializing to the data of the unitary case yields the following central limit theorem:
Theorem 2.8. For n E N let I?, be a random variable with values in U, whose distribution is the normalized Haar measure wu, o n U,. Fix r E N. Then, as n -+co, Y, := ( R e ( T r ( r , ) ) , I m ( ~ ( r , ) ) ,. . . ,Re(Tr(rL)),I m ( n ( r L ) ) )
converges in distribution to
N 0 References 1. A. D'Aristotile, P. Diaconis and C. M. Newman, Technical Report 2002-18, Stanford University, Dept. of Statistics 2. P. Billingsley, Probability and Measure, New York (Wiley) 31995 3. P. Diaconis, Proc. Symp. Appl. Math. 37 125-142 (1987) 4. P. Diaconis and M. Shahshahani, J . Appl. Probab. 31 A 49-62 (1994) 5. W. Feller, An Introduction to Probability Theory and its Applications, Vol. 2, New York (Wiley), '1971 6. M. de Jeu, Ann. of Prob. 31, 1205-1227 (2003) 7. 0. Kallenberg, Foundations of Modern Probability, New York (Springer) 22002 8. E. Lukacs, Characteristic Functions, London (Griffin) '1970. 9. K. R. Parthasarathy, Probability Measures on Metric Spaces, New York (Academic Press) 1967 10. L.-C. Petersen, Math. Scand. 51 361-366 (1982) 11. A. Ram, Paczfic J . Math. 169 173-200 (1995) 12. A. Ram, European J. Combin. 18 685-706 (1997) 13. J. A. Shohat and J . D. Tamarkin, The Problem of Moments, New York (AMS) 1943 14. M. Stolz, On the Diaconis-Shahshahani Method in Random Matrix Theory, Ph.D. thesis, University of Tubingen, 2004
Infinite Dimensional Harmonic Analysis I11 (pp. 325-351) Eds. H. Heyer et al. @ 2005 World Scientific Publishing Co.
ISOTROPY REPRESENTATION FOR HARISH-CHANDRA MODULES
HIROSHI YAMASHITA Department of Mathematics, Hokkaido University Sapporo 060-0810, JAPAN E-mail: yamasitaOmath.sci.hokudai.ac.jp We study the isotropy representation attached to an irreducible Harish-Chandra module with irreducible associated variety. It is shown that, under some assump tions, the dual of the isotropy representation in question can be characterized by means of the principal symbol of a differentialoperators of gradient type. By using this, the case of Harish-Chandra module of discrete series is more closely examined.
1. Introduction
Let g be a complex semisimple Lie algebra with a nontrivial involutive automorphism 8 of g. We write g = 0 @ p for the symmetric decomposition of g given by 8, where C and p denote the +1 and -1 eigenspaces for 8, respectively. Let Kc be a connected complex algebraic group with Lie algebra 0. We assume that the natural inclusion C v g gives rise to a group homomorphism from Kc to G;d through the exponential map. Here GEd denotes the adjoint group of g. Then, this homomorphism naturally induces the adjoint representation Ad of Kc on g. We say that a finitely generated g-module X is a (8, Kc)-module, or a Harish-Chandra module, if the action on X of the Lie subalgebra 0 is locally finite and if it liits up to a representation of Kc on X through the exponential map in such a way as ( k . X - k - l ) v = (Ad(k)X)v for X E g, Ic E Kc and v E X. It is a fundamental result of Harish-Chandra that the study of irreducible admissible representations of a real semisimple Lie group essentially reduces to that of irreducible (8, Kc)-modules. Let X be an irreducible (8,Kc)-module. A Kc-stable good filtration of X naturally gives rise to a graded, compatible (S(g), Kc)-module grX annihilated by 0 , where S(g) denotes the symmetric algebra of g. By BorhoBrylinski [l]and Vogan [22], [23], the associated cycle C ( X )of X is defined to be the support V ( X ) of g r X combined with the multiplicity at each
326
irreducible component of V ( X ) .The support V ( X )is called the associated variety of X. It is a Kc-stable affine algebraic cone contained in the set of nilpotent elements in p, and each irreducible component of V ( X )is the closure of a nilpotent Kc-orbit 0 in p. As we have shown in [6] and [27], the variety V ( X )controls some fundamental properties for X. The algebraic cycle C ( X ) describes a sort of asymptotic behavior of X (cf. [21]). Moreover, it is shown by Vogan [22, Theorem 2.131 that the multiplicity of X at an irreducible component of V ( X )can be interpreted as the dimension of a certain finite-dimensional representation ( w g ,W ) of the isotropy subgroup Kc(X) of KC at an X E 0. We call wg an isotropy representation attached to X . In terms of wo, the associated cycle C ( X ) of X is expressed as dim wg
C ( X )=
- [B].
0
Now, we assume that the associated variety V ( X ) of X is the closure of a single nilpotent Kc-orbit 0 in p. This assumption does not exclude important (g, Kc)-modules related to elliptic orbits. In reality, it is wellknown that the Harish-Chandra modules of discrete series (more generally Zuckerman derived functor modules) and also the irreducible admissible highest weight modules of hermitian Lie algebras satisfy this hypothesis. The purpose of this paper is to study the associated cycle C ( X ) and in particular the isotropy representation w g attached to a (g, Kc)-module X with irreducible associated variety, by developing our arguments in [30] for unitary highest weight modules and also those in [29] for discrete series. To do this, we first look at in Section 2 a relationship between the (S(g), Kc)-module g r X and the induced representation r ( W ) = Ind$(x,(wg,W) of Kc equipped with a natural S(g)-action. This amounts to a survey of some aspects of Vogan’s work ([22, Sections 2-41 and [23, Lectures 6 and 71) in a slightly modified and simplified form (but for limited X ’ s ) . A reciprocity law of Frobenius type for such an induced module (Proposition 2.2) plays an important role. In fact, it is effectively used to prove an irreducibility criterion for wg (Theorem 2.1). Also, we include a remarkable result (Theoren 2.2) on the isotropy representations for singular unitary highest weight modules, given in [30], [33] and [24]. In order to identify the isotropy representation w o , it is useful to consider not only gr X but also its Kc-finite dual realized as a space of certain (vector valued) polynomial functions on p. We present this idea in Section 3. A sufficient condition is given in Proposition 3.1 for grX being annihi-
327
lated by the whole prime ideal I of S(g) defining 6. In such a case, the isotropy representation, more precisely, its dual w & , can be described by means of the principal symbol of a differential operator on p of gradient type (see Propositions 3.2 and 3.3). In the last part of this paper, Section 4, we focus our attention on the irreducible Harish-Chandra modules X of discrete series. As is well known, such an X has irreducible associated variety (cf. [28], [29]). The multiplicities in the associated cycles for discrete series have been intensively studied by Chang [2], 131, by means of the localization theory of Harish-Chandra modules. He succeeds to describe C ( X )explicitly for the real rank one case. Taniguchi applies in [19] and [20] the results of Chang in order to specify Whittaker functions associated with discrete series for S U ( n ,l ) , Spin(n,1) and SOo(2n,2). Here in this paper, we exploit another way to identify C(X), by using a realization of X as the kernel of an invariant differential operator of gradient type on the Riemannian symmetric space (cf. [9], [MI; [26], [34]). Based on our results in Section 3 and also on the discussion in [29], we can construct a certain Kc(X)-submodule U,(QC)of the representation (w&,W ' ) contragredient to wo. Moreover some evidences are given for this subrepresentation being large enough in the whole w&. The gained results are summarized as Theorem 4.2 and Corollary 4.1. This article is an updated and enlarged version of the informal reports [31] and [32] appeared in RIMS Ktjkyiiroku.
2. Graded module g r X and induced representation r ( W )
As in Section 1, let X be an irreducible (g, Kc)-module with irreducible where 0 is a nilpotent Kc-orbit in p. For associated variety V ( X ) = later use, this section introduces some elementary aspects of Vogan's theory on the associated cycle and the isotropy representation attached to X . The results in this section may be read off from [22] and [23] with some effort. Nevertheless, we include the proofs for these important results in order to make this paper more accessible.
n,
2.1. Associated cycle and isotropy repwsentation First, we introduce our key notion precisely. Take an irreducible Kcsubmodule (.,Vr) of X , which yields a Kc-stable good filtration of X
328
in the following way:
X Oc X I C
c X, c
, with
x , := V,(g)V, ( n = 0,1,2,. ..).
(2.1)
Here U ( g )denotes the universal enveloping algebra of g, and we write Un(g) (n = 0,1,. ..) for the natural increasing filtration of U ( g ) . This filtration gives rise to a graded (S(g),&)-module M = g r X , annihilated by S(t), as follows: 00
M = grX = @ M n
with M, := X n / X n - 1 ( X - 1 := (0)).
(2.2)
n=O
We note that
M, = S"(g)V, = Sn(p)V. and MO= V,,
(2.3)
where S,(D)is the homogeneous component of the symmetric algebra S(a) of degree n. By definition, the associated variety V ( X ) of X is identified with the afEne algebraic variety of g given by the annihilator ideal Annqa)M in S(g) of M :
V ( X )= (2 E g 1 f(2)= 0 for all f E Anns(a)M) c p,
(2.4)
where S(g) is viewed as the ring of polynomial functions on g by identifying g with its dual space through the Killing form B of g. Throughout this section, we assume that V ( X ) is irreducible. The Hilbert Nullstellensatz tells us that the radical of Anns(,)M coincides with the prime ideal I = I ( V ( X ) )defining the irreducible variety V ( X ) : I = So we see I n M = (0) for some positive integer n, and we write no for the smallest n of this nature. Then, one gets a strictly decreasing filtration of the (S(g), &)-module M as
d w .
M =P M
2I'M 2 2 InoM = (0).
(2.5)
By the multiplicity multI(X) of X at I is meant the length as an S ( g ) p module of the localization MI of M = grX at the prime ideal I . Then, the associated cycle C ( X ) of X turns to be
C ( X )= multI(X) - [8] with V ( X )= 8.
(2.6)
Note that this cycle does not depend on the choice of a good filtration (2.1) of
x.
329
Now, let us explain how the multiplicity multI(X) can be interpreted as the dimension of an isotropy representation. For this, we take an element X in the open Kc-orbit 0 C V(X). Set Kc(X) := {k E KcI Ad(k)X = X}, the isotropy subgroup of Kc at X. We write m(X) for the maximal ideal of S(g) which defines the one point variety {X} in g: m ( x ) :=
C(Y- B(Y,x))s(~)
for x E 0.
(2-7)
y a
For each j = 0, . .. ,710 - 1, we introduce a finite-dimensional representation a o ( j ) of (S(d,&(X)) acting on
W ( j ):= IjM/m(X)IjM,
(2.8)
in the canonical way, and we set no-1
( a o , W := @
( ~ O ( j ) , W ) ) .
(2.9)
j=O
We call wo the isotropy representation attached to the data ( X , V , , U ) , where V, yields the filtration (2.1) of X. The following lemma (cf. [22, Corollary 2.71; see also [31, Remark 2.21) is essential for our succeeding discussion.
Lemma 2.1. Let N be a finitely generated (S(g), Kc)-module such that I N = (0). Then, the length of S(g)I-module NI is equal t o the dimension of the vector space N/m(X)N f o r every X E 0. This lemma tells us that the length of the localized S(g)I-module (IjM/Ij-+'M)1 equals d i m a o ( j ) , by noting that the ideal I annihilates the subquotient IjM/Ij+lM of M. Together with the exactness of localization, we immediately get the following
Proposition 2.1. One has multI(X) = dimmo. Moreover, the equality multI(X) = dim wo(0) = dim M/m(X)M
(2.10)
holds i f and only if the support of the S(g)-module I M i s contained in the boundary 3 0 = B\0. Remark 2.1. The representation ao(0)in (2.10) never vanishes because the annihilator ideal Anns(g)M/IM is equal to I (cf. [30, Lemma 3.41). Moreover, the equality (2.10) holds for a number of unitary (g, Kc)-modules
330
X with unique extreme Kc-types V,. See Example 2.1 and Theorem 4.2 (1). 2.2. Induced module
r(Z)
We consider a finite-dimensional (S(g), Kc(X))-module (a,2) with X E 0, where &(X) acts on 2 holomorphically. Let r(2)denote the space of all left Kc-finite, holomorphic functions f : Kc + 2 satisfying
f
(Yh)= w ( h ) - l f ( y ) (Y E Kc, h E Kc(X)).
Namely, r(2)consists of all &-finite, holomorphic cross sections of the &-homogeneous vector bundle Kc X K ~ ( X )2 on Kc/Kc(X) N 0. Then, r(2)has a structure of (S(g), &)-module by the following actions:
(D* f ) ( Y ) := d A d ( Y ) - l D ) f ( Y ) , (k * f ) ( Y ) := f ( k - l y ) , for D E S(g), k E Kc and f E r(2).We call I'(2) the (S(g),Kc)-module induced from w. We note that, if 2 is annihilated by the maximal ideal m(X), the S(g)-action on r(2)turns to be the multiplication of functions on the orbit 0: (2.11) In this case, the annihilator in S(g) of any nonzero function f E r(2) coincides with the prime ideal I defining 8. Let M be any (S(g),Kc)-module. If p is a homomorphism from M to 2 as (S(g),Kc(X))-modules, we define a function T, : Kc + 2 for each m E M by putting Tm(Y) := P ( Y - ~* m) (Y E Kc).
(2.12)
Then it is standard to verify that Tm lies in r(2)and that the map T : m H T, ( m E M ) gives an (S(g),Kc)-homomorphism from M to F(2). More precisely, one readily obtains the following reciprocity law of Frobenius type.
Proposition 2.2. Under the above notation, the assignment p up a linear isomorphism Homs(g),Kc(x)(M,2 ) = Homs(g),Kc(M,V)).
HT
sets
(2.13)
Here, for R-modules A and B , we denote by Homn(A, B ) the space of Rhomomorphisms from A to B.
33 1
2.3. Homomorphism T = @jf(j) We now return to our setting in Section 2.1, where M = g r X for an irreducible (g,Kc)-module X with V ( X ) = 8. Take an integer j such that 0 5 j 5 no - 1. Let p ( j ) denote the natural quotient map from I j M to W ( j )= I j M / m ( X ) I j M . Correspondingly, we get an (S(g), Kc)homomorphism T ( j ): I j M -+ I'(W(j))by Proposition 2.2. It follows that KerT(j) =
n
m ( Y ) I j M 3 Ij+'M,
(2.14)
YE0
by the definition of T ( j )together with m ( Y ) 3 I (Y E 0 ) .
Proposition 2.3. KerT(j) is the largest (S(g),&)-submodule of I j M among those N having the following two properties: (i) N 3 Ij+'M, and, (ii) the support of N / I j + l M is contained in ao. Proof. First, we show that KerT(j) have two properties (i) and (ii). The inclusion (2.14) assures (i). As for (ii), we consider a short exact sequence of (S(g), Kc)-modules: 0 + KerT(j)/Ij+lM
+ IjM/Ij+lM -+ IjM/KerT(j) + 0.
(2.15)
Each module is annihilated by I. In view of Lemma 2.1, we find that the multiplicity of IjM/KerT(j) at I is equal to the dimension of vector space
( I jM/Ker T ( j ) ) / m ( X(IjM/Ker ) T(j)) 21
+
I j M / ( m ( X ) I j M KerT(j)) = W ( j ) . (2.16)
Here, the last equality follows from KerT(j) C m ( X ) I j M (see (2.14)). This shows that the length of S(g)I-module IjM/Ij+'M and that of IjM/Ker T ( j ) coincide with one another. Hence (Ker T ( j ) / l j + ' M ) ~ vanishes by (2.15). This means that the support of KerT(j)/Ij+lM is contained in do. Second, let N be any (S(g), Kc)-submodule of I j M with two properties (i) and (ii) in question. (2.14) tells us that T ( j ) naturally induces an (S(g), Kc)-module map from I jM /I j+'M to I'(W(j)) which we denote by f'(j).Then, f ' ( j ) ( N / I j + ' M )must vanish by virtue of (2.11) together with the property (ii) for N . This proves N c KerT(j).
As for the injectivity of T ( j ) , one gets the following consequence of Proposition 2.3.
332
Corollary 2.1. The homomorphism T(j) : I j M + r ( W ( j ) )is injective if and only i f Annq,) m = I f o r all m E IjM\{O}. In this case, one has I j +‘ M = {0}, i.e., j = no - 1. Example 2.1. We encounter the situation in the above corollary with j = 0, for example, if X is a unitary highest weight module of a simple hermitian Lie algebra g, and V, in (2.1) is the extreme Kc-type of X.Note that the associated variety of such an X is the closure of a “holomorphic” nilpotent Kc-orbit in p. See [30, Section 3.21 for details. Summing up f’(j)’s on I j M / I j + l M ( j = 0,. . . ,no - l), we obtain an (S(g), Kc)-homomorphism T := @ j f ’ ( j ) :
&q~ :=)@ I ~ M / I ~ + 3 ~ M @ WW))= rw), j
(2.17)
j
where the support of the kernel KerT is contained in 80.
Remark 2.2. By using the “microlocalization technique”, Vogan constructed a new Kc-stable Z-gradation on X such that the corresponding graded module embeds into r ( W ) as a representation of Kc (see [22, Theorem 4.21). Thanks to this result, one always has X c) r ( W ) as Kcmodules. Noting that G(I)N X as Kc-modules, we find that the above T : G(I)+ r ( W )must be an isomorphism if T is surjective.
2.4. ImdUCibility of wo The results in Sections 2.1-2.3 lead us to prove the following natural criterion for the irreducibility of isotropy representation ( wg,W)of K c ( X ) (cf. [23, Proposition 7.61; see also [31, Section 51).
Theorem 2.1. T h e following two conditions o n X are equivalent t o each other. ( a ) ( w g , W ) i s irreducible as a K c ( X ) - m o d u l e . (b) If N i s any (S(g), Kc)-submodule of M = g r X , either the support of N or that of the quotient M / N i s contained in 80. In this case, we have wg = wg(O), or equivalently, the support of I M is contained in 8 0 by Proposition 2.1. Proof. The implication (a) + (b) is an easy consequence of the exactness of localization. In what follows let us prove (b) (a). First, we note that
+
333
the condition (b) together with Remark 2.1 implies that the support of I M is contained in 80. Thus one gets wg = wg(O), or,
W = W ( 0 )= M/m(X)M. Now, suppose by contraries that W is not irreducible. Then, there exists a Kc(X)-stable subspace C of M such that M 2 C 2 m(X)M and that 2 := M / C is irreducible as a Kc(X)-module. The condition C 3 m(X)M assures that C is S(g)-stable. Thus 2 becomes an (S(g), Kc(X))-module annihilated by m(X). Next, we consider two induced (S(g), Kc)-modules r ( W ) and r(Z). The quotient map W = M / m ( X ) M + 2 = M / C gives rise to an (S(g), Kc)-homomorphism, say y, from r ( W ) to r(2) in the canonical way. Set T' := y o T(O),where T ( 0 ) : M + r ( W ) is the (S(g),Kc)homomorphism defined in Section 2.3. Then, as shown in the proof of [23, Proposition 7.91, the image T ' ( M ) of T' is a finitely generated (S(g), Kc)submodule of r(2)whose isotropy representation is isomorphic to 2. This combined with T ' ( M ) N M/Ker T' tells us that the multiplicity of Ker T' at the prime ideal I is equal to dimW - d i m 2 > 0. By the assumption (b), we find that the support of M/KerT' N T ' ( M ) is contained in 80. This necessarily implies Ker T' = M , i.e., 7'' = 0, because the S(g)-module T ' ( M ) (C r ( 2 ) ) admits no embedded associated primes by (2.11). Finally, the resulting equality T' = 0 means that y-l. m
+ m(X)M = T ( o ) , ( ~E) C/m(X)M
for all y E Kc and m E M . This contradicts C
# M.
2.5. Case of unitary highest weight representations
Let X be an irreducible unitary highest weight (8, Kc)-module of a simple hermitian Lie algebra g, with extreme Kc-type V,.
Example 2.2. In [30, Section 51, we have described the isotropy representation wg = wo(0)explicitly, when X is the theta l i t of an irreducible representation of the compact groups G' = O(lc),U ( k ) and Sp(lc) with respect to the reductive dual pairs (G, G') = (Sp(n,R), O(lc)),(Sub,q), U(lc))and (SO*(2n),Sp(lc)),respectively. In particular, one finds that the representation a g is irreducible if the dual pair (G, G') is in the stable range with smaller member G'. In this case, X H a& essentially gives the Howe duality correspondence ([lo], [ll],[8,Part 111 etc.).
334
An irreducible highest weight (g, &-)-module X is called singular if the Gelfand-Kirillov dimension dimV(X) is strictly smaller than one half of the dimension of the corresponding hermitian symmetric space. Recently, we have described the isotropy representations by using the projection onto the PRV-component (cf. Proposition 3.3), for all singular unitary highest weight modules which can not be obtained by the Howe duality correspondence. This work is in collaboration with Wachi (see [24] and [33] for details). As a result, we establish the following remarkable theorem.
Theorem 2.2. The isotropy representation is irreducible and explicitly described f o r every irreducible singular unitary highest weight representation of a simple Lie group of hermitian type. The above result for DI and EVII gives a clear understanding of some multiplicity formulae obtained by Kato and Ochiai ([13], [14]). For EVII case, we get two distinguished series of isotropy representations which decompose the quasi-regular representations for the compact symmetric spaces S8 E S 0 ( 9 ) / S O ( S ) and P2(Cay) 7 Fq--52)/Spinn(9),respectively. These decompositions can be related to tensor products of singular unitary representations [4],by using the generalized Whittaker vectors [30].
3. Utility of the dual ( S ( g ) ,&-)-module In this section, we do not assume a priori that the associated variety V ( X ) of X is irreducible. Let M = gr X be, as in (2.2), the graded (S(g), Kc)module attached to an irreducible (8, Kc)-module X through the filtration (2.1). For any nilpotent element X E p, we can define the maximal ideal m(X) of S(g), the Kc(X)-module W ( 0 )= M / m ( X ) M , and (S(g),Kc)homomorphism T(0): M + I'(W(O)), just as in Section 2. Here I'(W(0))= I n d z ( x ,(W(0))is the (S(g), &)-module induced from W(0). The purpose of this section is to make some simple observations concerning the associated cycle of X,in connection with the Kc-finite dual M * of M realized as a space of V;-valued polynomial functions on p. This is done by developing our arguments in [29], [30] in full generality. A sufficient condition is given in Proposition 3.1 for the annihilator Arms(,) M of M being equal to the prime ideal of S(g) defined by the &-orbit 0 := Ad(&)X through X. Furthermore, we characterize M * as the kernel of a differential operator V on p of gradient type (Proposition 3.2). Then the principal symbol of V allows us to describe the Kc(X)-module W(O)*dual to W(0)(Proposition 3.3). The observations made in this section will be
335
used effectively in Section 4, in order to describe the associated cycles for (g, Kc)-modules X of discrete series. 3.1. Kc-finite dual M * and its submodule
*
First, the tensor product S(p)@V, admits a natural structure of (S(g), Kc)module so that t annihilates the whole S(p) €3 V,:
{
D1.(D€3 v) := D1D €3 v (DlE S(p)), D2.(D€3.) := 0 (D2 E S ( t ) ) , k . (D€3 V ) := Ad(k)D €3 kv (k E Kc),
(3.1)
where D €3 v E S(p) @ V, with D E S(p) and v E V,. Since M = S(p)V, with V, = Mo, there exists a unique surjective (S(g), Kc)-homomorphism 7r : S(p)
€3
v, +M
such that ~ ( 1 v) 8 = v for v E V,. We write N for the kernel of 7r. This is a graded (S(g), &)-submodule of S(p) €3 V,. On the other hand, we identify S(p*) €3V,* canonically with the space of polynomial functions on p with values in V,*,where U ' denotes the dual space of a vector space U.S(p*) €3 V,* also becomes an (S(g), Kc)-module on which g acts by directional differentiation through the quotient map g -b g/e
2(
p:
{
(Dl. f)(Y) := (Wl)f)(Y) (DlE S(P)), (D2 * f)(Y) := 0 (D2 E s(e)), (k * f)(Y) := k * f(Ad(k)-'Y) (k € Kc),
(34
for f E S(p*) €3 V,* and Y E p. Here D1 ++ a(&) denotes the algebra isomorphism from S(p) onto the algebra of constant coefficient differential operators on p, defined by d a(Z)f(Y) := zf(Y t Z ) ~ t = o for 2 E p. (3.3)
+
Note that the action of S(g) on S(p*) €3 V: is locally finite. The assignment
*
p3X X' E p* with X*(Y):= B(X,Y)(YE p). gives a Kc-isomorphism from p onto p*. Now it is standard to verify that (Sb) €3 VT) x (S(P*)€3 v3 3 (D€3 v, f )
(D@ % f := ) ((TD*f)(0),v)v;xV,,
* (D€3 v ,f ) E c,
(3.4)
(3.5)
336
gives a nondegenerate (S(g), K&invariant pairing, where denotes the principal automorphism of S(p) such that TY = -Y for Y E p, and ( - , - ) v ; ~ vis, the dual pairing on V !, x V,. Let M* denote the &-finite dual space of M , viewed as an (S(g), &)-module through the contragredient representation. We write N L for the orthogonal of N in S(p*)@V' with respect to (. , .). Then, (3.5) naturally induces a nondegenerate invariant pairing
( * , -)l:MxN'-+cC,
(3.6)
which gives an isomorphism of (S(g), Kc)-modules:
M* N N'-
c S(p*)@ V.:
(3.7)
For an integer n 2 0, we denote by ( N l ) , , the homogeneous component of N' of degree n: ( N l ) , , := N'- n (Sn(p*)@ V):. Let X be a nilpotent element in p. Noting that M = V, + m(X)M, we have a natural Kc(X)-homomorphism V,
+W(0)= M/m(X)M -+
0.
This induces an embedding of W(O)*into V,* as
W(O)*N (V,/(V, n m(X)M))*c)V,*,
(3.8)
by passing to the dual. In this way, we regard W(O)*as a Kc(X)-submodule of v,. For each integer n 2 0, let q,,be the Kc-submodule of Sn(p*) @ V,! generated by the vectors (X*),, €4 w* (w* E W(O)*): q n := ((X")" @ w* I v* E W ( O ) * ) K , .
Note that the V:-valued
(X')"
(3.9)
polynomial function ( X * ) n€4 w* on p is defined by @ w* : p 3
2 c)B ( X , Z)%* E
cc.
We set 00
9 := @ Qn
c S(p*)€4 v.:
n=O
Then, we can show the following Lemma 3.1. (1) 9 is an (S(g),&)-submodde of S(p*)8 V' contained in N'-.
337
*
(2) W e write L* for the orthogonal of in M with respect to the pairing ( . , . )1 on M x N I . Let T(0): M + r(W(0))be the (S(g), Kc)homomorphism defined in Section 2.3. Then, one gets KerT(0) n M,, = L9n M,,
for every integer n 2 0 ,
(3.10)
In particular, I* = @, L!P n M,, is contained in Ker T(0). Proof. (1) It is easy to see that q is (S(g), Kc)-stable by noting that
Ye (((Ad(y)X)*),,8 y v*) = nB(Ad(y)X,Y)((Ad(y)X)*)"-'@ y - v*
*
lies q n - 1 for Y E g, v* E W(O)*and y E Kc. To prove C N I , let &* denote the linear form on S(p) 8 V,, which is the pull back of v* E W(O)* through the quotient map
S(p)8 V, 1 ,M
+W ( 0 )= M/m(X)M.
+
Then $J"* is zero on the subspace m(X) 8 V, N c S(p) 8 V,. If f i = C j @ v j (y3 E p, vj E V,) is a homogeneous element of N of degree n, it follows that
qn
8 v*> = C(-l)nn!B(X,Yj)n(vj,v*)v,xV:
(fi,
j
= (-l)'h!& (6) = 0,
by noting that y - B ( X , Y j ) " E m(X). Hence one gets ((Ad(y)X)*)n8yv* E y N L = N' for all v* E W(O)*,y E Kc and n 2 0. Thus we obtain (1). (2) Let m = C j vj be an element of M,, with y j E p and vj E V,. Just as in the proof of (l),we see that rn E if and only if
-
y.
0 = ( m, ((Ad(Y)X)*)" 8 y
*
v*)1
= x(-l)nn!B(y3,Ad(y)X)n (9-l
- v j ,v*)v,~v:
j
= (-l)nn!(w)m(Y), v*)w(o)xw(o)* ,
for all v* E W(O)*and y E Kc. This means m E KerT(0).
0
3.2. A suflcient condition for I = Anns(@)M
Let 0 = Ad(Kc)X be the nilpotent Kc-orbit through X. We write I for the prime ideal of S(g) defining the Zariski closure of 0,and let (Oil i = 1 , . . . ,r } be a finite set of homogeneous elements of S(p) which generates the ideal I. We set n(i):= deg Di (1 5 i 5 r) and n ( 0 ) := 0.
338
Proposition 3.1. Assume that Q n ( i ) = ( N l ) , ( i ) f o r i = 0,. . . , r . T h e n we have I = Anns(,)M. Therefore, X has irreducible associated variety V ( X )= 8, and the corresponding associated cycle C ( X ) of X turns t o be
C ( X )= dimW(0). [a]. Proof. We see for every v E V, that T(0)Div(!l)=
( 0 2*
T(0)v)(!l)= oi(Ad(y)X)T(O)v(!l)
= 0 (Y E &),
since Di E I and Ad(y)X E 0. It then follows that Div = 0 by Lemma 3.1 = ( N L ) n ( i )which , is equivalent to (2) together with the assumption Qn(i) L!PflMn(i)= (0). Hence, Di annihilates V, and so the whole M = S(p)V,. This shows I c Anns(,)M. Now, W(O)*cannot vanish by the assumption QO= ( N L ) o N- V.: This implies m(X) 3 Anns(,)M. In reality, if m(X) $ Anns(,)M, one gets S(g) = m(X) Anns(,)M by the maximality of m(X). This yields
+
M = S(g)V, = m(X)V, Hence we deduce
I=
n
c m(X)M,
i.e., W ( 0 )= (0).
m(Ad(lc)X) 3 Anns(,)M,
kEKc
because Anns(,lM is stable under Ad(&). Thus we find I = Anns(,)M and in particular V ( X ) = 8. The last assertion follows immediately from Proposition 2.1. 0 Under the assumption in Proposition 3.1, the (S(g), &)-module !P is almost equal to N L N M*, in the sense that the support of the orthogonal l!P C KerT(0) is contained in the boundary 8 0 by Proposition 2.3 and Lemma 3.1.
3.3. Difleerential opemtor of gmdient type We wish to characterize N L N M* as the kernel of a certain differential operator on p of gradient type. For this, we first take an orthonormal basis (Xl,. . . ,X,) of p with respect to the Killing form B l p x p . Then, (X; ,... ,Xz) (cf. (3.4)) gives a basis of p*, dual to (XI ,... ,X,). For a multi-index (Y = ((~1,.. . ,a,)of nonnegative integers ( ~ (1 i 5 i 5 s), we set
Da := XFl.. .x;4,
(D*)* := (X;)al.. * ( X y .
339
Then, the elements D Q (resp. (D*)") with (a1= n form a basis of Sn(p) (resp. Sn(p*))for every integer n 2 1, where la1 := a1 a, denotes the length of a. Note that
+ +
a ( D a ) ( D * ) a= a! 6,,p (Kronecker's a(DQ)(Z*)= " n!(D*)a(Z)
and
(3.11)
--
for all a,/3 of length n, and 2 E p. Here, we set a! = al! . a,!.Note that the above functions 8 ( D Q ) ( D * ) pand a ( D a ) ( 2 * ) "are constant on p. We now introduce a gradient map V" of order n by
(V"f)(Y) :=
c
-+*)a a.
8 a ( D a )f (Y),
(3.12)
la[="
for f E S(p*) @ V,!. It is easy to observe that Vn f is independent of the choice of an orthonormal basis ( X i ) l l i l , . Furthermore, V" gives an (S(g), Kc)-homomorphism
S(p*) 8 v; 3 f
s V"f
E S(p*) 8 (Sn(p*)8 v;),
where S(p*) 8 (Sn(p*)@ V ): is looked upon as the space of polynomial .: functions on p with values in Sn(p*)8 V
Lemma 3.2. It holds that Vnf = 1 8 f for every f E Sn(p*)8VV,,where 1 denotes the identitp element of S(p*). Namely, V" f is the constant function on p with the value f E S" (p*) 8 V,* . Proof. It is enough to prove the lemma for f = (D*)a8 v* with 1/31 = n and 'u* E V .: In view of (3.11), Vnf turns to be 1 V" f (Y)= +*)" 8 (a(DQ)(D*)~)(Y)v* = (D*)P8 v* = f , a.
c
lal=n
which proves the lemma.
0
We note that our submodule N of S(p) 8 V, is finitely generated over
-
S(g), since the ring S(g) is Noetherian and since S(p) 8 V, = S(p) V,.
Hence, there exist a finite number of homogeneous Kc-submodules W, (u = 1 , . . . ,q ) which generate N over S(g):
N = S ( g ) . W l + . . . + S ( g ) . W , with W, c S i U ( p ) @ V ,
cN
(3.13)
for some integers i, 2 0 arranged as il < . .. < i,. For each u = 1 , . . . ,q, let P, denote the Kc-homomorphism from Si-(p*)@ V,! to W; defined by
Pu(h)(w) := ( w , h ) (wE Wu)
(3.14)
340
for h E Siu(p*) €3 V;. Here, (. , - ) is the dual pairing in (3.5). We now set a
and let us introduce an (S(g), Kc)-homomorphism
D : S(p*) €3 v: -+ S(p*) €3w*,
(3.15)
by putting Q
Pu(Viuf(Y)) (Y E p; f
( D f ) ( Y ):=
E S(p*) €3 V:).
(3.16)
u=l
Definition 3.1. We call D ' the differential operator of gradient type associated with (V,",W * ) . The space of solutions of the differential equation Df = 0 is characterized as follows.
Proposition 3.2. One gets N L = Ker'D. Hence, the kernel of the differential operator 2) is isomorphic to the Kc-finite dual M* of M N (S(p) €3 V T ) / N as , (S(g), Kc)-modules. Proof. Let f be a homogeneous element of S(p*) @V; of degree n. We are going to show that f lies in N L if and only if D f = 0. This will prove the proposition because both KerD and N L are graded (S(g), Kc)-submodules of S(p*) €3 v;. First, the condition f E N L is written as (Dwu, f ) = 0
(3.17)
for all D E Sn-iu(p) and w, E W, (u = 1 , . . . ,q), by noting that Sm(p)€3 V, is orthogonal to f E Sn(p*)€3 V; if m # n. Here Sn-iu(p) should be understood as ( 0 ) if n - i, < 0. Since the pairing ( - , is S(g)-invariant, (3.17) is equivalent to a )
P,(a(D)f) = 0 for all D
E
5 F i ~ ( p ) (u = 1 , . . . ,q).
(3.18)
We set 'D,f := Pu((Vif)(.)) E Sn-*,(p*) €3 W,*. Then, in view of Lemma 3.2, the left hand side of (3.18) turns to be
P u ( W > f )= Pu((ViUa(D)f)(Y)) = ( W ) ( D U f ) ) ( Y >(YE PI-
341
N L if and only if a(D)Vuf = o for all D E Sn+(p) (u = 1 , . . . ,q ) , or equivalently, V,f = 0 (u = 1 , . . . ,q). This means V f = 0 as desired.
We thus find that f lies in
0
Let us define a map u from p* x V,+to W *by 9
u ( X * , w * ):= x P u ( ( X * ) i u@v*) for (X*,w*) E p* x V * ,
(3.19)
u=l
which we call the symbol map of V. For any fixed X E p, we get the following characterization of W(O)*. Proposition 3.3. The Kc(X)-submodule W(O)*of V: (cf. (3.8)) is described as
W(O)*= Keru(X*, - ) := (w* E V,* I u(X*,w*) = 0},
(3.20)
where X * E p* corresponds to X E p by (3.4). This proposition can be proved just as in the proof of [30, Lemma 3.101 (see also the proof of Lemma 3.1 (1)). We omit the proof here. 4. Isotropy representation attached t o discrete series
In this section, we assume that g = t @ p is an equi-rank algebra (cf. [17]), i.e., rank g = rank t. By using our results in Section 3, we study the isotropy representations attached to irreducible (g, &)-modules of discrete series. This develops our work in [29]. 4.1. Discrete series
We begin with a quick review on the discrete series representations, and let us fix our notation. As is well known, the complex Lie algebra g has a &stable real form go such that go = to@ po
with to := t n go, po := p n go,
gives a Cartan decomposition of go. Such a real form go is unique up to Kc-conjugacy. Take a maximal abelian subalgebra of to, and we write t for the complexification ofto in t. Since g is an equi-rank algebra, t turns to be a Cartan subalgebra of g. We write A for the root system of (g, t). The subset of compact (resp. noncompact) roots will be denoted by Ac (resp. An).
342
Let G be a connected Lie group with Lie algebra of go such that Kc is the complexification of a maximal compact subgroup K of G. An irreducible unitary representation u of G is called a member of discrete series if the matrix coefficients of u are square-integrable on G. We are concerned with the irreducible (g, Kc)-modules X of discrete series, consisting of K-finite vectors for such u's. For example, we refer to [9], [18], and also [26, I, Section 11 for the parametrization and realization of discrete series representations. Now, let X = X Abe the (g, Kc)-module of discrete series with HarishChandra parameter A E t*. Since the parameter A is regular and real there exists a unique positive system A+ of A for which A is on dominant:
mt,,
A+ := { a E A I ( A , a ) > 0 ).
(4.1)
We denote by (T,V,) the unique lowest Kc-type of X which occurs in X with multiplicity one. Set A$ := A+ n A, (resp. A: := A+ n An).The A$-dominant highest weight X (say) for T is called the Blattner parameter of X.Then, X is expressed as X = A - p, pn with p, := (1/2) . CaEa:a and Pn := (1/2) * &.a;t P.
+
4.2. Results of Hotta-Parthasamthy
In what follows, we always assume that the Blattner parameter X of X is far from the walls (defined by compact roots) in the sense of [26, I, Definition 1.71. Let M = grX = en>oMn be the graded (S(g),&)-module defined through the lowest Kc-type V,. As in Section 3, we have a natural quotient map K : S(p) @ V, + M with N = Kerr. This subsection explains the structure of graded modules M , N , and M * N N L by interpreting the results of Hotta-Parthasarathy in [9]. For this, we first decompose the tensor product p €3 V, as
p @ V, = V,' @ V;
as Kc-modules,
where 'V denotes the sum of irreducible Kc-submodules of p 8 V, with highest weights of the form X f P (P E A:), respectively. The inclusion V,- L) p @ V, naturally induces a quotient map of Kc-modules:
P : p* €3 v: = (p €3 V,)*
+ (v,-)*.
(4.2)
Hereafter, we replace p' by p through the identification p = p* by the Killing form of g restricted to p x p.
343
Let B, be the Bore1 subgroup of Kc with Lie algebra 6, = t@CffEAcg a , where ga is the root subspace of g corresponding to a root a. We set
Then, we have p = p+ @ p- as vector spaces, and p- is stable under the action of B,. If U is a holomorphic representation of B,, the i-th cohomology space Hi(Kc/Bc; U )of Kc/B, with coefficients in the sheaf of holomorphic sections of the vector bundle Kc XB, U has a structure of Kc-module. The following theorem can be read off from the proof of [9, Theorem 11 by taking into account the Blattner multiplicity formula [7] for discrete series. (See also “1; [29].)
Theorem 4.1. (I) One has N = S(p)V,-. (2) The orthogonal N L of N in S(p)@V; coincides with the kernel of the diflerential operator 2) o n p of gradient-type associated with (V;, (VT-)*) :
for f E S(p) @V,*.Here {Xi}1siss is an orthonormal basis of p with respect to the Killing form. (3) For every integer n 2 0, the dual M i of M , is isomorphic to the cohomology space: Hq(Kc/Bc; Sn(p-) €3 (c--x--spc)
with q := dim Kc/B,,
as a Kc-module. Here ( c - ~ - z ~ =denotes the one dimensional B,-module corresponding to -A - 2p, E t*. Note that the claim (2) can be deduced from claim (1) by Proposition 3.2.
Remark 4.1. The maximal globalization of dual (g, Kc)-module X* of discrete series can be realized as the kernel space of an invariant differential operator V of gradient type (on the Riemannian symmetric space for (go, to)). The operator D in the above theorem gives the “polynomialization” of 8.
344
4.3. Description of associated cycle
We are going to apply Theorem 4.1 in order to describe the associated cycles for (g, Kc)-modules of discrete series. For a positive number c, we say that a linear form p on t satisfies the condition (FFW(c)) if (p, a) 3 c for all
a E A:.
(FFW(c))
Theorem 4.1 coupled with the Borel-Weil Bott theorem for the group Kc leads us to the following proposition(cf. [29, Section 6.1]),which is crucial to describe the associated cycle of X. Proposition 4.1. ( I ) Let vl; be a nonzero lowest weight vector of V,* of weight -A. Then, N L = Ker?) contains the Kc-submodule ( S ( p - - ) @ v ; ) ~ ~
generated by S(p-) @ vl;. (2) For any integer n 2 0, there exists a positive constant c, such that
holds if the Blattner parameter X satisfies the condition (FFW(c,)). Now, let 0 be the unique nilpotent Kc-orbit in p which intersects pdensely. Then one sees that 8 = Ad(Kc)p-. As before, we write I for the prime ideal of S(g) defining 8. It follows from the the claim (1) in Proposition 4.1 that Annq,)M C I, i.e., V ( X )3 8. Also, the same claim shows p- @ v; E Ker P,which can be easily verified by noting that -A - /3 (p E A:) cannot be a weight of (V;)*. Take an element X E 0 n p-. By Proposition 3.3, we find that the Kc(X)-module W(O)*= (M/m(X)M)* c V: consists exactly of all the vectors v* E V,* satisfying P(X 8 v*) = 0. Let N K ~ ( Xp-) , be the totality of elements k E Kc such that Ad(k)X E p- (cf. [3]). For any subset R of NK~(X, p-), we denote by Ux(R) the Kc(X)-submodule of V,* generated by R-' * v;: Ux(R) := ( R - l * ~ ; I ) ~ c ( x ) . Then, we readily find from p- 8 vl; E Ker P that
Ux(R)c W(O)*, and so ( Xn @UA(R))K,C 9, C ( N l ) , i , for every n 2 0. Moreover one gets the equality ( X n @ Ux(R) )K@ = (S"(p-)@v?, ) K ~ ,
(4.6)
(4.7)
if Ad(R)X c p- is Zariski dense in p-. This is true when R equals the whole N ~ ~ ( x , p -because ), Ad(NK,(X,p-))X = 0 0 p- is dense in p-.
345
As in Section 3, we take homogeneous generators Di (i = 1 , . .. ,T ) of the ideal I such that degDi = n(i). We set c ( I ) := maxi(c,(i)). By virtue of Proposition 3.1 together with (4.5), (4.6) and (4.7), we come to the following conclusion.
Theorem 4.2. Assume that the Blattner parameter X of discrete series X is far from the walls and that it satisfies the condition (FFW(c)) with c = c(I). (1) One gets I = Anns(,,)M and so V ( X ) = Ad(KC)p- = 8.Moreover, the K c ( X )-module W *contmgredient to the isotropy representation (wo,W ) is described as
w *= W(O)*= {v* E v; I P ( X €3 v * ) = O } ,
(4.8)
where X E 0 n p-, and P : p €3 V,* + (V;)* is the Kc-homomorphism in (4.2). (2) Let R be a subset of N K C ( X ,p-) such that Ad(R)X is Zariski dense in p-. Then, the Kc(X)-submodule Ux(R) = (R-' . V ; ) K ~ ( X )c W *is
exhaustive in the following sense: for every integer n 2 0, one has
(X"@ W * ) K=~(x"C3 % ( R ) ) K ~
(4.9)
i f X satisfies FFW(c,).
Remark 4.2. (1) The assertions I = Ann.q,lM and V ( X )= Ad(Kc)phave been obtained in [29]. But, in that paper, we did not discuss the possibility of applying the results to describe the isotropy representation. (2) One should get a result similar to Theorem 4.2, more generally for the derived functor modules d4(X). (3) Compare Theorem 4.2 (1) with Chang's result [3, Proposition 1.41 established by means of the localization theory of Harish-Chandra modules. 4.4. Submodule
Ux(Qc)
In this subsection, we give a natural choice of R c N K ~ ( X , ~ for - ) which we expect to have the property (4.9). Let ll be the set of simple roots in A+. We write S = 1T n A, for the totality of compact simple roots. Then, there exists a unique element Hs E t such that
346
The adjoint action of H s yields a gradation on the Lie algebra g as g=
@ g(j)
with g ( j ) := ( 2 E gl (adHs)Z = j Z } .
j
Here j runs through the integers such that Ijl 2 ~ ( H s with ) the highest root 6. Note that
e = @ dj), j:even
P = @ g(j) with j:odd
Pk =
@
g(fj).
j>O,odd
Now, we set
q := @ g ( j ) , I:= g(0)
ct
and u := @ g ( j ) .
j 50
j
Then, q = I@ u gives the Levi decomposition of the standard parabolic subalgebra q of g associated with the subset S of rI. We write Q (resp. Q,) for the parabolic subgroup of Gc := GEd (resp. of Kc) with Lie algebra q (resp. q n t). The group Q (resp. Q,) admits the Levi decomposition Q = LU (resp. Qc = L,U,), where L and U (resp. L, and Uc) are the connected subgroups of Q (resp. Q,) with Lie algebras I and u (resp. I and u n t) respectively. Note that Ad(Lc) = L. The parabolic subgroup Q acts on its nilradical u, and so Q , acts on p- = p nu by the adjoint action. Thus, Qc is contained in N K @ ( Xp-) , for all X E p-, and the corresponding Kc(X)-submodule Ux(Qc)of V: turns to be (4.10) Here, (Vfc)*= U(I)vf, denotes the irreducible L,-submodule of V: generated by the lowest weight vector vf,. We can now apply Theorem 4.2 to deduce Corollary 4.1. Under the assumption in Theorem 4.2, the Kc(X)submodule Ux(Qc) of W * is exhaustive in the sense of (4.9), i f p- i s a prehomogeneous vector space under the adjoint action of the group Q,, and if X E 0 n p- lies in the open Q,-orbit in p-. 4.5. Relation to the Richardson orbit
We end this article by looking at the condition for p- in Corollary 4.1, and also some related conditions, in relation to the Richardson Gc-orbit associated with the parabolic subalgebra q. First, let us recall some basic facts on the Richardson orbit (cf. [12, Chapter 51). The Gc-stable subset Gc u (C g) forms an irreducible affine
347
variety of g whose dimension is equal to 2 dimu. Noting that G c - u consists of nilpotent elements only, there exists a unique Gc-orbit d such that
by the finiteness of the number of nilpotent Gc-orbits in g. d is called the Richardson Gc-orbit associated with q. The parabolic subgroup Q acts on u prehomogenenously, and d n u turns to be a single Q-orbit in u. Moreover, the centralizer in g of any element X E d n u is contained in q. Now, we havetwo nilpotent Gc-orbits Gc . 0 and d with the closure relation G c - O C 6. By virtue of a result of Kostant-Rallis [15, Proposition 51, this relation implies that
1 1 dimO= -dimGc.O< -dimd=dimu. 2 2
(4.11)
In particular, we find that the Gelfand-Kirillov dimension dimV(X) = dim 0 of discrete series X cannot exceed dim u. The following proposition tells us when these two orbits turn to be equal. Proposition 4.2. The following three conditions (a), (b) and (c) o n the positive system A+ = (a1 ( h , a )> 0) are equivalent with each other:
(a) Gc . O = d,
(b) dim 0 = dim u,
(c)
d n p- # 0.
In this case, 0 n p- is a single open Q,-orbit in p-, and so one gets the conclusion of Corollary 4.1. Proof. The equivalence (a) H (b) is a direct consequence of (4.11). The condition (a) immediately implies (c), since 0 (C Gc .0 = 6 )contains an element of p-. Conversely, if d n p- # 0, this is a nonempty open subset of p-, since d n p- = (d n u) n p- with d n u open in u. Hence, 6 n pintersects 0. We thus get (c) 3 (a). This proves the equivalence of three conditions in question. Next, we assume the condition (b) (H (a) % (c)), and let X be any element of 0 n p-. We write a,(X) for the centralizer of X in a Lie subalgebra 5 of g. By noting that a,(X) c q, the dimension of the Q,-orbit Ad(Q,)X is calculated as
dimAd(Q,)X = dimq n t - dimasne(X) = (dime-dimunt) -dimJe(X) = d i m 0 - dimu n t = dimu - dimu n t = dimp-,
348
where we used the condition (b) for the forth equality. This shows that the orbit Ad(Q,)X is open in p- for every X E 0 n p-. We thus find that 0 n p- forms a single Q,-orbit, because of the uniqueness of the open Q,-orbit in p-. 0
Remark 4.3. Each of the conditions (a), (b) and (c) in Proposition 4.2is equivalent to Assumption 2.5 in [2]concerning the generically finiteness of the moment map defined on the conormal bundle Ti,(Gc/Q), where 21 is a closed Kc-orbit in Gc/Q through the origin eQ. Suggested by Corollary 4.1 and Proposition 4.2, let us consider the following three conditions on p- which depends on the choice of a positive system A+:
b n p - # 0 (w d i m U = d i m u 0 n p- is a single Q,-orbit,
Gc.U=b),
p- is a prehomogeneous vector space under Ad(Q,).
(C1) (C2) (C3)
Proposition 4.2 says (Cl)=+ (C2),and the implication (C2)=+ (C3)is obvious. As for the conditions (C2)and (C3),we can show the following
Proposition 4.3. One gets (C3)if 0 ng( -1) Ad(Q,)(O n g(-l)) = 0 n p- assures (C2).
# 0. Moreover, the equality
Proof. Let X E 0 n g(-l). Since 0 = Ad(Kc)X contains a nonempty open subset of p-, we find that [t,X]3 p-. We set t+ := @>0g(2j). Then t = t n q @ t+ is a direct sum of vector spaces. Then it follows from the assumption X E g(-1) that [e+,X] C p+ and [t n q,X] C p-. We thus obtain
p- = [e, X] n p- = [en4, XI. Hence Ad(Q,)X is open in p-, and one gets (C3). The above argument shows that any element X E 0 n g(-1) lies in the 0 unique open Q,-orbit in p-. This proves the latter claim, too. Following Gross-Wallach [5],we say that a discrete series (g, &)-module X is small if ~ ( H s 5) 2, or equivalently, g ( j ) = (0) if Ijl 2 3. Here 6 is the highest root of A+ as before. In this case, one has p- = g(-1), and so the above proposition implies
349
Corollary 4.2. The subspace p- corresponding t o a small discrete series admits the property (C2). Remark 4.4. By case-by-case analysis, Chang [3] proved the property (C2)for any discrete series representations of simple Lie groups of R-rank one.
4.6. Condition (Cl)f o r S U ( p , q )
It should be important to study when p- admits the properties (Cl),(C2) and (C3),respectively. Toward this direction, we end this paper by giving an explicit, combinatorial criterion for the condition (Cl)in case of G = S U ( p , q) with n = p q. In this case, we have g = sI(n,C), and let
+
a1
= € 1 - €2,
a2
= €2 - €3,
. . . , an-1
= €,-I
-&,
be the simple roots for (a, t) by the standard notation of Bourbaki. Then p-’s in question are in one-one correspondence to the set of noncompact .. . , where nl, . . . ,nt-1, and positive roots {a,,, a,,+,,, nt := n - (nl . . . nt-1) are positive integers such that
+ +
C
nj
= p or q.
j:odd
Theorem 4.3. p- satisfies the condition (Cl)if and only i f the corresponding partition (nl,n2,.. . ,nt) of n is unimodal, that is, there exists a positive integer k (15 k 5 t ) such that n1
5 - .- 5 n k - 1 5 n k 2 n k + 1 2
2 nt.
If p , q 5 3, every partition (nl,. . . ,nt) is unimodal, and we always get the property (Cl). On the contrary, the partition (2,1,2)for SU(4,l)is not unimodal, and (Cl)fails in this case, where d i m 0 = 7 < 8 = dimu. We will discuss the detail elsewhere. Acknowledgments The author thanks Hiroyuki Ochiai for stimulating discussion.
References 1. W. Borho and J.-L. Brylinski, Differential operators on homogeneous spaces. I11 Characteristic varieties of Harish-Chandra modules and of primitive ideals, Invent. Math., 80 (1986), 1-68.
350 2. J.-T. Chang, Characteristic cycles of holomorphic discrete series, Trans. Amer. Math. SOC.,334 (1992), 213-227. 3. J.-T. Chang, Characteristic cycles of discrete series for R-rank one groups, Trans. Amer. Math. SOC.,341 (1994), 603-622. 4. A. Dvorsky and S. Sahi, Tensor products of singular representations and an extension of the 0-correspondence, Selecta Math., New Series 4 (1998), 11-29. 5. B. Gross and N.R. Wallach, Restriction of small discrete series representations to symmetric subgroups, in: The mathematical legacy of HarishChandra (Baltimore, MD, 1998), 255-272, Proc. Sympos. Pure Math., 68, Amer. Math. SOC.,Providence, RI, 2000. 6. A. Gyoja and H. Yamashita, Associated variety, Kostant-Sekiguchi correspondence, and locally free U(n)-action on Harish-Chandra modules, J. Math. SOC.Japan, 51 (1999), 129-149. 7. H. Hecht and W. Schmid, A proof of Blattner’s conjecture, Invent. Math. 31 (1975), 129-154. 8. T. Hirai and H. Yamashita, Hyougenron nytimon seminar (Introductory Seminar on Representation Theory), in Japanese, Yiiseisha, Tokyo, 2003, pp. i-xi and 1-329, ISBN4-7952-6898-3. 9. R. Hotta and R. Parthasarathy, Multiplicity formulae for discrete series, Invent. Math. 26 (1974), 133-178. 10. R. Howe, Remarks on classical invariant theory, Trans. Amer. Math. SOC., 313 (1989), no. 2, 539-570. 11. R. Howe, Transcending classical invariant theory, J. Amer. Math. SOC.,2 (1989), no. 3, 535-552. 12. J.-E. Humphreys, Conjugacy classes in semisimple algebraic groups, Mathematical Surveys and Monographs, 43,American Mathematical Society, Providence, RI, 1995. 13. S. Kato, On the Gelfand-Kirillov dimension and Bernstein degree for unitary highest weight modules (in Japanese), Master Thesis, Kyushu University, 2000, 77 pages. 14. S. Kato and H. Ochiai, The degrees of orbits of the multiplicity-free actions, in: “Nilpotent orbits, associated cycles and Whittaker models for highest weight representations”, Astkrisque, 273 (2001), 139-158. 15. B. Kostant and S. Rallis, Orbits and representations associated with symmetric spaces, Amer. J. Math., 83 (1971), 753-809. 16. K. Nishiyama, H. Ochiai and K. Taniguchi, Bernstein degree and associated cycles of Harish-Chandra modules - Hermitian symmetric case, in: “Nilpotent orbits, associated cycles and Whittaker models for highest weight representations”, Astkrisque, 273 (2001), 13-80. 17. H. Rubenthaler, A new proof of the Borel-de Siebenthal theorem, the classification of equi-rank groups, and some compact and semi-compact dual pairs, to appear in Hokkaido Math. J., 33,no. 1, 2004. 18. W. Schmid, Some properties of square-integrable representations of semisimple Lie groups, Ann. of Math., 102 (1975), 535-564. 19. K. Taniguchi, Discrete series Whittaker functions of SU(n,1) and Spin(2n, l ) ,
351 J. Math. Sci. Univ. Tokyo, 3 (1996), no. 2, 331-377. 20. K. Taniguchi, On discrete series Whittaker functions SOo(2n, 2) (Japanese), in: ‘LAutomorphicforms on Sp(2; R) and SU(2,2), 11 (Kyoto, 1998)”, RIMS Kakytiroku, 1094 (1999), 11-27. 21. D.A. Vogan, Gelfand-Kirillovdimension for Harish-Chandra modules, Invent. Math. 48 (1978), 75-98. 22. D.A. Vogan, Associated varieties and unipotent representations, in: “Harmonic Analysis on Reductive Groups (W.Barker and P.Sally e&.)”, Progress in Math., Vol. 101, Birkhauser, 1991, 315-388. 23. D. A. Vogan, The method of coadjoint orbits for real reductive groups, an: “Representation Theory of Lie Groups (eds. J. Adams and D. Vogan)”, 177238, IAS/Park City Mathematics Series 8, AMS, 2000. 24. A. Wachi and H. Yamashita, Isotropy representations for singular unitary highest weight modules (tentative), in preparation. 25. H Yamashita, Finite multiplicity theorems for induced representations of semisimple Lie groups 11: Applications to generalized Gelfand-Graev representations, J. Math. Kyoto. Univ., 28 (1988), 383-444. 26. H. Yamashita, Embeddings of discrete series into induced representations of semisimple Lie groups, I: General theory and the case of S U ( 2 , 2 ) , Japan. J. Math.,l6 (1990), 31-95; 11: generalized Whittaker models for S U (2, 2), J . Math. Kyoto Univ., 3 1 (1991), 543-571. 27. H. Yamashita, Criteria for the finiteness of restriction of U(g)-modules to subalgebras and applications to Harish-Chandra modules: a study in relation to the associated varieties, J. Funct. Anal., 1 2 1 (1994), 296-329. 28. H. Yamashita, Associated varieties and Gelfand-Kirillov dimensions for the discrete series of a semisimple Lie group, Proc. Japan Acad., 70A (1994), 50-55. 29. H. Yamashita, Description of the associated varieties for the discrete series representations of a semisimple Lie group: An elementary proof by means of differential operators of gradient type, Comment. Math. Univ. St. Paul., 47 (1998), 35-52. 30. H. Yamashita, Cayley transform and generalized Whittaker models for irreducible highest weight modules, in: “Nilpotent orbits, associated cycles and Whittaker models for highest weight representations”, Asthrisque, 273 (2001), 81-137. 31. H. Yamashita, Associated cycles of Harish-Chandra modules and differential operators of gradient type, RIMS Kijkyiiroku, 1183 (2001), 157-167. 32. H. Yamashita, Isotropy representations attached to the associated cycles of Harish-Chandra modules, RIMS Kakyiiroku, 1238 (2001), 233-247. 33. H. Yamashita, Isotropy representation and projection to the PRVcomponent, RIMS Kakyiiroku, 1294 (2002), 62-71. 34. T. Yoshinaga and H. Yamashita, The embeddings of discrete series into principal series for an exceptional real simple Lie group of type Gal J. Math. Kyoto Univ., 36 (1996), 557-595.