Developments and Trends in Infinite-Dimensional Lie Theory (Progress in Mathematics, Volume 288)

Progress in Mathematics Volume 288 Series Editors Hyman Bass Joseph Oesterlé Alan Weinstein Karl-Hermann Neeb Arturo...

Author: Karl-Hermann Neeb | Arturo Pianzola

22 downloads 784 Views 3MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

Progress in Mathematics Volume 288

Series Editors Hyman Bass Joseph Oesterlé Alan Weinstein

Karl-Hermann Neeb Arturo Pianzola Editors

Developments and Trends in Infinite-Dimensional Lie Theory

Editors Karl-Hermann Neeb Department of Mathematics Friedrich-Alexander-Universität Erlangen-Nürnberg Bismarckstrasse 1 1/2 91054 Erlangen, Germany [email protected]

Arturo Pianzola Department of Mathematical Sciences University of Alberta Edmonton, Alberta T6G 2G1, Canada [email protected] and Instituto Argentino de Matemática Saavedra 15 1083 Buenos Aires, Argentina [email protected]

ISBN 978-0-8176-4740-7 e-ISBN 978-0-8176-4741-4 DOI 10.1007/978-0-8176-4741-4 Springer New York Dordrecht Heidelberg London Library of Congress Control Number: 2010938915 Mathematics Subject Classification (2010): 17Bxx, 20Gxx, 22Exx, 58Bxx, 58Dxx © Springer Science+Business Media, LLC 2011 All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights.

Printed on acid-free paper

www.birkhauser-science.com

Preface

Since the advent of Kac–Moody algebras in the late 1960s, infinite-dimensional Lie theory has become one of the core areas of modern mathematics. The global theory of infinite-dimensional Lie groups developed through several stages, starting with the class of Banach–Lie groups in the 1960s. In the 1980s it became apparent that more general classes of groups, such as smooth loop groups, were needed to meet the requirements of various kinds of applications. The core of infinite-dimensional Lie theory lies at the translation of the global group picture to the infinitesimal Lie algebra picture, where the means are often based on geometric structures such as fiber bundles on (possibly infinitedimensional) manifolds on which the groups act by transformations and on which the Lie algebra is realized by vector fields. The idea for the present book came up during an Oberwolfach meeting on infinite-dimensional Lie theory. Its purpose is to describe some of the current developments and trends in infinite-dimensional Lie theory. Many of these developments have been motivated by connections with other branches of mathematics. Accordingly, the book is divided into three main parts: (A) Infinite-dimensional Lie (super-)algebras; (B) Geometry of infinite-dimensional Lie (transformation) groups; (C) Representation theory of infinite-dimensional Lie groups. The different contributions clearly demonstrate the breadth of Lie theory in the infinite-dimensional context. It is also our intention to display the connectedness among the algebraic, geometric, and analytic branches, which are different faces of a general theory with many different aspects. Acknowledgments. All contributions to this book have been refereed, with some of them going through several stages of rewriting until they reached their final form. The editors are grateful to the referees for their efficient help. We thank the Mathematische Forschungsinstitut Oberwolfach for hosting the conference where this book was conceived.

vi

Preface

Last but not least we express our appreciation to Birkh¨ auser Boston, in particular Ann Kostant and Tom Grasso as acting representatives for a remarkably pleasant cooperation.

Karl-Hermann Neeb Arturo Pianzola Banff March 2008

Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

v

Part A Infinite-Dimensional Lie (Super-)Algebras Isotopy for Extended Affine Lie Algebras and Lie Tori Bruce Allison and John Faulkner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3

Remarks on the Isotriviality of Multiloop Algebras Philippe Gille and Arturo Pianzola . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 Extended Affine Lie Algebras and Other Generalizations of Affine Lie Algebras – A Survey Erhard Neher . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 Tensor Representations of Classical Locally Finite Lie Algebras Ivan Penkov and Konstantin Styrkas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 Lie Algebras, Vertex Algebras, and Automorphic Forms Nils R. Scheithauer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Kac–Moody Superalgebras and Integrability Vera Serganova . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 Part B Geometry of Infinite-Dimensional Lie (Transformation) Groups Jordan Structures and Non-Associative Geometry Wolfgang Bertram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 Direct Limits of Infinite-Dimensional Lie Groups Helge Gl¨ ockner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243

viii

Contents

Lie Groups of Bundle Automorphisms and Their Extensions Karl-Hermann Neeb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 Gerbes and Lie Groups Christoph Schweigert and Konrad Waldorf . . . . . . . . . . . . . . . . . . . . . . . . . . 339 Part C Representation Theory of Infinite-Dimensional Lie Groups Functional Analytic Background for a Theory of InfiniteDimensional Reductive Lie Groups Daniel Beltit¸a ˘ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367 Heat Kernel Measures and Critical Limits Doug Pickrell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393 Coadjoint Orbits and the Beginnings of a Geometric Representation Theory Tudor S. Ratiu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417 Infinite-Dimensional Multiplicity-Free Spaces I: Limits of Compact Commutative Spaces Joseph A. Wolf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483

Part A

Infinite-Dimensional Lie (Super-)Algebras

Isotopy for Extended Affine Lie Algebras and Lie Tori Bruce Allison1∗ and John Faulkner2 1

Department of Mathematics and Statistics, University of Victoria, Victoria BC V8W 3P4 Canada, [email protected] 2 Department of Mathematics, University of Virginia, Charlottesville VA 22904-4137 USA, [email protected]

Summary. Centreless Lie tori have been used by E. Neher to construct all extended affine Lie algebras (EALAs). In this article, we study isotopy for centreless Lie tori, and show that Neher’s construction provides a 1–1 correspondence between centreless Lie tori up to isotopy and families of EALAs up to isomorphism. Also, centreless Lie tori can be coordinatized by unital algebras that are in general nonassociative, and, for many types of centreless Lie tori, there are classical definitions of isotopy for the coordinate algebras. We show for those types that an isotope of a Lie torus is coordinatized by an isotope of its coordinate algebra, thereby connecting the two notions of isotopy. In writing the article, we have not assumed prior knowledge of the theories of EALAs, Lie tori or isotopy. Key words: isotope, extended affine Lie algebra, Lie torus, nonassociative algebra, Jordan algebra, alternative algebra. 2000 Mathematics Subject Classifications: Primary 17B65, 17A01. Secondary 17B60, 17B70, 17C99, 17D05.

1 Introduction If A is a unital associative algebra and u is invertible in A, one can define an algebra A(u) , called the u-isotope of A, which is equal to A as a vector space but has a new product x ·u y = xuy. This isotope A(u) is again unital and associative but with a shifted identity element u−1 . More generally there are definitions of isotope for several other classes of unital nonassociative algebras, notably Jordan algebras [Mc2, §I.3.2], alternative algebras [Mc1] and associative algebras with involution [Mc2, §I.3.4]. In each case, the isotope is obtained very roughly by shifting the identity element in the algebra, and two algebras are said to be isotopic if one is isomorphic to an isotope of the ∗

Bruce Allison gratefully acknowledges the support of the Natural Sciences and Engineering Research Council of Canada.

K.-H. Neeb and A. Pianzola (eds.), Developments and Trends in Infinite-Dimensional Lie Theory, Progress in Mathematics 288, DOI 10.1007/978-0-8176-4741-4_1, © Springer Basel AG 2011

3

4

Bruce Allison and John Faulkner

other. In the associative case, the u-isotope A(u) is isomorphic to A under left multiplication by u, and therefore isotopy has not played a role in associative theory. That is not true in general though, and in particular isotopy plays an important role in Jordan theory (see for example [Mc2, §II.7]). In contrast, isotopes and isotopy have not been defined for Lie algebras, for the evident reason that Lie algebras are not unital.2 In this article we study notions of isotope and isotopy, which were recently introduced in [ABFP2] for a class of graded Lie algebras called Lie tori. The point to emphasize here is that forming an isotope of a Lie torus does not change the multiplication at all, but rather it shifts the grading. We are primarily interested in the case in which the Lie torus is centreless, and there are two basic reasons why we are interested in isotopy in this case. First, centreless Lie tori arise naturally in the construction of families of extended affine Lie algebras (EALAs), and we see in this article that isotopes and isotopy play a natural and fundamental role in the theory of both of these classes of Lie algebras. In fact, we show in section 6 using a construction of E. Neher, that there is a 1–1 correspondence between centreless Lie tori up to isotopy and families of EALAs up to bijective isomorphism. (This terminology is explained in section 6.) Second, any Lie torus is by definition graded by the group Q × Λ, where Q is the root lattice of a finite irreducible root system ∆ and Λ is (in this article) a finitely generated free abelian group. Depending on the type of ∆, centreless Lie tori are coordinatized by unital algebras that are in general nonassociative. For example, in types A1 , A2 , Ar (r ≥ 3) and Cr (r ≥ 4) the coordinate algebras are respectively Jordan algebras, alternative algebras, associative algebras and associative algebras with involution. It turns out that an isotope of a centreless Lie torus is coordinatized by an isotope of its coordinate algebra, thereby connecting our notion of isotopy for Lie tori with the classical notions for unital nonassociative algebras. This fact can then be used to obtain necessary and sufficient conditions for two centreless Lie tori of the same type to be isotopic (in terms of their coordinate algebras). We describe these results in detail for types A1 , A2 , Ar (r ≥ 3) and Cr (r ≥ 4) in sections 8, 9, 10, and 11, and we remark very briefly on the other types in section 12. Section 12 also includes a brief discussion of an alternate approach, using multiloops, to constructing centreless Lie tori and studying isotopy. In writing this article, we have not assumed prior knowledge of the theories of EALAs, Lie tori or isotopy. In fact, we hope that this article will serve as a useful introduction to these theories and their interconnections. For this reason, we begin in sections 2, 3, 4 and 5 by recalling some of the important definitions and facts about EALAs, Lie tori, and isotopy for Lie tori. For the same reason, we have included a brief discussion of coordinatization in 2

There is a notion of isotopy that makes sense for nonunital algebras and hence Lie algebras [A], but it is quite different from the one discussed in this article.

Isotopy for EALAs and Lie Tori

5

section 7; and in sections 8 and 9 we have included some arguments that are familiar to experts in Jordan and alternative algebra theory, but may be less familiar to others. Acknowledgements: We thank Stephen Berman, Erhard Neher, and Arturo Pianzola for stimulating conversations about the topics of this paper. We also thank Erhard Neher for his helpful comments about a preliminary version of this article. Assumptions and notation: Throughout this work we assume that F is a field of characteristic 0. All algebras are assumed to be algebras over F, and all algebras (except Lie algebras) are assumed to be unital. We also assume that Λ is a finitely generated free abelian group. We denote the rank of Λ, which is an integer ≥ 0, by rank(Λ). If E is a Lie algebra and H is an ad-diagonalizable subalgebra of E, we let Eρ = { x ∈ E | [h, x] = ρ(h)x for h ∈ H } for ρ in the dual space H ∗ of H. Then, we define ∆(E, H) = { ρ ∈ H ∗ | Eρ = 0 } and we call elements of ∆(E, H) roots of E relative to H; so in particular, 0 is a root (if E = 0). Some terminology for graded algebras: If A = ⊕λ∈Λ Aλ is a Λ-graded algebra, the Λ-support of A is suppΛ (A) = { λ ∈ Λ | Aλ = 0 }. If Γ is a subgroup of Λ and A is a Γ -graded algebra or vector space, we regard A as Λ-graded by setting Aλ = 0 for λ ∈ Λ \ Γ . If A is a Λ-graded algebra and A is a Λ -graded algebra, we say that A and A are isograded-isomorphic, written A ig A , if there exists an algebra isomorphism η : A → A and a group isomorphism ηgr : Λ → Λ such that η(Aλ ) = Aηgr (λ) for λ ∈ Λ; in that case ηgr is uniquely determined by η if suppΛ (A) = Λ. (Here suppΛ (A) denotes the subgroup of Λ generated by suppΛ (A).) If A and A are Λ-graded algebras, we say that A and A are graded-isomorphic, written A Λ A , if there is an algebra isomorphism η : A → A that preserves the grading (that is ηgr = id). A Λ-graded algebra with involution is a pair (A, ι), where A is a Λ-graded algebra and ι is an involution (antiautomorphism of period 2) of A that is graded. There is an evident extension of the terms isograded-isomorphic and graded-isomorphic and the notations ig and Λ for graded algebras with involution. Finally, the group algebra F[Λ] = ⊕λ∈Λ Ftλ of Λ with tλ tλ = tλ+λ is naturally a Λ-graded algebra. If we choose a basis { λ1 , . . . , λn } for Λ, ±1 then F[Λ] is isograded-isomorphic to the algebra F[t±1 1 , . . . , tn ] of Laurent n polynomials with its natural Z -grading.

2 Extended affine Lie algebras The definition of an EALA has evolved from [HT], where these algebras were introduced under the name elliptic irreducible quasi-simple Lie algebra, and from [BGK] and [AABGP], where many of their properties were developed. We will use the definition given in [N2], which has the important advantage that it makes sense over any field of characteristic 0. Some of the facts about EALAs mentioned below were proved in the setting of [AABGP], but they can be verified in a similar fashion in our more general setting.

6


We recall that an extended affine Lie algebra (EALA) is a triple (E, ( | ), H), where E is a Lie algebra over F, ( | ) is a nondegenerate invariant symmetric bilinear form on E, and H is a finite dimensional nonzero self-centralizing ad-diagonalizable subalgebra of E, such that a list of axioms labeled as (EA3)– (EA6) are satisfied [N2]. For the purposes of this article we do not need the precise statement of these axioms, except to say that they are modeled after the properties of finite dimensional split simple Lie algebras and affine KacMoody Lie algebras. If (E, ( | ), H) is an EALA, we also say that E is an EALA with respect to ( | ) and H, or simply that E is an EALA. Roots of E relative to H are simply called roots of E. If E is an EALA, we can as usual transfer the restriction of ( | ) to H to a nondegenerate form ( | ) on the dual space H ∗ . If R = ∆(E, H), V = spanF (R), and rad(V ) is the radical of the restriction of ( | ) to V , then the image of R in V / rad(V ) is a finite irreducible root system (including 0) whose type is called the type of E. (See the beginning of §3 below for our conventions about finite irreducible root systems.) A root ρ of an EALA E is called isotropic if (ρ, ρ) = 0, and otherwise called nonisotropic. According to one of the axioms for an EALA, the group generated by the isotropic roots is a free abelian group of finite rank, and its rank is called the nullity of E. Then the extended affine Lie algebras of nullity 0 and 1 are precisely the finite dimensional split simple Lie algebras and the affine Kac-Moody Lie algebras, respectively. (See [ABFP2, Remark 1.2.4] for nullity 0 and [ABGP] for nullity 1.) If (E, ( | ), H) is an EALA, then so is (E, a( | ), H) for a ∈ F× , and it is sometimes convenient to adjust the form in this way. For this reason, the following is a natural notion of isomorphism for EALAs. Definition 2.1. If (E, ( | ), H) and (E , ( | ) , H ) are EALAs, an isomorphism from (E, ( | ), H) onto (E , ( | ) , H ) is a Lie algebra isomorphism χ : E → E such that (χ(x)|χ(y)) = a(x|y) for some a ∈ F× and all x, y ∈ E, and χ(H) = H . If such a map exists we say that (E, ( | ), H) and (E , ( | ) , H ) are isomorphic. For short we often say that χ is an EALA isomorphism from E onto E and that E and E are isomorphic as EALAs. Definition 2.2. If E is an EALA, the core of E is the subalgebra Ec of E that is generated by the root spaces of E corresponding to nonisotropic roots, and the centreless core of E is Ecc := Ec /Z(Ec ).3

3 Lie tori With the construction of families of EALAs in mind, Yoshii gave a definition of a Lie torus in [Y6]. An equivalent definition, which we recall next, was given by Neher in [N1]. 3

We denote the centre of a Lie algebra L by Z(L). L is called centreless if Z(L) = 0.


7

It will be convenient for us to work with root systems that contain 0. So by a finite irreducible root system we will mean a finite subset ∆ of a finite dimensional vector space X over F such that 0 ∈ ∆ and ∆× := ∆ \ { 0 } is a finite irreducible root system in X in the usual sense (see [B2, Chap VI, §1, no. 1, Definition 1]). With this convention we will assume for the rest of this section that ∆ is a finite irreducible root system in a finite dimensional vector space X over F. Recall that ∆ is said to be reduced if 2α ∈ / ∆× for α ∈ ∆× . If ∆ is reduced then ∆ has type A ( ≥ 1), B ( ≥ 2), C ( ≥ 3), D ( ≥ 4), E6 , E7 , E8 , F4 , or G2 , whereas if ∆ is not reduced, ∆ has type BC ( ≥ 1) [B2, Chapter VI, §4 and 14]. We let ∆ind := { 0 } ∪ { α ∈ ∆× | 12 α ∈ / ∆ }, in which case ∆ind is an irreducible reduced root system in X, and ∆ind = ∆ if ∆ is reduced. Also we let Q = Q(∆) := spanZ (∆) be the root lattice of ∆. If L = (α,λ)∈Q×Λ Lλα is a Q × Λ-graded algebra, then L = λ∈Λ Lλ is Λ-graded and L = α∈Q Lα is Q-graded, with Lλ =

α∈Q

Lλα for λ ∈ Λ

and

Lα =

λ∈Λ

Lλα for α ∈ Q.

Definition 3.1. A Lie Λ-torus of type ∆ is a Q × Λ-graded Lie algebra L = λ (α,λ)∈Q×Λ Lα over F which satisfies: (LT1) Lα = 0 for α ∈ Q \ ∆. (LT2) (i) If 0 = α ∈ ∆ind , then L0α = 0. (ii) If 0 = α ∈ Q, λ ∈ Λ and Lλα = 0, then there exist elements eλα ∈ Lλα and fαλ ∈ L−λ −α such that Lλα = Feλα , and

λ L−λ −α = Ffα ,

[[eλα , fαλ ], xβ ] = β, α∨ xβ

(1)

for xβ ∈ Lβ , β ∈ Q. (LT3) L is generated as an algebra by the spaces Lα , α ∈ ∆× . (LT4) Λ is generated as a group by suppΛ (L). In that case the Q-grading of L is called the root grading of L, and the Λgrading of L is called the external grading of L. If ∆ has type X , we also say that L has type X . Remark 3.2. Suppose L is a Lie Λ-torus of type ∆. (i) It is shown in [ABFP2, Prop. 1.1.10] that suppQ (L) equals either ∆ or ∆ind .

8


(ii) It is sometimes convenient to assume the following additional axiom: (LT5) suppQ (L) = ∆. Note that by (i), (LT5) holds automatically if ∆ is reduced. Also, if (LT5) does not hold, then we can replace ∆ by ∆ind , in which case (LT5) holds. Thus, when convenient there is no loss of generality in assuming (LT5). In the study of Lie tori it is not convenient to fix a particular realization of the root system ∆ or a particular identification of the group Λ with Zn . For this reason, the following is a natural notion of isomorphism for Lie tori. Definition 3.3. If L is a Lie Λ-torus of type ∆ with Q = Q(∆) and L is a Lie Λ -torus of type ∆ with Q = Q(∆ ), L and L are said to be bi-isogradedisomorphic or bi-isomorphic for short, if there is an algebra isomorphism ϕ : L → L , a group isomorphism ϕr : Q → Q, and a group isomorphism ϕe : Λ → Λ such that ϕ (λ) ϕ(Lλα ) = L ϕer (α) for α ∈ Q and λ ∈ Λ. In that case we write L bi L and we call ϕ a bi-isograded-isomorphism, or a bi-isomorphism for short. Since suppQ (L) = Q and suppΛ (L) = Λ, the maps ϕr and ϕe are uniquely determined. Remark 3.4. Suppose that ϕ : L → L is a bi-isomorphism as in Definition 3.3. We did not assume that ϕr carries ∆ onto ∆ . However, this holds automatically, if L and L satisfy (LT5). In particular, if L and L satisfy (LT5) and ∆ = ∆, then ϕr ∈ Aut(∆). Definition 3.5. Suppose that L is a centreless Lie Λ-torus of type ∆. We set g = L0

and

h = L00 .

Then g is a finite dimensional simple Lie algebra and h is a splitting Cartan subalgebra of g [N1, ABFP2]. Moreover, h acts ad-diagonally on L. Furthermore, we can and do identify ∆ with a root system in the dual space h∗ in such a way that ∆(g, h) = ∆ind , ∆(L, h) = ∆ or ∆ind , and, for α ∈ Q, the root space of L relative to α is Lα [ABFP2, Prop. 1.2.2]. In particular, the type of the split simple Lie algebra g is the type of the root system ∆ind . We call (g, h) the grading pair for L.4

4

In [N1], the grading pair is defined for not necessarily centreless Lie tori. For centreless Lie tori the two definitions are equivalent [ABFP2, Prop. 1.2.1].


9

Remark 3.6 ([BN, Prop. 3.13]). Suppose that L is a Λ-torus, and let Cent(L) be the centroid of L.5 Then, Cent(L) is Λ-graded, with Cent(L)λ = { c ∈ Cent(L) | c(Lµ ) ⊆ Lλ+γ for µ ∈ Λ }. Let Γ (L) := suppΛ (Cent(L)) be the support of Cent(L), in which case Γ (L) is a subgroup of Λ called the centroidal grading group of L. In that case, Cent(L) is graded-isomorphic to the group algebra F[Γ ] with its natural Γ -grading (and hence also with its Λ-grading). Example 3.7 (The untwisted centreless Lie torus). Let g be a finite dimensional split simple Lie algebra, let h be a splitting Cartan subalgebra of g, let g = ⊕α∈∆ gα be the root space decomposition of g relative to h, where ∆ = ∆(g, h). Set L = g ⊗ F[Λ] with Q × Λ-grading given by Lλα = gα ⊗ F[Λ]λ for α ∈ Q, λ ∈ Λ. Then L is a centreless Lie Λ-torus of type ∆ which we call the untwisted centreless Lie Λ-torus of type ∆.6 The grading pair for L is (g ⊗ 1, h ⊗ 1). The centroid of L consists of multiplications by elements of F[Λ], so Γ (L) = Λ. In later sections, we will need the following lemma about bi-isomorphisms of centerless Lie tori of the same type. Lemma 3.8. Let L be a Lie Λ-torus of type ∆ with grading pair (g, h), and let L be a Lie Λ -torus of type ∆ with grading pair (g , h ). Let Π = { α1 , . . . , α } be a base for the root system ∆, choose 0 = ei ∈ gαi , 0 = fi ∈ g−αi with [[ei , fi ], ei ] = 2ei , and choose 0 = ei ∈ gαi , 0 = fi ∈ g−αi with [[ei , fi ], ei ] = 2ei . Suppose there is a bi-isomorphism ϕ : L → L with ϕr in the Weyl group of ∆. (Here we are using the notation ϕr and ϕe of Definition 3.3.) Then, there exists a bi-isomorphism ϕ˜ : L → L such that ϕ˜e = ϕe , ϕ˜r = 1, ϕ(e ˜ i ) = ei and ϕ(f ˜ i ) = fi for all i. Proof. Since ϕr is in the Weyl group of ∆, there exists an automorphism η of g such that η(h) = h and η(gα ) = gϕr (α) for all α ∈ ∆ [B3, Theorem 2(ii), Ch. VIII, § 2, no. 2]. Moreover, η can be chosen in the form η = i exp(adg (xi )), where each xi is in a root space of g corresponding to a nonzero root. Thus each adL (xi ) is nilpotent and so we may extend η to η = i exp(adL (xi )). Then η : L → L is a bi-isomorphism with ηe = 1 and ηr = ϕr . So replacing ϕ by ϕη −1 , we can assume that ϕr = 1. Thus, ϕ(gαi ) = ϕ(g ∩ Lαi ) = g ∩ Lαi = gαi for all i. So ϕ(ei ) = ai ei and × × ϕ(fi ) = a−1 i fi for some ai ∈ F . Choose a group homomorphism ρ : Q → F such that ρ(αi ) = ai for all i, and define τ ∈ Aut(L) by τ (xα ) = ρ(α)xα for xα ∈ Lα , α ∈ Q. Then τ : L → L is a bi-isomorphism with τe = 1 and τr = 1. So ϕτ −1 is the required ϕ. ˜

5 6

Recall that the centroid of an algebra is the associative algebra of all endomorphisms of the algebra that commute with all left and right multiplications. The universal central extension of g⊗F[Λ] is called the toroidal Lie algebra [MRY], which is one of the origins of the term Lie torus.

10


4 The construction of EALAs from Lie tori In [BGK], Berman, Gao, and Krylyuk gave a construction of a family of EALAs starting from a centreless Lie tori of type Ar , r ≥ 3. In [N2], Neher simplified this construction and extended it to all types. Since we will be working with this construction in some detail, we give a careful description of it in this section. Facts that we note without reference are either straightforward or can be found in [N1, N2]. We first need to establish some notation and assumptions. Throughout the section, we assume that L is a centreless Lie Λ-torus of type ∆ and we let Q = Q(∆). Let Γ = Γ (L) be the centroidal grading group of L. We identify Cent(L) = F[Γ ] (see Remark 3.6) by fixing a basis { tγ }γ∈Γ for Cent(L) satisfying tγ tδ = tγ+δ . Next let Hom(Λ, F) be the group of group homomorphisms from Λ into F. Then for θ ∈ Hom(Λ, F), let ∂θ ∈ Der(L) be the degree derivation defined by ∂θ (xλ ) = θ(λ)xλ for xλ ∈ Lλ , λ ∈ Λ. Put D = { ∂θ | θ ∈ Hom(Λ, F) } and

CDer(L) = Cent(L)D.

Then CDer(L) is a subalgebra of the Lie algebra Der(L) with product given by [tγ1 ∂θ1 , tγ2 ∂θ2 ] = tγ1 +γ2 (θ1 (γ2 )∂θ2 − θ2 (γ1 )∂θ1 ), (2) and CDer(L) is Γ -graded with CDer(L)γ = Cent(L)γ D for γ ∈ Γ . Since L is a centreless Lie torus, there is a nondegenerate invariant Λgraded form ( | ) on L, and that form is unique up to multiplication by a nonzero scalar [Y6, Thm. 2.2 and 7.1]. We fix a choice of ( | ) on L. Then, since L is perfect and ( | ) is invariant we have (c(x), y) = (x, c(y)) for x, y ∈ L, c ∈ Cent(L). Now let SCDer(L) be the subalgebra of CDer(L) consisting of the derivations in CDer(L) that are skew relative to the form ( | ). (This subalgebra does not depend on the choice of the form ( | ).) Then, SCDer(L) is a Γ -graded subalgebra of CDer(L) with SCDer(L)γ = tγ { ∂θ ∈ D | θ(γ) = 0 }

(3)

for γ ∈ Γ . SCDer(L) is called the algebra of skew-centroidal derivations of L. Now suppose that D = ⊕γ∈Γ Dγ is a Γ -graded subalgebra of SCDer(L). The graded-dual space of D is the subspace Dgr∗ = ⊕γ∈Γ (Dγ )∗ of D∗ , where


11

(Dγ )∗ is embedded in D∗ by letting its elements act trivially on Dδ for δ = γ. We let C = Dgr∗ , and we give the vector space C a Γ -grading by setting Cγ = (D−γ )∗ .

(4)

With this grading, C is a Γ -graded D-module by means of the contragradient action given by (d · f )(e) = −f ([d, e]), for d, e ∈ D, f ∈ C. We also regard C as a L-module with trivial action. Now define σD : L × L → C by σD (x, y)(d) = (dx|y). Then σD is a Γ -graded 2-cocycle for L with values in the trivial L-module C.7 Finally, let ev : Λ → C0 denote the evaluation map defined by (ev(λ))(∂θ ) = θ(λ) for λ ∈ Λ and ∂θ ∈ D0 ⊆ SCDer(L)0 = D. With this background we now present the construction. Construction 4.1 ([N2]). Let L be a centreless Lie Λ-torus of type ∆. In order to construct an EALA from L we need two additional ingredients. We assume that (4.1a) D is a Γ -graded subalgebra of SCDer(L) such that the map ev : Λ → C0 is injective, where Γ = Γ (L) and C = Dgr∗ . (4.1b) τ : D × D → C is a Γ -graded invariant 2-cocycle of D with values in the D-module C and τ (D0 , D) = 0. Note that the assumption in (4.1b) that τ is invariant means that τ (d1 , d2 )(d3 ) = τ (d2 , d3 )(d1 ) for di ∈ D. Let

E = E(L, D, τ ) = D ⊕ L ⊕ C,

where C = D . We identify D, L and C naturally as subspaces of E. Define a product [ , ]E on E by gr∗ 8

7

8

In this section and in §6 we use the terminology (cochain, cocycle, coboundary and extension) from the cohomology of Lie algebras. See for example [B1, Chapter I, Exercise 12 for §3]. We have changed the order used in [N2] of the components of E. This is convenient in the proof of Corollary 6.3 below.

12


[d1 + l1 + f1 , d2 + l2 + f2 ]E = [d1 , d2 ] + ([l1 , l2 ] + d1 (l2 ) − d2 (l1 )) + (d1 · f2 − d2 · f1 + σD (l1 , l2 ) + τ (d1 , d2 )) for di ∈ D, li ∈ L, fi ∈ C. Then, E is a Λ-graded algebra with the direct sum grading. Next we extend the bilinear form ( | ) on L to a graded bilinear form ( | ) on E by defining: (d1 + l1 + f1 | d2 + l2 + f2 ) = (l1 | l2 ) + f1 (d2 ) + f2 (d1 ). Finally let H = D0 ⊕ h ⊕ C0

(5)

in E, where (g, h) = (L0 , L00 ) is the grading pair for L. Thus we have constructed a triple ((E(L, D, τ ), ( | ), H), which we often denote simply by E(L, D, τ ). Neher has announced the following fundamental result on this construction: Theorem 4.2 ([N2, Thm. 6]). (i) If L is a centreless Lie Λ-torus, D satisfies (4.1a) and τ satisfies (4.1b), then E(L, D, τ ) is an extended affine Lie algebra of nullity rank(Λ). (ii) If E is an EALA, then E is isomorphic as an EALA to E(L, D, τ ) for some L, D and τ as in (i). Remark 4.3. (i) It is easy to check that the EALA E(L, D, τ ) in part (i) of the theorem does not depend, up to EALA isomorphism, on the choice of the form ( | ) on L. (ii) Neher actually states more than we have stated in part (ii) of the theorem. Indeed, given an EALA E, then Ecc , with a suitable grading, is a Lie torus satisfying (LT5) and E is isomorphic as an EALA to E(Ecc , D, τ ) for some D and τ as in (i) [N2, Thm. 6(ii)]. (iii) In particular, we can always choose L in part (ii) of the theorem satisfying (LT5). In that case, the type of E and L are the same. (iv) The Lie torus L in part (ii) of the theorem is not uniquely determined up to bi-isomorphism. However, it is uniquely determined up to isotopy, a fact that has motivated our work on this topic. We discuss this in detail in §6 below. Remark 4.4. Suppose L is a centreless Lie torus, Γ = Γ (L), and D is a Γ graded subalgebra of SCDer(L). Let M := D ⊕ L and identify D and L naturally as subspaces of M . Then, M is a Λ-graded Lie algebra with product [ , ] given by [d1 + l1 , d + l2 ] = [d1 , d2 ] + ([l1 , l2 ] + d1 (l2 ) − d2 (l1 )) , and C is an M -module via (d + l) · f = d · f . The cocycles σD and τ can be extended to cocycles for the Lie algebra M with values in the M -module C


13

by letting σD (D, M ) = σD (M, D) = 0 and τ (L, M ) = τ (M, L) = 0. If D and τ satisfy (4.1a) and (4.1b), then as a Lie algebra E(L, D, τ ) is equal to the extension Ext(M, C, σD + τ ) of M by C using the cocycle σD + τ . If L is a centreless Lie torus, we let P(L) = { (D, τ ) | D satisfies (4.1a) and τ satisfies (4.1b) }. We note P(L) is nonempty since in particular (SCDer(L), 0) ∈ P(L). With this notation Theorem 4.2(i) tells us that Construction 4.1 builds a family { E(L, D, τ ) }(D,τ )∈P(L) of EALAs from L. Also, Theorem 4.2(ii) tells us that any EALA E occurs (up to isomorphism of EALAs) in the family { E(L, D, τ ) }(D,τ )∈P(L) constructed from some centreless Lie torus L.

5 Isotopy for Lie tori In this section, we recall the definition of isotopy from [ABFP2]. Throughout the section, we assume that L is a Lie Λ-torus of type ∆ and we let Q = Q(∆). If α ∈ Q, we let Λα = Λα (L) := { λ ∈ Λ | Lλα = 0 }. We call Λα the Λ-support of L at α. Note that 0 ∈ Λα if 0 = α ∈ ∆ind , by (LT2)(ii). Definition 5.1. Suppose that s ∈ Hom(Q, Λ), where Hom(Q, Λ) is the group of group homomorphisms from Q into Λ. We define a new Q × Λ-graded Lie algebra L(s) as follows. As a Lie algebra, L(s) = L. The grading on L(s) is given by (L(s) )λα = Lλ+s(α) (6) α for α ∈ Q, λ ∈ Λ. The following necessary and sufficient conditions for L(s) to be a Lie torus are obtained in [ABFP2, Prop. 2.2.3]. Proposition 5.2. Let s ∈ Hom(Q, Λ), and let Π be a base for the root system ∆. The following statements are equivalent: (a) L(s) is a Lie torus. (b) s(α) ∈ Λα for all 0 = α ∈ ∆ind . (c) s(α) ∈ Λα for all α ∈ Π. Definition 5.3 (Isotopes of Lie tori). Suppose that s ∈ Hom(Q, Λ). If s satisfies the equivalent conditions in Proposition 5.2, we say that s is admissible for L. In that case, we call the Lie torus L(s) the s-isotope of L.

14


Remark 5.4. Suppose that Π = { α1 , . . . , αr } is a base for ∆. By Proposition 5.2, to specify an admissible s ∈ Hom(Q, Λ) for L, and hence an isotope L(s) of L, one can arbitrarily choose λi ∈ Λαi for 1 ≤ i ≤ r, and then define s ∈ Hom(Q, Λ) with s(αi ) = λi for 1 ≤ i ≤ r. Definition 5.5. Suppose that L is a Lie Λ-torus of type ∆ and L is a Lie Λ -torus of type ∆ . We say that L is isotopic to L , written L ∼ L , if some isotope L(s) of L is bi-isomorphic to L . Using the facts that L(0) = L, (L(s) )(t) = L(s+t) and (L(s) )(−s) = L as Q × Λ-graded algebras for s, t ∈ Hom(Q, Λ), it is shown in [ABFP2, §2] that isotopy is an equivalence relation on the class of Lie tori.

6 Isotopy in the theory of EALAs As promised at the end of §4, we now describe the role that isotopy plays in the structure theory of EALAs. Our first theorem on this topic was actually the starting point of our investigation of isotopy. Theorem 6.1. Suppose L is a centreless Lie Λ-torus of type ∆ and L is a centreless Lie Λ -torus of type ∆ . If some member of the family { E(L, D, τ ) }(D,τ )∈P(L) is isomorphic as an EALA to some member of the family { E(L , D , τ ) }(D ,τ )∈P(L ) , then L is isotopic to L . Proof. Let (g, h) be the grading pair of L. Since ∆(L, h) = ∆ or ∆ind (see Definition 3.5), we can if necessary replace ∆ by ∆ind so that ∆(L, h) = ∆. Suppose now that E = E(L, D, τ ), where (D, τ ) ∈ P(L). We use the notation Q, C, ev, H from Construction 4.1. Recall that H = D0 ⊕ h ⊕ C, so we can identify H ∗ = (D0 )∗ ⊕ h∗ ⊕ (C0 )∗ = C0 ⊕ h∗ ⊕ D0 . We use ev to identify Λ with a subgroup of C0 ⊆ H ∗ . Let R be the set of roots of E and let R× be the set of nonisotropic roots in R. Then by [N2, §5], we have R× = ∪α∈∆× (Λα + α) ⊆ R ⊆ Λ ⊕ Q,

(7)

where Λα = Λα (L). It follows from this fact and (LT2)(i) that R = Λ ⊕ Q.

(8)

Eλ+α = Lλα

(9)

Now by [N2, §5], we have


15

for λ ∈ Λ, 0 = α ∈ Q. It is easy to show using this equation, (7) and (LT3) that the core of E (see Definition 2.1(ii)) is given by Ec = L ⊕ C.

(10)

(In fact this is implicit in [N2].) Hence, since L is centreless, we have Z(Ec ) = C.

(11)

Next, let E = E(L , D , τ ), where (D , τ ) ∈ P(L ), and we use all of the above notation (with primes added) for L and E . Suppose that χ : E → E is an EALA isomorphism. So χ(H) = H , and χ preserves the forms up multiplication by a nonzero scalar. It follows that χ(Ec ) = Ec . So, by (10) and (11) (and the corresponding primed equations), we have χ(L ⊕ C) = L ⊕ C and χ(C) = C . Thus, there exists a unique Lie algebra isomorphism ϕ : L → L with χ(l) ≡ ϕ(l)

(mod C )

(12)

for l ∈ L. It then follows that χ−1 (l ) ≡ ϕ−1 (l ) (mod C) for l ∈ L . Now, by (12), we have ϕ(h) ⊆ (H ⊕ C ) ∩ L = h , and similarly ϕ−1 (h ) ⊆ h. So ϕ(h) = h . (13) Hence, if h ∈ h, χ(h) − ϕ(h) ∈ H ∩ C = C0 , so χ(h) ≡ ϕ(h)

(mod C0 ).

(14)

Also, χ(C0 ) ⊆ H ∩ C = C0 and similarly χ−1 (C0 ) ⊆ C0 . So χ(C0 ) = C0 .

(15)

But then, by (13), (14) and (15), χ(h ⊕ C0 ) ⊆ h ⊕ C0 ; so using the corresponding property for χ−1 , we have χ(h ⊕ C0 ) = h ⊕ C0 . ∗

Next define χ ˆ : H ∗ → H by (χ(ρ))(χ(t)) ˆ = ρ(t)

(16)

16


for ρ ∈ H ∗ , t ∈ H. (χ ˆ is the inverse dual of χ|H : H → H .) Similarly, define ∗ ∗ ϕˆ : h → h by (ϕ(α))(ϕ(h)) ˆ = α(h) for α ∈ h∗ , h ∈ h. Then, using R = ∆(E, H) and ∆ = ∆(L, h) (and the corresponding primed equations), we see that χ(R) ˆ = R

and ϕ(∆) ˆ = ∆ .

Thus, since R = Λ ⊕ Q and ∆ = Q, we have χ(Λ ˆ ⊕ Q) = Λ ⊕ Q , ϕ(Q) ˆ = Q .

(17) (18)

Let α ∈ Q, and write, using (17), χ(α) ˆ = λ + α , where λ ∈ Λ , α ∈ Q . Then, (λ + α )(χ(h)) = α(h) for h ∈ h (in fact for h ∈ H). Thus, by (14), (λ + α )(ϕ(h)) = α(h), so α (ϕ(h)) = α(h) for h ∈ h. Hence, α = ϕ(α). ˆ Thus, since ϕˆ is invertible, χ(α) ˆ = s (ϕ(α)) ˆ + ϕ(α) ˆ (19) for α ∈ Q and some s ∈ Hom(Q , Λ ). Next let λ ∈ Λ, and as above write χ(λ) ˆ = λ + α , where λ ∈ Λ , α ∈ Q . Then, (λ +α )(χ(h+C0 )) = λ(h+C0 ) = 0. Hence, by (16), (λ +α )(h +C0 ) = −1 (Λ ) ⊆ Λ. So 0, so α = 0. Thus, χ(Λ) ˆ ⊆ Λ and similarly χ ˆ−1 (Λ ) = χ χ(Λ) ˆ = Λ .

(20)

Using (18) and (20), we let ϕr = ϕ| ˆ Q : Q → Q and ϕe = χ| ˆ Λ : Λ → Λ . Then, using (9) and (19), we have, for 0 = α ∈ Q, λ ∈ Λ, χ(Lλα ) = χ(Eλ+α ) = Eχ(λ+α) = Eχ(λ)+s (ϕ(α))+ ˆ ϕ(α) ˆ ˆ ˆ χ(λ)+s ˆ (ϕ(α)) ˆ

= (L )ϕ(α) ˆ So by (12), we have

ϕ (λ)

= ((L )(s ) )ϕer (α) .

ϕ (λ)

ϕ(Lλα ) = ((L )(s ) )ϕer (α) for 0 = α ∈ Q, λ ∈ Λ. But then, by (LT3), this equation holds for all α ∈ Q, λ ∈ Λ. Therefore, since L is a Lie Λ-torus of type ∆, so is (L )(s ) . Hence s is admissible for L , and ϕ is a bi-isomorphism of L onto (L )(s ) .

We next consider the relationship between the family of EALAs constructed from a Lie torus L and the family of EALAs constructed from an isotope of L. For this purpose, we assume now that L is a centreless Lie Λtorus of type ∆, Q = Q(∆), s ∈ Hom(Q, Λ) is admissible for L, and L(s) is the s-isotope of L. We fix a nondegenerate Λ-graded invariant symmetric bilinear form ( | ) on L and we use that same form on L(s) .


17

(s)

Let ∂θ be the degree derivation of L(s) determined by θ ∈ Hom(Λ, F). Define the Λ-graded Cent(L)-module isomorphism Ψ : CDer(L) → CDer(L(s) )

(s)

by Ψ (c∂θ ) = c∂θ .

Using (2) we see that Ψ is also a Lie algebra isomorphism. Let hθ ∈ h be given by α(hθ ) = θ(s(α)) for α ∈ Q, and define the Λ-graded Cent(L)-module monomorphism Ω : CDer(L) → L by λ+s(α)

for c ∈ Cent(L). For l ∈ (L(s) )λα = Lα (s) ∂θ (l) + [hθ , l]. Thus, (s)

∂θ

Ω(c∂θ ) = chθ , we have ∂θ (l) = θ(λ + s(α))l =

= ∂θ − adL (hθ ) and Ψ (d) = d − adL (Ω(d))

for d ∈ CDer(L). Since the derivations in adL (L) are skew, we see that Ψ maps SCDer(L) into SCDer(L(s) ). We also note using (2) that Ω([tγ1 ∂θ1 , tγ2 ∂θ2 ]) = tγ1 tγ2 (θ1 (γ2 )hθ2 − θ2 (γ1 )hθ1 ) = tγ1 ∂θ1 (tγ2 hθ2 ) − tγ2 ∂θ2 (tγ1 hθ1 ) = tγ1 ∂θ1 (Ω(tγ2 ∂θ2 )) − tγ2 ∂θ2 (Ω(tγ1 ∂θ1 )), for γi ∈ Γ and θi ∈ Hom(Λ, F), so Ω is a derivation of CDer(L) into L. If (D, τ ) ∈ P(L), let D (s) = Ψ (D), (21) in which case D(s) is a Γ -graded subalgebra of SCDer(L). Also let ψ : D → D(s)

and ω : D → L

be the restrictions of Ψ and Ω to D. Then ψ is a Λ-graded Lie algebra isomorphism, ω is a Λ-graded vector space monomorphism, and ψ(d) = d − adL (ω(d)) for d ∈ D. Finally let

(22)

C (s) = (D(s) )gr∗ ,

which is Γ -graded as in (3), and define a Λ-graded cochain τ (s) : D(s) × D(s) → C (s) for di ∈ D.

by τ (s) (ψd1 , ψd2 )(ψd3 ) = τ (d1 , d2 )(d3 )

(23)

18


Theorem 6.2. Suppose that L is a centreless Lie Λ-torus, s ∈ Hom(Q, Λ) is admissible for L, (D, τ ) ∈ P(L), and (D(s) , τ (s) ) is defined by (21) and (23). Then (D(s) , τ (s) ) ∈ P(L(s) ). Moreover, E = E(L, D, τ ) and E (s) = E(L(s) , D(s) , τ (s) ) are isomorphic as EALAs. To be more precise, define χ : E → E (s) by χ(x) = ψ(d) + (ω(d) + l) + η(x) for x = d + l + c ∈ E, d ∈ D, l ∈ L, f ∈ C, where η : E → C (s) is given by 1 η(x)(ψ(d )) = f (d ) − (l + ω(d) | ω(d )) 2 for d ∈ D. Then χ is an isometry and a Λ-graded Lie algebra isomorphism (s) (s) that maps H to H (s) = D0 ⊕ h ⊕ C0 . ˆ )(ψ(d)) = f (d) for d ∈ D, f ∈ C. Clearly, Proof. Define ψˆ : C → C (s) by ψ(f ˆ ψ is a Γ -graded vector space isomorphism. Let ev(s) be the evaluation map (s) (s) for L(s) . Then, for λ ∈ Λ and ∂θ ∈ D0 , we have (s) ˆ ˆ (ψ(ev(λ)))(∂ θ ) = (ψ(ev(λ)))(ψ(∂θ )) = (ev(λ))(∂θ ) = θ(λ),

so ψˆ |C0 ◦ ev = ev(s) . Hence, ev(s) is injective; i.e., D(s) satisfies (4.1a). It will be convenient to view E as Ext(M, C, σD + τ ), the extension of M = D ⊕ L by C using the cocycle σD + τ as in Remark 4.4. Let M (s) = D(s) ⊕ L and set ξ(d + l) = ψ(d) + ω(d) + l for d ∈ D, l ∈ L. We claim ξ : M → M (s) is a Lie algebra isomorphism.

(24)

Clearly, ξ is bijective since ψ is bijective. Since ψ : D → D(s) is an isomorphism and L is an ideal in M (s) , the map d + l → ψ(d) + L = ξ(d + l) + L is a homomorphism M → M (s) /L. Thus, for m1 , m2 ∈ M , z := ξ([m1 , m2 ]) − [ξ(m1 ), ξ(m2 )] ∈ L. On the other hand, ζ : M → Der(L) given by ζ(m) = adM (m) |L and similarly ζ (s) : M (s) → Der(L) are homomorphisms. Moreover, ζ (s) (ξ(d + l)) = ψ(d) + adL (ω(d)) + adL (l) = d + adL (l) = ζ(d + l), so ζ (s) ◦ ξ = ζ. Thus, adL (z) = ζ (s) (z) = ζ([m1 , m2 ]) − [ζ(m1 ), ζ(m2 )] = 0. Since L is centreless, z = 0 and ξ is an isomorphism.


19

We also note that for d, d ∈ D, l ∈ L, f ∈ C, ˆ ))(ψ(d )) = −ψ(f ˆ )([ψ(d), ψ(d )]) ˆ ))(ψ(d )) = (ψ(d) · ψ(f (ξ(d + l) · ψ(f ˆ = −f ([d, d ]) = (d · f )(d ) = ψ((d + l) · f )(ψ(d ), so ˆ ) = ψ(m ˆ · f) ξ(m) · ψ(f

(25)

for m ∈ M , f ∈ C. Recall that the group Z 2 (M, C) of 2-cocycles on M with values in C consists of all alternating bilinear maps µ : M × M → C such that (µ([mi , mj ], mk ) − mi · µ(mj , mk )) = 0 (i,j,k)

for m1 , m2 , m3 ∈ M where (i, j, k) means (i, j, k) is a cyclic permutation of (1, 2, 3). We can use ξ and ψˆ to transfer µ to µ ˜ : M (s) × M (s) → C (s) with ˆ µ ˜ (ξ(m1 ), ξ(m2 )) = ψ(µ(m 1 , m2 )). Using (24) and (25), we see that ρ(µ) = µ ˜ defines a group isomorphism ρ : Z 2 (M, C) → Z 2 (M (s) , C (s) ). Furthermore, the map ξ ⊕ ψˆ : Ext(M, C, µ) → Ext(M (s) , C (s) , ρ(µ)) is a Lie algebra isomorphism. ˆ (M, L)) = 0 and Let τ˜ = ρ(τ ). Then τ˜(M (s) , L) = τ˜(ξ(M ), ξ(L)) = ψ(τ thus also τ˜(L, M (s) ) = 0. Moreover, for d1 , d2 , d3 ∈ D, we have ˆ (d1 , d2 ))(ψ(d3 )) = τ (d1 , d2 )(d3 ), τ˜(ψ(d1 ), ψ(d2 ))(ψ(d3 )) = ψ(τ so τ˜ is the extension of τ (s) to M (s) and τ (s) is a cocycle. The invariance of τ (s) (s) is immediate from the invariance of τ , while the equation τ (s) (D0 , D(s) ) = (s) 0 is immediate from the definition of τ (s) and ψ(D0 ) = D0 . This shows (s) (s) (s) (D , τ ) ∈ P(L ). It remains to prove the statement about χ. To do this it will be helpful for us to write an element x = d + l + f ∈ E, where d ∈ D, l ∈ L, f ∈ C, as a column vector ⎡ ⎤ d x = ⎣ l ⎦. f Similarly we express elements of E (s) as column vectors. This notation allows us to write linear maps (for example from E to E (s) ) as matrices. Define ω # : L → C by ω # (l)(d) = (l | ω(d)). We see that for x = d + l + c ∈ E, d ∈ D, l ∈ L, f ∈ C and for d ∈ D, we have 1 ˆ − ω # (l + 1 ω(d)))(ψ(d )), η(x)(ψ(d )) = (f − ω # (l + ω(d)))(d ) = ψ(f 2 2

20


ˆ ) − (ψˆ ◦ ω # )(l) − 1 (ψˆ ◦ ω # ◦ ω)(d). Hence, in matrix form so η(x) = ψ(f 2 ⎡

ψ ⎣ ω χ= 1 ˆ − 2 ψ ◦ ω# ◦ ω Let

0 Id −ψˆ ◦ ω #

⎤ 0 0 ⎦. ψˆ

⎤ 0 0 0 ν = ⎣ω 0 0⎦, 0 −ω # 0 ⎡

so ν : E → E. The definition of ω # shows that ν is a skew transformation of E. Thus, exp(ν) is an isometry of E. We can write ⎡ ⎤⎡ ⎤ ⎡ ⎤ ψ 0 0 ψ 0 0 Id 0 0 ω Id 0 ⎦ = ⎣ 0 Id 0 ⎦ exp(ν). χ = ⎣ 0 Id 0 ⎦ ⎣ 1 # ˆ − 2 ω ◦ ω −ω # Id 0 0 ψ 0 0 ψˆ ˆ Thus, χ : E → The first factor is an isometry E → E (s) by the definition of ψ. (s) E is an isometry. ˆ ω # , η and χ. Also, ω(D0 ) ⊆ h, so Since ψ and ω are Λ-graded, so are ψ, (s) χ(H) ⊆ ψ(D0 ) ⊕ (ω(D0 ) + h) ⊕ η(E0 ) ⊆ H (s) . But since D0 = ψ(D0 ) and (s) ˆ 0 ), we have dim(H (s) ) = dim(H), so χ(H) = H (s) . C0 = ψ(C It remains to show that χ is a homomorphism of Lie algebras. Let ˆ (m1 , m2 )) σ = ρ−1 (σD(s) ) ∈ Z 2 (M, C). Then, σD(s) (ξ(m1 ), ξ(m2 )) = ψ(σ for m1 , m2 ∈ M , so σ (m1 , m2 )(d) = σD(s) (ξ(m1 ), ξ(m2 ))(ψ(d)) for d ∈ D. We will show that σ = σD +δ(κ) where κ : M → C is the 1-cocycle κ(d + l) =

1 # ω (ω(d)) + ω # (l) 2

and δ(κ) is the coboundary of κ; i.e., δ(κ)(m1 , m2 ) = m2 · κ(m1 ) − m1 · κ(m2 ) − κ([m1 , m2 ]). Indeed, using linearity and skew symmetry, it suffices to check this for the pairs (l1 , l2 ), (l1 , d2 ), and (d1 , d2 ). For this recall that L · C = 0, [ω(D), ω(D)] ⊆ [Cent(L)h, Cent(L)h] = 0, and that ω is a derivation of D into L. We have σ (l1 , l2 )(d) = σD(s) (l1 , l2 )(ψ(d)) = (ψ(d)(l1 ) | l2 ) = ((d − adL (ω(d))(l1 ) | l2 ) = (d(l1 ) − [ω(d), l1 ] | l2 ) = (d(l1 ) | l2 ) − (ω(d) | [l1 , l2 ]) = (σD (l1 , l2 ) − ω # ([l1 , l2 ]))(d),


21

while δ(κ)(l1 , l2 ) = l1 · κ(l2 ) − l2 · κ(l1 ) − κ([l1 , l2 ]) = −κ([l1 , l2 ]) = −ω # ([l1 , l2 ])). Also, σ (d1 , d2 )(d) = σD(s) (ψ(d1 ) + ω(d1 ), ψ(d2 ) + ω(d2 ))(ψ(d)) = (ψ(d)(ω(d1 )) | ω(d2 )) = ((d − adL (ω(d))(ω(d1 )) | ω(d2 )) = (d(ω(d1 )) | ω(d2 )) and 2(δ(κ)(d1 , d2 ))(d) = (d1 · ω # (ω(d2 )) − d2 · ω # (ω(d1 )) − ω # (ω([d1 , d2 ])))(d) = −(ω(d2 ) | ω([d1 , d])) + (ω(d1 ) | ω([d2 , d])) − (ω([d1 , d2 ])) | ω(d)) = −(ω(d2 ) | d1 (ω(d)) − d(ω(d1 ))) + (ω(d1 ) | d2 (ω(d)) − d(ω(d2 ))) − (d1 (ω(d2 )) − d2 (ω(d1 )) | ω(d)) = (ω(d2 ) | d(ω(d1 ))) − (ω(d1 ) | d(ω(d2 ))) = 2(d(ω(d1 )) | ω(d2 )). Finally, σ (d1 , l2 )(d) = σD(s) (ψ(d1 ) + ω(d1 ), l2 )(d) = (d · ω(d1 ), l2 ) and 1 (δ(κ)(d1 , l2 ))(d) = (d1 · ω # (l2 ) − l2 · ω # (ω(d1 )) − ω # ([d1 , l2 ]))(d) 2 = −(l2 | ω([d1 , d])) + (l2 | d1 · ω(d)) = (l2 | d · ω(d1 )). So σ = σD + δ(κ) as claimed. Since σ + τ = σD + τ + δ(κ), the map m + f → m + f − κ(m) for m ∈ M , f ∈ C is a Lie algebra isomorphism of E = E(L, D, τ ) with Ext(M, C, σ + τ ). We can write this isomorphism as ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ Id 0 0 D D ⎣ 0 Id 0 ⎦ : ⎣ L ⎦ → ⎣ L ⎦ . C C − 21 ω # ◦ ω −ω # Id Since σD(s) + τ (s) = ρ(σ + τ ), we also have that ⎡ ⎤ ψ 0 0 ξ ⊕ ψˆ = ⎣ ω Id 0 ⎦ 0 0 ψˆ

22


is a Lie algebra isomorphism of Ext(M, C, σ + τ ) with E (s) . Thus, ⎡ ⎤⎡ ⎤ ψ 0 0 Id 0 0 0 Id 0 ⎦ χ = ⎣ ω Id 0 ⎦ ⎣ − 1 ω # ◦ ω −ω # Id 0 0 ψˆ 2

is a Lie algebra isomorphism from E to E (s) .

Corollary 6.3. Let L be a centreless Lie Λ-torus of type ∆, and let L be a centreless Lie Λ -torus of type ∆ . Suppose that L is isotopic to L . Then, there is a bijection (D, τ ) → (D , τ ) from P(L) onto P(L ) such that E(L, D, τ ) is isomorphic as an EALA to E(L , D , τ ) for all (D, τ ) ∈ P(L). Proof. If L and L are bi-isomorphic, the result is clear. So we can assume that L = L(s) is an isotope of L. Define T : P(L) → P(L ) by T (D, τ ) = (D(s) , τ (s) ) using the notation of Theorem 6.2. Similarly, we have a map T : P(L ) → P(L (−s) ) = P(L), and it is clear that T and T are inverses of one another. So T is a bijection.

Less precisely (but more succinctly), Corollary 6.3 says that if the centerless Lie tori L and L are isotopic, then the families { E(L, D, τ ) }(D,τ )∈P(L) and { E(L , D , τ ) }(D ,τ )∈P(L ) of EALAs are bijectively isomorphic. Using this language, Theorem 4.2, Theorem 6.1, and Corollary 6.3 together tell us that Construction 4.1 provides a one-to-one correspondence between centreless Lie tori up to isotopy and families of EALAs up to bijective isomorphism. Remark 6.4. Suppose that L is a centreless Lie torus. To construct all of the members of the family { E(L, D, τ ) }(D,τ )∈P(L) of EALAs corresponding to L, one must do the following: (i) Calculate the centroid Cent(L) and identify it explicitly with F[Γ ] (in which case SCDer(L) is completely understood from (3)); (ii) Determine the Γ -graded subalgebras D of SCDer(L) which satisfy (4.1a); (iii) For each D in (ii), determine all 2-cocycles of D satisfying (4.1b). We note that (i) can be carried out for each type ∆ using the coordinatization theorems (see §7 below) and the results and methods in [BN, §5]. However, finding a general approach to (ii) and (iii) seems to be much more difficult. (See [BiN, p.3] for some remarks on problem (iii) and see [BGK, Remark 3.71(b)] for an example of a nontrivial choice of τ .) We do however note that the tasks (ii) and (iii) are independent of the type ∆; in fact they depend only on the rank of Γ .

7 Coordinatization of Lie tori In the classical study of finite dimensional split simple Lie algebras, one can construct the algebras using special linear, orthogonal and symplectic matrix


23

constructions along with some exceptional constructions starting from Jordan algebras and alternative algebras. (See for example [S, §III.8, IV.3 and IV.4].) Similar coordinatization theorems have been proved for centreless Lie tori. Very roughly speaking these results show that a centerless Lie torus of a given type can be constructed as a “matrix algebra” over a “coordinate torus”, which for some types is nonassociative. References for each type are: A1 : [Y1]; A2 : [BGKN, Y4]; Al (l ≥ 3), Dl (l ≥ 4), E6 , E7 and E8 : [BGK, Y2]; B2 = C2 : [AG, BY]; Bl (l ≥ 3): [AG, Y5, AB]; Cl (l ≥ 3): [AG, BY]; F4 and G2 : [AG]; BC1 [AFY]; BC2 : [F1]; and finally BCl (l ≥ 3): [AB]. We note that these theorems were in some cases proved before the notion of Lie torus had been introduced and were instead formulated in the language of EALAs. Also, in some cases the theorems were proved over the complex field. Nevertheless the translation to the language of Lie tori and the extension to arbitrary base fields of characteristic 0 are not difficult. In the following sections, we recall the coordinatization theorems for types A1 , A2 , Al (l ≥ 3) and Cl (l ≥ 4), which use respectively Jordan tori, alternative tori, associative tori and associative tori with involution as coordinates. First we recall the definitions of these types of tori. Recall that a Jordan algebra is an algebra with commutative product satisfying a2 (ab) = a(a2 b), and an alternative algebra is an algebra with product satisfying a(ab) = a2 b and (ba)a = ba2 (see for example [S] or [Mc2]). λ Definition 7.1. Suppose that A = λ∈Λ A is a Λ-graded Jordan algebra, alternative algebra or associative algebra. A is said to be a Jordan, alternative, or associative Λ-torus, respectively if: (i) For λ ∈ suppΛ (A), Aλ is spanned by a single invertible element of A.9 (ii) Λ is generated as a group by suppΛ (A). Note of course that any associative torus is an alternative torus. An associative Λ-torus with involution is a pair (A, ι), where A is an associative Λ-torus and ι is a graded involution of A. Remark 7.2. Suppose that A is a Jordan or alternative Λ-torus, and let S = suppΛ (A). Then, it is clear that −S = S. Moveover, if A is alternative, it is easy to see that S is closed under addition, so S = Λ. However, this is not true in the Jordan case, although S is closed under the operation (λ, µ) → λ + 2µ. Example 7.3. Let q = (qij ) be an n × n-matrix over F with qii = 1 and −1 ±1 qij = qji for all i, j. Let Fq = Fq [x±1 1 , . . . , xn ] be the associative algebra over −1 F presented by the generators xi , xi , i = 1, . . . , n, subject to the relations xi x−1 = x−1 i i xi = 1 and xj xi = qij xi xj , 9

1 ≤ i, j ≤ n.

We are using the standard definitions of invertibility for Jordan algebras [Mc2, p.210], alternative algebras [S, p.38], and associative algebras.

24


Then, Fq has a natural Zn -grading with (Fq )(l1 ,...,ln ) = Fxl11 . . . xlnn . The Zn graded algebra Fq is an associative Zn -torus called the quantum torus determined by q.10 Conversely, it is easy to see that any associative torus is isograded-isomorphic to a quantum torus. Recently, rational quantum tori, that is quantum tori Fq with each qij a root of unity, have been classified up to isograded-isomorphism by K.-H. Neeb in [Ne, Thm. III.4] as tensor products of rational quantum Z2 -tori and an algebra of Laurent polynomials. (In [H, Thm. 4.8], J.T. Hartwig gave a tensor product decomposition of rational quantum tori but he did not give a condition for isomorphism.) In the following sections, we will give some further examples of the tori defined in Definition 7.1.

8 Type A1 In this section we assume that ∆ = { α, 0, −α } is the root system of type A1 , and Q = Q(∆) = Zα. To construct a centreless Lie Λ-torus of type ∆, we use the Tits–Kantor– Koecher construction. (See [Mc2, p.13] for the facts mentioned below about this construction.) We begin with a Jordan Λ-torus A and we define operators Vx,y ∈ End(A) for x, y ∈ A by Vx,y z = {x, y, z} := 2((xy)z + (zy)x − (zx)y). These operators satisfy the identity [Vx,y , Vz,w ] = V{x,y,z},w − Vz,{y,x,w} , so their span VA,A is a Lie algebra under the commutator product. Moreover, the map ∗ defined by Vx,y → −Vy,x is a well-defined automorphism of VA,A . We now set L(A) = TKK(A) := A1 ⊕ VA,A ⊕ A−1 , where, for i = ±1, Ai is a copy of the vector space A under the linear bijection x → xi . Then L(A) is a Lie algebra, called the Tits–Kantor–Koecher Lie algebra of A, under the product defined by [T, x1 ] = (T x)1 , [T, y−1 ] = (T ∗ y)−1 , [xi , yi ] = 0 (i = ±1), [x1 , y−1 ] = Vx,y , [T, T ] = T T − T T, for x, y ∈ A, T, T ∈ VA,A . Finally, L(A) is a Q × Λ-graded algebra with VAµ ,Aν , L(A)λ−α = (Aλ )−1 , L(A)λα = (Aλ )1 , L(A)λ0 = µ+ν=λ

for λ ∈ Λ and L(A)λβ = 0 for λ ∈ Λ, β ∈ Q\∆. Yoshii has shown that L(A) is a centreless Lie Λ-torus of type ∆, and he proved the following coordinatization theorem. 10

If n = 0, we interpret q = ∅ and Fq = F.


25

Theorem 8.1 ([Y1, Thm. 1]). Any centreless Lie Λ-torus of type ∆ is graded-isomorphic to the Lie Λ-torus L(A) = TKK(A) constructed from a Jordan Λ-torus A. Yoshii went on to describe five families of Jordan Zn -tori and then he showed that every Jordan Λ-torus with rank(Λ) = n is isograded-isomorphic to a torus in one of the families [Y1, Thm. 2]. The simplest of these families + n consists of the tori F+ q , where Fq denotes the Z -graded algebra Fq with 1 multiplication x · y = 2 (xy + yx). Definition 8.2 (Isotopes of Jordan tori). Let A be Jordan Λ-torus and let u be a nonzero homogeneous element of A. So u ∈ A−ρ , where ρ ∈ suppλ (A). Let A(u) be the algebra with underlying vector space A and product ·u defined by 1 x ·u y = {x, u, y}. 2 Then, since u is invertible in A, A(u) is a Jordan algebra and the identity element of A(u) is 1(u) = u−1 [Mc2, p.86]. So 1(u) ∈ Aρ . We endow A(u) with a Λ-grading by setting (A(u) )λ = Aλ+ρ for λ ∈ Λ. It is easily checked that A(u) is a Jordan Λ-torus which we call the u-isotope of the Jordan torus A. It is also easily checked that up to gradedisomorphism A(u) does not depend on the choice of nonzero u in A−ρ . Suppose that u, v are nonzero homogeneous elements of a Jordan Λ-torus. It is well known that A(1) = A,

(A(u) )(v) = A(Uu v)

−2

and (A(u) )(u

)

= A,

(26)

as algebras [Mc2, Proposition 7.2.1], where Uu v = 12 {u, v, u}. It is easy to check that these are also equalities of Λ-graded Jordan algebras. Definition 8.3. Suppose that A is a Jordan Λ-torus and A is a Jordan Λ torus. We say that A is isotopic to A if some isotope of A is isogradedisomorphic to A . In that case, we write A ∼ A . It follows from (26) that ∼ is an equivalence relation on the class of Jordan tori. The following lemma is our key to understanding the connection between isotopy for Lie tori and isotopy for Jordan tori. In fact its proof (see (27) and (28) below) shows that the definition of isotope for Jordan tori is determined without foreknowledge from the definition of isotope for Lie tori. We will see the same phenomenon for alternative tori in §9 and for associative tori with involution in §11. Lemma 8.4 (Key lemma for type A1 ). Suppose that A is a Jordan Λ torus, A is an Jordan Λ-torus, and s ∈ Hom(Q, Λ) is admissible for L(A). Suppose that ϕ : L(A ) → L(A)(s) is a bi-isomorphism with ϕr = 1. Then,

26


−s(α) ∈ suppΛ (A) and for any nonzero u in A−s(α) there is an isogradedisomorphism η : A → A(u) such that ηgr = ϕe . Proof. Note first that since s is admissible, we have s(α) ∈ Λα (L(A)) = suppΛ (A), so −s(α) ∈ suppΛ (A). Thus we can choose nonzero u in A−s(α) as suggested. Let 1 denote the identity in A . Then −s(α)

ϕ(1−1 ) ∈ ϕ(L(A )0−α ) = ((L(A)(s) )0−α = L(A)−α

= (A−s(α) )−1 .

So replacing u by a scalar multiple we can assume that ϕ(1−1 ) = u−1 . Now define a linear bijection η : A → A by (ηx)1 = ϕ(x1 ) for x ∈ A . Then, for x, y ∈ A , we have (η(xy))1 = ϕ((xy)1 ) = =

1 ϕ({x, 1 , y}) = 12 ϕ([[x1 , 1−1 ], y1 ]) 2

1 [[(ηx)1 , u−1 ], (ηy)1 ] = ( 12 {ηx, u, ηy})1 = (η(x) ·u η(y))1 , 2

(27)

and, for λ ∈ Λ , λ

λ

ϕe (λ) (η(A ))1 = ϕ((A )1 ) = ϕ(L(A )λα ) = (L(s) )α

= (Aϕe (λ)+s(α) )1 = ((A(u) )ϕe (λ) )1 .

(28)

Proposition 8.5. Let A be a Jordan Λ-torus, and suppose that s ∈ Hom(Q, Λ) is admissible for L(A). Then, −s(α) ∈ suppΛ (A) and for any nonzero u in A−s(α) , we have L(A)(s) Q×Λ L(A(u) ). Proof. By Theorem 8.1, we have L(A)(s) Q×Λ L(A ) for some Jordan Λtorus A . So we have a bi-isomorphism ϕ : L(A ) → L(A)(s) with ϕr = 1 and ϕe = 1. Lemma 8.4 now tells us that A Λ A(u) , so L(A ) Q×Λ L(A(u) ). Hence L(A)(s) Q×Λ L(A(u) ).

Our main theorem in this section, gives necessary and sufficient conditions for two centreless Lie tori L(A) and L(A ) to be isotopic. To do this, we need to first give necessary and sufficient conditions for these Lie tori to be biisomorphic. Theorem 8.6. Suppose that A is a Jordan Λ-torus and A is a Jordan Λ torus. Let L(A) = TKK(A) and L(A ) = TKK(A ). Then (i) L(A) bi L(A ) ⇐⇒ A ig A . (ii) L(A) ∼ L(A ) ⇐⇒ A ∼ A .


27

Proof. Let L = L(A) and L = L(A ). (i): The proof of “⇐” is clear, so we prove “⇒”. Suppose ϕ : L → L is a bi-isomorphism. Then ϕr = ±1, so ϕr is in the Weyl group of ∆. Therefore, by Lemma 3.8, we can assume that ϕr = 1. Thus, by Lemma 8.4 with s = 0 and u = 1, we have A ig A. (ii): “⇒” By assumption, L bi L(s) for some s ∈ Hom(Q, Λ) that is admissible for L. So, choosing u as in Proposition 8.5, we have L bi L(A(u) ). By (i), A ig A(u) . “⇐” By (i) we can assume that A = A(u) , where u is a nonzero homogeneous element of A. Then u ∈ A−ρ , where ρ ∈ suppΛ L = Λα (L). So s : Q → Λ defined by s(α) = ρ is admissible for L, and hence L bi L(s) by Proposition 8.5.

We will next give an example of two centreless Lie tori that are isotopic but not bi-isomorphic. For this we need a little preparation. Remark 8.7. Suppose that A is a Jordan Λ-torus, and let Cent(A) be the centroid of A. It is clear that Cent(A) is Λ-graded and that the support Γ = Γ (A) of Cent(A) is a subgroup of Λ [Y1, §3]. Also, if S = suppΛ (A), it is clear that S is a union of cosets of Γ in Λ, and we let S/Γ = { λ+Γ | λ ∈ S } ⊆ Λ/Γ . If S/Γ is finite, we denote the sum of the elements of S/Γ in Λ/Γ by Σ(S/Γ ). Lemma 8.8. Suppose A is a Jordan Λ-torus with S/Γ finite and Σ(S/Γ ) = 0, where Γ = Γ (A) and S = suppΛ (A). If ρ ∈ S with |S/Γ | ρ ∈ / Γ and 0 = u ∈ A−ρ , then A(u) ig A. Proof. Let Γ (u) = Γ (A(u) ) and S (u) = suppΛ (A(u) ). It is easily checked that Cent(A(u) ) = Cent(A), so Γ (u) = Γ . Also, S (u) = S − ρ, so Σ(S (u) /Γ ) = Σ(S/Γ ) − |S/Γ | (ρ + Γ ) = − |S/Γ | (ρ + Γ ) = 0. But then Σ(S/Γ ) = 0 and Σ(S (u) /Γ ) = 0 imply our conclusion.

Our example uses Jordan tori of Clifford type [Y1, Example 5.2] as coordinate algebras. Example 8.9. Let Λ = Z3 and let R = F[2Λ] be the group algebra of 2Λ with its natural grading by 2Λ (and hence by Λ). Let { λ1 , λ2 , λ3 } be the 4 standard basis of Λ and let λ4 = λ1 + λ2 + λ3 . Let V = i=1 Rxi be the free R-module with base { x1 , x2 , x3 , x4 }, and give V the Λ-grading such that deg(tµ xi ) = µ + λi for µ ∈ 2Λ and 1 ≤ i ≤ 4. Define a Λ-graded R-bilinear form f : V × V → R by f (xi , xj ) = δi,j t2λi for 1 ≤ i, j ≤ 4. Finally, let A = R ⊕ V with the direct sum Λ-grading, and define a product in A by (r + v)(r + v ) = rr + f (v, v ) + rv + r v for r, r ∈ R, v, v ∈ V . Then A is a Jordan torus. (The algebra A is called a spin factor in [Mc2, p.74].) Further, using the notation of Lemma 8.8, Γ = 2Λ and S = (2Λ) ∪ (∪4i=1 (2Λ + λi )), so S/Γ is finite and Σ(S/Γ ) = 0. Also |S/Γ | ρ = 5ρ ∈ / Γ for ρ ∈ S \ Γ . So for any nonzero homogeneous element u in A \ R, we have, by Lemma 8.8, that A ig A(u) . Therefore, for any such u, we see by Theorem 8.6 that L(A) bi L(A(u) ) but L(A) ∼ L(A(u) ).

28


9 Type A2 Suppose in this section that ∆ = { 0 } ∪ { εi − εj | 1 ≤ i = j ≤ 3 } is the root system of type A2 (where ε1 , ε2 , ε3 is a basis for a space containing ∆), and Q = Q(∆). We let α1 = ε1 − ε2 and α2 = ε2 − ε3 , in which case { α1 , α2 } is a base for the root system ∆. Let A be an alternative Λ-torus. We construct a centreless Lie Λ-torus of type ∆ from A using J. Faulkner’s 3 × 3-projective elementary construction. (See [F1, Appendix, (A6)] for the properties mentioned below about this construction. Note however that we are using the notation eij (x) in place of the notation vij (x) that is used in [F1].) Recall that the 3×3 projective elementary Lie algebra is the Lie algebra L(A) = pe3 (A) := L0 ⊕ ( 1≤i=j≤3 eij (A)), (29) where L0 and eij (A) are subspaces of L satisfying: (a) (b) (c) (d) (e) (f) (g)

For i = j, there is a linear bijection x → eij (x) from A to eij (A), [eij (x), ejk (y)] = eik (xy) for x, y ∈ A and distinct i, j, k, [eij (x), eij (y)] = 0 for x, y ∈ A, i = j, [L0 , L0 ] ⊆ L0 , [L0 , eij (A)] ⊆ eij (A) for i = j, L0 = 1≤i<j≤3 [eij (A), eji (A)], { x ∈ L0 : [x, eij (A)] = 0 for i = j } = 0.

(30) These properties characterize the Lie algebra L(A) in the following sense: If L (A) = L0 ⊕ ( 1≤i=j≤3 eij (A)) is another Lie algebra satisfying the conditions in (30), then there is a unique isomorphism from L(A) → L (A) such that eij (x) → eij (x) for i = j, x ∈ A. Finally L(A) is a Q × Λ-graded algebra with L(A)λεi −εj = eij (Aλ ), L(A)λ0 = µ+ν=λ 1≤k 0 implies γ − α ∈ R and γ, α∨ < 0 implies γ + α ∈ R. We will now present some examples of (pre-)reflection systems.

62

Erhard Neher

2.7 Examples: Finite and locally finite root systems. A root system a la Bourbaki ([Bo2, VI, §1.1]) with 0 added will be called here a finite root ` system. Using the terminology introduced above, a finite root system is the same as a pre-reflection system (R, X) which is integral, coincides with its real part Re(R), and moreover satisfies the finiteness condition that R be a finite set. Replacing the finiteness condition by the local-finiteness condition defines a locally finite root system. Thus, a locally finite root system is a pre-reflection system (R, X) which is integral, coincides with its real part Re(R), and is locally-finite in the sense that |R∩Y | < ∞ for any finite-dimensional subspace Y of X. We mention some special features of locally finite root systems, most of them well-known from the theory of finite root systems. By definition, a locally finite root system R is integral and symmetric. One can show that it is also a reflection system whose root strings are unbroken and which is nondegenerate and coherent, but not necessarily reduced. Moreover, R is the direct sum of its connected components, see §2.4. In particular, R is connected if and only if R is indecomposable, in which case R is traditionally called irreducible. A root α ∈ R is called divisible or indivisible according to whether α/2 is a root or not. We put Rind = {0} ∪ {α ∈ R : α indivisible}

and Rdiv = {α ∈ R : α divisible}. (2.7.1) Both Rind and Rdiv are subsystems of R. A root basis of a locally finite root system R is a linearly independent set B ⊂ R such that every α ∈ R is a Z-linear combination of B with coefficients all of the same sign. A root system R has a root basis if and only if all irreducible components of R are countable ([LN1, 6.7, 6.9]). The set R∨ = {α∨ ∈ X ∗ : α ∈ R} is a locally finite root system in ∨ X = spanK (R∨ ), called the coroot system of R. The root systems R and R∨∨ are canonically isomorphic ([LN1, Th. 4.9]). 2.8 Classification of locally finite root systems. ([LN1, Th. 8.4]) Examples of possibly infinite locally finite root systems are the classical root systems ˙ I – BCI , defined as follows. Let I be an arbitrary set and let XI = A i∈I K i be the vector space with basis { i : i ∈ I}. Then ˙ I = { i − j : i, j ∈ I}, A BI = {0} ∪ {± i : i ∈ I} ∪ {± i ± j : i, j ∈ I, i = j} = {± i : i ∈ I} ∪ DI , CI = {± i ± j : i, j ∈ I} = {±2 i : i ∈ I} ∪ DI , DI = {0} ∪ {± i ± j : i, j ∈ I, i = j}, and BCI = {± i ± j : i, j ∈ I} ∪ {± i : i ∈ I} = BI ∪ CI ˙ I which only spans are locally finite root systems which span XI , except for A ˙ ˙ a subspace XI of codimension 1 in XI . The notation A is supposed to indicate

EALAs and Generalizations

63

this fact. For finite I, say |I| = n ∈ N, we will use the usual notation An = ˙ I , . . . , DI are ˙ {0,...,n} , but Bn , Cn , Dn , BCn for |I| = n. The root systems A A reduced, but R = BCI is not. Any locally finite root system R is the direct limit of its finite subsystems, which are finite root systems, and if R is irreducible it is a direct limit of irreducible finite subsystems. It is then not too surprising that an irreducible locally finite root system is isomorphic to exactly one of (i) the finite exceptional root systems E6 , E7 , E8 , F4 , G2 or ˙ I (|I| ≥ 1), BI (|I| ≥ 2), CI (|I| ≥ 3), DI (ii) one of the root systems A (|I| ≥ 4) or BCI (|I| ≥ 1). Conversely, all the root systems listed in (i) and (ii) are irreducible. The classification of locally finite root systems is independent of the base field K, which could therefore be taken to be Q or R. But since the root systems which we will later encounter naturally “live” in K-vector spaces, fixing the base field is not convenient. Locally finite root systems can also be defined via invariant nondegenerate symmetric bilinear forms: 2.9 Proposition ([LN2, 2.10]) For an integral pre-reflection system (R, X) the following conditions are equivalent: (i) R is a locally finite root system; (ii) there exists a nondegenerate strictly invariant form on (R, X), and for every α ∈ R the set R, α∨ is bounded as a subset of Z. In this case, (R, X) has a unique invariant bilinear form (·|·) which is normalized in the sense that 2 ∈ {(α|α) : 0 = α ∈ C} ⊂ {2, 4, 6, 8} for every connected component C of R. The normalized form is nondegenerate in general and positive definite for K = R. 2.10 Example: Reflection systems associated to bilinear forms. Let R be a spanning set of a vector space X containing 0, and let (·|·) be a symmetric bilinear form on X. For α ∈ X we denote the linear form x → (α|x) by α . Let Φ ⊂ {α ∈ R : (α|α) = 0}, define ∨ : R → X ∗ by 2α if α ∈ Φ, ∨ α = (α|α) (2.10.1) 0 otherwise and then define sα by 2.2.1. Thus sα is the orthogonal reflection in the hyperplane α⊥ if α ∈ Φ, and the identity otherwise. If sα (R) ⊂ R and sα (Φ) ⊂ Φ for all α ∈ Φ, then (R, X, s) is a coherent reflection system with the given subset Φ as set of reflective roots. The bilinear form (·|·) is invariant, and it is strictly invariant if and only if Rim = R ∩ Rad (·|·), i.e., (·|·) is affine in the sense of 3.8.

64

Erhard Neher

By 2.9 every locally finite root system is of this type. But also the not necessarily crystallographic finite root systems, see for example [Hu, 1.2], arise in this way. The latter are not necessarily integral reflection systems. 2.11 Example: Roots in Lie algebras with toral subalgebras. Let L be a Lie algebra over a ring k containing 12 (this generality will be used later). Following [Bo1, Ch.VIII, §11] we will call a nonzero triple (e, h, f ) of elements of L an sl2 -triple if [e, f ] = −h,

[h, e] = 2e

For example, in sl2 (k) = the elements esl2 =

01 , 00

and [h, f ] = −2f.

a b : a, b, c ∈ k c −a

hsl2 =

1 0 , 0 −1

0 0 , −1 0

fsl2 =

(2.11.1)

form an sl2 -triple in sl2 (k). In general, for any sl2 -triple (e, h, f ) in a Lie algebra L there exists a unique Lie algebra homomorphism ϕ : sl2 (k) → L mapping the matrices in (1) onto the corresponding elements (e, h, f ) in L. Now let k = K be a field of characteristic 0, and let L be a Lie algebra over K and T ⊂ L a subspace. We call T ⊂ L a toral subalgebra, sometimes also called an ad-diagonalizable subalgebra, if L = α∈T ∗ Lα (T ) where Lα (T ) = {x ∈ L : [t, x] = α(t)x for all t ∈ T } (2.11.2) for any α ∈ T ∗ . In this case, (2.11.2) is referred to as the root space decomposition of (L, T ) and R = {α ∈ T ∗ : Lα (T ) = 0} as the set of roots of (L, T ). We will usually abbreviate Lα = Lα (T ). Any toral subalgebra T is abelian, thus T ⊂ L0 and 0 ∈ R unless T = {0} = E (which is allowed but not interesting). We will say that a toral T is a splitting Cartan subalgebra if T = L0 . We denote by Rint the subset of integrable roots of (L, T ), i.e., those α ∈ R for which there exists an sl2 -triple (eα , hα , fα ) ∈ Lα × T × L−α such that the adjoint maps ad eα and ad fα are locally nilpotent. Let X = spanK (R) ⊂ T ∗ and for α ∈ R define sα ∈ GL(X) by x − x(hα )α, α ∈ Rint , sα (x) = x otherwise. It is a straightforward exercise in sl2 -representation theory to verify that with respect to the reflections sα defined above, (R, X) is an integral pre-reflection system, which is reduced in case T is a splitting Cartan subalgebra. A priori, sα will depend on hα , but this is not so for the Lie algebras which are of main interest here, the invariant affine reflection algebras of §6.


65

We note that several important classes of pre-reflection systems arise in this way from Lie algebras, e.g., locally finite root systems [Neh3], Kac–Moody root systems [Kac2], [MP] and extended affine root systems [AABGP], see 3.11. An example of the latter is described in detail in 2.12. We point out that in general R is not symmetric. For example, this is so for the set of roots of classical Lie superalgebras [Kac1] (they can be viewed in the setting above by forgetting the multiplication of the odd part). 2.12 Example: Affine root systems. Let S be an irreducible locally finite root system. By 2.9, S has a unique normalized invariant form, with respect to which we can introduce short roots Ssh = {α ∈ S : (α|α) = 2} and long roots Slg = {α ∈ S : (α|α) = 4 or 6} = S \ (Ssh ∪ Sdiv ), where Sdiv are the divisible roots of 2.7. Both Ssh ∪ {0} and Slg ∪ {0} are subsystems of S and hence locally finite root systems. Note that S = {0} ∪ Ssh if S is simply-laced, (Sdiv \ {0}) = ∅ ⇔ S = BCI , and if Slg = ∅ the set of long roots is given by Slg = {α ∈ S : (α|α) = k(S)} for 2, S = F4 or of type BI , CI , BCI , |I| ≥ 2, (2.12.1) k(S) = 3 S = G2 . Let X = span(S) ⊕ Kδ and let (·|·) be the symmetric bilinear form which restricted to span(S) is the normalized invariant form of S and which satisfies (X | Kδ) = 0. Furthermore, choose a tier number t(S) = 1 or t(S) = k(S), with t(S) = 1 in case S = BCI , |I| ≥ 2. Finally, put

Φ= α ⊕ t(S)Zδ ∪ α + (1 + 2Z)δ , (α ⊕ Zδ) ∪ α∈Ssh

α∈Slg

0=α∈Sdiv

R = Zδ ∪ Φ. Of course, if Slg = ∅ or Sdiv = {0} the corresponding union above is to be interpreted as the empty set. Note (β|β) = 0 for any β ∈ Φ. One can easily verify that every reflection sβ defined in 2.10 leaves Φ and Zδ invariant, hence R = R(S, t(S)) is a symmetric reflection system, called the affine root system associated to S and t(S). The following table lists all possibilities for (S, t(S)) where S is locally finite. For finite S of type XI , say |I| = l, we also include the corresponding labels of the affine root system R(S, t(S)) as defined [MP] (third column) and [Kac2] (fourth column). If S is reduced, then R = S ⊕ Zδ, i.e., t(S) = 1 in the definition of Φ, is a called an untwisted affine root system. 2.13 Affine Lie algebras. Every affine root system appears as the set of roots of a Lie algebra E with a splitting Cartan subalgebra H in the sense of 2.11 (the letters (E, H) are chosen in anticipation of the definition of an extended affine Lie algebra 6.11). This is easily verified in the untwisted case.

66

Erhard Neher

S

t(S)

affine label [MP]

affine label [Kac2]

reduced

1

S (1)

S (1)

BI (|I| ≥ 2)

2

Bl

(2)

Dl+1

CI (|I| ≥ 3)

2

Cl

(2)

A2l−1

F4

2

F4

G2

3

G2

BC1

−

BC1

BCI (|I| ≥ 2)

1

BCl

(2)

(2)

(2)

E6

(2)

(3)

D4

(3)

(2)

A2

(2)

(2)

A2l

(2)

Indeed, let g be a finite-dimensional split simple Lie K-algebra, e.g., a finite-dimensional simple Lie algebra over an algebraically closed field K, or let g be a centreless infinite rank affine algebra [Kac2, 7.11]. Equivalently, g is a locally finite split simple Lie algebra, classified in [NS] and [Stu], or g is the Tits–Kantor–Koecher algebra of a Jordan pair spanned by a connected grid, whose classification is immediate from [Neh3]. In any case, g contains a splitting Cartan subalgebra h such that the set of roots of (g, h) is an irreducible locally finite root system S, which is finite if and only if g is finite dimensional. Moreover, g carries an invariant nondegenerate symmetric bilinear form κ, unique to a scalar (in the finite-dimensional case κ can be taken to be the Killing form). The Lie algebra E will be constructed in three steps: (I) L = g⊗K[t±1] is the untwisted loop algebra of g. Its Lie algebra product [·, ·]L is given by [x ⊗ p, y ⊗ q]L = [x, y]g ⊗ pq for x, y ∈ g and p, q ∈ K[t±1 ]. Although L could be viewed as a Lie algebra over K[t±1 ], we will view L as a Lie algebra over K. (II) K = L ⊕ Kc with Lie algebra product [., .]K given by [l1 ⊕ s1 c, l2 ⊕ s2 c]K = [l1 , l2 ]L ⊕ σ(l1 , l2 )c for li ∈ L, si ∈ K and σ : L × L → K the K-bilinear map determined by σ(x1 ⊗ tm , x2 ⊗ tn ) = δm+n,0 m κ(x1 , x2 ). (III) E = K ⊕ Kd. Here d is the derivation of K given by d(x ⊗ tm ) = mx ⊗ tm , d(c) = 0, and E is the semidirect product of K and Kd, i.e., [k1 ⊕ s1 d, k2 ⊕ s2 d]E = [k1 , k2 ]K + s2 d(k2 ) − s1 d(k1 ). One calls E the untwisted affine Lie algebra associated to g.


67

Denote the root spaces of (g, h) by gα , so that h = g0 . Then H = (h ⊗ K1) ⊕ Kc ⊕ Kd is a toral subalgebra of E whose set of roots in E is the untwisted affine root system associated to the root system S of (g, h): E0 = H, Emδ = h ⊗ Ktm for 0 = m ∈ Z and Eα⊕mδ = gα ⊗ tm for 0 = α ∈ S and m ∈ Z. One can also realize the twisted affine root systems, i.e., those with t(S) > 1, by replacing the untwisted loop algebra L in (I) by a twisted version. To define it in the case of a finite-dimensional g, let σ be a nontrivial diagram automorphism of g of order r. Hence g is of type Al (l ≥ 2, r = 2), Dl (l ≥ 4, r = 2), D4 (r = 3) or E6 (r = 2). For an infinite-dimensional g (where diagrams need not exist) we take as σ the obvious infinite-dimensional analogue of the corresponding matrix realization of σ for a finite-dimensional g, e.g., x → −xt for type A. The twisted loop algebra L is the special case n = 1 of the multi-loop algebra Mr (g, σ) defined in 4.7. The construction of K and E are the same as in the untwisted case. Replacing the untwisted h in the definition of H above by the subspace hσ of fixed points under σ defines a splitting Cartan subalgebra of E in the twisted case, whose set of roots is a twisted affine root systems and all twisted affine root systems arise in this way, see [Kac2, 8.3] for details. In the later sections we will generalize affine root systems and affine Lie algebras. Affine root systems will turn out to be examples of extended affine root systems and locally extended affine root systems 3.11, namely those of nullity 1. Extended and locally extended affine root systems are in turn special types of affine reflection systems, which we will describe in the next section. In the same vein, untwisted affine Lie algebras are examples of (locally) extended affine Lie algebras, which in turn are examples of affine reflection Lie algebras which we will study in §6. From the point of view of extended affine Lie algebras, §6.11, the Lie algebras K and L are the core and centreless core of the extended affine Lie algebra (E, H). As Lie algebras, they are Lie tori, defined in 5.1. The construction from L to E described above is a special case of the construction in 6.13, see the example in 6.14. 2.14 Notes. With the exceptions mentioned in the text and below, all results in this section are proven in [LN1] and [LN2]. Many of the concepts introduced here are well known from the theory of finite root systems. The notion of tameness (§2.2) goes back to [ABY]. It is equivalent to the requirement that R not have isolated roots, where β ∈ R is called an isolated root (with respect to Rre ), if α + β ∈ R for all α ∈ Rre . In the setting of the Lie algebras considered in §6, tameness of the root system is a consequence of tameness of the algebra, see 6.8. Locally finite root systems are studied in [LN1] over R. But as already mentioned in [LN1, 4.14] there is a canonical equivalence between the categories of locally finite root systems over R and over any field K of characteristic

68

Erhard Neher

0. In particular, the classification is the “same” over any K. The classification itself is proven in [LN1, Th. 8.4]. A substitute for the non-existence of root bases are grid bases, which exist for all infinite reduced root systems ([Neh1]). Our definition of an sl2 -triple follows Bourbaki ([Bo1, Ch.VII, §11]). It differs from the one used in other texts by a sign. Replacing f by −f shows that the two notions are equivalent. Bourbaki’s definition is more natural in the setting here and avoids some minus signs later. The concept of an integrable root (2.11) was introduced by Neeb in [Nee1]. For a finite S, affine root systems (2.12) are determined in [Kac2, Prop. 6.3]. The description given in 2.12 is different, but of course equivalent to the one in [Kac2]. It is adapted to viewing affine root systems as extended affine root systems of nullity 1. In fact, the description of affine root systems in 2.12 is a special case of the Structure Theorem for extended affine root systems [AABGP, II, Th. 2.37], keeping in mind that a semilattice of nullity 1 is a lattice by [AABGP, II, Cor. 1.7]. The table in 2.12 reproduces [ABGP, Table 1.24]. Other examples of reflection systems are the set of roots associated to a “root basis” in the sense of Hée [He]. Or, let R be the root string closure of the real roots associated to root data à la Moody–Pianzola [MP, Ch. 5]. Then R is a symmetric, reduced and integral reflection system. The Weyl group of a pre-reflection system is in general not a Coxeter group. For example, this already happens for locally finite root systems, see [LN1, 9.9]. As a substitute, one can give a presentation by conjugation, essentially the relation (ReS4), see [LN1, 5.12] for locally finite root systems and [Ho] for Weyl groups of extended affine root systems (note that the Weyl groups in [Ho] and also in [AABGP] are defined on a bigger space than the span of the roots and hence our Weyl groups are homomorphic images of the Weyl groups in [Ho] and [AABGP]).

3 Affine reflection systems In this section we describe extensions of pre-reflection systems and use them to define affine reflection systems, which are a generalization of extended affine root systems. The roots of the Lie algebras, which we will be studying in §6, will turn out to be examples of affine reflection systems. Hints to references are given in 3.12. 3.1 Partial sections. Let f : (R, X) → (S, Y ) be a morphism of prereflection systems with f (R) = S, and let S ⊂ S be a subsystem spanning Y . A partial section of f over S is a morphism g : (S , Y ) → (R, X) of prereflection systems such that f ◦ g = IdY . The name is (partially) justified by the fact that a partial section leads to a partial section of the canonical epimorphism W (R) → W (S), namely a section defined over W (S ).


69

Let f : (R, X) → (S, Y ) be a morphism of pre-reflection systems satisfying f (R) = S. In general, a partial section of f over all of S need not exist. However, partial sections always exist in the category of reflection systems. 3.2 Extensions. Recall 2.2: f (Rim ) ⊂ S im and f (Rre ) ⊂ {0} ∪ S re for any morphism f : (R, X) → (S, Y ) of pre-reflection systems. We call f an extension if f (Rim ) = S im and f (Rre ) = S re . We will say that R is an extension of S if there exists an extension f : R → S. We mention some properties of an extension f : R → S. By definition, f is surjective. Also, R is coherent if and only if S is so, and in this case f induces a bijection C → f (C) between the set of connected components of Re(R) and of Re(S). In particular, Re(R) is connected ⇐⇒ Re(S) is connected ⇐⇒ Re(R) is indecomposable. Moreover, R is integral if and only if S is so. Finally, f maps a root string S(β, α), β ∈ R, α ∈ Rre , injectively to S f (β), f (α) . If R is an extension of a nondegenerate S, e.g., a locally finite root system, then S is unique up to isomorphism and will be called the quotient pre-reflection system of R. On the other hand, if R is nondegenerate every extension f : R → S is injective, hence an isomorphism. In particular, a locally finite root system R does not arise as a nontrivial extension of a prereflection system S. But as we will see below, a locally finite root system does have many interesting extensions. A first example of an extension is the canonical projection R → S for R an affine root system associated to a locally finite irreducible root system S, 2.12. In this case, S is the quotient pre-reflection system = quotient root system of R. 3.3 Extension data. Let (S, Y ) be a pre-reflection system, let S be a subsystem of S with span(S ) = Y and let Z be a K-vector space. A family L = (Λξ )ξ∈S of nonempty subsets of Z is called an extension datum of type (S, S , Z) if (ED1) for all ξ, η ∈ S and all λ ∈ Λξ , µ ∈ Λη we have µ − η, ξ ∨ λ ∈ Λsξ (η) , (ED2) 0 ∈ Λξ for all ξ ∈ S , and (ED3) Z is spanned by the union of all Λξ , ξ ∈ S. The only condition on Λ0 is (ED2): 0 ∈ Λ0 . Moreover, Λ0 is related to the other Λξ , 0 = ξ ∈ S, only via the axiom (ED3). As (ED3) only serves to determine Z and can always be achieved by replacing Z by the span of the Λξ , ξ ∈ S, it follows that one can always modify a given extension datum by replacing Λ0 by some other set containing 0. However, as 3.6 shows, this may change the properties of the associated pre-reflection system. Any extension datum L has the following properties, where WS ⊂ W (S) is the subgroup generated by all sξ , ξ ∈ S .

70

Erhard Neher

Λ−ξ = −Λξ

for all ξ ∈ S re ,

2Λξ − Λξ ⊂ Λξ for all ξ ∈ S re , Λη = Λw (η) for all η ∈ S and w ∈ WS , Λη − η, ξ ∨ Λξ ⊂ Λη for ξ ∈ S , η ∈ S Λξ = Λ−ξ for ξ ∈ S .

and

In particular, Λξ is constant on the WS -orbits of S. However, in general, the Λξ are not constant on all of S, see the examples in 3.7. One might wonder if the conditions (ED1)–(ED3) are strong enough to force the Λξ to be subgroups of (Z, +). But it turns out that this is not the case. For example, this already happens for the extension data of locally finite root systems, 3.7. On the other hand, if R is an integral pre-reflection system and Λ a subgroup of (Z, +) which spans Z, then Λξ ≡ Λ is an example of an extension datum, the so-called untwisted case, cf. the definition of an untwisted affine root system in 2.12. The conditions above have appeared in the context of reflection spaces (not to be confused with a reflection system). Recall ([Loo]) that a reflection space is a set S with a map S × S → S : (s, t) → s · t satisfying s · s = s, s · (s · t) = t and s · (t · u) = (s · t) · (s · u) for all s, t, u ∈ S. A reflection subspace of a reflection space (S, ·) is a subset T ⊂ S such that t1 · t2 ∈ T for all t1 , t2 ∈ T . In this case, T is a reflection space with the induced operation. Any abelian group (Z, +) is a reflection space with respect to the operation x · y = 2x − y for x, y ∈ Z. Correspondingly, a reflection subspace of the reflection space (Z, ·) is a subset A ⊂ Z satisfying 2a − b ∈ A for all a, b ∈ A, symbolically 2A − A ⊂ A. Hence all Λξ , ξ ∈ S re , are reflection subspaces of (Z, ·). Moreover, for any subset A of Z it is easily seen that A − 2A ⊂ A ⇔ A = −A and 2A + A ⊂ A ⇔ A = −A and 2A − A ⊂ A.

(3.3.1)

A subset A satisfying (3.3.1) will be called a symmetric reflection subspace. We will consider 0 as the base point of the reflection space Z. Also, we denote by Z[A] the subgroup of (Z, +) generated by A ⊂ Z. Then the following are equivalent ([NY, 2.1]): (i) 0 ∈ A and A − 2A ⊂ A, (ii) 0 ∈ A and 2A − A ⊂ A, (iii) 2Z[A] ⊂ A and 2Z[A] − A ⊂ A, (iv) A is a union of cosets modulo 2Z[A], including the trivial coset 2Z[A]. In this case, A will be called a pointed reflection subspace. It is immediate from the above that every Λξ , ξ ∈ S re is a pointed reflection subspace. We note that pointed reflection subspaces are necessarily symmetric. It is obvious from (iv) above that a pointed reflection subspace is in general not a subgroup of (Z, +).


71

The following theorem characterizes extensions in terms of extension data. 3.4 Theorem. Let (S, Y ) be a pre-reflection system. (a) Let L = (Λξ )ξ∈S be an extension datum of type (S, S , Z). Put X := Y ⊕ Z, denote by π : X → Y the projection with kernel Z, and define

R := ξ ⊕ Λξ ⊂ X and sα (x) := sξ (y) ⊕ z − y, ξ ∨ λ , (3.4.1) ξ∈S

for all α = ξ⊕λ ∈ ξ⊕Λξ ⊂ R and all x = y⊕z ∈ X. Then R is a pre-reflection system in X, denoted E = E(S, S , L). Moreover, π : (R, X) → (S, Y ) is an extension of pre-reflection systems, and the canonical injection ι : Y → X is a partial section of π over S . (b) Conversely, let f : (R, X) → (S, Y ) be an extension and let g : S → R be a partial section of f , cf. 3.1. For every ξ ∈ S define Rξ ⊂ R and Λξ ⊂ Z := Ker(f ) by Rξ = R ∩ f −1 (ξ) = g(ξ) ⊕ Λξ . (3.4.2) Then L = (Λξ )ξ∈S is an extension datum of type (S, S , Z), and the vector space isomorphism ϕ : Y ⊕ Z ∼ = X sending y ⊕ z to g(y) ⊕ z is an isomorphism E(S, S , L) ∼ = R of pre-reflection systems making the following commutative diagram: S E EE g ι yyy EE EE yy y " |y ϕ /R. E EE y EE y y E yy π EE " |yy f S (c) In the setting of (b), the following are equivalent for g ∈ HomK (Y, Z): (i) g : S → R is another partial section of f , (ii) there exists ϕ ∈ HomK (Y, Z) such that g = g + ϕ and ϕ(ξ ) ∈ Λξ for all ξ ∈ S . In this case, the extension datum L = (Λξ )ξ∈S defined by (3.4.2) with respect to g is related to the extension datum L by Λξ = Λξ − ϕ(ξ)

for ξ ∈ S.

(3.4.3)

In the context of root-graded Lie algebras, the partial sections g and g in (c) lead to isotopic Lie algebras, see 5.2 and 6.4. 3.5 Corollary. Let (S, Y ) be a pre-reflection system and let L = (Λξ )ξ∈S be an extension datum of type (S, S , Z). Let R = E(S, S , Z) be the pre-reflection system defined in 3.4(a).

72

Erhard Neher

1. R is reduced if and only if, for all 0 = ξ ∈ S, ξ, cξ ∈ S for c ∈ K \ {0, ±1}

=⇒

Λcξ ∩ cΛξ = ∅.

2. R is symmetric if and only if S is symmetric and Λ−ξ = −Λξ for all ξ ∈ S im . 3. R is a reflection system if and only if S is a reflection system. 3.6 Affine reflection systems. A pre-reflection system is called an affine reflection system if it is an extension of a locally finite root system. A morphism between affine reflection systems is a morphism of the underlying pre-reflection systems. The name affine reflection system, as opposed to affine pre-reflection system, is justified in view of 3.5(c), since a locally finite root system is a reflection system and hence so is any extension of it. A morphism between affine reflection systems is a morphism of the underlying reflection systems (2.2). Isomorphisms have the following characterization. A vector space isomorphism f : X → X is an isomorphism of the reflection systems (R, X) and (R , X ) if and only if f (Rim ) = Rim and f (Rre ) = Rre .

(3.6.1)

By definition, the nullity of an affine reflection system R is the rank of the torsion-free abelian group Rim generated by Rim , i.e., nullity = dimQ ( Rim ⊗Z Q). It is clear from the definitions that a locally finite root system is an affine reflection system of nullity 0. The affine root systems R(S, t(S)) of 2.12 are affine reflection systems of nullity 1, see also 3.11(a). In the following let (R, X) be an affine reflection system and let f : (R, X) → (S, Y ) be an extension where (S, Y ) is a locally finite root system. As mentioned in 3.2, R is then coherent and integral. Also, nondegeneracy of S implies that S is unique, up to a unique isomorphism. We will call S the quotient root system of R in this context and refer to f as the canonical projection. Let Z = Ker(f ). One can show that the extension f has a partial section g over Sind . Let L = (Λξ )ξ∈S be the extension datum of type (S, Sind , Z) associated to f and g in Th. 3.4(b). Thus, up to an isomorphism we may assume that R is given by (3.4.1). The extension datum L appearing there has some special properties besides the ones mentioned in 3.5. Namely, for 0 = ξ, η ∈ S and w ∈ W (S) we have Λξ = Λ−ξ = −Λξ = Λw(ξ) for all w ∈ W (S), Λξ ⊃ Λξ − ξ, η ∨ Λη , and Λξ ⊃ Λ2ξ whenever 2ξ ∈ S.

Moreover, define Λdiff := 0=ξ∈S Λξ − Λξ . Then ZΛdiff = Λdiff and we have: (a) R is symmetric if and only if Λ0 = −Λ0 .


73

(b) All root strings S(β, α), β ∈ R, α ∈ Rre , are unbroken if and only if Λdiff ⊂ Λ0 . (c) R is tame (cf. 2.2) if and only if Λ0 ⊂ Λdiff . (d) |S(β, α)| ≤ 5 for all β ∈ R and α ∈ Rre . We now describe extension data in the irreducible case in more detail. 3.7 Extension data for irreducible locally finite root systems. Let S be an irreducible locally finite root system. Recall the decomposition S = Ssh ∪ Slg ∪ Sdiv of 2.12. Let L = (Λξ )ξ∈S be an extension datum of type (S, Sind , Z). One knows that W (S) operates transitively on the roots of the same length ([LN1, 5.6]). By 3.6 we can therefore define Λsh , Λlg and Λdiv by ⎧ ⎪ ⎨Λsh for α ∈ Ssh , Λα = Λlg (3.7.1) for α ∈ Slg , ⎪ ⎩ Λdiv for 0 = α ∈ Sdiv . By convention, Λlg and Λdiv are only defined if the corresponding set of roots Slg and Sdiv \ {0} are not empty (we always have Ssh = ∅). To streamline the presentation, this will not be specified in the following. It will (hopefully) always be clear what is meant from the context. (a) We have already seen in 3.3 that in general Λsh and Λlg are pointed reflection subspaces and that Λdiv is a symmetric reflection subspace of Λ. In the setting of this subsection we have in addition the following: (i) Λsh is a subgroup of Z, if Ssh ∪ {0} contains a subsystem of type A2 , (ii) Λlg is a subgroup of Z, if Slg ∪ {0} contains a subsystem of type A2 , and (iii) the following relations hold for S as indicated and k = k(S) defined in 2.12: Λsh + Λlg ⊂ Λsh , Λsh + Λdiv ⊂ Λsh , Λlg + Λdiv ⊂ Λlg ,

Λlg + kΛsh ⊂ Λlg , Λdiv + 4Λsh ⊂ Λdiv , Λdiv + 2Λlg ⊂ Λdiv ,

(Slg = ∅) (S = BC1 ) (S = BCI , |I| ≥ 2).

(b) Conversely, given a subset Λ0 with 0 ∈ Λ0 , pointed reflection subspaces Λsh and Λlg (if Slg = ∅) and a symmetric reflection subspace Λdiv ⊂ Λ (if {0} Sdiv ) satisfying (i) – (iii) of (a), then (3.7.1) defines an extension datum of type (S, Sind , Z). Moreover, the following hold where, we recall k is defined in (2.12.1): (i) Λsh is a subgroup of Z if S = A1 , BI or BCI for any I, (ii)’ Λlg is a subgroup of Z if S = BI or BCI and |I| ≥ 3, or if S = F4 or G2 , (iv) kΛsh ⊂ Λlg ⊂ Λsh , 4Λsh ⊂ Λdiv ⊂ Λsh and Λdiv ⊂ Λlg ⊂ Λsh . (v) The inclusions Λsh + Λdiv ⊂ Λsh and Λdiv + 4Λsh ⊂ Λdiv hold for all S = BCI . One therefore obtains the following description of L× := (Λsh , Λlg , Λdiv ) for the various types of irreducible root systems. Note that the only condition on Λ0 is 0 ∈ Λ0 because of (ED2).

74

Erhard Neher

˙ I , DI (|I| ≥ 4), E6 , E7 or E8 : L× = (Λsh ), (I) S is simply laced, i.e., S = A where Λsh is a pointed reflection subspace for S = A1 and a subgroup of Z otherwise. (II) S = BI (|I| ≥ 2), CI (|I| ≥ 3) or F4 : L× = (Λsh , Λlg ), where Λsh and Λlg are pointed reflection subspaces satisfying Λsh + Λlg ⊂ Λsh and Λlg + 2Λsh ⊂ Λlg . Moreover, Λsh is a subgroup of Z if S = CI or F4 , while Λlg is a subgroup if S = BI , |I| ≥ 3 or F4 . (III) S = G2 : L× = (Λsh , Λlg ), where Λsh and Λlg are subgroups of Z satisfying Λsh + Λlg ⊂ Λsh and Λlg + 3Λsh ⊂ Λlg . (IV) S = BC1 : L× = (Λsh , Λdiv ), where Λsh is a pointed reflection subspace, Λdiv is a symmetric reflection subspace and Λsh + Λdiv ⊂ Λsh and Λdiv + 4Λsh ⊂ Λdiv . (V) S = BCI , |I| ≥ 2: L× = (Λsh , Λlg , Λdiv ) where Λsh and Λlg are pointed reflection subspaces, Λdiv is a symmetric reflection subspace and Λsh + Λlg ⊂ Λsh , Λlg + 2Λsh ⊂ Λlg , Λlg + Λdiv ⊂ Λlg , Λdiv + 2Λlg ⊂ Λdiv . Moreover, if |I| ≥ 3 we require that Λlg is a subgroup of Z. 3.8 Affine forms. Our definition of affine reflection systems follows the approach of [Bo2] where root systems are defined without reference to a bilinear form. In the literature, it is customary to define affine root systems and their generalizations, the extended affine root systems (EARS), in real vector spaces using positive semidefinite forms. We will therefore give a characterization of affine reflections systems in terms of affine invariant forms where, by definition, an affine form of a pre-reflection system (R, X) over K is an invariant form b satisfying Rim = R ∩ Rad b, cf. (2.5.2). In particular, affine forms are strictly invariant in the sense of 2.5. As an example, the forms used in the theory of EARS are affine forms in our sense. In particular, the form (·|·) in 2.12 is an affine form of the affine root system R. 3.9 Theorem. Let (R, X) be a pre-reflection system. Then (R, X) is an affine reflection system if and and only if it satisfies the following conditions: (i) (R, X) is integral, (ii) (R, X) has an affine form, and (iii) R, α∨ is bounded for every α ∈ Rre . In this case: (a) Let b be an affine form for (R, X) and let f : X → X/ Rad b be the canonical map. Then (S, Y ) = (f (R), X/ Rad b) is the quotient root system of R and f its canonical projection. Moreover, Re(R) is connected if and only if S is irreducible. (b) There exists a unique affine form (·|·)a on (R, X) that is normalized in the sense of 2.9, i.e., for every connected component C of Re(R) we have 2 ∈ {(α|α)a : 0 = α ∈ C} ⊂ {2, 3, 4, 6, 8}.


75

The form (·|·)a satisfies {(α|α)a : 0 = α ∈ C} ∈ {2}, {2, 4}, {2, 6}, {2, 8}, {2, 4, 8} . Its radical is Rad (·|·)a = Ker f . If K = R then (·|·)a is positive semidefinite. 3.10 Corollary. A pre-reflection system over the reals is an affine reflection system if and only if it is integral and has a positive semidefinite affine form. 3.11 Special types of affine reflection systems. As usual, the rank of a reflection system (R, X) is defined as rank(R, X) = dim X. If R has finite rank and K = R, we will say that R is discrete if R is a discrete subset of X. (a) Let (R, X) be an affine reflection system over K = R with the following properties: R has finite rank, Re(R) is connected, R = −R and R is discrete. Then R is called (i) an EARS, an abbreviation for extended affine root system, if R is reduced, tame (see (3.6.c)), and all root strings are unbroken; (ii) a SEARS, an abbreviation of Saito’s extended affine root system, if R = Re(R). It was shown in [Az3, Th. 18] that every reduced SEARS can be uniquely extended to an EARS and, conversely, the reflective roots of an EARS are the nonzero roots of a SEARS. This is now immediate. Indeed, by the results mentioned in 3.6, an affine reflection system is tame and has unbroken root strings if and only Rim = Λdiff . The affine root systems R(S, t(S)) of 2.12 with S finite are precisely the EARSs of nullity 1 ([ABGP]). (b) In [MY], Morita and Yoshii define a LEARS, an abbreviation for locally extended affine root system. In our terminology, this is a symmetric affine reflection system R over K = R such that R = Re(R) is connected. The equivalence of this definition with the one in [MY] follows from 3.10. (c) In [Az4], Azam defines a GRRS, an abbreviation of a generalized reductive root system. In our terminology, this is a symmetric real reduced, discrete affine reflection system R which has finite rank and unbroken root strings. 3.12 Notes. With the exceptions mentioned in the text and below, all results in this section are proven in [LN2]. If S is an integral reflection system, our definition of an extension datum in 3.3 makes sense for any abelian group Z instead of a K-vector space. This generality is however not needed for extension data arising from Lie algebras with toral subalgebras, 2.11, like the generalizations of affine Lie algebras to be discussed later. But we note that the results on extension data stated in 3.6 and 3.7 are true for an abelian group Z. In this generality, but for S a finite irreducible root system, extension data without Λ0 were defined by Yoshii in [Yo4] as “root systems of type S extended by Z”. The results stated in 3.7 extend [Yo4, Th. 2.4] to the setting of locally finite root systems.

76

Erhard Neher

We point out that our definition of an EARS is equivalent to the one given by Azam, Allison, Berman, Gao, and Pianzola in [AABGP, II, Def. 2.1]. That our definition of a SEARS is the same as Saito’s definition of an “extended affine root system” in [Sa], follows from 3.10. Our characterization of extensions in terms of extension data 3.4 generalizes the Structure Theorem for extended affine root systems [AABGP, II, Th. 2.37] as well as [MY, Prop. 4.2] and — modulo the limitations mentioned above — the description of root systems extended by an abelian group in [Yo4, Th. 3.4]. The isomorphism criterion (3.6.1) is proven in [Yo6] for the case of LEARSs (3.11(c)). The extension data arising in the theory of extended affine root systems are studied in detail in [AABGP, Ch. II] and [Az1]. These references are only a small portion of what is presently known on extended affine root systems and their Weyl groups. Many of these results likely have generalizations to the setting of affine reflection systems.

4 Graded algebras In the previous sections 2 and 3 we reviewed the “combinatorics” needed to describe extended affine Lie algebras and their generalizations in §6. In this and the following section we will introduce the necessary algebraic concepts. All of them have to do with algebras graded by an arbitrary abelian group, usually denoted Λ and written additively. (The reader will notice that at least in the first few subsections Λ need not be abelian.) 4.1 Algebras per se. Throughout, we will consider algebras A over a unital commutative associative ring of scalars k. Unless explicitly stated otherwise, we do not assume that A belongs to some special variety of algebras, like associative algebras or Lie algebras. Therefore, an algebra or a k-algebra if we want to be more precise, is simply a k-module A together with a k-bilinear map A × A → A, (a, b) → ab, the product of A. As is customary, if A happens to be a Lie algebra, its product will be denoted [a, b] or sometimes just [a b] instead of ab. Let A be an algebra. The k-module of all its k-linear derivations is always a Lie algebra, denoted Derk A. A symmetric bilinear form (·|·) : A × A → k is called invariant if (ab|c) = (a|bc) holds for all a, b, c ∈ A; it is nondegenerate if (a|A) = 0 implies a = 0. We put AA = spank {ab : a, b ∈ A}. The algebra A is perfect if A = AA, and is simple if AA = 0 and if {0} and A are the only ideals of A. Any unital algebra is perfect, and any simple algebra is perfect since AA is an ideal of A. 4.2 The centroid of an algebra. The centroid of an algebra A is the subalgebra Centk (A) of the endomorphism algebra Endk (A) consisting of the


77

k-linear endomorphisms of A, which commute with the left and right multiplications of A. One always has k IdA ⊂ Centk (A), but in general this is a proper inclusion. If the canonical map k → Centk (A) is an isomorphism, A is called central , and one says that A is central-simple if A is central and simple, see §4.4 for a characterization of central-simple algebras. The centroid is always a unital associative algebra, but not necessarily commutative in general. For example, the centroid of the null algebra (all products are 0) is the full endomorphism algebra. On the other hand, the centroid of a perfect algebra is always commutative. For example, let A be a unital associative algebra and denote by Z(A) = {z ∈ A : za = az for all a ∈ A} its centre. Then Z(A) → Centk (A),

z → Lz

(A associative)

(4.2.1)

is an isomorphism of algebras, where Lz denotes the left multiplication by z. However, even if the names are quite similar, one should not confuse the centroid with the centre of A! A nonzero Lie algebra L may well be centreless in the sense that Z(L) = {0}, but always has 0 = k IdL ∈ Cent(L), see 4.12(a) for an example. Since Centk (A) is a subalgebra of Endk (A) one can consider A as a left module over Centk (A). This change of perspective is particularly useful if Centk (A) is commutative since then A is an algebra over Centk (A). In general, an algebra A is called fgc (for finitely generated over its centroid) if A is a finitely generated Centk (A)-module. Fgc algebras are more tractable than general algebras, and we will characterize various classes of fgc algebras throughout this paper, see 4.6, 4.7, 4.13 and 5.8. 4.3 Graded algebras. Recall that Λ is an abelian group. A Λ-graded algebra is an algebra A together with a family (Aλ : λ ∈ Λ) of submodules Aλ of A satisfying A = λ∈Λ Aλ and Aλ Aµ ⊂ Aλ+µ for all λ, µ ∈ Λ. We will usually indicate this by saying “Let A = λ Aλ be a Λ-graded algebra ...” or simply “Let A be graded algebra...”. We will encounter many example of graded algebras. But an immediate and important example is k[Λ], the group algebra of Λ over k. By definition, k[Λ] is a free k-module with a k-basis in bijection with Λ, say by λ → z λ , and product determined by z λ z µ = z λ+µ for λ, µ ∈ Λ. It is graded by (k[Λ])λ = k z λ and is a unital associative algebra. We will need some more terminology for a graded algebra A. The submodules Aλ of A are referred to as homogeneous spaces. We will occasionally use subscripts to describe them, in particular whenever we consider an algebra with two gradings, say A = λ∈Λ Aλ and A = q∈Q Aq . The two gradings are called compatible if, putting Aλq := Aλ ∩Aq , we have Aλ = q∈Q Aλq for all λ ∈ Λ. We will indicate compatible gradings by saying that A is (Q, Λ)-graded. The support of a Λ-graded algebra A is the set suppΛ A = {λ ∈ Λ : Aλ = 0}. In general suppΛ A is a proper subset of Λ. The subgroup of Λ generated

78

Erhard Neher

by suppΛ A will be denoted suppΛ A. One says that A has full support if

suppΛ A = Λ. Since the grading of A only depends on suppΛ A, it is of course always possible, and sometimes even useful, to assume that a graded algebra has full support. Two graded algebras A and A with grading groups Λ and Λ , respectively, are called isograded-isomorphic if there exists an isomorphism f : A → A of the underlying k-algebras and a group isomorphism ϕ : suppΛ A →

suppΛ A satisfying f (Aλ ) = Aϕ(λ) for all λ ∈ suppΛ A. Note that ϕ is uniquely determined by f . A more restrictive concept is that of a gradedisomorphism between two Λ-graded algebras A and A , which by definition is an isomorphism f : A → A of the ungraded algebras satisfying f (Aλ ) = Aλ for all λ ∈ Λ. A Λ-graded algebra is called graded-simple if AA = 0 and {0} and A are the only Λ-graded ideals of A. For example, k[Λ] is graded-simple if and only if k is a field. But of course in this case k[Λ] may have many nontrivial ungraded ideals. A bilinear form (·|·) : A × A → k on A is Λ-graded if (Aλ |Aµ ) = 0 for λ + µ = 0. For example, the Killing form of a finite-dimensional Λ-graded Lie algebra is Λ-graded. The radical of an invariant Λ-graded symmetric bilinear form (·|·) is a graded ideal of A, namely Rad(·|·) = {a ∈ A : (a|A) = 0}. A Λ-graded bilinear form (·|·) is nondegenerate, i.e. Rad(·|·) = {0}, if and only if for all λ ∈ Λ the restriction of (·|·) to Aλ × A−λ is a nondegenerate pairing. We denote by GIF(A, Λ) the k-module of all Λ-graded invariant symmetric bilinear forms on A. If L is a perfect Λ-graded Lie algebra, then Rad(·|·) ⊃ Z(L) = {z ∈ L : [z, L] = 0}, the centre of L. Since Z(L) is also Λ-graded, the quotient algebra inherits a canonical Λ-grading, and every (·|·) ∈ GIF(L, Λ) induces an invariant symmetric bilinear form on L/ Z(L) by (¯ x|¯ y ) = (x|y), where x → x ¯ denotes the canonical map. This gives rise to an isomorphism (Lie algebras). (4.3.1) GIF(L, Λ) ∼ = GIF(L Z(L), Λ) If A is a unital algebra, say with identity element 1, then necessarily 1 ∈ A0 . If in addition A is associative, then GIF(A, Λ) ∼ (associative algebras) (4.3.2) = (A0 [A, A]0 )∗ by assigning to any linear form ϕ of A0 [A, A]0 the bilinear form (a | b)ϕ = ϕ((ab)0 + [A, A]0 ). Here [A, A] = spank {ab − ba : a, b ∈ A}, and (ab)0 and [A, A]0 denote the 0-component of ab and [A, A]. For example, the group algebra k[Λ] has up to scalars just one invariant symmetric bilinear form, which assigns to (a, b) the z 0 -coefficient of ab with respect to the canonical k-basis. An endomorphism f of the underlying k-module of A is said to have degree λ if f (Aµ ) ⊂ Aλ+µ for all µ ∈ Λ. The submodules (Endk A)λ , consisting of


79

all endomorphisms of degree λ, are the homogeneous spaces of the Λ-graded subalgebra grEndk (A) = λ∈Λ (Endk A)λ of the associative algebra Endk (A). We put grDerk (A) = grEndk (A) ∩ Derk (A) =

grCentk (A) = grEndk (A) ∩ Centk (A) =

λ∈Λ (Derk

λ∈Λ

A)λ ,

and

(Centk A)λ .

Then (Derk A)λ and (Centk A)λ are, respectively, the derivations and centroidal transformations of A of degree λ. In general, grDerk (A) or grCentk (A) are proper subalgebras of Derk (A) and Centk (A). However, grDerk (A) = Derk (A) if A is a finitely generated algebra ([Fa, Prop. 1]). Similarly, one has that grCentk (A) = Centk (A) if A is finitely generated as ideal. For example, any unital algebra is finitely generated as ideal (namely by 1 ∈ A). We will discuss some examples of (gr) Derk A and (gr) Centk A later, see 4.12, 4.13, 7.9 and 7.12. We call A graded-central if the canonical map k → (Centk A)0 is an isomorphism, and graded-central-simple if A is graded-central and graded-simple. It is well known that an (ungraded) algebra A over a field k is central-simple if and only if A ⊗k K is simple for every extension field K of k. The following theorem extends this to the case of graded algebras. 4.4 Theorem. Let A be a graded-simple k-algebra, where k is a field. Then the following are equivalent: (i) A is graded-central-simple, (ii) for any extension field K of k the base field extension A⊗k K is a gradedcentral-simple K-algebra, (iii) for any extension field K of k the algebra A ⊗k K is a graded-simple K-algebra, (iii) A ⊗k (Centk A)0 is graded-simple. In this theorem, “base field extension” means that we consider A ⊗k K as the K-algebra with product (a1 ⊗ x1 ) (a2 ⊗ x2 ) = (a1 a2 ) ⊗ (x1 x1 ) for ai ∈ A and xi ∈ K and the Λ-grading determined by (A ⊗k K)λ = Aλ ⊗k K. We will describe the centroids of graded-simple algebras in 4.6, after we have introduced special classes of associative graded algebras. 4.5 Associative (pre)division-graded algebras and tori. A Λ-graded unital associative algebra A = λ∈Λ Aλ is called • predivision-graded , if every nonzero homogeneous space contains an invertible element; • division-graded if every nonzero homogenous element is invertible; • an associative torus if A is predivision-graded, k is a field and dimk Aλ ≤ 1.

80

Erhard Neher

For example, any group algebra k[Λ] is predivision-graded. Moreover, k[Λ] is division-graded if and only if k is a field, and in this case k[Λ] is a torus. Group algebras cover only a small part of all possible predivision-graded algebras. If A is such an algebra, it is easily seen that the support of A is a subgroup of the grading group Λ. To simplify the notation, we assume suppΛ A = Λ in this subsection. Let (uλ : λ ∈ Λ) be a family of invertible elements uλ ∈ Aλ and put B = A0 . Then (uλ ) is a B-basis of the left B-module A, and the product of A is uniquely determined by the two rules uλ uµ = τ (λ, µ) uλ+µ

and uλ b = bσ(λ) uλ

(b ∈ B)

(4.5.1)

where τ : Λ × Λ → B × and σ : Λ → Autk (B) are functions. Associativity of A leads to the two identities τ (λ, µ)σ(ν) τ (ν, λ + µ) = τ (ν, λ) τ (νλ, µ), and bσ(λ) )σ(ν) τ (ν, λ) = τ (ν, λ) bσ(ν+λ)

(4.5.2)

for ν, λ, µ ∈ Λ and b ∈ B. Conversely, given any unital associative k-algebra B and functions τ , σ as above satisfying (4.5.2), one can define a Λ-graded kalgebra by (4.5.1). It turns out to be an associative predivision-graded algebra. Algebras arising in this way are called crossed product algebras and denoted (B, (uλ ), τ, σ). To summarize, any associative predivision-graded algebra A with support Λ is graded-isomorphic to a crossed product algebra (B, (uλ ), τ, σ). A is division-graded iff B is a division algebra, and A is a torus iff B = k is a field.

(4.5.3)

The description of predivision-graded algebras as crossed product algebras may not be very illuminating, except that it (hopefully) demonstrates how general this class is. The reader may be more comfortable with the subclass of twisted group algebras defined by the condition that σ is trivial, i.e., σ(λ) = IdB for all λ ∈ Λ, and denoted B t [Λ]. For example, any commutative crossed product algebra is a twisted group algebra. Or, up to graded isomorphism, a Λ-torus with full support is the same as a twisted group algebra k t [Λ]. In the torus case, (4.5.2) says that τ is a 2-cocycle of the group Λ with values in k × . We will describe the special case of associative Zn -tori in 4.13. Twisted group algebras arise naturally as centroids of graded-simple algebras. 4.6 Proposition. Let A = λ∈Λ Aλ be a graded-simple k-algebra. (a) Then Centk (A) is a commutative associative division-graded k-algebra,


Centk (A) = grCentk (A) =

γ∈Γ

(Centk A)γ

where

81

(4.6.1)

Γ = suppΛ Centk A is a subgroup of Λ, called the central grading group of A. Hence Centk (A) is a twisted group algebra over the field K = (Centk A)0 . (b) A is a graded-central-simple K-algebra. (c) A is a free Centk (A)-module with Γ + suppΛ A ⊂ suppΛ A. Moreover, A has finite rank as Centk (A)-module, i.e., A is fgc, if and only if [suppΛ A : Γ ] < ∞

and

dimK (Aλ ) < ∞ for all λ ∈ Λ.

(4.6.2)

The interest in algebras which are graded-central-simple and also fgc comes from the following important realization theorem. 4.7 Theorem. ([ABFP1, Cor. 8.3.5]) Let k be an algebraically closed field of characteristic 0 and let Λ be free abelian of finite rank. Then the following are equivalent for a Λ-graded k-algebra B: (i) B is graded-central-simple, fgc and [Λ : Γ ] < ∞ where Γ is the central grading group of B; (ii) B is isograded-isomorphic to a multiloop algebra Mm (A, σ) for some finite-dimensional simple k-algebra A, described below. The multiloop algebras referred to in (ii) provide an interesting class of examples of graded algebras. They are defined as follows (A can be arbitrary, but we assume that k is a field with enough roots of unity). We let m = (m1 , . . . , mn ) ∈ Nn+ be an n-tuple of positive integers and let π : Zn → Z/(m1 ) ⊕ · · · ⊕ Z/(mn ) =: Ξ be the canonical map. Furthermore, σ = (σ1 , . . . , σn ) is a family of n pairwise commuting automorphisms of the (ungraded) algebra A such that σimi = Id for 1 ≤ i ≤ n. For λ = (λ1 , . . . , λn ) ∈ Zn we define λi Aπ(λ) = {a ∈ A : σi (a) = ζm a for 1 ≤ i ≤ n}, i

where for each l ∈ {m1 , . . . , mn } we have chosen a primitive lth -root of unity ζl ∈ k. Then A = ξ∈Ξ Aξ is a Ξ-grading of A. Denoting by k[z1±1 , . . . , zn±1 ] the Laurent polynomial algebra over k in the variables z1 , . . . , zn , Mm (A, σ) = (λ1 ,...,λn )∈Zn Aπ(λ1 ,...,λn ) ⊗ k z1λ1 · · · znλn is a Zn -graded algebra, called the multiloop algebra of σ based on A and relative to m. Thus, the theorem in particular says that Mm (A, σ) is graded-centralsimple. In fact, the centroid can be determined precisely.

82

Erhard Neher

4.8 Proposition. ([ABP3, Cor. 6.6]) The centroid of a multiloop algebra Mm (A, σ) based on a finite-dimensional central-simple algebra A is gradedisomorphic to the Laurent polynomial ring k[z1±m1 , . . . , zn±mn ]. 4.9 Centroidal derivations. There is a close connection between the centroid and the derivations of an algebra A: [d, χ] ∈ Centk (A) for d ∈ Derk (A) and χ ∈ Centk (A). In other words, Centk (A) is a submodule of the Derk (A)module Endk (A). In the graded case one can define special derivations using the centroid. Let A = λ∈Λ Aλ be a Λ-graded algebra. We abbreviate C = Centk (A), λ grC = grCentk (A) and Cλ = (Centk A)λ , thus grC = λ C . We denote by Hom(Λ, C) ∼ = HomC (Λ ⊗Z C, C) the (left) C-module of all abelian group homomorphisms θ : Λ → (C, +). Every θ ∈ Hom(Λ, C) gives rise to a so-called centroidal derivation, defined by ∂θ (aλ ) = θ(λ) (aλ )

for aλ ∈ Aλ .

(4.9.1)

We put CDer(A, Λ) = {∂θ : θ ∈ Hom(Λ, C)}, sometimes also denoted CDer(A) if Λ is clear from the context. We emphasize that CDer(A, Λ) not only depends on C but also on Λ. For example, any algebra A can be graded by Λ = {0} in which case CDer(A, Λ) = {0}. We will also consider the Λ-graded submodule grCDer(A, Λ) of CDer A, grCDer(A, Λ) = λ∈Λ (CDer A)λ where (CDer A)λ = CDer(A) ∩ (Endk A)λ = { ∂θ : θ ∈ Hom(Λ, Cλ )}. Suppose now that Centk (A) is commutative, e.g., that A is perfect. Then any ∂θ is C0 -linear and CDer(A) is a C0 -subalgebra of the Lie algebra DerC0 (A) of all C0 -linear derivations of A. Indeed, [∂θ , ∂ψ ] = ∂θ∗ψ−ψ∗θ for θ ∗ ψ ∈ Hom(Λ, C) defined by (θ ∗ ψ) (λ) = µ∈Λ θ(µ) ψ(λ)µ , (4.9.2) where ψ(λ)µ denotes the µ-component of ψ(λ) ∈ C. The sum in (4.9.2) converges in the finite topology. In particular, for θ ∈ Hom(Λ, Cλ ) and ψ ∈ Hom(Λ, Cµ ) one has [∂θ , ∂ψ ] = θ(µ)∂ψ − ψ(λ)∂θ ∈ (CDer A)λ+µ ;

(4.9.3)

in particular (CDer A)0 is abelian. Suppose A is perfect and (·|·) is a Λ-graded invariant nondegenerate symmetric bilinear form on A. We denote by grSCDer(A, Λ) = λ∈Λ (SCDer A)λ


83

the graded skew centroidal derivations of A, i.e., the graded subalgebra of grCDer(A, Λ) consisting of those graded centroidal derivations of A which are skew-symmetric with respect to (·|·). Since A is perfect, any centroidal transformation is symmetric with respect to (·|·). This implies that (SCDer A)λ = {∂θ ∈ (CDer A)λ : θ(λ) = 0}, in particular (SCDer A)0 = (CDer A)0 . Because of (4.9.3) we have [(SCDer A)λ , (SCDer A)−λ ] = 0 for any λ ∈ Λ. Hence grSCDer(A, Λ) is the semidirect product of the ideal spanned by the nonzero homogeneous spaces and the abelian subalgebra (CDer A)0 . 4.10 Degree derivations. Again let A = λ Aλ be a Λ-graded k-algebra. Recall that k IdA ⊂ (Centk A)0 . We can therefore consider the submodule D = D(A, Λ) = {∂θ : θ ∈ Hom(Λ, k IdA )} ⊂ (CDer A)0 of so-called degree derivations of (A, Λ). For example, if Λ = Zn we have D = spank {∂i : 1 ≤ i ≤ n} where ∂i (aλ ) = λi aλ for λ = (λ1 , . . . , λn ) ∈ Zn . Observe that [∂θ , ∂ψ ] = θ(µ) ∂ψ for ∂θ ∈ D and ψ ∈ Hom(Λ, Centk (A)µ ). So the action of D on grCDer A is diagonalizable. In order to actually get a toral action, we will assume in this subsection and in 4.11 that k is a field and that A has full support. In this case, we can identify Hom(Λ, k IdA ) ≡ HomZ (Λ, k). Moreover, the canonical map Hom(Λ, k) → D = {∂θ : θ ∈ Hom(Λ, k)} is an isomorphism. We therefore have an evaluation map ev : Λ → D∗ , defined by evλ (∂θ ) = θ(λ). Observe that (4.9.1) says that Aλ ⊂ {a ∈ A : d(a) = evλ (d) a for all d ∈ D}.

(4.10.1)

The following lemma specifies a condition under which (4.10.1) becomes an equality. 4.11 Lemma. As in 4.10 we suppose that k is a field and A is a Λ-graded k-algebra with full support. We let T ⊂ D be a subspace satisfying the condition that the restricted evaluation map Λ → T ∗ , λ → evλ |T , is injective.

(4.11.1)

Then Aλ = {a ∈ A : d(a) = evλ (t) for all t ∈ T }, in particular equality holds in (4.10.1). Moreover, if Centk (A) is commutative, T is a toral subalgebra of

84

Erhard Neher

grCDerk (A), and if A is perfect and has a Λ-graded invariant nondegenerate symmetric bilinear form, then T is also a toral subalgebra of grSCDerk (A). We mention that (4.11.1) always holds for T = D, Λ torsion-free and k a field of characteristic 0. 4.12 Examples. (a) Let g be a finite-dimensional split simple Lie algebra over a field K of characteristic 0, e.g., a finite-dimensional simple Lie algebra over an algebraically closed K. Hence g has a root space decomposition g = α∈R gα with respect to some splitting Cartan subalgebra h = g0 of g, where R ⊂ h∗ is the root system of (g, h). We consider this decomposition as a Q(R)-grading. It is well known that g is central (hence central-simple): Any centroidal transformation χ leaves the root space decomposition of g invariant and, by simplicity, is uniquely determined by χ|gα for some 0 = α ∈ R. For any θ ∈ Hom(Q(R), K) ∼ = HomK (h∗ , K) there exists a unique hθ ∈ h such that α(hθ ) = θ(α). Therefore ∂θ = ad hθ and CDer(g, Q(R)) = grCDer(g, Q(R)) = grSCDer(g, Q(R)) = ad h. (b) Let g be as in (a) and let C be a unital commutative associative Λ-graded K-algebra. We view L = g ⊗ C as a Λ-graded Lie algebra over K with product [x ⊗ c, x ⊗ c ] = [x, x ] ⊗ cc for x, x ∈ g and c, c ∈ C and with homogeneous spaces Lλ = g ⊗ C λ (here and in the following all tensor products are over K). The centroid of L is determined in [ABP3, Lemma 2.3] or [BN, Cor. 2.23]: CentK (L) = K Idg ⊗C. (4.12.1) (We note in passing that (4.12.2) also holds for the infinite-dimensional versions of g defined in 2.13, as well as for the version of g over rings which we will introduce in 5.5.) By [Bl2, Th. 7.1] or [BM, Th. 1.1], the derivation algebra of L is

DerK (L) = Der(g)⊗C ⊕ K Idg ⊗ DerK (C) = IDer(L)⊕ K Idg ⊗ DerK (C) (4.12.2) where IDer L is the ideal of inner derivations of L. It is then immediate that CDer(L) and grCDer(L) are given by the formulas (gr)CDer(L, Λ) = K Idg ⊗(gr)CDer(C, Λ), where here and in the following (gr) indicates that the formula is true for the graded as well as the ungraded case. Let κ be the Killing form of g and suppose that is a Λ-graded invariant nondegenerate symmetric bilinear form on C. The bilinear form κ ⊗ of L, defined by (κ ⊗ )(x ⊗ c, x ⊗ c ) = κ(x, x ) (c, c ), is Λ-graded, invariant, nondegenerate and symmetric (if g is simple, any such form arises in this way


85

by [Be, Th. 4.2]). With respect to κ ⊗ , the skew-symmetric derivations and skew-symmetric centroidal derivations are

(gr)SDerK (L) = IDer L ⊕ K Idg ⊗(gr)SDerK (C) , (gr)SCDerK (L) = K Idg ⊗(gr)SCDerK (C). In particular, let C = K[t±1 ] be the K-algebra of Laurent polynomials over K, which we view as Z-graded algebra with C i = Kti . Its derivation algebra is (isomorphic to) the Witt algebra,

DerK K[t±1 ] = i∈Z Kd(i) , where d(i) acts on K[t±1 ] by d(i) (tj ) = jti+j . Hence d(i) is the centroidal derivation given by the homomorphism θ(i) : Z → C i , θ(i) (j) = jti . It follows that

Der K[t±1 ] = grDer K[t±1 ] = CDer K[t±1 ] = grCDer K[t±1 ] . Up to scalars, the algebra K[t±1 ] has a unique Z-graded invariant nondegenerate symmetric given by (ti , tj ) = δi+j,0 . With respect to this

bilinear form ±1 (0) form SCDer K[t ] = Kd and hence the skew derivations of the untwisted loop algebra L = g ⊗ K[t±1 ] are SCDer(g ⊗ K[t±1 ]) = Kd(0). It is important for later that SCDer(L) are precisely the derivations needed to construct the untwisted affine Lie algebra in terms of L. (c) Let E be an affine Lie algebra (see 2.13). Then (b) applies to L = [E, E] Z([E, E]) ∼ = g ⊗ K[t±1 ] (the centreless core when E is considered as an extended affine Lie algebra, see 6.11). In particular, L has an infinite-dimensional centroid. But the centroid of the core K := [E, E] is 1-dimensional: Cent(K) = K IdK , while the centroid of E is 2-dimensional, namely isomorphic to K IdE ⊕ HomK (Kc, Kd) where Kc is the centre of the derived algebra K and E = K Kd ([BN, Cor. 3.5]). 4.13 Quantum tori. In short, a quantum torus is an associative Zn -torus. As in 4.5 we may assume that a quantum torus has full support. In this case it can be described as follows. Let k = F be a field and let q = (qij ) ∈ Matn (F ) be a quantum matrix , i.e., qij qji = 1 = qii for all 1 ≤ i, j ≤ n. The quantum torus associated to q is the associative F -algebra Fq presented by generators ti , ti −1 and relations ti ti −1 = 1 = ti −1 ti

and ti tj = qij tj ti

for 1 ≤ i, j ≤ n.

±1 Observe that if all qij = 1, then Fq = F [t±1 1 , . . . , tn ] is the Laurent polynomial ring in n variables. Hence, for a general q, a quantum torus is a

86

Erhard Neher

non commutative version (a quantization) of the Laurent polynomial algebra ±1 F [t±1 1 , . . . , tn ], which is the coordinate ring of an n-dimensional algebraic torus. This explains the name “quantum torus” for Fq . Let A = Fq be a quantum torus. It is immediate that Fq =

λ∈Zn

F tλ ,

for tλ = tλ1 1 · · · tλnn .

The multiplication of Fq is determined by the 2-cocycle τ : Zn × Zn → F × where λ µ tλ tµ = τ (λ, µ) tλ+µ , τ (λ, µ) = 1≤j k. The empty partition is denoted by 0. For any partition λ we denote by 2λ the partition determined by (2λ)i = 2λi for all i, and by λ the partition determined by (λ )i = #{j | λj ≥ i} for all i. The Young diagram for λ is obtained by transposing the Young diagram for λ. Irreducible representations of the symmetric group Sd are parameterized by partitions λ satisfying |λ| = d, and can be realized inside the regular representation k[Sd ]. We denote def

Hλ = k[Sd ] cλ , where cλ ∈ k[Sd ] is the Young projector corresponding to the standard Young tableau of shape λ (see e.g., [FH] for details). The left regular action makes Hλ an irreducible Sd -module, and any irreducible representation of Sd can be obtained in this way. For any vector space V , the symmetric group Sd acts in V ⊗d by permuting the tensor factors. For any partition λ with |λ| = d we denote ⊗d ⊗d Sλ V = im cλ : V . →V The correspondence V Sλ V is called the Schur functor corresponding to λ. The Littlewood–Richardson coefficients Nνλ,µ are nonnegative integers, ν determined for any partitions λ, µ, ν by the relation Sλ Sµ = ν Nλ,µ Sν , where Sλ denotes the Schur symmetric polynomial corresponding to the partition λ. In particular, Nνλ,µ = 0 unless |ν| = |λ| + |µ|.

130

Ivan Penkov and Konstantin Styrkas

2 Tensor representations of gl∞ and sl∞ Let V, V∗ be countable dimensional vector spaces, and let ·, · : V ⊗ V∗ → k be a nondegenerate pairing. The Lie algebra gl∞ is defined as the space V⊗V∗ , equipped with the Lie bracket [u ⊗ u∗, v ⊗ v ∗ ] = u∗ , v u ⊗ v ∗ − v ∗ , u v ⊗ u∗,

u, v ∈ V, u∗ , v ∗ ∈ V∗ . (1)

The kernel of the map ·, · is a Lie subalgebra of gl∞ , which we denote sl∞ . As observed by G. Mackey [M], there always exist dual bases {ξi }i∈I of V and {ξi∗ }i∈I of V∗ , indexed by a countable set I, so that we have ξj∗ , ξi = δi,j for i, j ∈ I. This gives another, more straightforward coordinate definition of gl∞ as the Lie algebra with a linear basis {Ei,j = ξi ⊗ ξj∗ }i,j∈I satisfying the usual commutation relations [Ei,j , Ek,l ] = δj,k Ei,l − δi,l Ek,j . We call V the natural representation of gl∞ and sl∞ , and V∗ its restricted dual. For any nonnegative integers p, q we define the tensor representation V⊗(p,q) as the vector space V⊗p ⊗ V∗⊗q , equipped with the following gl∞ module structure: (u ⊗ u∗ ) · (v1 ⊗ . . . ⊗ vp ⊗ v1∗ ⊗ . . . ⊗ vq∗ ) =

p

u∗ , vi v1 ⊗ . . . ⊗ vi−1 ⊗ u ⊗ vi+1 ⊗ . . . ⊗ vp ⊗ v1∗ ⊗ . . . ⊗ vq∗

i=1

−

q

∗ vj∗ , u v1 ⊗ . . . ⊗ vp ⊗ v1∗ ⊗ . . . ⊗ vj−1 ⊗ u∗ ⊗ vj+1 ⊗ . . . ⊗ vq∗ ,

j=1

for u, v1 , . . . , vp ∈ V and u∗ , v1∗ , . . . , vq∗ ∈ V∗ . The product of symmetric groups Sp × Sq acts in V⊗(p,q) by permuting the factors, and this action commutes with the action of gl∞ . We express this by saying that V⊗(p,q) is a (gl∞ , Sp × Sq )-module, and use similar notation throughout the paper. Our main goal in this section is to reveal the structure of the tensor representations as a gl∞ -module, and in particular to identify the Jordan–H¨ older constituents of V⊗(p,q) . We describe these modules explicitly as appropriately defined highest Consider the direct sum decomposition weight gl∞ -modules. gl gl∞ = hgl ⊕ α∈∆gl k Xα , where hgl =

i∈I

k Ei,i ,

∆gl

= εi − εj i, j ∈ I, i = j ,

Xεgli −εj = Ei,j ,

and εi denotes the functional on hgl determined by εi (Ej,j ) = δi,j . We have [H, Xαgl ] = α(H)Xαgl for any α ∈ ∆gl and H ∈ hgl . As usual, we refer to elements of ∆gl as roots, and to functionals on hgl as weights. Remark 1. For an algebraically closed field k the general notion of a Cartan subalgebra of gl∞ has been defined and studied in [NP, DPS]. The subalgebra

Tensor Representations of Classical Locally Finite Lie Algebras

131

hgl is an example of a splitting Cartan subalgebra, and all splitting Cartan subalgebras of gl∞ are conjugated. From now on we identify the index set I with Z \ {0}, and consider the + polarization of the root system ∆gl = ∆+ −∆+ gl gl , where the set ∆gl of positive roots is given by ∆+ {εi − εj | i < j < 0} {εi − εj | j < 0 < i}. gl = {εi − εj | 0 < i < j} gα and define bgl as the Lie subalgebra of gl∞ , We denote ngl = α∈∆+ gl generated by ngl and hgl . It is clear that ngl is a Lie subalgebra of gl∞ , that + h , and that [b , b ] = n . bgl = ngl ⊂ gl gl gl gl Let V be a gl∞ -module, and let v ∈ V . We say that v is a highest weight vector if it generates a one-dimensional bgl -module. Any such v must satisfy ngl v = 0,

H v = χ(H) v

∀H ∈ hgl ,

for some χ ∈ h∗gl . We say that V is a highest weight module if V is generated by a highest weight vector v as above; the functional χ is then called the highest weight of V . As in the finite-dimensional case, it is easy to prove that for each χ ∈ h∗gl there exists a unique irreducible highest weight gl∞ -module with highest weight χ. Remark 2. The subalgebra bgl is an example of a splitting Borel subalgebra of gl∞ , see [DP2] for a classification of the splitting Borel subalgebras in gl∞ over algebraically closed fields. In this paper we only consider the notion of a highest weight module, associated with bgl . This choice has the important property that all Jordan–H¨ older constituents of all tensor representations are highest weight modules. In particular, the natural representation V is a highest weight module with highest weight ε1 , generated by the highest weight vector ξ1 ; similarly, V∗ is a highest weight module with highest weight −ε−1 , generated by the highest ∗ weight vector ξ−1 . We now describe the Weyl construction for gl∞ . For any pair of indices I = (i, j) with i ∈ {1, 2, . . . , p} and j ∈ {1, 2, . . . , q}, define the contraction ΦI : V⊗(p,q) → V⊗(p−1,q−1) ,

v1 ⊗ . . . ⊗ vp ⊗ v1∗ ⊗ . . . ⊗ vq∗ →

→ vj∗ , vi v1 ⊗ . . . ⊗ vî ⊗ . . . ⊗ vp ⊗ v1∗ ⊗ . . . ⊗ vˆj∗ ⊗ . . . ⊗ vq∗ , and consider the (gl∞ , Sp × Sq )-submodule V{p,q} of V⊗(p,q) , def ker ΦI : V⊗(p,q) → V⊗(p−1,q−1) . V{p,q} = I

132


Set also V{p,0} = V⊗p and V{0,q} = V∗⊗q . For any partitions λ, µ such that |λ| = p and |µ| = q, define the gl∞ -submodule of V⊗(p,q) def

def

Γλ;µ = V{p,q} ∩ (Sλ V ⊗ Sµ V∗ ). def

Theorem 2.1 For any p, q there is an isomorphism of (gl∞ , Sp × Sq )modules V{p,q} ∼ Γλ;µ ⊗ (Hλ ⊗ Hµ ). (2) = |λ|=p |µ|=q

For any partitions λ, µ, the gl∞ -module Γλ;µ is an irreducible highest weight def module with highest weight χ = i∈N λi εi − i∈N µi ε−i . Furthermore, Γλ;µ is irreducible when regarded by restriction as an sl∞ -module. Proof. Pick an enumeration I = {k1 , k2 , k3 , . . . } of the index set I = Z \ {0}, such that ki = i for 1 ≤ i ≤ p, and kp+i = −i for 1 ≤ i ≤ q. For each n denote by Vn the subspace of V spanned by ξk1 , . . . , ξkn , and by Vn∗ the subspace of V∗ spanned by ξk∗1 , . . . , ξk∗n ; the pairing between V and V∗ restricts to a nondegenerate pairing between Vn and Vn∗ . Denote by gn the Lie subalgebra of gl∞ generated by {Eki ,kj }1≤i,j≤n , def def and set bn = bgl ∩ gn , hn = hgl ∩ gn . It is clear that gn ∼ = gln , and that hn (respectively, bn ) is a Cartan (respectively, Borel) subalgebra of gn . Moreover, the inclusion gn → gn+1 restricts to inclusions hn → hn+1 and bn → bn+1 . The tensor representations (Vn )⊗(p,q) and the contractions (n) ΦI : (Vn )⊗(p,q) → (Vn )⊗(p−1,q−1) are defined by analogy with their infinitedimensional counterparts. We set def (n) (n) def (Vn ){p,q} = ker ΦI , Γλ;µ = (Vn ){p,q} ∩ (Sλ Vn ⊗ Sµ Vn∗ ). I

A version of the finite-dimensional Weyl construction (see Appendix for (n) more details) implies that for n ≥ p + q the gln -module Γλ;µ is an irreducible highest weight module with highest weight χ, regarded by restriction as a functional on hn . Futhermore, we have (gn , Sp × Sq )-module isomorphisms (n) Γλ;µ ⊗ (Hλ ⊗ Hµ ). (3) (Vn ){p,q} ∼ = |λ|=p |µ|=q

It is clear from the definitions that we have the following diagram of inclusions: (Vp+q ){p,q} ⊂ (Vp+q+1 ){p,q} ⊂ . . . ⊂ (Vn ){p,q} ⊂ . . . ⊂ ∪ ∪ ∪ (p+q) (p+q+1) . . . ⊂ Γ ⊂ ⊂ Γ(n) ⊂ . . . ⊂ Γ λ;µ

λ;µ

λ;µ

{p,q} n∈N (Vn )

∪

(n) n∈N Γλ;µ

= V{p,q} ∪ = Γλ;µ . (4)


133

(n)

The irreducibility of the gln -module Γλ;µ for each n implies the irreducibility of the gl∞ -module Γλ;µ . Similarly, the isomorphism (2) follows from the isomorphisms (3), and it remains to show that Γλ;µ is a highest weight gl∞ module with highest weight χ. (n) For each n ≥ p + q the highest weight subspace Γλ;µ [χ] is one-dimensional, and therefore (n) (p+q) (p+q+1) (n) Γλ;µ [χ] = Γλ;µ [χ] = · · · = Γλ;µ [χ] = · · · = Γλ;µ [χ] = Γλ;µ [χ]. n∈N

We also have bgl · Γλ;µ [χ] =

n∈N

(n) (n) bn · Γλ;µ [χ] = n∈N Γλ;µ [χ] = Γλ;µ [χ],

which means that any nonzero v ∈ Γλ;µ [χ] is a highest weight vector generating the gl∞ -module Γλ;µ [χ]. This completes the proof of the theorem. Applying Theorem 2.1 to the special cases p = 0 and q = 0, we obtain a version of Schur–Weyl duality for gl∞ : Γλ;0 ⊗ Hλ , V∗⊗q ∼ Γ0;µ ⊗ Hµ . (5) V⊗p ∼ = = |λ|=p

|µ|=q

Next, we study the structure of the gl∞ -module V⊗(p,q) . The main feature and the crucial difference from the finite-dimensional case is that V⊗(p,q) fails to be completely reducible. Theorem 2.2 Let p, q be nonnegative integers, and let = min(p, q). Then the Loewy length of the gl∞ -module V⊗(p,q) equals + 1, and (r) ⊗(p,q) ⊗(p,q) ⊗(p−r,q−r) soc V , r = 1, . . . , . = ker ΦI1 ,...,Ir : V →V I1 ,...,Ir

(6) Moreover, the socle filtration of V⊗(p,q) , regarded as an sl∞ -module, coincides with (6). Proof. Denote by F(r) the subspaces of V⊗(p,q) on the right side of (6), and for each n set (r) ⊗(p,q) ⊗(p−r,q−r) , r = 1, . . . , , Fn = ker Φ{I1 ,...,Ir } : (Vn ) → (Vn ) I1 ,...,Ir

where {Vn }n∈N is the exhaustion of V used in the proof of Theorem 2.1. Then we have the commutative diagram of (gl∞ , Sp × Sq )-modules in which the vertical arrows are obtained as restrictions of the inclusions in the rightmost column, and yield exhaustions of F(r) . Denote for convenience (0) (+1) F(0) = 0, F(+1) = V⊗(p,q) , and similarly Fn = 0, Fn = (Vn )⊗(p,q) for all (r+1) (r) (r+1) (r) n. It is easy to check that the induced maps Fn /Fn → Fn+1 /Fn+1 are

134


injective for all n, r, and therefore for each r = 0, 1, . . . , the layer F(r+1) /F(r) (r+1) (r) is the union of quotients Fn /Fn .

(1) Fn

⊂

(2) Fn

⊂

...

⊂

() Fn

⊂

(Vn )⊗(p,q)

(1) Fn+1

⊂

(2) Fn+1

⊂

...

⊂

() Fn+1

⊂

(Vn+1 )⊗(p,q)

F(1)

⊂

F(2)

⊂

...

⊂

F()

⊂

V⊗(p,q) ,

It is a standard exercise (see Appendix for more details) to show that for each r there exists a (gln , Sp × Sq )-module isomorphism (n) Γλ;µ ⊗ H(λ, µ; r), (7) Fn(r+1) /Fn(r) ∼ = |λ|=p−r |µ|=q−r

where H(λ, µ; r) are some Sp × Sq -modules. As in the proof of Theorem 2.1, it follows that F(r+1) /F(r) ∼ Γλ;µ ⊗ H(λ, µ; r). = |λ|=p−r |µ|=q−r

older series In particular, this shows that V⊗(p,q) has a finite Jordan–H¨ with irreducible constituents of the form Γλ;µ for appropriate λ, µ. Moreover, F(r) is characterized as the unique submodule of V⊗(p,q) such that for any λ, µ [V⊗(p,q) : Γλ;µ ] if |λ| > p − r and |µ| > q − r, (r) [F : Γλ;µ ] = 0 otherwise . We use induction on r to prove the main statement of the theorem: F(r) = soc(r) V⊗(p,q) . The base of induction r = 0 is trivial since F(0) = soc(0) V⊗(p,q) = 0. Suppose that F(r) = soc(r) V⊗(p,q) for some r. The quotient F(r+1) /F(r) is a semisimple gl∞ -module, hence F(r+1) ⊂ soc(r+1) V⊗(p,q) . Now let U be a simple submodule of V⊗(p,q) /F(r) ; then U ∼ = Γλ;µ with λ, µ satisfying |λ| = p − s and |µ| = q − s for some s ≥ r. In particular U ⊂ F(s+1) /F(r) . Our goal is to show that in fact s = r; indeed, we would then conclude that soc(r+1) V⊗(p,q) ⊂ F(r+1) , proving the induction step. Fix a vector u ∈ V⊗(p,q) of weight χ = i λi εi − i µi ε−i , such that the image of u under the projection V⊗(p,q) → V⊗(p,q) /F(r) generates U. Fix a


135

large enough m ∈ N such that u ∈ (Vm )⊗(p,q) . We may further assume without loss of generality that u generates a glm -submodule of (Vm )⊗(p,q) isomorphic (m) to Γλ;µ . (s+1)

(n)

→ Γλ;µ ⊗ H(λ, µ; s) the proFor each n > m denote by πn : Fn jection corresponding to the decomposition (7). The map πn can be described explicitly, and we show in Proposition 6.3 in the appendix that (s−1) u − πn (u) ∈ / Fn for infinitely many n. On the other hand, u generates (s+1) (r) (n) (r) a submodule of Fn /Fn isomorphic to Γλ;µ , therefore u − πn (u) ∈ Fn for all n. We conclude that s ≤ r, which implies that in fact s = r. The induction is now complete, and thus F(r) = soc(r) V⊗(p,q) for all r. Finally, we need to verify that the filtration {F(r) } is also the socle filtration of V⊗(p,q) regarded as an sl∞ -module. The irreducibility of Γλ;µ as an sl∞ -module implies that the layers of the filtration F(r) remain semisimple. Therefore F(r) ⊂ soc(r) V⊗(p,q) for all r. The proof of the opposite inclusion, given above for gl∞ , works without any alterations for sl∞ as well, and this completes the proof of the theorem. Next, we describe the indecomposable constituents of the tensor representations of gl∞ . Theorem 2.3 For any partitions λ, µ the gl∞ -module Γλ;0 ⊗ Γ0;µ is indecomposable, and ⎛ ⎞ ⎝ soc(r+1) (Γλ;0 ⊗ Γ0;µ ) ∼ Nλλ ,γ Nµµ ,γ ⎠ Γλ ;µ . (8) = λ ,µ

|γ|=r

The same applies to Γλ;0 ⊗ Γ0;µ regarded as an sl∞ -module. Proof. Schur–Weyl Duality (5) for gl∞ implies that (Γλ;0 ⊗ Γ0;µ ) ⊗ (Hλ ⊗ Hµ ). V⊗(p,q) ∼ = |λ|=p |µ|=q

Hence the gl∞ -module Γλ;0 ⊗ Γ0;µ is realized as the direct summand (Sλ V ⊗ Sµ V∗ ) of the gl∞ -module V⊗(p,q) . Therefore soc(r) Γλ;µ = Γλ;µ ∩ soc(r) V⊗d . It is known that for all partitions λ, µ, λ , µ we have [Γλ;0 ⊗ Γ0;µ : Γλ ;µ ] = λ µ γ Nλ ,γ Nµ ,γ provided n is large enough, see e.g. [HTW]. This yields (n)

[Γλ;0 ⊗ Γ0;µ : Γλ ;µ ] =

γ

Nλλ ,γ Nµµ ,γ ,

(n)

(n)

(9)

136


and combining (9) with the description of the socle filtration of V⊗(p,q) , we obtain (8). In particular, soc(Γλ;0 ⊗ Γ0;µ ) ∼ = Γλ;µ , and the simplicity of the socle implies the indecomposability of the gl∞ -module Γλ;0 ⊗ Γ0;µ . Corollary 2.4 The decomposition of V⊗(p,q) into indecomposable gl∞ -modules is given by the isomorphism V⊗(p,q) ∼ (dim Hλ dim Hµ ) Γλ;0 ⊗ Γ0;µ . = |λ|=p |µ|=q

Examples. We begin with the description of tensors of rank 2. Purely covariant and purely contravariant tensor representations of gl∞ and sl∞ are completely reducible, and are decomposed by the Schur–Weyl duality (2). For tensors of rank 2 these decompositions correspond to the isomorphisms V⊗V∼ = S 2 V ⊕ Λ2 V,

V∗ ⊗ V∗ ∼ = S 2 V∗ ⊕ Λ2 V∗ .

The tensor representation of mixed type V⊗V∗ is the adjoint representation of gl∞ , for which we have the non-splitting short exact sequence of gl∞ -modules 0 → sl∞ → gl∞ → k → 0. In other words, the socle filtration of the adjoint gl∞ -module has length 2, with a simple socle isomorphic to sl∞ , and a simple top isomorphic to k. We summarize the structure of these modules graphically by V⊗(2,0) ∼ Γ(2);(0) ⊕ Γ(1,1);(0) , V(1,1) ∼

Γ(0);(0) , Γ(1);(1)

V⊗(0,2) ∼ Γ(0);(2) ⊕ Γ(0);(1,1) . In these diagrams each tower represents an indecomposable direct summand of a module, and the vertical arrangement of boxes in each tower represents the structure of the layers of the socle filtration, with the bottom box corresponding to the socle. Similarly, Theorem 2.3 yields the following diagrams for tensor representations of rank 3: V⊗(3,0) ∼ Γ(3);(0) ⊕ 2 Γ(2,1);(0) ⊕ Γ(1,1,1);(0) , V⊗(2,1) ∼

Γ(1);(0) Γ(1);(0) ⊕ , Γ(2);(1) Γ(1,1);(1)

V⊗(0,3) ∼ Γ(0);(3) ⊕ 2 Γ(0);(2,1) ⊕ Γ(0);(1,1,1) , and for tensor representations of rank 4:


137

V⊗(4,0) ∼ Γ(4);(0) ⊕ 3 Γ(3,1);(0) ⊕ 2 Γ(2,2);(0) ⊕ 3 Γ(2,1,1);(0) ⊕ Γ(1,1,1,1);(0) , V⊗(3,1) ∼

Γ(2);(0) Γ ⊕ Γ(1,1);(0) Γ(1,1);(0) ⊕ 2 (2);(0) ⊕ , Γ(3);(1) Γ(2,1);(1) Γ(1,1,1);(1)

Γ(0);(0) Γ(0);(0) V⊗(2,2) ∼ Γ(1);(1) ⊕ Γ(1);(1) ⊕ Γ(1);(1) ⊕ Γ(1);(1) , Γ(2);(1,1) Γ(1,1);(2) Γ(2);(2) Γ(1,1);(1,1) V⊗(1,3) ∼

Γ(0);(2) Γ ⊕ Γ(0);(1,1) Γ(0);(1,1) ⊕ 2 (0);(2) ⊕ , Γ(1);(3) Γ(1);(2,1) Γ(1);(1,1,1)

V⊗(0,4) ∼ Γ(0);(4) ⊕ 3 Γ(0);(3,1) ⊕ 2 Γ(0);(2,2) ⊕ 3 Γ(0);(2,1,1) ⊕ Γ(0);(1,1,1,1) .

3 Tensor representations of sp∞ Let V be a countable dimensional vector space, and let Ω : V ⊗ V → k be a non-degenerate anti-symmetric bilinear form on V. We realize the Lie algebra gl∞ as in Section 2 by taking V∗ = V and Ω as the pairing ·, · . The Lie algebra sp∞ is defined as the Lie subalgebra of gl∞ , which preserves the form Ω, i.e.,

sp∞ = X ∈ gl∞ Ω(Xu, v) + Ω(u, Xv) = 0 for all u, v ∈ V . It is always possible to pick a basis {ξi } of V, indexed as before by the set I = Z\{0}, such that Ω(ξi , ξj ) = sign(i) δi+j,0 . In the coordinate realization of gl∞ , the Lie algebra sp∞ has a linear basis {sign(j)Ei,j − sign(i)E−j,−i }i,j∈I . Since the dual basis {ξi∗ }i∈I is given by ξi∗ = sign(i) ξ−i , it follows that sp∞ = S 2 V, and the Lie bracket is induced by (1). We call V, regarded as an sp∞ -module by restriction, the natural representation of sp∞ . It is easy to see that the sp∞ -action on the gl∞ -module V⊗(p,q) coincides with the sp∞ -action on the gl∞ -module V⊗(p+q) . Therefore it suffices to study the tensor representations V⊗d . To define the notion of a highest weight sp∞ -module, we consider the sp direct sum decomposition sp∞ = hsp ⊕ α∈∆sp k Xα , where hsp =

k (Ei,i − E−i,−i ),

i∈I

∆sp = {±(εi + εj ) | i, j ∈ N}

{εi − εj | i, j ∈ N and i = j},

sp = Ei,−j + Ej,−i , X−ε = E−i,j + E−j,i , Xεsp i +εj i −εj

Xεsp = Ei,j − E−j,−i . i −εj

as before using the polarizaThe subalgebra bsp and its ideal nsp

are defined tion of the root system ∆sp = ∆+ −∆+ sp sp , where {εi − εj | i, j ∈ N and i < j}. ∆+ sp = {εi + εj | i, j ∈ N and i ≤ j}

138


An sp∞ -module V is called a highest weight module with highest weight χ ∈ h∗sp , if it is generated by a vector v ∈ V satisfying nsp v = 0 and H v = χ(H) v for all H ∈ hsp . In particular, the natural representation V is a highest weight sp∞ -module with highest weight ε1 , generated by the highest weight vector ξ1 . For any pair I = (i, j) of integers such that 1 ≤ i < j ≤ d, the form Ω determines a contraction ΦI : V⊗d → V⊗(d−2) , v1 ⊗ . . . ⊗ vd → Ω(vi , vj ) v1 ⊗ . . . ⊗ vî ⊗ . . . ⊗ vˆj ⊗ . . . ⊗ vd . def

def

We set V0 = k, V1 = V and def Vd = ker ΦI : V⊗d → V⊗(d−2) I

for d ≥ 2. For any partition λ with |λ| = d we define the sp∞ -submodule Γλ of V⊗d , def Γλ = Vd ∩ Sλ V. Theorem 3.1 For any nonnegative integer d there is an isomorphism of (sp∞ , Sd )-modules Vd ∼ Γλ ⊗ Hλ . (10) = |λ|=d

Γλ is an irreducible highest weight For every partition λ, the sp∞ -module module with highest weight ω = i∈N λi εi . Proof. For each n, denote by Vn the 2n-dimensional subspace of V spanned by ξ±1 , . . . , ξ±n , and identify the Lie subalgebra of End(Vn ) which preserves the restriction of the form Ω to Vn with the Lie algebra sp2n . The tensor rep(n) resentations (Vn )⊗d of sp2n and the contractions ΦI : (Vn )⊗d → (Vn )⊗(d−2) are defined as before. We set def (n) (n) def (Vn )d = ker ΦI , Γλ = (Vn )d ∩ Sλ Vn . I

The Weyl construction for sp2n implies that there are isomorphisms of (sp2n , Sd )-modules (n) (Vn )d ∼ Γλ ⊗ Hλ , (11) = |λ|=d

(n)

and that for n ≥ d the module Γλ is an irreducible highest weight module with highest weight i∈N λi εi . As in the proof of Theorem 2.1, the diagram of inclusions


(Vd )d ⊂ (Vd+1 )d ⊂ . . . ⊂ (Vn )d ⊂ . . . ⊂ ∪ ∪ ∪ (d) (d+1) (n) ⊂ . . . ⊂ Γλ ⊂ ... ⊂ Γλ ⊂ Γλ

d n∈N (Vn )

∪ (n) n∈N Γλ

139

= Vd ∪ = Γλ

(12) establishes both the irreducibility of the sp∞ -module Γλ and the isomorphism (10), and the equalities (n) (d) (d+1) (n) Γλ [χ] = Γλ [χ] = · · · = Γλ [χ] = · · · = Γλ [χ] = Γλ [χ] n∈N

yield the characterization of Γλ [χ] as a highest weight sp∞ -module.

Next, we describe the socle filtration of the tensor representations of sp∞ . Theorem 3.2 For any nonnegative integer d the gl∞ -module V⊗d , regarded d as an sp∞ -module, has Loewy length 2 + 1, and d . r = 1, . . . , ker ΦI1 ,...,Ir : V⊗d → V⊗(d−2r) , soc(r) V⊗d = 2 I1 ,...,Ir

(13) Proof. The proof is obtained as a minor variation of the proof of Theorem 2.2. Denoting by F(r) the module on the right side of (13), we argue as before that Γλ ⊗ H(λ; r) F(r+1) /F(r) ∼ = |λ|=p−2r

for some Sd -modules H(λ; r). The semisimplicity of these layers shows that F(r) ⊂ soc(r) V⊗d for all r, and the opposite inclusion is proved by an obvious modification of Corollary 6.3. Theorem 3.3 For any partition λ the sp∞ -module Γλ;0 is indecomposable, and ⎛ ⎞ d (r+1) λ ⎝ ⎠ . (14) soc Γλ;0 = Nµ, (2γ) Γµ , r = 1, . . . , 2 µ |γ|=r

Proof. By construction, the gl∞ -module Γλ;0 is realized as the direct summand Sλ V of the gl∞ -module V⊗d . It remains a direct summand when V⊗d is regarded as an sp∞ -module, therefore soc(r) Γλ;0 = Γλ;0 ∩ soc(r) V⊗d . (2n) (n) It is known that for any partitions λ, µ we have [Γλ : Γµ ] = γ Nλµ, (2γ) provided n is large enough, see e.g. [HTW]. This implies that [Γλ;0 : Γµ ] = λ γ Nµ, (2γ) , and combining it with the description of the socle filtration of ⊗d V , we get (14). In particular, socΓλ;0 ∼ = Γλ , and the simplicity of the socle implies the indecomposability of Γλ;0 as an sp∞ -module.

140


Examples. We begin by describing the structure of V⊗V = Λ2 V⊕S 2 V as an sp∞ -module. The symmetric square S 2 V = Γ(2);(0) is the irreducible adjoint sp∞ -module, isomorphic to Γ(2) . For the exterior square Λ2 V = Γ(1,1);(0) one has the short exact sequence of sp∞ -modules Ω

0 → Γ(1,1) → Λ2 V → k → 0 which does not split. Therefore the structure of V ⊗ V is graphically represented as Γ V⊗V ∼ ⊕ (0) . Γ(2) Γ(1,1) Similarly, the structure of the tensor representation of rank 3 is represented as Γ Γ(1) V⊗3 ∼ ⊕ 2 (1) ⊕ , Γ(3) Γ(2,1) Γ(1,1,1) and the structure of the tensor representation of rank 4 as V⊗4

∼ Γ(4)

Γ(0) Γ(0) ⊕3 Γ(2) ⊕2 Γ(1,1) ⊕3 Γ(2) ⊕ Γ[(1,1)] ⊕ Γ(1,1) . Γ(3,1) Γ(2,1,1) Γ(2,2) Γ(1,1,1,1)

4 Tensor representations of so∞ Let V be a countable dimensional vector space, and let Q : V ⊗ V → k be a nondegenerate symmetric bilinear form on V. We realize the Lie algebra gl∞ as in Section 2 by taking V∗ = V and Q as the pairing ·, · . The Lie algebra (Q) so∞ is defined as the Lie subalgebra of this gl∞ , which preserves the form Q, i.e.,

so(Q) Q(Xu, v) + Q(u, Xv) = 0 for all u, v ∈ V . = X ∈ gl ∞ ∞ It is customary when dealing with the orthogonal groups to work over algebraically closed fields. Throughout this section we assume that k is algebraically closed. Then all nondegenerate forms Q on V are equivalent (e.g., they can all be transformed to the standard sum of squares), and as a (Q) consequence the corresponding Lie algebras so∞ are isomorphic. Hence we drop Q from the notation, and denote our Lie algebra simply by so∞ . It is convenient to pick a basis {ξi } of V, indexed as before by the set I = Z \ {0}, such that Q(ξi , ξj ) = δi+j,0 . In the coordinate realization of gl∞ , the Lie algebra so∞ has a linear basis {Ei,j − E−j,−i }i,j∈I . Since the dual 2 basis {ξi∗ }i∈I is given by ξi∗ = ξ−i , it follows that so∞ = V, and the Lie bracket is induced by (1).


141

We call V, regarded as an so∞ -module by restriction, the natural representation of so∞ . It is easy to see that the so∞ -action on the gl∞ -module V⊗(p,q) coincides with the so∞ -action on the gl∞ -module V⊗(p+q) . Therefore it suffices to study the tensor representations V⊗d . To define the notion of a highest weight so∞ -module, we consider the so direct sum decomposition so∞ = hso ⊕ , where k X α α∈∆sp hso =

k (Ei,i − E−i,−i ),

∆so = {±εi ± εj ) | i, j ∈ N and i = j},

i∈I so = E−j,i − E−i,j . Xεsoi +εj = Ei,−j − Ej,−i , Xεsoi −εj = Ei,j − E−j,−i , X−ε i −εj

The subalgebras bso and its ideal nso defined as before using the polariza are + tion of the root system ∆so = ∆+ −∆ so so , where ∆+ so = {εi ± εj | i, j ∈ N and i < j}. An so∞ -module V is called a highest weight module with highest weight χ ∈ h∗so , if it is generated by a vector v ∈ V satisfying nso v = 0 and H v = χ(H) v for all H ∈ hso . In particular, the natural representation V is a highest weight module with highest weight ε1 , generated by the highest weight vector ξ1 . Remark 3. Our choice of a splitting Cartan subalgebra hso leads to a root decomposition of type D∞ , corresponding to the “even” infinite orthogonal series of Lie algebras so2n . It is also possible to pick another splitting Cartan ˜so which leads to a root decomposition of type B∞ , corresponding subalgebra h to the “odd” infinite orthogonal series so2n+1 . These two types of splitting Cartan subalgebras are clearly not conjugated, and in [DPS] it is proved that any splitting Cartan subalgebra of so∞ is either “even” or “odd”. For any pair I = (i, j) of integers, satisfying 1 ≤ i < j ≤ d, define the contraction Φ[I] : V⊗d → V⊗(d−2) , v1 ⊗ . . . ⊗ vd → Q(vi , vj ) v1 ⊗ . . . ⊗ vî ⊗ . . . ⊗ vˆj ⊗ . . . ⊗ vd . def

(15)

def

We set V[0] = k, V[1] = V and V

[d] def

=

⊗d ⊗(d−2) ker Φ[I] : V → V

I

for d ≥ 2. For any partition λ, define the so∞ -submodule Γ[λ] of V⊗d , def

Γ[λ] = V[d] ∩ Sλ V.

(16)

142


Theorem 4.1 For any d ∈ N there is an isomorphism of (so∞ , Sd )-modules Γ[λ] ⊗ Hλ . (17) V[d] ∼ = |λ|=d

For every partition λ, the so∞ -module Γ[λ] is an irreducible highest weight module with highest weight ω = i∈N λi εi . Proof. For each n denote by Vn the 2n-dimensional subspace of V spanned by ξ±1 , . . . , ξ±n . Let gn be the Lie subalgebra of End(Vn ) which preserves the restriction of Q to Vn . It is clear that gn ∼ = so2n . The tensor representations (n) (Vn )⊗d of gn and the contractions Φ[I] : (Vn )⊗d) → (Vn )⊗(d−2) are defined as before. We set def (n) (n) def ker Φ[I] , Γ[λ] = (Vn )[d] ∩ Sλ Vn . (Vn )[d] = I

The Weyl construction for so2n implies that there are isomorphisms of (so2n , Sd )-modules (n) (Vn )[d] ∼ Γ[λ] ⊗ Hλ , (18) = |λ|=d (n)

and that for n ≥ d the module Γ[λ] is an irreducible highest weight module with highest weight i∈N λi εi . As in the proof of Theorem 2.1, the diagram of inclusions (V1 )[d] ⊂ (V2 )[d] ⊂ . . . ⊂ (Vn )[d] ⊂ . . . ⊂ n∈N (Vn )[d] = V[d] ∪

∪

(1) Γ[λ]

(2) Γ[λ]

⊂

∪ ⊂ ... ⊂

(n) Γ[λ]

⊂ ... ⊂

∪

(n) n∈N Γ[λ]

∪

(19)

= Γ[λ]

establishes both the irreducibility of the so∞ -module Γλ and the isomorphism (17), and the equalities (n) (d) (d+1) (n) Γλ [χ] = Γλ [χ] = · · · = Γλ [χ] = · · · = Γλ [χ] = Γλ [χ] n∈N

yield the characterization of Γ[λ] [χ] as a highest weight so∞ -module.

⊗d Theorem 4.2 For any nonnegative d integer d the Loewy length of V , regarded as an so∞ -module, equals 2 + 1, and d . r = 1, . . . , ker Φ[I1 ,...,Ir ] : V⊗d → V⊗(d−2r) , soc(r) V⊗d = 2 I1 ,...,Ir

(20)


143

Proof. The proof is a minor variation of the proof of Theorem 2.2. Denoting by F(r) the module on the right side of (20), we argue as before that Γ[λ] ⊗ H(λ; r) F(r+1) /F(r) ∼ = |λ|=p−2r

for some Sd -modules H(λ; r). The semisimplicity of these layers shows that F(r) ⊂ soc(r) V⊗d for all r, and the opposite inclusion is proved by an obvious modification of Corollary 6.3.

Theorem 4.3 Let λ be a partition of a positive integer d. Then the so∞ module Γλ;0 is indecomposable, and the layers of its socle filtration satisfy ⎛ ⎞ ⎝ soc(r+1) Γλ;0 ∼ Nλµ, 2γ ⎠Γ[µ] . (21) = µ

|γ|=r

Proof. By construction, the gl∞ -module Γλ;0 is realized as the direct summand Sλ V of the gl∞ -module V⊗d . It remains a direct summand when V⊗d is regarded as an so∞ -module, therefore soc(r) Γλ;0 = Γλ;0 ∩ soc(r) V⊗d . (2n) (n) It is known that [Γλ : Γ[µ] ] = γ Nλµ, 2γ for any partitions λ, µ, provided λ n is large enough, see e.g., [HTW]. Hence [Γλ;0 : Γ[µ] ] = γ Nµ, 2γ , and combining this with the description of the socle filtration of V⊗d , we get (21). In particular, socΓλ;0 ∼ = Γ[λ] , and simplicity of the socle implies the indecomposability of Γλ;0 as an so∞ -module. Corollary 4.4 The decomposition of V⊗d into indecomposable so∞ -modules is given by the isomorphism V⊗d ∼ (dim Hλ ) Γλ;0 . = |λ|=d

Examples. We begin by describing the structure of V ⊗ V = Λ2 V ⊕ S 2 V as an so∞ -module. The exterior square Λ2 V = Γ(1,1);(0) is the irreducible adjoint so∞ -module, isomorphic to Γ[(1,1)] . For the symmetric square S 2 V = Γ(2);(0) one has the short exact sequence of so∞ -modules Q

0 → Γ[(2)] → S 2 V → k → 0 which does not split. Therefore, the structure of V ⊗ V is graphically represented as Γ[(0)] V⊗V ∼ ⊕ . Γ[(1,1)] Γ[(2)]

144


Similarly, the structure of the tensor representation of rank 3 is represented as V⊗3

∼

Γ[(1)] Γ ⊕ 2 [(1)] ⊕ , Γ[(1,1,1)] Γ[(3)] Γ[(2,1)]

and the structure of the tensor representation of rank 4 as V⊗4

∼

Γ[(0)] Γ[(0)] Γ[(2)] ⊕3 Γ[(2)] ⊕ Γ[(1,1)] ⊕2 Γ[(2)] ⊕3 Γ[(1,1)] ⊕ Γ[(1,1,1,1)] . Γ[(3,1)] Γ[(2,1,1)] Γ[(4)] Γ[(2,2)]

5 Tensor representations of root-reductive Lie algebras It is known that over an algebraically closed field any infinite-dimensional simple locally finite Lie algebra which admits a root decomposition is classical, i.e., isomorphic to sl∞ , sp∞ or so∞ , see [PS, NS]. Here we discuss a generalization of our results for sl∞ , so∞ and sp∞ to a more general class of infinite-dimensional Lie algebras. Let k be one of the Lie algebras sl∞ , sp∞ or so∞ , and let hk denote its splitting Cartan subalgebra introduced in previous sections. Let V denote the natural representation of k. Suppose that a Lie algebra g is the semidirect sum of k and some Lie algebra m, + m, g = k⊂ and suppose furthermore that g has a subalgebra h such that h = hk ⊕ m. Theorem 5.1 The socle filtration of the tensor representation of V⊗(p,q) as a g-module coincides with the socle filtration of V⊗(p,q) as a k-module. Proof. Since m commutes with hk , the action of m preserves the hk -weight subspaces of any k-module. The hk -weight subspaces of V are one-dimensional, hence any element of m acts in any V[χ] as a scalar. Thus V admits a hweight subspace decomposition, and the same is true for the weight subspaces of tensor representations V⊗(p,q) . This shows that each k-submodule of V⊗(p,q) is automatically a g-module, and thus the socle filtration of V⊗(p,q) as a k-module is a filtration by gsubmodules. Moreover, the layers of the socle filtration for k remain semisimple as g-modules, and the statement follows. Our main application of Theorem 5.1 is to the class of infinite-dimensional root-reductive Lie algebras, studied in [DP1, PS]. Recall the following structural theorem from [DP1].


145

Theorem 5.2 Let g be a root reductive Lie algebra. Set s = [g, g] and a = g/[g, g]. Then s= s(i) , i∈I

where each s(i) is isomorphic either to sl∞ , so∞ , sp∞ , or to a simple finitedimensional Lie algebra, and I is an at most countable index set. Moreover, the short exact sequence of Lie algebras 0→s→g→a→0

(22)

+ a. splits. In other words, g ∼ = s⊂

To apply Theorem 5.1 to a root-reductive Lie algebra g, we identify g + a, pick an infinite-dimensional direct summand with a semidirect sum g ∼ = s⊂ (j) (i) + k = s of s, and set m = ⊂ a. The Lie algebra g acts in the i =j s natural representation V of s(j) : s(i) annihilate V for i = j, and a acts by scalars in each hk -weight subspace. In contrast with the finite-dimensional reductive Lie algebras, there exist root-reductive Lie algebras g such that g s ⊕ a. For example, for g = gl∞ one has s = sl∞ and a = k, but gl∞ sl∞ ⊕ k. Another interesting example is the Lie algebra ˜ g, constructed via the following root injections ⎛ Tr(A) ⎞ gln → gln+2 ,

A → ⎝

n

A ⎠. 0

Then g˜ is not isomorphic to gl∞ , although it can still be included in a short exact sequence of Lie algebras 0 → sl∞ → ˜ g → k → 0, see [DPS]. However, Theorem 5.1 still applies and describes the socle filtration of the tensor representations of ˜ g.

6 Appendix For completeness, we discuss the details of Weyl’s duality approach, see [W, FH]. Let p, q, n be nonnegative integers such that n > p + q. Let Vn = kn be the natural representation of the Lie algebra gln . For partitions λ, µ such (n) that |λ| = p and |µ| = q, we denote by Γλ;µ the irreducible highest weight gln -module with highest weight ω = (λ1 , . . . , λp , 0, . . . , 0, −µq , . . . , −µ1 ). Proposition 6.1 For any n > p + q there is an isomorphism (3) (n) (Vn ){p,q} ∼ Γλ;µ ⊗ (Hλ ⊗ Hµ ). = |λ|=p |µ|=q

146


Proof. For any I = (i, j) with i ∈ {1, 2, . . . , p} and j ∈ {1, 2, . . . , q}, we define the inclusion (n)

ΨI

∗ : (Vn )⊗(p−1,q−1) → (Vn )⊗(p,q) , v1 ⊗ . . . ⊗ vp−1 ⊗ v1∗ ⊗ . . . ⊗ vq−1 → n 1 ∗ ∗ → . . . ⊗ vi−1 ⊗ ζk ⊗ vi+1 ⊗ . . . ⊗ vj−1 ⊗ ζk∗ ⊗ vj+1 ⊗ ..., n k=1

where {ζk } and {ζk∗ } are any dual bases of Vn and Vn∗ respectively. Set θI = (n) (n) (n) (n) ΨI ΦI . The operators θI are idempotent, and (Vn ){p,q} = I ker θI . ⊗(p,q) Let A denote the subalgebra of endomorphisms of (Vn ) , generated by the images of elements of gn , and let B denote its commutator subalgebra, i.e., the set of all endomorphisms of (Vn )⊗(p,q) , commuting with the action (n) of gn . From invariant theory it is known that B is generated by θI for various I, and by permutation maps corresponding to elements from Sp × Sq . A general result from the theory of semisimple finite-dimensional algebras In other words, implies that, conversely, A is the commutator subalgebra of B. (n) ⊗(p,q) any endomorphism of (Vn ) , commuting with all θI and all permutations from Sp × Sq , must lie in A. Let A denote the subalgebra of endomorphisms of (Vn ){p,q} , generated by the images of elements of gn , and let B denote the subalgebra of endomorphisms of (Vn ){p,q} , generated by the permutations from Sp × Sq. We claim that A and B are each other’s commutator subalgebras in End (Vn ){p,q} . Indeed, suppose that L lies in the commutator subalgebra of B. Using the (gn , Sp × Sq )-module isomorphism (see e.g., [FH] for details) (n) (Vn )⊗(p,q) = (Vn ){p,q} ⊕ im θI , (n)

I

of (Vn )⊗(p,q) by extending L trivially on the we construct an endomorphism L commutes with all permutations second direct summand. It is clear that L (n) from Sp × Sq , and also with the operators θI , all of which act on (Vn ){p,q} belongs to the commutator subalgebra of B, i.e., L ∈ A, by zero. Hence L and by restriction L ∈ A. Thus A is the commutator subalgebra of B, and it follows that B is also the commutator subalgebra of A. The general theory of dual pairs and the fact that {Hλ ⊗ Hµ }|λ|=p,|µ|=q is a complete list of irreducible Sp × Sq -modules imply the existence of an isomorphism (Vn ){p,q} ∼ Γ (λ, µ) ⊗ (Hλ ⊗ Hµ ) = |λ|=p |µ|=q

for some irreducible gln -modules Γ (λ, µ). To identify these modules explicitly, we note that Schur–Weyl Duality yields (n) (n) (Vn )⊗p ∼ Γλ;0 ⊗ Hλ , (Vn∗ )⊗q ∼ Γ0;µ ⊗ Hµ , = = |λ|=p

|µ|=q

Tensor Representations of Classical Locally Finite Lie Algebras (n)

147

(n)

and therefore Γ (λ, µ) must be a submodule of Γλ;0 ⊗ Γ0;µ . On the other hand, (n) Γλ;µ

of this tensor product does not occur as a submodule of the submodule (Vn )⊗(p−1,q−1) , and thus lies in the kernel of all operators ΦI . We conclude that (n) Γλ;µ ⊂ Γ (λ, n), and the irreducibility of Γ (λ, µ) yields the desired statement. Finally, to prove the technical statement used in the proof of Theorem 2.2, we need a preparatory lemma. Define the contractions ΦI1 ,...,Ir : V⊗(p,q) → V⊗(p−r,q−r) as the r-fold convolutions between copies of V and V∗ indicated by the pairwise disjoint collection of index pairs I1 , . . . , Ir . For r = 1, . . . , , define the inclu(n) sions ΨI1 ,...,Ir : (Vn )⊗(p−r,q−r) → (Vn )⊗(p,q) , by analogy with ΦI1 ,...,Ir , as the r-fold insertions of the canonical element of Vn ⊗ Vn∗ into positions specified by disjoint pair of indices I1 , . . . , Ir . Set also (n) (Vn ){p,q} = im ΨI1 ,...,Ir : (Vn )⊗(p−r,q−r) → (Vn )⊗(p,q) . r I1 ,...,Ir

It is a standard exercise to show that for each n one has the direct sum decomposition {p,q}

(Vn )⊗(p,q) = (Vn )0

{p,q}

⊕ (Vn )1

{p,q}

⊕ (Vn )2

{p,q}

⊕ · · · ⊕ (Vn )

.

For any I = (i, j) we consider the linear map (n)

ΞI

: (Vn )⊗(p,q) → (Vn )⊗(p−1,q−1) ,

v1 ⊗ . . . ⊗ vp ⊗ v1∗ ⊗ . . . ⊗ vq∗ →

→ n ξn∗ , vi vj∗ , ξn v1 ⊗ . . . ⊗ vî ⊗ . . . ⊗ vp ⊗ v1∗ ⊗ . . . ⊗ vˆj∗ ⊗ . . . ⊗ vq∗ . Lemma 6.2 For any v ∈ V{p,q} we have v, if {I1 , . . . , Ir } = {J1 , . . . , Jr } as sets (n) (n) (n) . lim Ξ ΦJ2 ,...,Jr ΨI1 ,...,Ir v = n→∞ J1 0, otherwise. Proof. We use induction on s. The base of induction s = 1 states that v, if I = J (n) (n) lim Ξ ΨI v = n→∞ J 0, otherwise, which is clear from the definition. Assume now that r ≥ 2, and let Jr = (i, j). Case 1. There exists k such that Ik = (i, j); we may assume that k = r. Then (n)

(n)

(n)

(n)

(n)

(n)

(n)

(n)

(n)

ΞJ1 ΦJ2 ,...,Jr ΨI1 ,...,Ir v = ΞJ1 ΦJ2 ,...,Jr−1 ΨI1 ,...,Ir−1 ΦIr ΨJr v = ΞJ1 ΦJ2 ,...,Jr−1 ΨI1 ,...,Ir−1 v, and the desired statement follows from the induction hypothesis.

148


Case 2. There exist k, l such that Ik = (i, a) and Il = (b, j); we may assume that k = r and l = r − 1. Setting I = (b, a) and using the identity (n) (n) (n) Φ(i,j) Ψ(i,a),(b,j) = n1 Ψb,a , we get (n)

(n)

(n)

(n)

(n)

(n)

(n)

(n)

ΞJ1 ΦJ2 ,...,Jr ΨI1 ,...,Ir v = ΞJ1 ΦJ2 ,...,Jr−1 ΨI1 ,...,Ir−2 ΦJr ΨIr−1 ,Ir v 1 (n) (n) (n) Ξ Φ Ψ v. n J1 J2 ,...,Jr−1 I1 ,...,Ir−2 ,I

=

Applying the induction hypothesis, in both cases we obtain (n)

(n)

(n)

lim ΞJ1 ΦJ2 ,...,Jr ΨI1 ,...,Ir v = 0.

n→∞

Case 3. There exists k such that Ik = (i, a), but j never occurs in the second position of any Il ; we may assume that k = r. Using the identity (n) (n) (m) (m) Φ(i,j) Ψ(i,a) v = m n Φ(i,j) Ψ(i,a) v, we get (n)

(n)

(n)

(n)

(n)

(n)

(n)

(n)

lim ΞJ1 ΦJ2 ,...,Jr ΨI1 ,...,Ir v = lim ΞJ1 ΦJ2 ,...,Jr−1 ΨI1 ,...,Ir−1 ΦJr ΨIr v n→∞ m (n) (n) (n) (m) (m) ΞJ1 ΦJ2 ,...,Jr−1 ΨI1 ,...,Ir−1 ΦJr ΨIr v . = lim n→∞ n

n→∞

Applying the induction hypothesis to ΦJr ΨIr v ∈ V{p,q} , we see that the desired statement holds. Case 4. The index i never occurs in the first position of any Ik , and j never occurs in the second position of any Il . Then for all n we have (m)

(n)

(n)

(n)

(n)

(n)

(m)

(n)

ΞJ1 ΦJ2 ,...,Jr ΨI1 ,...,Ir v = ΞJ1 ΦJ2 ,...,Jr−1 ΨI1 ,...,Ir ΦIr v = 0.

We are now ready to prove the following assertion, used in the proof of Theorem 2.2. (s−1)

/ Fn Proposition 6.3 There exist infinitely many n such that u − πn (u) ∈ (s−1)

.

for all n m. Proof. Assume that, on the contrary, u − πn (u) ∈ Fn Let J1 , . . . , Js be any collection of pairwise disjoint indices. Since for all n ΦJ2 ,...,Jr (u − πn (u)) = 0, we obtain (n)

(n)

ΞJ1 ΦJ2 ,...,Jr πn (u) = ΞJ1 ΦJ2 ,...,Jr u = 0. On the other hand, the vector πn (u) can be represented as πn (u) = ΨI1 ,...,Is ζI1 ,...,Is I1 ,...,Is


149

for some collection {ζI1 ,...,Is } of vectors from (Vn ){p,q} , and according to Lemma 6.2 (n) lim ΞJ1 ΦJ2 ,...,Jr πn (u) = ζJ1 ,...,Jn . n→∞

It follows that ζJ1 ,...,Jn = 0, and thus πn (u) = 0. This contradicts the assump(s+1) (r) (n) tion that πn (u) generates a submodule of Fn /Fn isomorphic to Γλ;µ .

References [B]

A. Baranov, Finitary simple Lie algebras. J. Algebra 219 (1999), no. 1, 299–329. [BB] Y. Bakhturin, G. Benkart, Weight modules of direct limit Lie algebras, Comm. Algebra 27 (1998), 2249–2766. [BS] Y. Bakhturin, H. Strade, Some examples of locally finite simple Lie algebras. Arch. Math. (Basel) 65 (1995), no. 1, 23–26. [DP1] I. Dimitrov, I. Penkov, Weight modules of direct limit Lie algebras. Internat. Math. Res. Notices 1999, no. 5, 223–249. [DP2] I. Dimitrov, I. Penkov, Borel subalgebras of gl(∞). Resenhas 6 (2004), no. 2-3, 153–163. [DPW] I. Dimitrov, I. Penkov, J. Wolf, A Bott-Borel-Weil theory for direct limits of algebraic groups. Amer. J. Math. 124 (2002), no. 5, 955–998. [DPS] E. Dan-Cohen, I. Penkov, N. Snyder, Cartan subalgebras of root-reductive Lie algebras. J. Algebra 308 (2007), no. 2, 583–611. [FH] W. Fulton, J. Harris, Representation Theory. A First Course, Graduate Texts in Mathematics, 129. Readings in Mathematics. Springer-Verlag, New York, 1991. [HTW] R. Howe, E. Tan, J. Willenbring, Stable branching rules for classical symmetric pairs. Trans. Amer. Math. Soc. 357 (2005), no. 4, 1601–1626. [M] G. Mackey, On infinite-dimensional linear spaces. Trans. Amer. Math. Soc. 57 (1945), 155–207. [N] K.-H. Neeb, Holomorphic highest weight representations of infinitedimensional complex classical groups, J. Reine Angew. Math, 497 (1998), 171–222. [NP] K.-H. Neeb, I. Penkov, Cartan subalgebras of gl∞ , Canad. Math. Bull. 46 (2003), no. 4, 597–616. [NS] K.-H. Neeb, N. Stumme, The classification of locally finite split simple Lie algebras. J. Reine Angew. Math. 533 (2001), 25–53. [Na] L. Natarajan, Unitary highest weight-modules of inductive limit Lie algebras and groups, J. Algebra 167 (1994), no. 1, 9–28. [O1] G. Olshanskii, Unitary representations of the group SO0 (∞, ∞) as limits of unitary representations of the groups SO0 (n, ∞) as n → ∞. Funct. Anal. Appl. 20 (1998), 292–301. [O2] G. Olshanskii, Representations of infinite-dimensional classical groups, limits of enveloping algebras, and Yangians. In: Topics in Representation Theory (A. A. Kirillov, ed.). Advances in Soviet Math., vol. 2. Amer. Math, Soc., Providence, R.I., 1991, 1–66. [PS] I. Penkov, H. Strade, Locally finite Lie algebras with root decomposition, Archiv Math 80 (2003), 478–485.

150


[PZ]

I. Penkov, G. Zuckerman, A construction of generalized Harish-Chandra modules for locally reductive Lie algebras, Transformation Groups 13 (2008), 799–817. H. Weyl, The Classical Groups: Their Invariants and Representations, Princeton, 1939.

[W]

Lie Algebras, Vertex Algebras, and Automorphic Forms Nils R. Scheithauer Fachbereich Mathematik, Technische Universit¨ at Darmstadt, Schloßgartenstraße 7, 64289 Darmstadt, Deutschland, [email protected] Summary. Generalized Kac–Moody algebras are natural generalizations of the finite dimensional simple Lie algebras. In important cases their denominator identities are automorphic forms on orthogonal groups. The generalized Kac–Moody algebras with this property can probably be classified and realized as strings moving on suitable spacetimes. In this paper we describe these ideas in more detail. Key words: generalized Kac–Moody algebras, vertex algebras, automorphic forms. 2000 Mathematics Subject Classifications: 17B67, 17B68, 11F55.

1 Introduction The Fock space V of a bosonic string moving on a 26-dimensional torus carries a rich algebraic structure, i.e., V is a vertex algebra. There is a BRST-operator Q with Q2 = 0 acting on V . The physical states of the bosonic string are described by the cohomology group g of Q. The vertex algebra structure on V induces a Lie bracket on g. The Lie algebra g is called the fake monster algebra. It is almost a Kac–Moody algebra but not quite. The Lie algebra g has imaginary simple roots and therefore is a generalized Kac–Moody algebra. The theory of generalized Kac–Moody algebras is still similar to the Kac–Moody case. In particular there is a character formula and a denominator identity. The denominator identity of g is ∞ ρ α [1/∆](−α2 /2) ρ mρ 24 e . (1 − e ) = det(w) w e (1 − e ) + α∈II25,1

w∈W

m=1

Borcherds showed that the infinite product defines an automorphic form for an orthogonal group. The automorphism group of the Leech lattice acts naturally on g. Since the denominator identity is a cohomological identity this action gives twisted denominator identities. One of the nicest conjectures in this field is that these identities are also automorphic forms on orthogonal groups. K.-H. Neeb and A. Pianzola (eds.), Developments and Trends in Infinite-Dimensional Lie Theory, Progress in Mathematics 288, DOI 10.1007/978-0-8176-4741-4_5, © Springer Basel AG 2011

151

152

Nils R. Scheithauer

This example shows that generalized Kac–Moody algebras, vertex algebras and automorphic forms on orthogonal groups are closely connected. However we are far away from a good understanding of this relation. In the following sections we explain the above terms and ideas in more detail and state some of the open problems.

2 Generalized Kac–Moody algebras Generalized Kac–Moody algebras are natural generalizations of the finite dimensional simple Lie algebras. In this section we motivate the definition of these Lie algebras and sketch some of their properties. Let g be a finite-dimensional simple complex Lie algebra with Cartan subalgebra h and Killing form ( , ). Then g decomposes as g=h⊕ gα , α∈∆

where ∆ is the set of roots and the gα are 1-dimensional root spaces. We can choose simple roots α1 , . . . , αn , n = dim h. They have the property that each root α can be written as a linear combination of the simple roots αi with integral coefficients which are either all nonnegative or nonpositive. The symmetrized Cartan matrix A = (aij ) of g is defined by aij = (αi , αj ) . The entries of A are rational. Furthermore A has the following properties: 1. 2. 3. 4. 5.

aii > 0 aij = aji aij ≤ 0 for i = j 2aij /aii ∈ Z A is positive definite.

Serre found out that g can be recovered from its Cartan matrix in the following way. The Lie algebra with generators {ei , hi , fi | i = 1, . . . , n} and relations 1. 2. 3.

[ei , fj ] = δij hi [hi , ej ] = aij ej , [hi , fj ] = −aij fj ad(ei )1−2aij /aii ej = ad(fi )1−2aij /aii fj = 0 for i = j

is isomorphic to g. Killing and Cartan classified the finite-dimensional simple Lie algebras over the complex numbers. There are 4 infinite families: the classical Lie algebras

Lie Algebras, Vertex Algebras, and Automorphic Forms

153

An = sln+1 (C), Bn = so2n+1 (C), Cn = sp2n (C) and Dn = so2n (C), and 5 exceptional Lie algebras G2 , F4 , E6 , E7 and E8 . Let g be a finite-dimensional simple Lie algebra. Then the finite-dimensional representations of g decompose into irreducible components. Their characters can be calculated by Weyl’s character formula. A special case of this formula is the denominator identity of g eρ (1 − eα ) = det(w)w(eρ ) . α>0

w∈W

Outer automorphisms of g give twisted denominator identities. Serre’s construction can also be applied to matrices which are not positive definite. In this way we obtain Kac–Moody algebras [K1]. These Lie algebras are in general no longer finite dimensional but their theory is still similar to the finite-dimensional theory. Let A be an indecomposable real n × n-matrix satisfying the conditions 1.-4. above. Vinberg has shown that A is either positive definite or positive semidefinite and of rank n − 1 or indefinite. We say that A is of finite, affine or indefinite type. The Kac–Moody algebra corresponding to A is then finitedimensional simple, affine or indefinite. Let A be of affine type. Then Vinberg’s trichotomy implies that any proper principal submatrix of A decomposes into a sum of matrices of finite type. Using this result it is easy to classify the affine Kac–Moody algebras. The outcome is that to each finite dimensional simple Lie algebra corresponds an untwisted affine Kac–Moody algebra, and furthermore there are 6 twisted affine Kac–Moody algebras. The affine Kac–Moody algebras have become very important in many areas of mathematics. The main reason for this is that they admit realizations as central extensions of current algebras. Let g be a finite dimensional simple Lie algebra and ˆ g the corresponding untwisted affine Kac–Moody algebra. Then ˆ g = g ⊗ C[t, t−1 ] ⊕ Ck where C[t, t−1 ] is the algebra of Laurent polynomials, k a central element and [ x ⊗ tm, y ⊗ tn ] = [x, y] ⊗ tm+n + mδm+n (x, y)k . This realization implies for example that the denominator identities of the affine Kac–Moody algebras are Jacobi forms and that their vacuum modules carry a vertex algebra structure. Borcherds found out that the conditions on the Cartan matrix can be weakened further. We can also drop the condition that the elements on the diagonal of A are positive, i.e., allow imaginary simple roots, and still get nice Lie algebras from Serre’s construction. These Lie algebras are called generalized Kac–Moody algebras [B1], [B3], [B4], [J]. We describe them in more detail.

154

Nils R. Scheithauer

Let A = (aij )i,j∈I be a real square matrix, where I is some finite or countable index set, satisfying the following conditions. 1. 2. 3.

aij = aji aij ≤ 0 for i = j 2aij /aii ∈ Z if aii > 0

Let ˆ g be the Lie algebra with generators {ei , hij , fi | i ∈ I} and relations 1. 2. 3. 4.

[ei , fj ] = hij [hij , ek ] = δij aik ek , [hij , fk ] = −δij aik fk ad(ei )1−2aij /aii ej = ad(fi )1−2aij /aii fj = 0 [ei , ej ] = [fi , fj ] = 0 if aij = 0.

if aii > 0 and i = j

Then ˆ g has the following properties. The elements hij are 0 unless the i-th and j-th columns of A are equal. The nonzero hij are linearly independent and span a commutative subalgebra ˆ of ˆ h g. The elements hii are usually denoted hi . Every nonzero ideal of ˆg ˆ The center of ˆg is in h ˆ and contains all the has nonzero intersection with h. elements hij with i = j. ˆ of gˆ is the free abelian group generated by elements The root lattice Q αi , i ∈ I, with the bilinear form given by (αi , αj ) = aij . The elements αi are called simple roots. The Lie algebra ˆ g is graded by the root lattice if we define the degree of ei as αi and the degree of fi as −αi . We have the usual definitions of roots and root spaces. A root is called real if it has positive norm and imaginary otherwise. There is a unique invariant symmetric bilinear form on ˆg satisfying ˆ to (hi , hj ) = aij . We have a natural homomorphism of abelian groups from Q ˆ sending αi to hi . h There is also a character formula. The denominator identity takes the following form eρ (1 − eα )mult(α) = det(w) w eρ ε(α)eα α>0

w∈W

α

where ε(α), for α in the root lattice, is (−1)n if α is the sum of n pairwise orthogonal imaginary simple roots and 0 otherwise. A Lie algebra g is a generalized Kac–Moody algebra if g is isomorphic to ˆ g/c ⊕ d, where c is a subspace of the center of ˆ g and d an abelian subalgebra of g such that the elements ei and fi are eigenvectors of d and [d, hij ] = 0. ˆ ⊕ d is an abelian subalgebra of g called Cartan subalgebra. We Then h/c ˆ to h ˆ to transfer notations from ˆg to g. use the natural homomorphism from Q The denominator identity of g is usually given as a specialization of the denominator identity of ˆ g.


155

Borcherds showed [B5] that a real Lie algebra g satisfying the following conditions is a generalized Kac–Moody algebra. 1. g has a nondegenerate invariant symmetric bilinear form. 2. g has a selfcentralizing subalgebra h such that g decomposes into finite dimensional eigenspaces of h. 3. The bilinear form is Lorentzian on h. 4. The norms of the roots are bounded above. 5. If two roots are positive multiples of the same norm 0 vector then their root spaces commute. This characterization applies in most cases of interest. We will see that similar to the affine case in Kac–Moody theory there seems to be a class of generalized Kac–Moody algebras which can be classified and admit realizations.

3 Vertex algebras Vertex algebras give a mathematically rigorous definition of 2-dimensional quantum field theories. In this section we give a short introduction into the theory of these algebras. Nice references are [K2] and [FB]. A vertex algebra is a vector space V with a state-field correspondence which associates to each state a ∈ V a field a(z) = n∈Z an z −n−1 satisfying certain conditions. The precise definition is as follows. Let V be a vector space and V → End(V )[[z, z −1 ]] an z −n−1 a → a(z) = n∈Z

a state-field correspondence, i.e., an b = 0 for n sufficiently large. V is a vertex algebra if it satisfies the following conditions. 1. There is an element 1 ∈ V such that 1(z)a = a and a(z)1|z=0 = a. 2. The operator D on V defined by Da = a−2 1 satisfies [D, a(z)] = ∂a(z). 3. The locality condition (z − w)n [a(z), b(w)] = 0 holds for n sufficiently large. V is called a vertex operator algebra of conformal weight c if in addition V contains an element ω such that the operators Lm = ωm+1 give a representation of the Virasoro algebra of central charge c, i.e. [Lm , Ln ] = (m − n)Lm+n + L−1 = D and V = of L0 and V0 = C.

n≥0

m3 − m δm+n c , 12

Vn , where the Vn are finite-dimensional eigenspaces

156

Nils R. Scheithauer

The above definitions easily carry over to the Z2 -graded case. The most important property of vertex operators is Borcherds’ identity which follows from locality. For k, m, n ∈ Z we have ∞ m (−1)j am+k−j bn+j − (−1)m bm+n−j ak+j j j=0

=

∞ k (am+j b)k+n−j . j j=0

A consequence of this identity is that the space V /DV is a Lie algebra under [a, b] = a0 b. We will see that a slight modification of this construction relates vertex algebras to generalized Kac–Moody algebras. The simplest and best understood vertex algebras are lattice vertex algebras and Wess–Zumino–Witten models. A lattice vertex algebra describes the Fock space of a quantized chiral bosonic string moving on a torus. We give a short sketch of the construction. Let L be an even lattice, h = L ⊗ C the underlying complex vector space and ˆ = h ⊗ C[t, t−1 ] ⊕ Ck h the bosonic Heisenberg algebra with central element k and commutation relations [ x ⊗ tm, y ⊗ tn ] = mδm+n (x, y)k . ˆ− = h ⊗ t−1 C[t−1 ] is an abelian subalgebra of h. ˆ We denote by S(h ˆ− ) Then h − ˆ the symmetric algebra of h . ˆ →L→0 The lattice L has a unique central extension 0 → {±1} → L by a group of order 2 such that the commutator of the inverse images of α, β in L is (−1)(α,β) . Let ε : L × L → {±1} be the corresponding 2-cocycle and C[L]ε the twisted group algebra with basis {eα | α ∈ L} and products eα eβ = ε(α, β)eα+β . Then the vector space ˆ− ) ⊗ C[L]ε V = S(h carries a natural action of the Heisenberg algebra which can be used to define a vertex algebra structure. Now let ˆ g be an untwisted affine Kac–Moody algebra. Then the irreducible highest weight module LkΛ0 , where k is a positive integer, has a vertex operator algebra structure. These vertex algebras are called WZW models. They describe strings moving on compact Lie group manifolds. Modules of vertex algebras and vertex operator algebras are defined in a natural way. A vertex operator algebra is rational if it has only finitely many nonisomorphic simple modules and every module decomposes into a finite direct sum of simple modules. Examples of such vertex operator algebras are


157

vertex algebras of positive definite lattices and WZW models. A vertex operator algebra is called meromorphic if it has only one simple module, namely itself. Let V be a rational vertex operator algebra and M a simple V -module of conformal weight h. Then the character χM (τ ) = q −c/24 tr q L0 = q h−c/24

∞

dim(Mn )q n

n=0

is well defined. Zhu showed that the characters of a rational vertex operator algebra satisfying suitable finiteness conditions are holomorphic on the upper halfplane and the vector space generated by the characters is invariant under SL2 (Z). For example the character of a meromorphic vertex operator algebra of central charge 24 satisfying the finiteness conditions is modular invariant and therefore equal to j(τ ) = q −1 + 744 + 196884q + 21493760q 2 + 864299970q 3 + . . . up to an additive constant. Zhu’s result holds for vertex algebras of positive definite lattices and WZW models. Let V be a vertex operator algebra. Then V1 is a finite dimensional subalgebra of the Lie algebra V /DV . Schellekens shows in [ANS] that if V is a meromorphic vertex operator algebra of central charge 24 and nonzero V1 , then either dim V1 = 24 and V1 is commutative or dim V1 > 24 and V1 is semisimple. In the first case V is the vertex operator algebra of the Leech lattice. In the second case V can be written as a sum of modules over the affinization Vˆ1 of V1 and Schellekens shows that there are at most 69 possibilities for V1 . For each of these possibilities he finds a modular invariant partition function and describes V explicitly as a module over Vˆ1 . If the monster vertex operator algebra is unique and for each of the 69 candidates there exists a unique vertex operator algebra, Schellekens’ result implies that there are 71 meromorphic vertex operator algebras of central charge 24. Up to now these conjectures are open.

4 Automorphic forms on orthogonal groups Borcherds’ singular theta correspondence is a map from modular forms transforming under the Weil representation of SL2 (Z) to automorphic forms on orthogonal groups. In this section we give a short description of this lifting. References are [B6], [Br], [S3], and [S4]. Let L be an even lattice of even rank and L the dual lattice of L. The discriminant form of L is the finite abelian group D = L /L with quadratic

158

Nils R. Scheithauer

form γ 2 /2 mod 1. The level of D is the smallest positive integer N such that N γ 2 /2 = 0 mod 1 for all γ ∈ D. We define a scalar product on the group ring C[D] which is linear in the first and antilinear in the second variable by (eγ , eβ ) = δ γβ . Then there is a unitary action of SL2 (Z) on C[D] defined by ρD (T ) eγ = e(−γ 2 /2) eγ ρD (S) eγ =

e(sign(D)/8)

e(γβ) eβ , |D| β∈D

and T = ( 10 11 ) are the standard generators of SL2 (Z). This where S = 01 −1 0 representation is called the Weil representation of SL2 (Z) corresponding to D. A holomorphic function F (τ ) = Fγ (τ )eγ γ∈D

on the upper halfplane with values in C[D] is called a modular form for ρD of weight k if F (M τ ) = (cτ + d)k ρD (M )F (τ ) a b for all M = c d ∈ SL2 (Z) and F is meromorphic at i∞. Classical examples of modular forms transforming under the Weil representation are theta functions. Suppose L is a positive definite even lattice of even rank 2k. For γ ∈ D define 2 θγ+L (τ ) = q α /2 . α∈γ+L

Then θ(τ ) =

θγ+L (τ )eγ

γ∈D

is a modular form for the dual Weil representation of weight k which is holomorphic at i∞. Suppose D has level N . Then we can construct modular forms transforming under the Weil representation by lifting scalar valued modular forms on Γ0 (N ). The identity e0 in C[D] is up to a character value invariant under Γ0 (N ). Hence if f is a modular form on Γ0 (N ) of weight k and character χD then F (τ ) = f |M (τ ) ρD (M −1 )e0 M∈Γ0 (N )\SL2 (Z)

is a modular form for ρD of weight k. We can use the map for example to construct Eisenstein series for the Weil representation by lifting scalar valued Eisenstein series on Γ0 (N ).


159

Now we assume that L has signature (n, 2) with n ≥ 2. Let V = L ⊗ R and G be the Grassmannian of 2-dimensional negative definite subspaces of V . Then the orthogonal group O(V ) acts naturally on G. The Grassmannian G can be realized as a tube domain H ⊂ Cn as follows. Let z be a primitive norm 0 vector in L and z ∈ L such that (z, z ) = 1. Define K = L ∩ z ⊥ ∩ z ⊥ . Then V = K ⊗ R ⊕ Rz ⊕ Rz. Since K is a Lorentzian lattice the vector space K ⊗ R has two cones of negative norm vectors. We choose one of these cones and denote it by C. Let Z = X + iY ∈ K ⊗ C with Y ∈ C. Then ZL = Z + z − (Z 2/2 + z 2/2)z = XL + iYL is a norm 0 vector in V ⊗ C and {XL , YL } is an orthogonal basis of a negative definite subspace of V . This map gives a bijection between K ⊗ R + iC and G. The natural action of O+ (V ) on G induces an action by fractional linear transformations on H. A meromorphic function f on H is called an automorphic form of weight k for a subgroup Γ of finite index in O+ (L) if f (M Z) = j(M, Z)k f (Z) for all M ∈ Γ. Here j(M, Z) is an automorphy factor for O+ (V ). If n > 2 and f is holomorphic on H then by the Koecher principle f is also holomorphic at the cusps. We say that f has singular weight if f has weight n/2 − 1. If n > 2 this is the smallest possible weight of a holomorphic automorphic form on H. If f is a holomorphic automorphic form of singular weight and n > 2 then the only nonzero Fourier coefficients of f correspond to norm 0 vectors of K. Borcherds’ singular theta correspondence is a map from modular forms with poles at cusps transforming under the Weil representation of SL2 (Z) to automorphic forms on orthogonal groups. Since these automorphic forms have nice product expansions they are called automorphic products. We describe Borcherds’ construction in more detail. Let F be a modular form for ρD of weight 1 − n/2. The Siegel theta function θ(Z, τ ) of L is a function that is invariant under O+ (L) in Z ∈ H and transforms under the dual Weil representation in the second variable. Then the integral

dx dy Φ(Z) = F (τ )θ(Z, τ )y 2 , y F where F is the standard fundamental domain of SL2 (Z) on the upper halfplane, is formally invariant under the subgroup O+ (L, F ) preserving F . However the integral does not converge because F has a pole for τ → i∞. It can be regularized by taking the constant term in the Laurent expansion at s = 0 of the meromorphic continuation in s of

dx dy lim F (τ )θ(Z, τ )y 1+s 2 u→∞ F y u which converges for sufficiently large real part of s. Here Fu = {τ ∈ F | y ≤ u} denotes the truncated fundamental domain.

160

Nils R. Scheithauer

The automorphic product Ψ corresponding to F is obtained by exponentiating Φ. It is an automorphic form for O+ (L, F ). The zeros and poles of Ψ lie on rational quadratic divisors λ⊥ = {Z ∈ H | aZ 2/2−(Z, λk )−az 2/2−b = 0 }, where λ = λK + az + bz is a primitive vector of positive norm in L, and their orders are determined by the principal part of F . In the theory of generalized Kac–Moody algebras reflective automorphic products [S3] play an important role. A root of L is a primitive vector α in L of positive norm such that the reflection σα (x) = x − 2(x, α)α/α2 is an automorphism of L. If α is a vector in L of norm 2, then α is a root. In general L also has roots of norm greater than 2. We say that Ψ is reflective if its divisors are orthogonal to roots of L and are zeros of order one. The automorphic product Ψ is called symmetric reflective if Ψ is reflective and all roots of L of a given norm correspond to zeros of order one, and Ψ is completely reflective if Ψ is reflective and all roots of L give zeros of order one.

5 Moonshine for Conway’s group The largest sporadic simple group, the monster, acts on the monster algebra. This is a generalized Kac–Moody algebra describing the physical states of a bosonic string moving on a 2-dimensional orbifold. Borcherds [B4] shows in his proof of Conway and Norton’s moonshine conjecture for the monster [CN], [G] that the twisted denominator identities under this operation are automorphic forms of weight 0 on O2,2 (R). Conway’s group Co0 acts by diagram automorphisms on the fake monster algebra. We conjecture that the corresponding twisted denominator identities are automorphic forms of singular weight on orthogonal groups. The vertex algebras in this section are defined over the real numbers. The vertex algebra VII25,1 of the even unimodular Lorentzian lattice II25,1 describes a bosonic string moving on the torus R25,1 /II25,1 . In order to determine the physical states of the string we multiply VII25,1 with the vertex superalgebra VZσ of the b,c -ghost system. We obtain the vertex superalgebra V = VII25,1 ⊗ VZσ . Since VII25,1 has a conformal structure of central charge 26 and VZσ of central charge −26, the vertex superalgebra V carries an action of the Virasoro algebra of central charge 0. There is also a BRST-operator Q with Q2 = 0 acting on V and this operation commutes with the action of the Virasoro algebra. We define the ghost b = eσ and the ghost number operator j0N = σ(0). Then C = V ∩ ker b1 ∩ ker L0 is invariant under Q and graded by the ghost number Cαn . C= α∈II25,1 n∈Z


161

We have a sequence Q

Q

Q

Q

. . . −→ Cαn−1 −→ Cαn −→ Cαn+1 −→ . . . with cohomology groups Hαn . Let H=

Hαn .

α∈II25,1 n∈Z

Then H 1 = α∈II25,1 Hα1 is the space of physical states of the compactified bosonic string. It is easy to see that Hα1 is isomorphic to II25,1 ⊗ R if α = 0. Frenkel et al. [FGZ] have shown that for α = 0 the cohomology groups Hαn are trivial unless n = 1. The Euler–Poincaré principle implies dim Hα1 = [1/∆](−α2 /2) for α = 0, where 1/∆(τ ) =

∞ 1 1 = q −1 + 24 + 324q + 3200q 2 + . . . . q n=1 (1 − q n )24

For example if α has norm 2, then Hα1 is 1-dimensional. We define a product [LZ] on C by [u, v] = (b0 u)0 v. This product projects down to H and to g = H 1 . It defines a Lie bracket on g. It is not difficult to see that g satisfies all the conditions of Borcherds’ characterization. Hence g is a generalized Kac–Moody algebra. Using the singular theta correspondence we can show that the denominator identity of g is given by ∞ 24 2 eρ . 1 − emρ (1 − eα )[1/∆](−α /2) = det(w) w eρ + α∈II25,1

w∈W

m=1

Here ρ is a primitive norm 0 vector in II25,1 corresponding to the Leech lattice and W is the reflection group of II25,1 . The real simple roots of g are the norm 2 vectors in II25,1 such that (ρ, α) = −1 and the imaginary simple roots are the positive multiples mρ of ρ, each with multiplicity 24. The Lie algebra g has been constructed first in a slightly different way in [B2]. It is now called the fake monster algebra. Conway’s group Co0 is the automorphism group of the Leech lattice Λ. The characteristic polynomial of an element g in O(Λ) n can be written of order as k|n (xk − 1)bk . The eta product ηg (τ ) = η(kτ )bk is a modular form, possibly with poles at cusps, for a group of level N . We call N the level of g. The Leech lattice has a unique central extension 0 → {±1} → Λˆ → Λ → 0 such that the commutator of the inverse images of α, β in Λ is (−1)(α,β) . The ˆ = 224 .O(Λ) of automorphisms preserving the inner product acts group O(Λ) naturally on the fake monster algebra [B4]. Since II25,1 = Λ ⊕ II1,1 we can

162

Nils R. Scheithauer

write g =

e

ρ

ga with ga =

v∈Λ g(v,a) . Then for a = 0 the space ga is ˆ isomorphic as O(Λ)-module to the L0 -eigenspace VΛ,1−a2 /2 of degree 1 − a2 /2 of the vertex algebra of the Leech lattice. If a = 0 then ga is isomorphic to ˆ VΛ,1 ⊕ R2 as O(Λ)-module. The orthogonal decomposition Λ ⊗ R = Λg ⊗ R ⊕ Λg ⊥ ⊗ R gives a natural projection π : Λ ⊗ R → Λg ⊗ R which sends Λ to Λg . ˆ which acts trivially on the Each element g in O(Λ) has a lift gˆ to O(Λ) g inverse image of the fixed point lattice Λ . If g has order n, then the order n ˆ of gˆ is either n or 2n. The denominator identity of the fake monster algebra can be written as a cohomological identity. Taking the trace of gˆ over this identity we obtain a∈II1,1

α mult(α)

(1 − e )

=

det(w) w e

w∈W g

α∈L+

ρ

∞

1−e

mkρ bk

k|n m=1

where mult(α) =

dk|((α,L),ˆ n)

v∈Λ π(v)=r/dk

µ(k) d tr gˆ VΛ,1−a2 /2d2 k2 ,v dk

for α = (r, a) in L = Λg ⊕ II1,1 . The vector ρ is a primitive norm 0 vector in II1,1 and W g is the subgroup of elements in W mapping L into L. This identity is independent of the choice of gˆ. It is the denominator identity of a generalized Kac–Moody superalgebra whose real simple roots are the simple roots of W g , i.e., the roots α of L such that (ρ, α) = −α2 /2, and imaginary simple roots are the positive multiples mρ of the Weyl vector with multiplicity mult(mρ) = k|(m,n) bk . We conjecture that this identity defines an automorphic product of singular weight k/2, where k = dim Λg . We have Theorem 5.1. The conjecture is true for elements of squarefree level. The assertion is proved in [S1] for elements of squarefree level and nontrivial fixed point lattice and in [S2] for elements of squarefree level and trivial fixed point lattice. We give an outline of the proof for nontrivial fixed point lattices. Let g be an element in O(Λ) of squarefree level N and nontrivial fixed point lattice Λg . Then g has order N . We lift the modular form fg = 1/ηg to a vector valued modular form Fg on the lattice Λg ⊕ II1,1 ⊕ II1,1 (N ) . Then we apply the singular theta correspondence to obtain an automorphic form Ψg . We can represent this by the following diagram


163

g → 1/ηg → Fg → Ψg . Explicit calculation shows that Ψg has singular weight. The theta correspondence gives the product expansions of Ψg at the different cusps. We can determine the corresponding sum expansions because Ψg has singular weight so that the nonzero Fourier coefficients correspond to norm 0 vectors. The expansion of Ψg at the level N cusp is the twisted denominator identity of the fake monster algebra corresponding to g eρ

2

(1 − eα )[ck fg |Wk ](−α

k|N α∈(L∩kL )+

=

/2k)

∞ b 1 − emkρ k . det(w) w eρ k|N m=1

w∈W

Here L = Λg ⊕ II1,1 and ck is a constant such that ck fg |Wk has constant term bk . The conjecture is also proved for some elements which do not have squarefree level [S4]. Recall that the Mathieu group M23 acts on Λ. Theorem 5.2. Let g be an element in M23 of order N . Then the twisted denominator identity of g defines a reflective automorphic product Ψ of singular weight on Λg ⊕II1,1 ⊕II1,1 (N ). If N is squarefree then Ψ is completely reflective and the expansion of Ψ at any cusp is given by e((ρ, Z))

[1/ηg ](−α2 /2k) 1 − e((α, Z))

k|N α∈(L∩kL )+

=

det(w) ηg ((wρ, Z)),

w∈W

where W is the full reflection group of L = Λg ⊕ II1,1 . This identity is the denominator identity of a generalized Kac–Moody algebra whose real simple roots are the simple roots of W , which are the roots α of L with (ρ, α) = −α2 /2, and imaginary simple roots are the positive multiples mρ of the Weyl vector with multiplicity 24 σ0 ((N, m))/σ1 (N ). We remark that if g is an element in M23 of squarefree order N , then bk = 24/σ1 (N ) for all k|N . The theorem gives 10 generalized Kac–Moody algebras similar to the fake monster algebra. We describe them in more detail. We take simple roots αi in L according to their multiplicity as stated in the theorem and form the symmetrized Cartan matrix aij = (αi , αj ). Then we define ˆg by generators and relations as above. Let g be the quotient of ˆg by its center. Then g is a simple real generalized Kac–Moody algebra with Cartan subalgebra naturally isometric to L ⊗ R and root lattice L. The denominator identity of g is given in the theorem.

164

Nils R. Scheithauer

6 Classification results In this section we describe some classification results for automorphic products and generalized Kac–Moody algebras [S3]. Reflective automorphic products are theta lifts of reflective modular forms. These are modular forms transforming under the Weil representation which have very mild singularities corresponding to roots of the underlying lattice. We obtain a necessary condition for the existence of a reflective modular form by pairing it with an Eisenstein series. We describe this in more detail. Let an even lattice of even signature L be and discriminant form D. Let F = Fγ eγ be a modular form for the Weil representation ρD of weight 2 − k and E = Eγ eγ the Eisenstein series for ρD of weight k. Then by the unitarity of ρD the inner product F E = Fγ Eγ is a scalar valued modular form on SL2 (Z) of weight 2 with a pole at the cusp i∞. Hence F Edτ defines a meromorphic 1-form on the Riemann sphere with a pole at i∞. By the residue theorem its residue has to vanish. This implies that the constant term in the Fourier expansion of F E is 0. If L has squarefree level we can calculate this condition explicitly. The fast growth of the Bernoulli numbers implies the following result. Theorem 6.1. The number of symmetric reflective automorphic products of singular weight on lattices of signature (n, 2) with n > 2, squarefree level and p-ranks at most n + 1 is finite. In the theorem we consider symmetric automorphic products in order to exclude oldforms which are induced from sublattices. The condition on the p-ranks prevents that we have to discuss rescalings of lattices. We can work out a bound on the level and determine the solutions of the necessary condition by computer search. We find that the necessary condition has very few solutions and that in the case of completely reflective forms they come from existing forms. Theorem 6.2. Let L be an even lattice of signature (n, 2) with n > 2 and squarefree level N . Suppose L splits II1,1 ⊕ II1,1 (N ). Let Ψ be a completely reflective automorphic product of singular weight on L. Then Ψ can be constructed from an element of order N in M23 . If L has even p-ranks np the equation parametrizing the completely reflective automorphic products is given by n /2 k 1 1 −1 p k−np /2 np /2 p − 2 =1 + p p k − 2 Bk pk − 1 p p|N

where k = 1+n/2 . . . and Bk is the k-th Bernoulli number. This is an equation in the indeterminates N, k, p and np . It has exactly 8 solutions. Since the automorphic products in Theorem 6.2 are the denominator identities of generalized Kac–Moody algebras we obtain the following classification result.


165

Theorem 6.3. Let L be an even lattice of signature (n, 2) with n > 2 and squarefree level N . Suppose L splits II1,1 ⊕ II1,1 (N ). Let g be a real generalized Kac–Moody algebra whose denominator identity is a completely reflective automorphic product of singular weight on L. Then g corresponds to an element of order N in M23 .

7 Construction as strings In this section we show that some of the above generalized Kac–Moody algebras describe strings moving on suitable orbifolds (cf. [HS], [CKS]). There are exactly 10 generalized Kac–Moody algebras whose denominator identities are completely reflective automorphic products of singular weight on lattices of squarefree level. These Lie algebras correspond naturally to the elements of squarefree order N in M23 . In the case N = 1 the generalized Kac–Moody algebra can be realized as the cohomology group of the BRSToperator Q acting on the vertex superalgebra VII25,1 ⊗ VZσ . We will see that a similar result also holds in 4 other cases. Recall that Schellekens has constructed 69 vector spaces as modules over affine Kac–Moody algebras which are conjecturally meromorphic vertex operator algebras of central charge 24. Theorem 7.1. Suppose the meromorphic vertex operator algebra V of central charge 24 and spin-1 algebra Aˆrq,p , where p = 2, 3, 5 or 7, q = p − 1 and r = 48/q(p + 1), exists and has a real form. Then the cohomology group of the BRST-operator Q acting on V ⊗ VII1,1 ⊗ VZσ gives a natural realization of the generalized Kac–Moody algebra corresponding to the elements of order p in M23 . For p = 2 the existence of V and of a real form is proved. We sketch the proof of the theorem. As a vector space V is the sum of tensor products ⊗ri=1 VΛi where VΛi is an irreducible highest weight module over Aˆq of level p. The character of VΛi can be expressed in terms of string and theta functions. Using the transformation properties of the string functions under SL2 (Z) we show that the character of V as Aˆrq -module can be written as Fγ θγ . χV = F θ = γ∈N /N

Here N is the lattice of the largest possible minimal norm in the genus II2m,0 pp (m+2) with m = 24/(p + 1), θ = γ γ∈N /N Fγ e the lift of

γ∈N /N

f (τ ) =

θγ eγ the theta function of N and F =

1 η(τ )m η(pτ )m

166

Nils R. Scheithauer

to N . If V has the structure of a vertex algebra and admits a real form, we can proceed analogously as in the construction of the fake monster algebra. We only have to replace the vertex algebra of the Leech lattice VΛ in VII25,1 = VΛ ⊗ VII1,1 by V . Then the cohomology group of ghost number one of the BRST-operator Q acting on V ⊗ VII1,1 ⊗ VZσ is a generalized Kac–Moody algebra which we denote by g. Let Λp be the sublattice of the Leech lattice fixed by an element of order p in M23 . The Lie algebra g is graded by the lattice N ⊕ II1,1 which becomes isomorphic to L = Λp ⊕ II1,1 after rescaling by p. Using the singular theta correspondence we show that the denominator identity of g is given by eρ

2

(1 − eα )[f ](−α

α∈L+

/2)

α∈pL+

=

w∈W

2

(1 − eα )[f ](−α det(w) w e

ρ

/2p)

∞

nρ m

(1 − e ) (1 − e

pnρ m

)

n=1

so that g is the generalized Kac–Moody algebra corresponding to the elements of order p in M23 .

8 Open problems In this section we describe some open problems in the theory of generalized Kac–Moody algebras, vertex algebras and automorphic products. 1. One of the most important open problems in the theory of vertex algebras is to classify the meromorphic vertex operator algebras of central charge 24. There are conjecturally 71 such vertex operator algebras. This problem includes the proof that the monster vertex algebra is the only meromorphic vertex operator algebra of central charge 24 and trivial V1 and the verification of the existence and uniqueness of the potential theories in [ANS]. For the 24 lattice vertex algebras corresponding to the 24 positive definite even unimodular lattices in dimension 24 the existence and uniqueness is proved and 15 other vertex algebras can be constructed as Z2 -orbifolds [DGM1], [DGM2]. The remaining candidates can possibly be constructed by using higher twists and further orbifolding of some of the twisted theories [M]. However a more natural construction would be desirable. 2. The automorphism group of the Leech lattice acts naturally on the fake monster algebra. Show that the corresponding twisted denominator identities are automorphic forms of singular weight on orthogonal groups. So far this is


167

proved for the elements of squarefree level and some additional classes [S1], [S2], [S4]. 3. Classify reflective automorphic products, which are not assumed to be symmetric, of singular weight on arbitrary lattices. Then derive classification results for generalized Kac–Moody algebras whose denominator identities are reflective automorphic products of singular weight. Both problems have been solved for completely reflective automorphic products on lattices of squarefree level [S3]. The pairing argument used there can also be applied to the general case. For some classification results in signature one, see [GN]. 4. Can the generalized Kac–Moody algebras, whose denominator identities are reflective automorphic products of singular weight, be characterized by their root system? 5. Show that the generalized Kac–Moody algebras, whose denominator identities are reflective automorphic products of singular weight, can be realized as bosonic strings moving on suitable target spaces. This is proved in 2 cases [B2], [HS] and proved up to the existence of certain vertex operator algebras in 3 more cases [CKS]. 6. The denominator identities of some generalized Kac–Moody algebras are automorphic forms on orthogonal groups. Is the representation theory of these Lie algebras related to automorphic forms?

References [B1] [B2] [B3] [B4] [B5] [B6] [Br] [CKS] [CN] [DGM1]

R. E. Borcherds, Generalized Kac–Moody algebras, J. Algebra 115 (1988), 501–512. R. E. Borcherds, The monster Lie algebra, Adv. Math. 83 (1990), 30–47. R. E. Borcherds, Central extensions of generalized Kac–Moody algebras, J. Algebra 140 (1991), 330–335. R. E. Borcherds, Monstrous moonshine and monstrous Lie superalgebras, Invent. Math. 109 (1992), 405–444. R. E. Borcherds, A characterization of generalized Kac–Moody algebras, J. Algebra 174 (1995), 1073–1079. R. E. Borcherds, Automorphic forms with singularities on Grassmannians, Invent. Math. 132 (1998), 491–562. J. H. Bruinier, Borcherds products on O(2, l) and Chern classes of Heegner divisors, Lecture Notes in Mathematics, Vol.1780, Springer, Berlin, 2002. T. Creutzig, A. Klauer, N. R. Scheithauer, Natural constructions of some generalized Kac–Moody algebras as bosonic strings, Comm. Number Theory Phys. 1 (2007), 453–477. J. H. Conway and S. P. Norton, Monstrous moonshine, Bull. London Math. Soc. 11 (1979), 308–339. L. Dolan, P. Goddard, P. Montague, Conformal field theory of twisted vertex operators, Nuclear Phys. B 338 (1990), 529–601.

168

Nils R. Scheithauer

[DGM2] [FB] [FGZ] [G] [GN] [HS] [J] [K1] [K2] [LZ] [M] [S1] [S2] [S3] [S4] [ANS]

L. Dolan, P. Goddard, P. Montague, Conformal field theory, triality and the Monster group, Phys. Lett. B 236 (1990), 165–172. E. Frenkel, D. Ben-Zvi, Vertex Algebras and Algebraic Curves, 2nd ed., Mathematical Surveys and Monographs 88, Amer. Math. Soc, 2004. I. B. Frenkel, H. Garland and G. Zuckerman, Semi-infinite cohomology and string theory, Proc. Natl. Acad. Sci. USA 83 (1986), 8442–8446. T. Gannon, Monstrous moonshine: the first twenty-five years, Bull. London Math. Soc. 38 (2006), 1–33. V. A. Gritsenko, V. V. Nikulin, On the classification of Lorentzian Kac– Moody algebras, Russian Math. Surveys 57 (2002), 921–979. G. H¨ ohn and N. R. Scheithauer, A natural construction of Borcherds’ Fake Baby Monster Lie Algebra, Am. J. Math. 125 (2003), 655–667. E. Jurisich, Generalized Kac–Moody Lie algebras, free Lie algebras and the structure of the Monster Lie algebra, J. Pure Appl. Algebra 126 (1998), 233–266. V. Kac, Infinite Dimensional Lie Algebras, 3rd ed., Cambridge University Press, 1990. V. Kac, Vertex Algebras for Beginners, 2nd ed., University Lecture Series 10, Amer. Math. Soc, 1998. B. H. Lian, G. J. Zuckerman, New perspectives on the BRST-algebraic structure of string theory, Comm. Math. Phys. 154 (1993), 613–646. P. S. Montague, Conjectured Z2 -orbifold constructions of self-dual conformal field theories at central charge 24 – the neighborhood graph, Lett. Math. Phys. 44 (1998), 105–120. N. R. Scheithauer, Generalized Kac–Moody algebras, automorphic forms and Conway’s group I, Adv. Math. 183 (2004), 240–270. N. R. Scheithauer, Generalized Kac-Moody algebras, automorphic forms and Conway’s group II, J. Reine Angew. Math. 625 (2008), 125–154. N. R. Scheithauer, On the classification of automorphic products and generalized Kac–Moody algebras, Invent. Math. 164 (2006), 641–678. N. R. Scheithauer, The Weil representation of SL2 (Z) and some applications, Int. Math. Res. Not. 2009, no. 8, 1488–1545. A. N. Schellekens, Meromorphic c=24 conformal field theories, Comm. Math. Phys. 153 (1993), 159–185.

Kac–Moody Superalgebras and Integrability Vera Serganova Dept. of Mathematics, University of California at Berkeley, Berkeley, CA 94720 [email protected]

Summary. The first part of this paper is a review and systematization of known results on (infinite-dimensional) contragredient Lie superalgebras and their representations. In the second part, we obtain character formulae for integrable highest weight representations of sl (1|n) and osp (2|2n); these formulae were conjectured by Kac–Wakimoto. Key words: Kac–Moody superalgebras, affine superalgebras, Cartan matrices, integrable representations, character formulae. 2000 Mathematics Subject Classifications: Primary 17B67. Secondary 17B20.

1 Introduction The principal goal of this paper is to study a special class of Lie superalgebras which, in our opinion, plays the same role in the theory of Lie superalgebras as the Kac–Moody Lie algebras play in the theory of Lie algebras. Since the terminology is not completely uniform even in the case of Lie algebras, we start with brief discussion of this case. Given an arbitrary matrix A, one can define a contragredient Lie algebra g (A) by generators and relations (see [2] or Section 2 for precise definition). People are usually interested in the subclass of so-called Kac–Moody Lie algebras; by definition those are contragredient Lie algebras whose matrices satisfy the conditions: aii = 2, aij ∈ Z≤0 , and aij = 0 implies aji = 0. The main property which distinguishes Kac–Moody Lie algebras among all contragredient Lie algebras is the local nilpotency of generators in the adjoint representation. That allows one to define the Weyl group and the notion of an integrable representation whose character has a large group of symmetries. Finite-dimensional semisimple Lie algebras and affine Lie algebras can be characterized as the only Kac–Moody superalgebras which are of polynomial growth.


169

170

Vera Serganova

The first indication of usefulness of the notion of a contragredient Lie superalgebra appeared in [1]: a significant part of the list of finite-dimensional simple superalgebras is contragredient. Many properties of contragredient Lie algebras are shared by contragredient Lie superalgebras; one notable exception is the uniqueness of representation: contragredient Lie superalgebras g (A) and g (B) might be isomorphic even if A is essentially different from B. Our understanding of contragredient Lie superalgebras advanced a lot in the last three decades. In [10] Kac classified finite growth contragredient superalgebras whose Cartan matrices do not have zeros on the diagonal. Superalgebras of this class have the Weyl group and nice character formulas. But this class does not cover all finite-dimensional simple contragredient Lie superalgebras, and that makes one look for further generalization. Van de Leur in [8] studied a class of contragredient superalgebras with symmetrizable matrices. The condition that A is symmetrizable is equivalent to the existence of an even invariant symmetric form on the superalgebra g (A). He classified such superalgebras of finite growth; as in the usual case, those superalgebras are either finite dimensional, or central extensions of (twisted) loop superalgebras (the latter are called affine superalgebras). The representation theory of affine superalgebras has interesting applications in physics and number theory (see [11] and [12]). (These applications are based on conjectural character formulae for highest-weight representations; we prove these formulae for algebras sl (1|n)(1) and osp (2|n)(1) .) In this paper we discuss some aspects of structure theory and representations of contragredient Lie superalgebras. In particular, we suggest a definition of a Kac–Moody Lie superalgebra as a contragredient superalgebra with locally nilpotent adXi and adYi for any choice of a set of contragredient generators Xi , Yi , hi . The observation that all finite-growth contragredient Lie superalgebras are in this class led to a classification of finite-growth contragredient Lie superalgebras ([9]); this gives evidence that such a definition is reasonable. Defined in this way, Kac–Moody superalgebras have a very interesting replacement of the notion of Weyl group; it is not a group but a groupoid which acts transitively on the set of all Cartan matrices describing the given algebra. (This groupoid was the tool used in [9] to classify contragredient superalgebras of finite-growth; also see Section 7.) [9] also incorporates the classification of Kac–Moody superalgebras with indecomposable Cartan matrices satisfying certain nondegeneracy conditions; this classification was obtained by C. Hoyt in her thesis [3]. In the first part of the paper we review the classification results mentioned above, and we develop the structure theory similar to [2]. The main new result of this part is Theorem 4.14, which provides an algorithm listing all Cartan matrices defining the same Kac–Moody superalgebra. This result is parallel to the transitivity of the Weyl group action on the set of bases (see [2] Proposition 5.9). In Section 2 we define contragredient Lie superalgebras, introduce the notions of admissibility and type, and apply restrictions of quasisimplicity and

Kac–Moody Superalgebras and Integrability

171

regularity. The purpose of the restriction of quasisimplicity is, essentially, to simplify the discussion; on the one hand, the regularity is a key restriction required by the methods of this paper; on the other hand, in the examples we know, similar results seem to hold without this restriction. In Section 3 odd reflections and Weyl groupoid for a contragredient Lie superalgebra g are defined. We use odd reflections to construct a “large contragredient Lie subalgebra” g (B) of g and to investigate the span of roots of g (B), as well as the cone spanned by positive roots of g (B). In Section 4 we introduce notions of Kac–Moody superalgebra, Weyl group, real and imaginary roots, and discuss some geometric properties of roots of Kac–Moody superalgebras. Next, we provide useful examples of quasisimple Kac–Moody superalgebras: in Section 5, of rank 2, and in Section 6, what would turn out to be all quasisimple regular Kac–Moody Lie superalgebras. In Section 7 we collect together all known classification results for contragredient Lie superalgebras. Section 8 contains amplifications of several results from Sections 3 and 4. In Section 9 we revisit the examples and provide a more detailed description of their structure. In the rest of the paper we use the Weyl groupoid to study highest weight representations of Kac–Moody superalgebras. The main results there are Corollary 14.5 and Theorem 14.7, where we prove Kac–Wakimoto conjecture of [11, 12] for completely integrable highest weight representations. In Section 10 we define highest weight modules and integrable modules, and classify integrable highest weight modules over regular quasisimple Kac–Moody superalgebras. Section 11 collects results about highest weight representations which are true for any quasisimple regular Kac–Moody superalgebra. In Section 12 we study in detail highest weight representations for the Lie superalgebras sl (1|n)(1) , osp (2|2n)(1) , and S (1, 2, b). Those are the only infinite-dimensional Kac–Moody superalgebras of finite growth which have nontrivial integrable highest weight representations and a nontrivial Weyl groupoid. The invariant symmetric form starts playing its role in Section 13; this excludes S (1, 2, b) from the discussion; here we obtain “the non-obvious” results on geometry of odd reflections. Finally, in Section 14 the collected information allows us to obtain character formulae for all highest weight (1) (1) integrable representations over sl (1|n) and osp (2|2n) . The first part of the paper is mostly a review of results obtained somewhere else, we skip most technical proofs and refer the reader to the original papers. In the second part all proofs are written down as most of results there are new. The author thanks I. Zakharevich for careful and belligerent reading of a heap of versions of the manuscript and for providing innumerable useful suggestions and comments. The author also thanks M. Gorelik for pointing out defects in the initial version of the manuscript.

2 Basic definitions Our ground field is C. By p (x) we denote the parity of an element x in a vector superspace V = V¯0 ⊕ V¯1 .

172

Vera Serganova

Let I be a finite set of indices, p : I → Z2 , and A = (aij ) , i, j ∈ I, be a matrix. Fix an even vector space h of dimension 2|I| − rk (A). It was shown in [2] that one can choose linearly independent αi ∈ h∗ and hi ∈ h such that αj (hi ) = aij for all i, j ∈ I, and this choice is unique up to a linear transformation of h. Define a superalgebra ¯ g (A) by generators Xi , Yi , i ∈ I and h with relations [h, Xi ] = αi (h) Xi ,

[h, Yi ] = −αi (h) Yi ,

[Xi , Yj ] = δij hi ,

[h, h] = 0. (1)

Here we assume that h ∈ h, p (Xi ) = p (Yi ) = p (i). By g (A) (or g when the Cartan matrix is fixed) denote the quotient of ¯ g (A) by the (unique) maximal ideal which intersects h trivially. It is clear that if B = DA for some invertible diagonal D, then g (B) ∼ = g (A). Indeed, an isomorphism can be obtained by mapping hi and Xi to dii hi and dii Xi respectively. Therefore, without loss of generality, we may assume that aii = 2 or 0. We call such matrices normalized. The action of h on the Lie superalgebra g = g (A) is diagonalizable and defines a root decomposition g=h⊕ gα , ∆ ⊂ h∗ ; α∈∆

by linear independence of αi , every root space gα is either purely even or purely odd. Therefore one can define p : ∆ → Z2 by putting p (α) = 0 or 1 whenever gα is even or odd, respectively. By ∆0 and ∆1 we denote the set of even and odd roots, respectively. The relations (1) imply that every root is a purely positive or purely negative integer linear combination i∈I mi αi . According to this, we call a root positive or negative, and have the decomposition ∆ = ∆+ ∪ ∆− . The triangular decomposition is by definition the decomposition g = n+ ⊕h⊕n− , where n± are subalgebras generated by Xi , i ∈ I, respectively, Yi , i ∈ I. The roots αi , i ∈ I, are called simple roots. Sometimes instead of aij we write aαβ with α = αi , β = αj ; likewise for Xα , Yα , hα instead of Xi , Yi , hi . Remark 2.1. Obviously, given a simple root α = αi , one of the following possibilities holds: 1. if aαα = 0 and p (α) = 0, then Xα , Yα and hα generate a Heisenberg subalgebra; 2. if aαα = 0 and p (α) = 1, then [Xα , Xα ] = [Yα , Yα ] = 0 and Xα , Yα , and hα generate a subalgebra isomorphic to sl (1|1); in this case Xα2 = Yα2 = 0 in any representation of g (A); 3. if aαα = 2 and p (α) = 0, then Xα , Yα , and hα generate a subalgebra isomorphic to sl (2); 4. if aαα = 2 and p (α) = 1, then [Xα , Xα ] = [Yα , Yα ] = 0, and Xα , Yα , and hα generate a (3|2)-dimensional subalgebra isomorphic to osp (1|2); in this case 2α ∈ ∆.


173

Consider a simple root α. We call α isotropic iff aαα = 0, otherwise α is called non-isotropic; α is called regular if for any other simple root β, aαβ = 0 implies aβα = 0, otherwise a simple root is called singular. A superalgebra g = g (A) has a natural Z-grading with 0-component being h, 1-component being i∈I gαi . We say that g is of finite growth if the dimension of n-component grows not faster than a polynomial in n. We say that A is admissible if adXi is locally nilpotent in g (A) for all i ∈ I. In this case adYi is also locally nilpotent. One can check (see, for instance, [9]) that A is admissible if and only if, after normalization, A satisfies the following conditions: 1. If aii = 0 and p (i) = 0, then aij = 0 for any j ∈ I; 2. If aii = 2 and p (i) = 0, then aij ∈ Z≤0 for any j ∈ I, i = j and aij = 0 ⇒ aji = 0; 3. If aii = 2 and p (i) = 1, then aij ∈ 2Z≤0 for any j ∈ I, i = j and aij = 0 ⇒ aji = 0; Note that if aii = 0 and p (i) = 1, there is no condition on the entries aij . We call g (A) quasisimple if for any ideal i ⊂ g (A) either i ⊂ h, or i + h = g (A). This is equivalent to saying that every ideal of g (A) is either in the center of g (A), or contains the commutator [g (A) , g (A)]. Recall that for usual Kac–Moody Lie algebras, g (A) is quasisimple iff A is indecomposable; g (A) is simple if, in addition, A is nondegenerate. For finite-dimensional Lie algebras quasisimplicity is equivalent to simplicity, but in general this is not true. For example, affine Lie algebras are quasisimple but not simple. In supercase even finite-dimensional contragredient superalgebras can be quasisimple but not simple (for example, gl (n|n), see Section 6). The following theorem is proven in [9]. Theorem 2.2. (Hoyt, Serganova) If g (A) is quasisimple and has finite growth, then A is admissible. One can describe an arbitrary contragredient Lie superalgebra as a sequence of extensions of quasisimple contragredient superalgebras and Heisenberg superalgebras (see [9] for details). Lemma 2.3. If g (A) is quasisimple, then A is indecomposable and does not have zero rows. In particular, for admissible A, aii = 0 implies p (i) = 1, so all simple isotropic roots are odd. Proof. If after suitable permutation of indices A = B ⊕ C is the sum of two nontrivial blocks, then g (A) is a direct sum of ideals g (B) ⊕ g (C). Hence A must be indecomposable. If A has a zero row, i.e., aij = 0 for some fixed i ∈ I and all j ∈ I, then it is not hard to see that [Xi , b] + Chi is an ideal in g (A), here b = h ⊕ n+ . Hence A cannot have zero rows. We call A regular if for any i, j ∈ I, aij = 0 implies aji = 0. The following is a straightforward generalization of Proposition 1.7 in [2] to the supercase.

174

Vera Serganova

Lemma 2.4. Let A be regular. Then g (A) is quasisimple iff A is indecomposable and not zero. Denote by Q the lattice generated by ∆, and by Q0 the lattice generated by even roots ∆0 . Theorem 2.5. If g (A) is quasisimple and Q0 = Q, then either Q/Q0 ∼ = Z or Q/Q0 ∼ = Z2 . In the first case, g (A) has a Z-grading g (A) = g−1 ⊕ g0 ⊕ g1 such that g0 is the even part of g (A). Proof. Consider the Q/Q0 grading g (A) = i∈Q/Q0 gi , where gi = s(α)=i gα , s : Q → Q/Q0 being the natural projection. Obviously, g0 = g¯0 , and gi ⊂ g¯1 for any i = 0. Choose i ∈ Q/Q0 , i = 0. Let l = gi ⊕ [gi , g−i ] ⊕ g−i . We claim that l is an ideal of g. Indeed, [gi , gi ] = [g−i , g−i ] vanish if 2i = 0, and coincide with [gi , g−i ] if 2i = 0. For j ∈ / {−i, 0, i}, [gj , g±i ] = 0, thus [gj , [gi , g−i ]] = 0 by the Jacobi identity; [g0 , [gi , g−i ]] ⊂ [gi , g−i ], again by the Jacobi identity. Now quasisimplicity of g (A) implies l + h = g (A). Therefore g (A) = gi ⊕ g0 ⊕ g−i ; hence i is a generator of Q/Q0 . If 2i = 0, then Q/Q0 ∼ = Z2 . If 2i = 0, then g (A) has Z-grading with gi being degree 1. Hence Q/Q0 must be Z.

Following [1], we call g (A) of type I, if Q/Q0 ∼ = Z, and of type II if Q/Q0 ∼ = Z2 . In the case when g (A) is finite dimensional and quasisimple, g (A) is of type I if it is isomorphic to sl (m|n) , m = n, gl (n|n), or osp (2|2n), and of type II if it is isomorphic to osp (m|2n) with m = 2, G3 or F4 . For more examples, see Corollary 7.4. Remark 2.6. It is useful to note that if g (A) is of type I, then all its odd simple roots are isotropic.

3 Odd reflections, Weyl groupoid, and principal roots A linearly independent set Σ of roots of a Lie superalgebra g (A) is a base if one can find Xβ ∈ gβ and Yβ ∈ g−β for each β ∈ Σ such that Xβ , Yβ , β ∈ Σ, and h generate g (A), and for any β, γ ∈ Σ, β = γ [Xβ , Yγ ] = 0.

(2)

If we put hβ = [Xβ , Yβ ], then Xβ , Yβ and hβ satisfy the relations [h, Xβ ] = β (h) Xβ ,

[h, Yβ ] = −β (h) Yβ ,

[Xβ , Yγ ] = δβγ hβ .

(3)

¯ (AΣ ) → g (A), here AΣ = This induces a natural surjective mapping g (β (hγ )), β, γ ∈ Σ; since g (AΣ ) and g (A) have the same Cartan subalgebra h, the kernel must coincide with Ker ¯ g (AΣ ) → g (AΣ ). Therefore, g (A) is isomorphic to g (AΣ ). The matrix AΣ is called the Cartan matrix of a base Σ.


175

The original set Π = {αi } is called a standard base. If g is a Kac–Moody Lie algebra, then every base can be obtained from the standard one by the Weyl group action (and, maybe, for infinite roots systems, by multiplication by −1). It is crucial for our discussion that this is not true for superalgebras. Let Σ be a base, α ∈ Σ, aαα = 0 and p (α) = 1. For any β ∈ Σ, define rα (β) by rα (α) = −α, if α = β and aαβ = aβα = 0, rα (β) = β rα (β) = β + α

if aαβ = 0 or aβα = 0.

Lemma 3.1. The set rα (Σ) = {rα (β) | β ∈ Σ} is linearly independent. Moreover, if α is regular, then rα (Σ) is a base. Proof. The linear independence of rα (Σ) is obvious. To prove the second statement, set X−α = Yα ,

Yrα (β) = [Yα , Yβ ] (4) for any β ∈ Σ such that r (β) = β + α. First, we have to check that α Xrα (β) , Yrα (γ) = 0 if β, γ ∈ Σ and β = γ. If β, γ = α, then rα (β) − rα (γ) = β − γ or β − γ ± α ∈ / ∆, hence Xrα (β) , Yrα (γ) = 0. So assume that β = α. A simple calculation shows that if aαγ = aγα = 0, then [Yα , Yγ ] = [Xα , Xγ ] = 0. Therefore if rα (γ) = γ, then [X−α , Yγ ] = [Yα , Yγ ] = 0. If rα (γ) = γ + α, then

Y−α = Xα ,

Xrα (β) = [Xα , Xβ ] ,

1 X−α , Yrα (γ) = [Yα , [Yα , Yγ ]] = [[Yα , Yα ] , Yγ ] = 0, 2

since [Yα , Yα ] = 0. Similarly, one can deal with the case α = γ. Now we have to check that h, Xrα (β) and Yrα (β) generate g (A). Note that if rα (β) = α + β, then 1 1 [Yα , [Xα , Xβ ]] = X−α , Xrα (β) . β (hα ) β (hα ) Similarly, Yβ = β(h1 α ) Y−α , Yrα (β) . Hence every generator Xβ , Yβ , β ∈ Σ, can be expressed in terms of Xrα (β) , Yrα (β) . Xβ =

Remark 3.2. It is useful to write down the generators hrα (β) of the reflected base rα (Σ): hrα (β) = hβ if aαβ = aβα = 0, hrα (β) = aαβ hβ + aβα hα

if aαβ = 0 or aβα = 0.

If α is singular, then rα (Σ) generates a subalgebra in g (A) (see Section 6, examples D (2, 1; 0), S (1, 2, ±1)). We see from Lemma 3.1 that, given a base Σ with the Cartan matrix AΣ and a regular isotropic odd root α ∈ Σ, one can construct another base Σ

176

Vera Serganova

with the Cartan matrix AΣ such that g (AΣ ) ∼ = g (AΣ ). Using (4), one can construct an isomorphism g (AΣ ) → g (AΣ ); denote it by the same symbol rα . We say that AΣ is obtained from AΣ and a base Σ is obtained from Σ by an odd reflection with respect to α. If Σ = rα (Σ) and Σ = rβ (Σ ) one can define a composition rβ rα : g (AΣ ) → g (AΣ ). It is clear from the definition of rα that r−α rα (Σ) = Σ. For any base Π denote by ∆± (Π) the set of positive (negative) roots with respect to the base Π. If α ∈ Π, then −α ∈ rα (Π) ⊂ ∆+ (rα (Π)). Lemma 3.3. Let Σ = rα (Π), where α ∈ Π is an odd isotropic regular root. Then ∆+ (Π) {α} = ∆+ (Σ) {−α}. Proof. Assume γ ∈ ∆+ (Π) {α}. Then γ = β∈Π mβ β for some nonnegative integers mβ ; moreover, mβ > 0 for at least one β = α. Note that rα (β) = β + nβ α for some integer nβ ; hence ⎛ ⎞ mβ rα (β) + ⎝mα − mβ n β ⎠ α γ= β=α

=

β=α

⎛ mβ rα (β) + ⎝

β=α

⎞

mβ nβ − mα ⎠ rα (α) .

β=α

At least one of mβ is positive, β = α, therefore γ ∈ ∆+ (Σ). Hence ∆+ (Π) {α} ⊂ ∆+ (Σ) {−α}, and the lemma follows by symmetry. Definition 3.4. g (A) is called a regular contragredient superalgebra if any A obtained from A by odd reflections is regular. Consider a category C, which has an object g (A) for every square matrix A. A C-morphism f : g (A) → g (A ) is by definition an isomorphism of superalgebras which maps a Cartan subalgebra of g (A) to the Cartan subalgebra of g (A ). A Weyl groupoid C (A) of a contragredient superalgebra g (A) is a connected component of C which contains g (A). Let g (A) be a regular superalgebra. With each base Σ of g (A) we associate a Dynkin graph ΓΣ whose vertices are elements of Σ, and α and β are connected iff aαβ = 0. If one denotes even, odd non-isotropic, and odd isotropic roots by white, black and gray circles, respectively, and one writes aαβ and aβα above and under the edge joining α and β, then one can reconstruct the Cartan matrix of AΣ from ΓΣ . Definition 3.5. Call an even root α ∈ ∆0 principal if there exists a base Π obtained from a standard base Π by odd reflections such that α or α/2 belongs to Π . Let B denote the set of all principal roots. For each β ∈ B choose Xβ ∈ gβ , Yβ ∈ g−β , set hβ = [Xβ , Yβ ], and define the matrix B = bαβ = β (hα ), α, β ∈ B.


177

Note that by Lemma 3.3 a positive even root remains positive for any base obtained from Π by odd reflections. Hence all principal roots are positive. Lemma 3.6. [Xα , Yβ ] = δαβ hβ for any α, β ∈ B. Proof. It is sufficient to check that if α = β, then α − β is not a root. Assume that γ = α − β is a root. Without loss of generality, one may assume that γ is positive. Then α = β + γ. Choose a base Π such that α or α/2 belongs to Π , then α can not be a sum of two positive roots. Generally speaking, B may be infinite. We define a contragredient Lie algebra with infinite Cartan matrix as an inductive limit of usual contragredient Lie algebras. Let g denote the Lie subalgebra of g¯0 (A) generated by Xβ , Yβ for all principal β. Lemma 3.7. There exist a subalgebra c in the center of g (B) and a homomorphism q : g → g (B)/c which maps Xβ to the corresponding generator of g (B). Im q = [g (B) , g (B)]/c. ¯ β , Y¯β be the generators of ¯ Proof. Let X g (B). Lemma 3.6 implies that there exists a surjective homomorphism s : [¯ g (B) , ¯g (B)] → g defined by ¯ β = Xβ , s Y¯β = Yβ . Since all principal roots are positive, one can s X find h in the Cartan subalgebra of g (A) such that β (h) > 0 for all β ∈ B. Hence g has an R-grading such that g0 = h ∩ g , [¯g (B) , ¯g (B)] has the similar grading and s is a homomorphism of R-graded Lie algebras. Then Ker s is a graded ideal, in particular, Ker s = m− ⊕m0 ⊕m+ , where m0 lies in the Cartan subalgebra of ¯ g (B), and m± are the positive (negative) graded components. + 0 Since simple roots are not weights of m , m is central and Y¯β , m+ ⊂ m+ ; ¯ β , m− ⊂ m− . Therefore, m− ⊕ m+ generates the ideal in ¯g (B) likewise X which intersects the Cartan subalgebra trivially. If p : ¯g (B) → g (B) denotes the natural projection, then m− ⊕ m+ ⊂ Ker p. Let c = p m0 , and p : ¯ g (B)/m0 → g (B)/c be the surjection induced by p. Then q : g → g (B)/c given by q (Xβ ) = p s−1 Xβ , q (Yβ ) = p s−1 Yβ is well defined. It is straightforward to check that q satisfies the conditions of the lemma. Remark 3.8. The subalgebra g is the best approximation to “the largest almost contragredient Lie subalgebra” of g¯0 we can construct explicitly (compare with Lemma 3.10). In many examples q is an embedding and we suspect that is so in general. It would be nice to prove it. For any base Σ consider a closed convex cone

C + (Σ) = aα α | a α ≥ 0 . α∈Σ

178

Vera Serganova

+ Denote by CΠ the intersection of C + (Σ) for all bases Σ obtained from Π by + odd reflections. By Lemma 3.3, CΠ contains all positive even roots.

Remark 3.9. Note that Lemma 3.1 used only the following relation of g (A) which is not a relation in ¯ g (A) : if aαγ = aγα = 0, then [Xα , Xγ ] = 0. Thus one could redo our discussion for any quotient of ¯g (A) in which these relations hold. Lemma 3.10. Assume that there are only finitely many Σ obtained from Π + by odd reflections. Then CΠ is the convex cone generated by principal roots. Proof. We prove the statement by induction on the number of simple roots n. + Note that CΠ is an intersection of simplicial cones, hence a closed convex set. + Let C + (B) be the cone generated by principal roots; obviously C + (B) ⊂ CΠ . + To prove inverse inclusion it suffices to show that the boundary of CΠ belongs + to C + (B). If γ is a point on the boundary of CΠ , then there is Σ such that γ lies on the boundary of C + (Σ); in other words, γ belongs to one of the faces of C + (Σ). Hence there exists α ∈ Σ such that γ = β∈Σ{α} mβ β for some nonnegative mβ . Let k be the contragredient subalgebra of g with simple roots Σ = Σ {α} (as proven in [9], the corresponding Xγ , Yγ , hγ generate a subalgebra k which is contragredient; alternatively, one could use Remark 3.9, and work with k directly). Since odd reflections of Σ are induced by those of + Σ, γ ∈ CΣ . By inductive assumption γ is a nonnegative linear combination of principal roots of k. Clearly, any principal root of k is a principal root of g. Thus γ ∈ C + (B). Remark 3.11. We suspect that Lemma 3.10 is true without any finiteness assumption. Denote by Q the sublattice of Q0 generated by all principal roots. Lemma 3.12. For any regular contragredient superalgebra g (A) with n simple roots and indecomposable A, the rank of Q is at least n − 1. Proof. Choose a base Π of g (A), let Γ be its graph. We say that a subgraph α1 , . . . , αk is a string if it is connected, the vertices α1 and αk are of valence 1 and odd, and all vertices α2 , . . . , αk−1 are of valence 2 and even. We claim that, for each string α1 , . . . , αk , either 2 (α1 + · · · + αs ) is principal for some s < k, or α1 + · · · + αk is principal. Indeed, note that either α1 is non-isotropic, in which case 2α1 is principal, or α1 is isotropic, in which case rα1 (α2 ) = α1 +α2 is an odd root. Again either α1 + α2 is non-isotropic, in which case 2 (α1 + α2 ) is principal, or α1 + α2 is isotropic, in which case rα1 +α2 (α3 ) = α1 + α2 + α3 is an odd root. Proceeding in this manner, one gets either a principal root 2 (α1 + · · · + αs ) for some s < k, or a principal root α1 + · · · + αk . Since all even simple roots are principal, one can see that for any odd roots α and β ∈ Π joined by a string, at least one of 2α and α + β belongs to Q . Consider all odd roots α such that 2α ∈ / Q . Since Γ is connected, the sum or difference of any two of them is in Q . Therefore Q generates a subspace of dimension at least n − 1. Hence the rank of Q is at least n − 1.


179

4 Kac–Moody superalgebras In this section we, finally, introduce the main character of our story. Definition 4.1. g (A) is a Kac–Moody Lie superalgebra if any A obtained from A by odd reflections is admissible. Lemma 4.2. Let g (A) be a regular Kac–Moody superalgebra. Then g (A) is a direct sum of quasisimple regular Kac–Moody superalgebras and even or odd Heisenberg superalgebras. Proof. Follows from Lemma 2.4. Remark 4.3. Note that the theory of highest-weight representations of Heisenberg (super)algebras can be formulated in a very similar way to what we do in this paper. For “typical” representations, the Weyl character formula (22) holds. For “atypical” representations, the character can be easily described. Therefore, although this paper concentrates on regular quasisimple Kac– Moody superalgebras, one could easily rewrite the theory for all regular Kac– Moody superalgebras. We start with defining the Weyl group. Let g be a quasisimple Kac–Moody superalgebra. Then an even simple root is not isotropic. Therefore, for any principal root α, there exists an sl (2)-triple {Xα ∈ gα , hα ∈ h, Yα ∈ g−α }. Define an even reflection rα : h∗ → h∗ by rα (µ) = µ − µ (hα ) α. Since the sl (2)–subalgebra spanned by {Xα , hα , Yα } acts locally finitely on g (A), rα is dual to the restriction of the automorphism exp adXα exp ad−Yα exp adXα to h. Therefore rα permutes the roots and maps a base to a base. If β is an odd root such that 2β is principal, then we put rβ = rα . The group W generated by all even reflections is called the Weyl group of g (A). The main difference between even and odd reflections is that an even reflection depends only on a root and does not depend on a base to which the root belongs. The following straightforward identity is very important: wrα (Π) = rw(α) w (Π)

(5)

for a base Π, a regular isotropic α ∈ Π, and w ∈ W . Note that one can define exp adXα exp ad−Yα exp adXα for any even α such that adXα and adYα are locally nilpotent. We do not know any example where α is not in the W -orbit of a simple root (in which case the corresponding automorphism of h∗ is, obviously, in W ).

180

Vera Serganova

Lemma 4.4. Let Σ = rα (Π), where α ∈ Π is any root (even or odd). Then ∆+ (Π) {α, 2α} = ∆+ (Σ) {−α, −2α}. Proof. The same as for Lemma 3.3. Corollary 4.5. Let ∆ be finite, Π and Π be bases. Then Π can be obtained from Π by even and odd reflections. Proof. Assume Π = Π . There exists α ∈ Π ∩ ∆− (Π ). Then |∆+ (rα (Π)) ∩ ∆− (Π ) | < |∆+ (Π) ∩ ∆− (Π ) |. Repeating this process, one can get Π such that ∆+ (Π ) ∩ ∆− (Π ) = ∅. But then Π = Π. Recall notations from Definition 3.5. Lemma 4.6. If g is a quasisimple Kac–Moody Lie algebra, then B is a Cartan matrix of a Kac–Moody Lie algebra. Proof. By quasisimplicity of g, α (hα ) = 0 for any α ∈ B. We normalize hα so that α (hα ) = 2. First, we have to show that α (hβ ) ∈ Z≤0 for any α, β ∈ B, α = β. By definition of a Kac–Moody superalgebra, the adjoint action of sl2 subalgebra spanned by Xβ ∈ gβ , hβ ∈ h, Yβ ∈ g−β on g is locally finite, and Xα is a lowest vector, hence α (hβ ) ∈ Z≤0 . Finally, α (hβ ) = 0 implies [Xβ , Xα ] = 0, therefore β (hα ) = 0. Remark 4.7. Recall that Serre’s relations in a Kac–Moody Lie algebra g (B) are the relations 1−bβγ 1−bβγ adXβ Xγ = 0, adYβ Yγ = 0. If B is symmetrizable, then Serre’s relations together with the “contragredient” relations [h, Xβ ] = β (h) Xβ ,

[h, Yβ ] = β (h) Yβ ,

[Xβ , Yγ ] = δβγ hβ ,

[h, h] = 0

define the Kac–Moody Lie algebra g (B). Since adXβ , adYβ , β ∈ B, are locally nilpotent operators in g, Serre’s relations hold in g . Thus, if B is symmetrizable, the homomorphism q constructed in Lemma 3.7 is injective. Remark 4.8. Later, in Section 9, we will see that for all quasisimple regular Kac–Moody superalgebras the set B of principal roots is finite. It would be interesting to obtain a proof avoiding the use of the classification. Remark 4.9. As the preceding remark shows, the group W is manifestly finitely generated. Later the fact that we restrict our attention to (a finite set of) principal roots leads to an effective criterion of integrability of a weight (see Theorem 10.5). On the other hand, the group W is also “sufficiently large”. Indeed, there are the following indications:


181

1. It induces the “even” Weyl group when reduced to the “even contragredient part” g (B) of g (Corollary 4.10); 2. When the character formula for integrable infinite-dimensional highest weight irreducible representations is known for weights in general position, it coincides with the Weyl character formula (Corollary 14.5). Corollary 4.10. Let g be a quasisimple Kac–Moody superalgebra. Then the Weyl group W is isomorphic to the Weyl group of g (B). Define real and imaginary roots following [2]. A root α is real iff there exists a base Π obtained from Π by even and odd reflections such that one of α or α/2 belongs to Π . By definition, a principal root is real. If α is not real, then we call it imaginary. Lemma 4.11. (a) The set of real roots is W-invariant. (b) If α is real, then gα is one-dimensional. (c) If α is a real odd isotropic root, then mα ∈ ∆ implies m = ±1. (d) If α is a real odd non-isotropic root, then mα ∈ ∆ implies m = ±1, ±2. (e) If α is an even real root, then mα ∈ ∆ implies m = ±1, ±1/2. Proof. W -invariance follows from the fact that if Π is a base, then w (Π) is a base. The other statements easily follow from the fact that α belongs to some base Π and Remark 2.1. + Denote by DΠ the intersection of C + (Σ) for all Σ obtained from Π by even and odd reflections. It follows from (5) that + + DΠ . = w CΠ w∈W + . Corollary 4.12. If α is an imaginary root, then α or −α belongs to DΠ + Proof. Assume that α is a positive root, and α ∈ / DΠ . By definition α ∈ + + C (Π), but since α ∈ / DΠ there exists some base Π obtained from Π by even and odd reflection such that α ∈ / C + (Π ). Write Π = r1 . . . rs (Π), and + choose i such that α ∈ C (ri+1 . . . rs (Π)) and α ∈ / C + (ri . . . rs (Π)). Then Lemma 4.4 implies ri = rα , hence α or α/2 ∈ ri+1 . . . rs (Π). Thus α is real. Similarly for α being a negative root.

Corollary 4.13. If ∆ is finite, then every root is real; therefore g is finitedimensional. Proof. By Corollary 4.5, every base can be obtained from a standard base Π by even and odd reflections. In particular, −Π can be obtained from Π in this + way. Therefore DΠ = {0}. Now the statement follows from Corollary 4.12. Theorem 4.14. Let g (A) and g (A ) be regular quasisimple Kac–Moody superalgebras which belong to the same connected component of C. Then A is obtained from A by a composition of odd reflections, permutation of indices, and multiplication by an invertible diagonal matrix. This theorem will be proven in Section 8. (See Remark 8.4.)

182

Vera Serganova

5 Regular Kac–Moody superalgebras with two simple roots It is easy to classify rank 2 regular quasisimple Kac–Moody superalgebras. We assume that a superalgebra g (A) has at least one simple isotropic root; otherwise, by definition, there are no restrictions on a matrix A (except the admissibility conditions). Theorem 5.1. ([9]) Let g (A) be a regular quasisimple Kac–Moody superalgebra with 2 simple roots such that at least one root is isotropic. Then g is isomorphic to one of the following: 0 1 2 −1 ; with p (1) = 0, or 1. g ∼ sl (1|2), A = = 1 0 1 0 2 −2 with p (1) = 0, or p (1) = 1. 2. g ∼ = osp (3|2), A = 1 0 Proof. If g has a base with two roots, then, without loss of generality, isotropic 01 (since multiplication by a diagonal matrix one may assume that A = 10 does not change the algebra). The reflection with respect to α2 gives α + 1 = α1 2 −1 . α2 , α2 = −α2 , h1 = h1 + h2 , h2 = h2 (see Remark 3.2). Hence A = 1 0 Now assume that g has a base with one non-isotropic odd root α1 and one isotropic odd root α2 . One may assume without loss of generality that 2 2b for some negative integer b. Again by Remark 3.2, the reflection A= 1 0 1 2b rα2 gives α1 = α1 +α2 , α2 = −α2 , h1 = 1+2b h1 + 1+2b h2 (after normalization), −2b h2 = h2 . Since α2 (h1 ) ∈ Z≤0 , we obtain 1+2b ∈ Z≤0 , which is possible only if b = −1. Finally assume that g has a base with one even root and one isotropic root. In this case the odd reflection will move the base to one of the above cases. Hence theorem is proven.

The following hereditary principle is obvious but very important. If A is a Cartan matrix of some regular Kac–Moody superalgebra, then any main minor of A is also a Cartan matrix of some regular Kac–Moody superalgebra. Corollary 5.2. Let A be a Cartan matrix of a quasisimple regular Kac–Moody superalgebra. If α is an even simple root and β is an isotropic root, then aαβ = 0, −1 or −2. If α is an odd non-isotropic root and β is an isotropic root, then aαβ = 0 or −2.

6 Examples of regular quasisimple superalgebras Finite-dimensional superalgebras. Finite-dimensional quasisimple Kac–Moody superalgebras were classified in [1]. In most cases, they have nondegenerate


183

Cartan matrices. However the superalgebra gl (n|n) is quasisimple, but not simple, and its Cartan matrix has corank 1. (To get the simple superalgebra psl (n|n) one should factor the commutator sl (n|n) = [gl (n|n) , gl (n|n)] by the center consisting of scalar matrices.) For example, the superalgebra gl (2|2) has the following Cartan matrices: ⎞ ⎞ ⎛ ⎞ ⎛ ⎛ 0 1 0 0 1 0 2 −1 0 ⎝−1 2 −1⎠ ⎝1 0 −1⎠ , ⎝1 0 −1⎠ , 0 1 0 0 1 0 0 −1 2 Another unusual example in this class is a family D (2, 1; a) depending on the parameter a. If we start with the matrix ⎞ ⎛ 2 −1 0 ⎝ 1 0 a⎠ 0 −1 2 with a = 0, −1, then reflection rα2 transforms it to the matrix ⎛ ⎞ 0 1 −1 − a ⎝ 1 0 a ⎠, −1 − a1 1 0 and the odd reflections of the latter matrix give the matrices of D (2, 1; −1 − a) and D 2, 1; −1 − a1 (after suitable permutations of indices). Thus, the group S3 generated by a → a1 , and a → −1 − a acts on the space of parameters, and the points on the same orbit of this action correspond to isomorphic superalgebras. Hence one can describe the moduli space of such algebras as CP1 {0, −1, ∞} modulo the above S3 -action. The cases a = 0, −1, ∞ correspond to a non-regular Kac–Moody superalgebra. For instance, if a = 0, then the singular odd reflection rα2 maps the generators into an ideal isomorphic to psl (2|2). It is not hard to show that D (2, 1; 0) is isomorphic to the algebra of ∼ all derivations of psl (2|2), and D (2, 1; 0)/ psl (2|2) = sl (2). Another 3-element 1 S3 -orbit a = 1, −2, − 2 corresponds to the algebra osp (4|2). All Cartan matrices of regular finite-dimensional quasisimple Kac–Moody superalgebras are symmetrizable, hence they have non-degenerate invariant bilinear symmetric forms. Furthermore, ∆ is a finite set, hence any root is real. Principal roots are the simple roots of g = [g¯0 , g¯0 ]. Affine Kac–Moody superalgebras. Let s be a finite-dimensional simple Lie superalgebra from the previous class (so s = gl (n|n)), (·, ·) be a nondegenerate invariant symmetric form on s. Define an infinite-dimensional superalgebra s(1) as s ⊗ C t, t−1 ⊕ CD ⊕ CK, here D, K are even elements and the bracket is defined by X ⊗ tk , Y ⊗ tl = [X, Y ] ⊗ tk+l + kδk,−l (X, Y ) K, [D, K] = 0, D, X ⊗ tk = kX ⊗ tk .

(6)

184

Vera Serganova

It is not difficult to show (as in the Lie algebra case) that s(1) is a regular quasisimple Kac–Moody superalgebra. To construct the set of generators choose some generators X1 , . . . , Xn , Y1 , . . . , Yn of s, and add X0 = x ⊗ t, Y0 = y ⊗ t−1 , where x is a lowest weight vector, and y is a highest weight vector in the adjoint module. Quasisimplicity follows from simplicity of s. Regularity follows from existence of an invariant symmetric form defined by X ⊗ tk , Y ⊗ tl = (X, Y ) δk,−l , (D, K) = 1, (7) k k K, X ⊗ t = D, X ⊗ t = (D, D) = (K, K) = 0. Indeed, existence of the form implies that every Cartan matrix is symmetrizable; therefore, there are no singular roots. Now let us define the affinization of gl (n|n). Let s = psl (n|n) be the quotient of sl (n|n) by the center, for simplicity of s we need n ≥ 2. Then s(1) is defined as s ⊗ C t, t−1 ⊕ CD ⊕ CK ⊕ CD ⊕ CK , the bracket with additional even elements D and K is given by p(X) D , X ⊗ tk = 1 − (−1) X ⊗ tk , K , X ⊗ tk = 0, X ⊗ tk , Y ⊗ tl = [X, Y ] ⊗ tk+l + kδk,−l (X, Y ) K + δk,−l tr [X, Y ] K ; here tr is the usual trace (not the supertrace), which is not zero only if both (1) X and Y are odd. The Cartan matrix of affine superalgebra psl (n|n) has corank 2, therefore psl (n|n)(1) is a regular quasisimple Kac–Moody superalgebra with two dimensional center. For example, one of Cartan matrices of (1) psl (2|2) is ⎞ ⎛ 0 1 0 −1 ⎜1 0 −1 0⎟ ⎟. ⎜ ⎝0 1 0 −1⎠ 1 0 −1 0 What follows is applicable for any finite-dimensional simple s. To describe the roots of s(1) , define δ ∈ h∗ by conditions δ (hi ) = 0, δ (D) = 1. Denote by ∆◦ the roots of s. Then the roots of g are of the form α+kδ with k ∈ Z, α ∈ ∆◦ , or kδ with k ∈ Z {0}. The standard base Π = {α0 , . . . , αn } is obtained from the standard base {α1 , . . . , αn } of s by adding the root α0 = θ + δ, where θ is the lowest root of s. Twisted affine superalgebras. Choose an automorphism φ of s of finite order p, which preserves the invariant form on s. Let ε be a p-th primitive root of 1. One can extend the action of φ on s(1) by φ X ⊗ tk = εk φ (X) ⊗ tk , φ (D) = D, φ (K) = K. Then sφ is defined as the set of elements fixed by φ. It was proven in [5] that the construction does not depend on ε, and if two automorphisms are


185

in the same connected component of Aut s, then the corresponding algebras are isomorphic. Up to isomorphism, there are the following twisted affine (2) (4) superalgebras: (p) sl (m|n) if mn is even, (m, n) = (2, 2); (p) sl (m|n) if (2) mn is odd; osp (2m|2n) ; the upper index denotes the minimal possible order of an automorphism. In the case of twisting of psl (n|n) the corank of a Cartan matrix is 1, and the center is one-dimensional. For example, ⎞ ⎛ 2 −2 0 ⎝−1 0 1⎠ 0 −2 2 with parity functions (0,1,0) or (1,1,1) is a Cartan matrix of psl (3|3)(4) . All twisted affine superalgebras are regular quasisimple Kac–Moody superalgebras. A base in a twisted affine superalgebra can be obtained by the following procedure. Define a Zp -grading s = s0 ⊕ s1 ⊕ · · · ⊕ sp−1 by sk = s ∈ s | φ (s) = εk s . In all cases φ can be chosen so that s0 is a simple finite-dimensional superalgebra, s1 is an irreducible s0 -module and s0 + s1 generates s. For example, for (p) sl (m|2n)(2) , s0 = osp (m|2n), s1 is a unique nontrivial irreducible subquotient of the symmetric square of the standard (m|2n)-dimensional s0 -module; (4) for (p) sl (2m + 1|2n + 1) , s0 = osp (2m + 1|2n), and s1 is the standard (2) module with inverted parity; for osp (2m|2n) , s0 = osp (2m − 1|2n), s1 is the standard module. Let ∆0 be the set of roots of s0 and ∆j denote the set of weights of sj with respect to a Cartan subalgebra of s0 . The roots of s(p) are of the form α + kδ, with α ∈ ∆j , k ∈ j + pZ, and kδ with k ∈ pZ {0}. The standard base Π = {α0 , . . . , αn } can be obtained from a base {α1 , . . . , αn } of s by adding the weight α0 = θ + δ, where θ is the lowest weight of s1 . Finally, let us mention that the condition that φ preserves the invariant form on For example, an automorphism π of psl (n|n) such s is nontrivial. DC AB does not preserve the invariant form on psl (n|n), = that π BA CD and psl (n|n)π is not contragredient. It also explains why the Lie superalgebra psl (2|2) does not have a twisted affinization. The subgroup of automorphisms of psl (2|2) preserving the invariant form is connected due to existence of nontrivial derivations, see above example for D (2, 1; 0). A strange twisted affine superalgebra. The Lie superalgebra s = q (n) ⊂ psl (n|n) is the subalgebra of all elements fixed by the automorphism π defined above. The superalgebra q (n) is simple for n ≥ 3. The involution φ such that φ (x) = (−1)p(x) x does not belong to the connected component of unity (1) in Aut q (n). Although the Lie superalgebras q (n) and q (n) are not contragredient, twisting by φ gives a regular quasisimple Kac–Moody superalgebra

186

Vera Serganova (2)

(2)

which we denote by q (n) . As a vector space q (n) is isomorphic to s¯0 ⊗ C t2 , t−2 ⊕ s¯1 ⊗ tC t2 , t−2 ⊕ CK ⊕ CD ∂ with D = t ∂t and K being a central element. For any x, y ∈ s, the commutator is defined by the formula

[x ⊗ tm , y ⊗ tn ] = [x, y] ⊗ tm+n + δm,−n (1 − (−1)m ) tr (xy) K. A Cartan matrix of q (n)(2) has size n × n. Identify the set of indices with the abelian group Zn . Fix a parity function p : Zn → {0, 1} so that the number of odd indices is odd. Set aij = 0 if j = i or i ± 1 (modulo n); set aii = 2, aii = 0,

ai,i±1 = −1 if p (i) = 0; ai,i+1 = −1, ai,i−1 = 1 if p (i) = 1.

The graph of this Cartan matrix is a cycle with odd number of isotropic vertices. Any two such Cartan matrices are related by a chain of odd reflections; it is a simple exercise to check that they are not symmetrizable. The corank of a Cartan matrix is 1. (2) Roots of g = q (n) are of the form α + mδ, where α is a root of sl (n) and m ∈ Z or mδ with m ∈ Z {0}. The parity of a root equals the parity of m. Non-symmetrizable superalgebra S (1, 2; b). This superalgebra appears in the list of conformal superalgebras classified by Kac and van de Leur [7]. By definition, it is the Kac–Moody superalgebra with Cartan matrix ⎞ ⎛ 0 b 1−b ⎝ −b 0 1 + b⎠ −1 −1 2 and parity (1,1,0). Obviously, S (1, 2; b) ∼ = S (1, 2; −b). If b = 0, then the matrix is a matrix of gl (2|2); if b = −1, 1, the matrix has singular roots. In all other cases the odd reflection rα2 transforms the matrix of S (1, 2; b) to the matrix of S (1, 2; 1 + b). Hence an isomorphism S (1, 2; b) ∼ = S (1, 2; 1 + b). 1 The moduli space of S (1, 2; b) can be identified with CP /G where G is the 11 −1 0 . A Cartan matrix and subgroup of PGL (2, Z) generated by 01 0 1 of S (1, 2; b) has corank 1, hence the algebra has one-dimensional center. The Lie superalgebra S (1, 2; b) is a quasisimple Kac–Moody superalgebra. It is regular if b ∈ / Z. Realization of S (1, 2; b). Consider a supercommutative algebra R = C t, t−1 , ξ1 , ξ2 with an even generator t and two odd generators ξ1 , ξ2 . Denote by W (1, 2) the Lie superalgebra of derivations of R, in other words W (1, 2) is the superspace of all linear maps d : R → R such that d (f g) = d (f ) g + (−1)p(d)p(f ) f d (g) .


187

An element d ∈ W (1, 2) can be written as d=f

∂ ∂ ∂ + f1 + f2 ∂t ∂ξ1 ∂ξ2

for some f, f1 , f2 ∈ R. It is easy to see that the subset of all d ∈ W (1, 2) satisfying the condition ∂f1 ∂f ∂f2 p(d) − (−1) bf t−1 + ≡ const, + ∂t ∂ξ1 ∂ξ2 form a subalgebra S¯b of W (1, 2). Let E = ξ1 ∂ξ∂ 1 + ξ2 ∂ξ∂2 . Then adE is a diagonalizable operator in S¯b which defines the grading ⊕ S¯b ⊕ S¯b ⊕ S¯b , S¯b = S¯b

such that S¯b

2

−1

0

1

2

∂ = 0 for b ∈ / Z, S¯b 2 = Cξ1 ξ2 t−b ∂t for b ∈ Z. Define Sb = S¯b −1 ⊕ S¯b 0 ⊕ S¯b 1 .

One can see that Sb is an ideal in S¯b when b ∈ Z. The commutator [Sb , Sb ] is simple and has codimension 1 in Sb . It consists of all derivations d ∈ Sb satisfying the condition ∂f1 ∂f ∂f2 p(d) −1 − (−1) = 0. + bf t + ∂t ∂ξ1 ∂ξ2 Set X1 =

∂ ∂ξ1 ,

∂ Y1 = (b + 1) ξ1 ξ2 ∂ξ∂2 + ξ1 t ∂t , ∂ h1 = (b + 1) ξ2 ∂ξ∂ 2 + t ∂t ,

∂ X2 = −bξ1 ξ2 t−1 ∂ξ∂ 1 + ξ2 ∂t ,

X3 = ξ1 ∂ξ∂ 2 ,

Y2 = t ∂ξ∂ 2 ,

Y3 = ξ2 ∂ξ∂1 ,

∂ h2 = bξ1 ∂ξ∂1 + ξ2 ∂ξ∂ 2 + t ∂t , h3 = ξ1 ∂ξ∂ 1 − ξ2 ∂ξ∂ 2 .

Then Xi , Yi , hi , i = 1, 2, 3, generate [Sb , Sb ] if b = 0. They satisfy the relations (1) with Cartan matrix S (1, 2; b). The contragredient Lie superalgebra S (1, 2; b) can be obtained from Sb by a suitable central extension. Divide the fist row of the Cartan matrix S (1, 2; b) by b and the second by −b, let a = 1b . Then the renormalized matrix ⎞ ⎛ 0 1 −1 − a ⎝ 1 0 −1 + a⎠ −1 −1 2 (1)

clearly is a deformation of a Cartan matrix of the affine superalgebra sl (1|2) . (1) Thus, S (1, 2, ∞) ∼ = sl (1|2) . The roots and the root multiplicities of S (1, 2; b) are the same as the roots and the root multiplicities of sl (1|2)(1) .

188

Vera Serganova

The family Q± (l, m, n). This family was discovered by C. Hoyt (see [3]). To define it let us classify 3 × 3 matrices with zeros on the diagonal and nonzeros anywhere else, which define regular quasisimple Kac–Moody superalgebras. Using multiplication by a diagonal matrix, any such matrix can be reduced to the form ⎞ ⎛ 0 a 1 ⎝1 0 b ⎠ . c 1 0 The odd reflection rα1 transforms it to the matrix (parity= (1, 0, 0)) ⎛ ⎞ 0 a 1 ⎝−1 2 1 + b + a1 ⎠ . 1 −1 1 + a + c 2 The latter matrix must be admissible, therefore a+ 1c , b+ a1 ∈ Z 0 and δ (h ) > 0. We claim that ∆+ (Π) ∩ ∆− (Π ) is finite. Indeed, let C be the maximum of {|s (h) |, |s (h ) | | s ∈ S}. Let α ∈ ∆+ (Π) ∩ ∆− (Π ). Since α = mδ + s for some s ∈ S, the conditions α (h) > 0, α (h ) < 0 imply |m| ≤ C. Thus there are only finitely many α in ∆+ (Π) ∩ ∆− (Π ). If δ (h) > 0, δ (h ) < 0, then ∆+ (Π) ∩ ∆− (Π ) is finite by a similar argument. Now let g satisfy (a) or (d). Then the matrix B is indecomposable, and the conditions of Lemma 3.10 hold (in the first case B is obtained from A by dividing all odd rows by 2 and multiplying all odd columns by 2, in the second case it follows from the direct computation done in Section 6). Moreover,


191

+ Lemma 3.10 implies that DΠ = w∈W w (C + (B)). It was shown in [2], Lemma 5.8, that the set of rays spanned by positive imaginary roots of g (B) is dense + in DΠ in any metric topology. We choose this metric so that the roots of Π form an orthonormal basis in Q. + + First, we claim that for any base Π , either DΠ , or −DΠ is contained in + C (Π ). Indeed, if this is not true, there exists h ∈ h such that α (h) ≥ 1 for + all α ∈ Π and γ (h) = 0 for some γ ∈ DΠ . Since the set of rays spanned by + positive imaginary roots of g (B) is dense in DΠ , one can choose a sequence γ . But each βi of positive imaginary roots of g (B) such that |ββii | approaches |γ| βi is a root of g, hence βi = α∈Π miα α where all miα are nonnegative (or i (h) i (h) nonpositive). Therefore β|β ≥ 1, or β|β ≤ −1, and we obtain contradiction i| i| with the assumption γ (h) = 0. + Without loss of generality, one may assume that DΠ ⊂ C + (Π ) (if necessary one can use −Π instead of Π ). Let us prove that in this case ∆+ (Π)∩∆− (Π ) is finite. Assume the opposite. Then one can find an infinite sequence of roots α1 , . . . , αn , · · · ∈ ∆+ (Π) ∩ ∆− (Π ). Let γ be a limit point αi of the sequence |α . By Lemma 4.4, after any chain of reflections r1 . . . rk ini| + . On the other finitely many of αi remain in ∆+ (r1 . . . rk (Π)). Hence γ ∈ DΠ + hand, γ ∈ −C (Π ). Contradiction. Remark 8.4. Theorem 8.3 implies Theorem 4.14.

9 Description of g (B) and g in examples Let us recall (see [1]) that if s is a simple classical Lie superalgebra with a nondegenerate even invariant symmetric form, then s 0 is either semisimple, or reductive with one-dimensional center. Every such s is contragredient. Now let us consider the case of (twisted) affine superalgebras. Theorem 9.1. Let s be a simple classical Lie superalgebra with nondegenerate even invariant symmetric form (·, ·), g = s(1) , [s¯0 , s¯0 ] = s1 ⊕ · · · ⊕ sk be the sum of simple ideals s1 , . . . , sk , q : g → g (B)/c be as defined in Lemma 3.7. Then (1) (1) (a) g (B) = s1 ⊕ · · · ⊕ sk ; (b) dim c = k − 1, Ker q = 0; (c) If s 0 is semisimple, then g¯0 = g ⊕ CD. If s¯0 has one-dimensional center Cc, then g¯0 = (g + i) ⊕ CD and g ∩ i = CK, where i is the subalgebra generated by c ⊗ tk . In particular, ∆0 (g) = ∆ (g (B)). (1) (1) (d) Let K1 , . . . , Kk denote the canonical central elements of s1 , . . . , sk , (·, ·)i denote the Killing form on si and (x, y) = bi (x, y)i for x, y ∈ si ; then c = a 1 K 1 + · · · + ak K k | ai b i = 0 . (8) (e) ∆0 (g) = ∆ (g ) = ∆ (g (B)), and imaginary roots are roots of the form mδ, where m ∈ Z {0}.

192

Vera Serganova

Proof. (a) follows from the fact that even roots of g are the same as the roots (1) (1) of s1 ⊕ · · · ⊕ sk . This fact can be also used to prove (c). Indeed, (g¯0 )α = gα for any even root α of the form α◦ + mα with α◦ ∈ ∆0 (s), and (g¯0 )mδ = hs ⊗ tm ,

gmδ = h[s,s] ⊗ tm ,

where hs , h[s,s] are Cartan subalgebras of s and [s, s], respectively. If s¯0 is semisimple, then hs = h[s,s] and gmδ = (g¯0 )mδ . If s¯0 is not semisimple, then (g¯0 )mδ = gmδ ⊕ C (c ⊗ tm ) , and hence g¯0 = g + i. Next we will prove (d). Pick elements x1 ⊗ t, . . . , xk ⊗ t,

y1 ⊗ t−1 , . . . , yk ⊗ t−1

¯ i be such that xi , yi belong to the Cartan subalgebra of si , (xi , yi ) = 1. Let K the images of Ki under the natural projection g (B) → g (B)/c. Then q (K) = q

¯i = 1 K ¯ i. xi ⊗ t, yi ⊗ t−1 = (xi , yi )i K bi

Thus, c is generated by b1i Ki − b1j Kj , that implies (d). To prove (b) note that dim c = k − 1 by (d) and Ker q = 0 since g (B) is symmetizable (see Remark 4.7); this implies ∆ (g ) = ∆ (g (B)). Roots of i coincide with roots of h ⊗ tZ ; this implies ∆0 (g) = ∆ (g ). Since Weyl groups of g (B) and g coincide, even real roots of g coincide with real roots of g (B). Thus the required description of real even roots follows from the theory of affine Lie algebras, see [2]. The absence of odd imaginary roots is not used in this paper; note only that it follows easily from the fact that the Weyl group of g (B) acts transitively on the subset of isotropic and nonisotropic odd roots in the set ∆◦1 + Zδ of odd roots of g. The latter fact can be checked case by case (see [11] for details); therefore, all odd roots are real. Remark 9.2. Observe that if g = s(1) , then k = 1 only for s = sl (1|n), osp (1|2n) or osp (2|2n). Example 9.3. Let s = D (2, 1; a). Then s¯0 ∼ = sl (2) ⊕ sl (2) ⊕ sl (2) ,

g (B) ∼ = sl (2)

(1)

(1)

⊕ sl (2)

(1)

⊕ sl (2)

,

and c is spanned by K2 − aK1 and (1 + a) K1 + K3 , where K1 , K2 and K3 are the standard central elements of the components in g (B) isomorphic to (1) sl (2) . In this case g¯0 = g ⊕ CD. Corollary 9.4. In the notations of Theorem 9.1 let k = 2. Then c is spanned by K1 + uK2 for some positive rational u.


193

Proof. Follows from the fact that b1 , b2 are rational and have different signs. The latter fact can be found in [1]. Below we list g (B) for all twisted affine superalgebras. g sl (1|2n)(2) , n≥2 sl (2|2n)(2) , n≥2 (2) (p) sl (m|2n) , m ≥ 3, n ≥ 2 (2) sl (2m + 1|2) , m≥1 (4) sl (1|2n + 1) , n≥1 (p) sl (2m + 1|2n + 1)(4) , n ≥ m ≥ 1 (2) osp (2|2n) , n≥1 (2) osp (4|2n) , n≥1 (2) osp (2m|2n) , m ≥ 3, n ≥ 1

B) g (B sl (2n)(2) (1) sl (2) ⊕ sl (2n)(2) (2) (2) sl (m) ⊕ sl (2n) (2) (1) sl (2m + 1) ⊕ sl (2) (2) sl (2n + 1) sl (2m + 1)(2) ⊕ sl (2n + 1)(2) (1) sp (2n) (1) (1) sl (2) ⊕ sp (2n) (2) (1) o (2m) ⊕ sp (2n)

Remark 9.5. From this table one can see that g (B) is either a (twisted) affine Lie algebra, or a direct sum of two (twisted) affine Lie algebras. In the former case c = 0, and in the latter case c is one dimensional, and generated by K1 + uK2 ; here by K1 and K2 we denote the canonical central elements of (twisted) affine superalgebras which appear as direct summands of g (B), and u is an appropriate positive rational number. It is not difficult to see that g¯0 = g ⊕ CD for the twisted affinization of osp (2m|2n), m ≥ 2, and of psl (n|n), n ≥ 3. If g is the twisted affinization of sl (m|n) with m = n, or osp (2|2n), then g¯0 = (g + i) ⊕ CD, i ∩ g = CK; here i is generated by c ⊗ t2k+1 for k ∈ Z if the order of the twisting automorphism is 2, and by c ⊗ t4k+2 for k ∈ Z if the order of the twisting automorphism is 4 (as before, c denotes a central element of s¯0 ). (1) Strange twisted affine superalgebra q (n)(2) . g (B) ∼ = sl (n) , c coincides (1) with the center of sl (n) , the homomorphism q : g → g (B)/c is injective, and g¯0 = g ⊕ CD ⊕ CK. (9)

Non-symmetrizable superalgebra S (1, 2; b), b ∈ / Z. To describe g¯0 , one has to use the realization Sb = S (1, 2; b)/CK; here K denotes a central element of S (1, 2; b). It is a straightforward calculation that Sb = Ct

∂ ⊕ [Sb , Sb ] , ∂t

and [Sb , Sb ]¯0 is a semidirect sum of the subalgebra L spanned by tm

∂ + (m + b) tm−1 E, ∂t

194

Vera Serganova

and the ideal g /CK spanned by tm i,j=1,2 cij ξi ∂ξ∂ j with c11 + c22 = 0. Note −1 that L is isomorphic to the algebra of derivations of C t, t , and g /CK is −1 isomorphic to the loop algebra sl (2)⊗C t, t . Going to the central extension g = S (1, 2; b) of Sb , one obtains g¯0 = L + g ⊕ Ct

∂ , ∂t

g ⊕ Ct

∂ ∼ (1) = sl (2) , ∂t

L ∩ g = CK;

here L is isomorphic to the Virasoro algebra. The family Q± (l, m, n). Due to the lack of an explicit description, we can say only a few things about the roots and the structure of g and g (B) for g = Q± (l, m, n). For example, we would like to calculate the multiplicities of imaginary roots of these algebras; however, we have not yet succeeded. Let α1 , α2 , α3 be the base with three isotropic roots. Then there are three linearly independent principal roots β1 = α2 + α3 , β2 = α1 + α3 , β3 = α1 + α2 . Since [Q : Q ] = 2 and [Q : Q0 ] = 2, one has Q0 = Q . Note that the matrix B ⎞ ⎛ 2 −k −k ⎝ −l 2 −l ⎠ . −m −m 2 has a negative determinant (as easily follows from the condition k, l, m ≥ 1 and klm > 1). Therefore g (B) is a simple Kac–Moody Lie algebra of indefinite type. In this case we do not know if q is injective.

10 Integrable modules and highest weight modules Let g = g (A) be a regular Kac–Moody superalgebra with a standard base Π. A g-module M is called a weight module if h acts semisimply on M ; in other words M has a weight decomposition M= Mµ , Mµ = {m ∈ M | hm = µ (h) m, ∀h ∈ h} , µ∈h∗

and dim Mµ < ∞ for all µ ∈ h∗ . The set P (M ) = {µ ∈ h∗ | Mµ = 0} is called the set of weights of M . The formal character ch M of M is defined by the formula ch M = dim Mµ eµ . µ∈P (M)

A module M is integrable if M is a weight module, and Xβ and Yβ act locally nilpotently on M for every principal root β of g. Note that if α is isotropic, then Xα2 = 0, hence Xα acts locally nilpotently on any module. If


195

β is a principal root, then exp Xβ exp (−Yβ ) exp Xβ is a well-defined linear operator on an integrable module M ; Thus the Weyl group W of g acts on M ; therefore ch M is W -invariant. This implies also that Xα and Yα act locally nilpotently on M for any real root α. By definition of a Kac–Moody superalgebra, its adjoint module is integrable. Our next step is to define the category O of highest weight modules. Define a Verma module M (λ) with the highest weight λ as the induced module U (g) ⊗U (b) C (λ), where b = h ⊕ n+ , C (λ) is a one-dimensional bmodule with generator v such that hv = λ (h) v, Xα v = 0 for all simple roots α. A vector v is called a highest vector of M (λ). Any quotient V of M (λ) is an indecomposable module generated by the image of v under the natural projection M (λ) → V . The following lemma can be proven exactly as in the Lie algebra case. Lemma 10.1. Let m (α) = dim gα . (a) ch M (λ) = eλ

+ α∈∆1 + α∈∆0

m(α)

(1+e−α )

(1−e−α )m(α)

.

(b) M (λ) has a unique irreducible quotient which we denote by L (λ). The category O is a full subcategory of the category of g-modules, whose objects are weight modules M such that P (M ) ⊂ si=1 P (M (µi )) for some finite set {µ1 , . . . , µs } ∈ h∗ . In what follows we will often need to change the base Π to a base Σ by odd reflections. So we will write n+ Σ , bΣ , OΣ , MΣ (λ), and LΣ (λ) if we mean the corresponding object for a non-standard base. If the subindex is omitted, then Σ = Π. Lemma 10.2. Let Σ is obtained from Σ by an odd reflection rα . (a) MΣ (λ) and LΣ (λ) are objects of OΣ . (b) If λ (hα ) = 0, then MΣ (λ − α) ∼ = MΣ (λ), and LΣ (λ − α) ∼ = LΣ (λ). ∼ (c) Let λ (hα ) = 0, then LΣ (λ) = LΣ (λ). If p = CYα ⊕ b, then the b-module structure on C (λ) and C (λ − α) extends uniquely to a p-module structure, and the following exact sequences hold: 0 → U (g) ⊗U (p) C (λ − α) → MΣ (λ) → U (g) ⊗U(p) C (λ) → 0, 0 → U (g) ⊗U (p) C (λ) → MΣ (λ − α) → U (g) ⊗U (p) C (λ − α) → 0. Proof. The first statement of the lemma follows from the identity ch MΣ (λ − α) = ch MΣ (λ) , which is a straightforward consequence of Lemma 10.1 (a). Now prove (b). Let v be a highest vector of MΣ (λ). Slightly abusing notations, we denote the image of v in LΣ (λ) by v. A simple calculation shows that Yα v is n+ Σ -invariant. If λ (hα ) = 0, then Xα Yα v is proportional to v with a nonzero coefficient. Hence

196

Vera Serganova

MΣ (λ) ∼ = U (g) ⊗U (bΣ ) (CYα v) ∼ = MΣ (λ − α) , and LΣ (λ − α) ∼ = LΣ (λ) . To show (c), assume that λ (hα ) = 0. Then Yα v generates a proper submodule in MΣ (λ). Therefore Yα v = 0 in LΣ (λ), thus v is a highest vector with respect to both Σ and Σ . Hence LΣ (λ) ∼ = LΣ (λ). Finally, note that the submodule of MΣ (λ) generated by Yα v is isomorphic to U (g) ⊗U(p) C (λ − α) (indeed, choose a subalgebra m ⊂ n− of codimension one, such that g = m ⊕ p. Then U (n− ) = U (m) ⊕ U (m) Yα , hence U (g) Yα v = U (m) Yα v). That implies the first exact sequence. The second follows by symmetry. Corollary 10.3. If Σ is obtained from Π by odd reflections, then OΣ = OΠ . Remark 10.4. Corollary 10.3 guarantees that the category O does not change if we change a base by odd reflection. In what follows we omit the subindex in the notation for category O. A weight λ is called integrable if L (λ) is an integrable module. One can use Lemma 10.2 to describe integrable weights: it implies that for any Σ obtained from Π by odd reflections there is exactly one weight λΣ such that LΠ (λ) ∼ = LΣ (λΣ ). Theorem 10.5. Let g be a regular quasisimple Kac–Moody superalgebra. A weight λ is integrable iff for any principal root β and any Σ obtained from Π by odd reflections such that β ∈ Σ, the condition λΣ (hβ ) ∈ Z≥0 holds. If it holds for one Σ such that β ∈ Σ, then it holds for any Σ such that β ∈ Σ. ∼ L (λ); Proof. If λ is integrable, then Yβ acts locally nilpotently on LΣ (λΣ ) = in particular, sufficiently large power of Yβ annihilates a highest vector vΣ of LΣ (λΣ ). The standard sl (2)-calculation implies that λΣ (hβ ) ∈ Z≥0 . λ (h )+1 On the other hand, if λΣ (hβ ) ∈ Z≥0 , then Yβ Σ β vΣ = 0. Since LΣ (λ) = U (g) vΣ , and adYβ is locally nilpotent on U (g), Yβ acts locally nilpotently on LΣ (λ). If the condition λΣ (hβ ) ∈ Z≥0 holds for one Σ containing β, then Yβ is locally nilpotent; hence the condition holds for any other Σ containing β. Hence, in the absence of odd reflections, Corollary 10.6. Let g be a regular Kac–Moody superalgebra without isotropic simple roots. Then λ is integrable iff λ (hα ) ∈ Z≥0 for any even simple root α, and λ (hα ) ∈ 2Z≥0 for any odd simple root α. Indeed, if α is odd, then hα /2 is hβ for β = 2α ∈ B. Since B is finite (see Remark 4.8), Theorem 10.5 is an explicit test for integrability. Using it, one can recover the description of integrable weights given in [1] for finite-dimensional g (see appendix in [13]). Let us illustrate the method with examples.


197

Example 10.7. Let g = D (2, 1; a); choose the base Π = {α1 , α2 , α3 } such that all simple roots are isotropic. All principle roots are β1 = α2 + α3 , β2 = α1 + α3 and β3 = α1 + α2 . They are linearly independent. Therefore one can parameterize λ by setting λ = (c1 , c2 , c3 ), where c1 = λ (hβ1 ), c2 = λ (hβ2 ) and c3 = λ (hβ3 ). Using Remark 3.2 one gets hβ 1 =

h2 + ah3 , a

hβ 2 = −

h1 + ah3 , a+1

hβ 3 = h1 + h2 .

(10)

Let Σk = rαk (Π), k = 1, 2, 3. If λ (h1 ) = 0, then λΣ1 = λ − α1 = (c1 + 1, c2 − 1, c3 − 1) . Hence c2 , c3 ∈ Z>0 . If λ (h2 ) = 0, then λΣ2 = λ − α2 = (c1 − 1, c2 + 1, c3 − 1) and hence c1 ∈ Z>0 . We do not have to make the third odd reflection rα3 since Σ1 = {−α1 , β2 , β3 } and Σ2 = {β1 , −α2 , β3 } already contain all principal roots. By symmetry, if λ (hi ) , λ (hj ) = 0 for some i = j, then c1 , c2 , c3 ∈ Z>0 . On the other hand if λ (hi ) = λ (hj ) = 0 for some i = j, then two odd reflections do not change λ and hence c1 , c2 , c3 ∈ Z≥0 . Thus, λ = (c1 , c2 , c3 ) is integrable iff c1 , c2 , c3 ∈ Z≥0 and, in addition, one of the following conditions holds: 1. c1 , c2 , c3 > 0; 2. c1 = (a + 1) c2 + c3 = 0; 3. c2 = −ac1 + c3 = 0; 4. c3 = −ac1 + (a + 1) c2 = 0. Note that for λ = 0, the conditions (2) − (4) imply a ∈ Q. Thus, there are more integrable weights for rational a than for irrational a. Now, do similar calculations for Q± (k, l, m), which also has a base with all isotropic roots. Example 10.8. Let g = Q± (m, k, l). Again Π = {α1 , α2 , α3 }, the principal roots are β1 = α2 +α3 , β2 = α1 +α3 and β3 = α1 +α2 . Since the principal roots are linearly independent, we again can use parameterization λ = (c1 , c2 , c3 ), where c1 = λ (hβ1 ), c2 = λ (hβ2 ) and c3 = λ (hβ3 ). It is easy to check that hβ 3 =

h1 + h2 , a

hβ 1 =

h2 + h3 , b

hβ 2 =

h3 + h1 . c

(11)

We notice that, as in the previous example, integrability implies ci ∈ Z≥0 for i = 1, 2, 3. Moreover, if at least for two hi , hj , λ (hi ) and λ (hj ) = 0, then c1 , c2 , c3 ∈ Z>0 . Assume that λ (h1 ) = λ (h2 ) = 0. Then by (11), c1 = λ (h3 ), c2 c = λ (h3 ). It was shown in [3] that a, b, c ∈ / Q. Hence c1 = c2 = c3 = 0. Similarly, λ (h1 ) = λ (h3 ) = 0 and λ (h3 ) = λ (h2 ) = 0 imply λ = 0. Thus, all integrable weights are (c1 , c2 , c3 ) ∈ Z3>0 and 0.

198

Vera Serganova

The following theorem was proven in [11] for untwisted affine superalgebras. Other cases were done in [3]. (2)

(1)

(2)

Theorem 10.9. Let g = q (n) , (p) sl (m|n) or (p) sl (m|n) with (1) (1) (1) (2) m, n ≥ 2, osp (m|2n) or osp (m|2n) with m ≥ 3, G3 or F4 ; then any integrable module L (λ) is one-dimensional and trivial over [g, g]. (2)

Proof. Let g = q (n)

. As follows from Section 9, (1) g ⊕ CD ∼ = sl (n) /c,

where c is the center of sl (n)(1) . Thus, λ is an integrable highest weight of (1) sl (n) with zero central charge. A nontrivial integrable highest weight module over an affine Lie algebra must have a positive central charge (see [2]). Therefore λ (hβ ) = 0 for all principal β. Recall from (9) that Chβ ⊕ CK ⊕ CD, h ∩ [g, g] = Chβ ⊕ CK. h= β∈B

β∈B

We claim next that λ (K) = 0. Indeed, choose a base Π = {α1 , . . . , αn } such that α1 is odd (automatically isotropic), and αi is even for any i ≥ 2 (and therefore principal); the Dynkin diagram is a loop, so α1 is connected with α2 and αn . Let Σ = rα1 (Π). A direct calculation shows that the subspace of h spanned by hβ (for principal β) has codimension 1 in the space spanned by hα , α ∈ ∆, and K and hα1 are not in this subspace. (This does not contradict the algebra being of type II, since in the absence of duality this does not relate to the geometry of {hβ | β ∈ B} ⊂ h.) If λ (K) = 0, then λ (hα1 ) = 0, therefore λΣ = λ − α1 . Note that then λΣ (hαi ) = 0 for αi connected with α1 in ΓΠ . But λΣ (hβ ) = 0 for any principal β. Contradiction. Thus λ (h ∩ [g, g]) = 0, hence L (λ) is onedimensional. For other superalgebras in the list g (B) is a sum of two or three components. If there are two components, i.e., g = D (2, 1; a)(1) , one can use Corollary 9.4 or Remark 9.5, g ∼ = [g (B) , g (B)]/c, where c is spanned by K1 + uK2 for some positive u. Since L (λ) is integrable over g , the g submodule Lg (λ) ⊂ L (λ) generated by a highest vector is a highest weight g -integrable module. Therefore it is a g (B)-integrable module, and thus K1 and K2 have nonnegative eigenvalues (see [2]). But K1 + uK2 acts by zero on L (λ), therefore both K1 and K2 are zero. That implies λ (hβ ) = 0 for any principal β and, by the same argument as for q (n)(2) , λ (h ∩ [g, g]) = 0, and L (λ) is one-dimensional. (1) In the case D (2, 1; a) we have again g ∼ = [g (B) , g (B)]/c. The submodule Lg (λ) is the g (B)-module Lg(B) (µ); here µ is an integrable weight for g (B) such that µ (c) = 0. The description of c in Example 9.3 implies µ (K2 ) = aµ (K1 ) ;

µ (K3 ) = − (1 + a) µ (K1 ) ;

µ (Ki ) ∈ Z≥0 ;


199

therefore µ (Ki ) = 0. Apply the representation theory of affine Lie algebras again; hence Lg(B) (µ) is one-dimensional, and λ (hβ ) = 0 for all principal β. Since hβ generate h ∩ [g, g], as above, L (λ) is one-dimensional. The above theorem, Theorem 7.1, and Theorem 7.3 imply the following corollary. Corollary 10.10. Let g be an infinite-dimensional quasisimple regular Kac–Moody superalgebra, and g has infinite-dimensional integrable highest weight modules. Then g has no simple isotropic roots, or g is isomorphic (1) (1) to sl (1|n) , osp (2|2n) , S (1, 2; b) or Q± (k, l, m). (1)

(1)

We describe the set of integrable weights for sl (1|n) , osp (2|2n) S (1, 2; b) in Section 12 (Q± (m, k, l) is done in Example 10.8).

,

Remark 10.11. Theorem 10.9 shows that most affine superalgebras do not have an interesting integrable highest weight representation. [11] and [12] propose a weaker condition of integrability by requiring integrability over one of the components of g (B) and a finite-dimensional part of another component. These papers contain a conjecture about character formulae of such representations; these formulae have interesting applications in number theory and combinatorics. Although here we do not discuss partially integrable modules, we will prove the Kac–Wakimoto conjecture for sl (1|n)(1) and osp (2|2n)(1) in Theorem 14.7 (in this case k = 1, hence their condition of integrability coincides with our condition).

11 General properties of category O Throughout this section g = g (A) is a regular quasisimple Kac–Moody superalgebra with the standard base Π. Here we investigate the properties of the category O which we need to calculate characters, and which are valid in the context of such g. Let ρ = ρΠ be any element of h∗ , satisfying the condition ρ (hα ) =

aαα 2

for all α ∈ Π; the choice of ρ is unique unless A is degenerate. When Σ is obtained from Π by odd reflections, define ρΣ = ρ + β. β∈∆+ (Π)∩∆− (Σ)

Lemma 11.1. Let Σ be a base obtained from Π by odd reflections. Then ρΣ (hα ) = aαα 2 for any α ∈ Σ.

200

Vera Serganova

Proof. It is sufficient to check that if Σ = rβ (Σ) for some isotropic β ∈ Σ, rβ (α) hrβ (α) α(hα ) then ρΣ (hα ) = 2 for any α ∈ Σ implies ρΣ hrβ (α) = . Note 2 that, by Lemma 4.4, ρΣ = ρΣ + α. The statement can be checked by a direct calculation, which is not trivial only if rβ (α) = α + β. Then, by Remark 3.2, hα+β = α (hβ ) hα + β (hα ) hβ , so ρΣ (hα+β ) =

α (hβ ) α (hα ) α + β, hα+β + α (hβ ) β (hα ) = . 2 2

Remark 11.2. It is convenient to reformulate Lemma 10.2 in the following way. Let Σ = rα (Σ) for an odd isotropic α ∈ Σ. If λ + ρ, hα = 0, then λΣ + ρΣ = λΣ + ρΣ .

(12)

λΣ + ρΣ = λΣ + ρΣ + α.

(13)

If λ + ρ, hα = 0, then

A weight λ ∈ h∗ is called typical if λ + ρ, hα = 0 for any real isotropic root α. (12) implies that if λ is typical, then λΣ +ρΣ is invariant with respect to odd reflections. Hence the notion of typicality can be defined for an irreducible highest weight module L (λ) independent of a choice of a base, i.e., if λ is typical, then λΣ is typical. If A is symmetrizable, then g admits an invariant symmetric nondegenerate even form (·, ·); this form induces an isomorphism η : h → h∗ such 2α that η (hα ) = (α,α) for every real even root α. If α is a real isotropic root, then η (hα ) is proportional to α. In this case the typicality condition can be rewritten as (λ + ρ, α) = 0 for any real isotropic root α. Note also that in this case (12) and (13) imply that for any Σ obtained from Σ by odd reflections (λΣ + ρΣ , λΣ + ρΣ ) = (λΣ + ρΣ , λΣ + ρΣ ) .

(14)

The following statement can be found in [11]; it is a straightforward generalization of the corresponding statement for Lie algebras. Lemma 11.3. Let g be a Kac–Moody superalgebra with an invariant nondegenerate symmetric even form. For each α ∈ ∆+ , let e1α , . . . , esαα be a basis in gα , fα1 , . . . , fαsα be the dual basis in g−α , u1 , . . . , ur and let u1 , . . . , ur be dual bases in h. The operator Ω = 2η −1 (ρ) +

r i=1

ui ui + 2

α∈∆+

(−1)p(α)

sα

fαj ejα

j=1

is well defined on any module from the category O and commutes with the action of g. Ω acts on M (λ) as a scalar operator with eigenvalue (λ + 2ρ, λ).


201

For any base Σ introduce a partial order on h∗ by putting iff λ−µ= mα α µ ≤Σ λ α∈∆+ (Σ)

for some mα ∈ Z≥0 . For the standard base Π, we omit the subindex in ≤Π . The next three statements below are straightforward generalizations of the similar statements for Kac–Moody Lie algebras (which can be found in [2]). The proofs are omitted. Lemma 11.4. Let g be an arbitrary Kac–Moody superalgebra, M ∈ O and ν ∈ h∗ . There exists a filtration 0 = F 0 (M ) ⊂ F 1 (M ) ⊂ F 2 (M ) ⊂ · · · ⊂ F s (M ) = M i (M ) = L (λ i ) for every and a subset of indices J such thatF i+1 (M )/F j i ∈ J and for every µ ≥ ν, µ ∈ / j ∈J P F (M )/F j−1 (M ) . For every / λ ≥ ν, define the multiplicity [M : L (λ)] as # {i | λi = λ}. The multiplicity [M : L (λ)] depends neither on a choice of filtration, nor on ν ≤ λ. Hence [M : L (λ)] is well defined for any λ.

Lemma 11.5. (a) [M (λ) : L (λ)] = 1; (b) [M (λ) : L (µ)] > 0 implies λ ≥ µ; (c) If g has an invariant form (·, ·), then [M (λ) : L (µ)] > 0 implies (λ + ρ, λ + ρ) = (µ + ρ, µ + ρ). Corollary 11.6. Let V = 0 be a quotient of M (λ). There is a unique way to write the character of V as an (infinite) linear combination ch V = aµ ch M (µ) . µ≤λ

Furthermore, aλ = 1, and if T = {µ | aµ = 0}, then for any µ ∈ T there are ν1 , . . . , νk ∈ T such that [M (λ) : L (ν1 )] > 0,

[M (ν1 ) : L (ν2 )] > 0, . . . , [M (νk ) : L (µ)] > 0.

If g has an invariant form (·, ·), then all µ ∈ T satisfy the additional condition (λ + ρ, λ + ρ) = (µ + ρ, µ + ρ). Lemma 11.7. Let Σ be obtained from Π by odd reflections. Then [M (λ) : L (µ)] = [MΣ (λ + ρ − ρΣ ) : L (µ)] for any µ. Proof. Sufficient to check that [MΣ (λ) : L (µ)] = [MΣ (λ + ρΣ − ρΣ ) : L (µ)] for any µ if Σ = rα (Σ). In this case it follows directly from Lemma 10.2 (b) and (c).

202

Vera Serganova

Corollary 11.8. Let V be a subquotient of M (λ). For any weight µ of V and any Σ obtained from Π by odd reflections, µ + ρΣ ≤Σ λ + ρ. Proof. If µ is a weight of M (λ), then µ is a weight of MΣ (λ + ρ − ρΣ ). Corollary 11.9. If [M (λ) : L (µ)] > 0, then µΣ + ρΣ ≤Σ λ + ρ for any Σ obtained from Π by odd reflections. Proof. If [M (λ) : L (µ)] > 0, then [MΣ (λ + ρ − ρΣ ) : L (µ)] > 0. But L (µ) = LΣ (µΣ ). Hence µΣ is a weight of MΣ (λ + ρ − ρΣ ). Now the statement follows from the previous corollary.

12 Lie superalgebras sl (1|n)(1) , osp (2|2n)(1) , and S (1, 2; b) (1)

(1)

In this section g is sl (1|n) , osp (2|2n) , or S (1, 2; b) (b ∈ / Z). The importance of these particular Lie superalgebras for our discussion is that that they are the only finite growth regular quasisimple Kac–Moody superalgebras with isotropic roots which have nontrivial integrable modules, and it is easy to describe all bases obtained by odd reflections. (1) If g = sl (1|n) or S (1, 2; b), a graph of any base is a cycle of length n + 1 with exactly two isotropic roots (which are neighbors); here n = 2 for the case of S (1, 2; b). If g = osp (2|2n), then we will see that all possible Dynkin graphs are ◦ ⇒ ◦ − · · · − ◦ − ⊗ − ⊗ − ◦ − · · · − ◦ ⇐ ◦,

⊗

\

| / ◦ − · · · − ◦ ⇐ ◦.

⊗

Here we write ⊗ instead of gray nodes. For some of our calculation we need an explicit description of roots of g. One can choose linearly independent ε, ε1 , . . . , εn , δ ∈ h∗ so that the even (1) roots of sl (1|n) are εi − εj + mδ, m ∈ Z, i = j, and mδ, m ∈ Z {0}; odd (1) (1) roots of sl (1|n) are ± (ε + εi ) + mδ, m ∈ Z. If g = osp (2|2n) , then the even roots of g are ±εi ± εj + mδ, ±2εi + mδ, m ∈ Z, and mδ, m ∈ Z {0}; odd roots of g are ±ε ± εi + mδ, m ∈ Z. Finally, the roots of S (2, 1; b) are the (1) same as the roots of sl (1|2) , as explained in Section 6. (1) For the standard base in case of sl (1|n) (or S (1, 2; b)) choose Π = {ε1 − ε2 , . . . , εn−1 − εn , εn + ε, −ε − ε1 + δ} ; for osp (2|2n)(1) choose Π = {−2ε1 , ε1 − ε2 , . . . , εn−1 − εn , εn + ε, εn − ε + δ} . Since g is of type I, it has a grading g−1 ⊕ g0 ⊕ g1 ; it is


∆ (g1 ) = {ε + εi + mδ}, ∆ (g1 ) = {ε ± εi + mδ},

(1)

if g = sl (1|n)

203

or S (1, 2; b) , (1)

if g = osp (2|2n)

,

and ∆ (g−1 ) = −∆ (g1 ). This grading induces a Z-grading on the root lattice Qi . (15) Q= i∈Z

Set Q+ =

Qi ,

i>0

Q− =

Qi .

(16)

i 0. Moreover, every odd root is real and isotropic, and ∆+ (g1 ) = {α1 , . . . , αk , . . . } ,

∆+ (g−1 ) = {γ1 , . . . , γk , . . . } .

Imaginary even roots are (by Theorem 9.1) {mδ | m ∈ Z {0}}. Lemma 12.1. Let α1 and γ1 be isotropic roots of Σ, and γ ∈ Qi ∩ C + (Σ). Then γ = iα1 + mβ β, if i ≥ 0, γ = −iγ1 + mβ β, if i ≤ 0, β∈B

β∈B

for some mβ ∈ Z≥0 ; the representation is unique. Proof. Uniqueness is obvious. Express γ as a linear combination of simple roots γ = n1 γ1 + n2 α1 + mβ β. β∈B∩Σ

Note that i = n2 − n1 . If i ≥ 0, one can write it as mβ β. γ = iα1 + n1 (γ1 + α1 ) + β∈B∩Σ

If i < 0, one can write it as

γ = −iγ1 + n2 (γ1 + α1 ) +

mβ β.

β∈B∩Σ

Since γ1 + α1 ∈ B, this finishes the proof. Lemma 12.2. Let k ≥ 0. Then C + (rαs . . . rα1 (Π)) = ∅, Q+ ∩

Q− ∩

s≥k

C + (rγs . . . rγ1 (Π)) = ∅.

s≥k

Proof. Let γ ∈ Q+ ∩ Σ = rαs . . . rα1 (Π))

s≥k

C + (rαs . . . rα1 (Π)). By Lemma 12.1 (applied to γ = iαs +

msβ β,

β∈B

for every s ≥ k. Let βs = αs+1 − αs ; note that βs ∈ B. Then ms+1 = msβ β

if β = βs ,

ms+1 = msβ − i β

if β = βs .

Since msβ ≥ 0 for all β ∈ B and all s ≥ k, i ≤ 0. Contradiction. The proof of the second statement is similar.


205

Lemma 12.3. Let [M (λ) : L (µ)] > 0. Then there exists a base Σ obtained from Π by odd reflections, and numbers mβ ∈ Z≥0 such that λ + ρ − µΣ − ρΣ = mβ β. β∈B

Proof. Let γΣ = λ + ρ − µΣ − ρΣ . By Corollary 11.9, 0 ≤Σ γΣ for any Σ obtained from Π by odd reflections. Let Σ = rα (Σ) for some odd α ∈ Σ. Then by (12) and (13) γΣ = γΣ

if µΣ + ρΣ , hα = 0,

γΣ = γΣ − α if µΣ + ρΣ , hα = 0. (18) Suppose that γΣ1 ∈ Q+ for some Σ1 , and γΣ2 ∈ Q− for some Σ2 . Since Σ1 and Σ2 are connected by odd reflections, and every odd root α is in Q±1 , there exists Σ such that γΣ ∈ Q0 . In this case the statement is true by Lemma 12.1. Now assume that γΣ ∈ Q+ for all Σ. Then one can find i > 0 and k > 0 such that γrαs ...rα1 (Π) ∈ Qi for all s ≥ k. Hence, by (18), γrαs ...rα1 (Π) = γrα

s

...rα1 (Π)

=γ

for all s, s > k. Thus γ ∈ Q+ ∩ s≥k C + (rαs . . . rα1 (Π)), which is impossible by Lemma 12.2. Similarly, the case when γΣ ∈ Q− for all Σ is impossible. Lemma 12.4. Let λ ∈ h∗ . LΣ (λ) is integrable iff λ + ρΣ , hβ ∈ Z>0 for any β ∈ B ∩ Σ, and one of the following two conditions holds: 1. λ + ρΣ , hα1 +γ1 ∈ Z>0 , 2. λ + ρΣ , hα1 = λ + ρΣ , hγ1 = 0. Proof. There is only one β ∈ B Π. To check the conditions of Theorem 10.5 make an odd reflection rα1 . The details are left to the reader. Remark 12.5. Call µ ∈ h∗ regular if µ, hβ = 0 for any even real root β. As follows from the lemma above, if LΣ (λ) is integrable, then λ + ρΣ is regular iff the first condition holds. Lemma 12.6. Let LΣ (λ) be integrable, λ + ρΣ be not regular, and λ (hβ ) = 0 at least for one β ∈ B. Then one can find a base Σ obtained from Σ by odd reflections such that LΣ (λ) ∼ = LΣ (λ), and λ + ρΣ is regular. Moreover, one can choose this base Σ equal to rγk . . . rγ1 (Σ), and find w ∈ W such that λ + ρΣ = λ + ρΣ + γ1 + · · · + γk = w (λ + ρ − kα1 ) , w (α1 ) = −γk . The choice of k is unique.

w

k

(−1) = (−1) ,

206

Vera Serganova

Proof. If λ + ρΣ is not regular, then λ + ρΣ , hα1 = λ + ρΣ , hγ1 = 0. Let Σk = rγk . . . rγ1 (Σ). We claim that LΣk (λ) is not isomorphic to LΣ (λ) at least for one k. Indeed, otherwise (λ, γi ) = 0 for all i, but that would imply λ, hβ = 0 for all β ∈ B. Choose the last k such that LΣk (λ) and LΣ (λ) are isomorphic. Then λ + ρΣk = λ + ρΣ + γ1 + · · · + γk , λ + ρΣk , hγk+1 = 0. By Remark 12.5 applied to Σ = Σk , λ + ρΣk must be regular. Next consider even roots β1 = α1 + γ1 ,

β2 = γ2 − γ1 ,

. . . , βk = γk − γk−1 .

Note that for any i < k, βi+1 ∈ B Σi , and that λ + ρΣ + γ1 + · · · + γi , hβi+1 = 0. Furthermore, rβi+1 (γi+1 ) = γi (recall γ0 = −α1 ). Therefore λ + ρΣ + γ1 + · · · + γk = rβk (λ + ρΣ + γ1 + · · · + γk−1 + γk−1 ) = rβk (λ + ρΣ + γ1 + · · · + 2γk−1 ) = rβk−1 rβk (λ + ρΣ + γ1 + · · · + 3γk−2 ) = · · · = rβ1 . . . rβk (λ + ρΣ + kγ0 ) . Remark 12.7. Recall that g ⊕ CD is isomorphic to g (B) which is sl (n) for (1) (1) (1) (1) g = sl (1|n) , sl (2) for g = S (1, 2; b), and sp (2n) for g = osp (2|2n) . Principal roots B form a base of g . Let (1)

C = {µ ∈ h∗ | µ (hβ ) ∈ Z≥0 , β ∈ B} . It follows from Lemma 12.4, that if LΣ (λ) is integrable, then λ ∈ C. We need two followingfacts about affine Weyl group action, see [2]. If λ ∈ C, then w (λ) = λ − β∈B mβ β for some mβ ∈ Z≥0 . If µ (hβ ) ∈ Z for all β ∈ B and µ (K) > 0, then W µ intersects C in exactly one point. Consider the ring R of all formal expressions µ∈P cµ eµ for all P satisfying the condition that there is a finite set L ⊂ h∗ such that P ⊂ L − Z≥0 α. (19) α∈Π

It is not difficult to check that indeed such R is a commutative ring without zero divisors. Let R be a subring satisfying the additional condition that (19) holds for w (P ) for any w ∈ W . Then R enjoys the natural action of W ; this action preserves multiplication and addition. For an arbitrary µ ∈ h∗ and Σ obtained from Π by odd reflections, define m(α) e−ρΣ α∈∆+ (Σ) (1 + e−α ) w 1 UΣ (µ) = (−1) ew(µ) . m(α) −α ) w∈W α∈∆+ (Σ) (1 − e 0


207

It is an immediate calculation that for any Σ obtained from Σ by odd reflections, and for any w ∈ W w

UΣ (µ) = UΣ (µ) ,

UΣ (w (µ)) = (−1) UΣ (µ) .

(20)

Hence one can drop the index Σ in UΣ (µ). Assume now that µ ∈ C. Then w (µ) ≤ µ and hence UΣ (µ) ∈ R. It easily follows from Lemma 10.1 (a) that w U (µ) = (−1) ch M (w (µ) − ρ) . (21) w∈W

If eν appears in ch M (κ) with nonzero coefficient, then ν ≤ κ. Therefore, if eν appears in U (µ) with nonzero coefficient, then ν ≤ w (µ) − ρ ≤ µ − ρ. Lemma 12.8. If µ is regular and integrable, then U (µ) ∈ R . Moreover, w (U (µ)) = U (µ) for any w ∈ W . Proof. Let ρ0 ∈ h∗ be such that ρ0 (hβ ) = Introduce the expressions

D0 = eρ0

1 − e−α

m(α)

aββ 2

for all β ∈ B, and let ρ1 = ρ0 −ρ.

D1 = eρ1

,

α∈∆+ 0

1 + e−α

m(α)

.

α∈∆+ 1

Then U (µ) =

D1 (−1)w ew(µ) . D0 w∈W

Note that ∆ (g ) = ∆0 , and the multiplicity of kδ in g is m (kδ) − 1 (see Theorem 9.1 (c), the case of s with one-dimensional center) D0 = D0

∞

1 − e−nδ ,

n=1

where D0 the corresponding expression for the Lie subalgebra g + h. The expression w (−1) ew(µ) S (µ) = w∈W D0 gives a character of the simple integrable module over Kac–Moody Lie algebra g +h with highest weight µ−ρ0 (see [2]). (Note that regularity and integrability of µ implies µ − ρ0 ∈ C.)Therefore S (µ) is W -invariant, and S (µ) ∈ R . ∞ On the other hand, n=1 1/ 1 − e−nδ is W -invariant, since W δ = δ. Therefore, it is sufficient to show that D1 is W -invariant. One has to check that rβ (D1 ) = D1 for all β ∈ B; assume first that β ∈ Π; then ρ1 (hβ ) = 0, so rβ (ρ1 ) = ρ1 ; moreover, by Lemma 4.4, rβ permutes roots of ∆+ 1 . Hence rβ (D1 ) = D1 . If β ∈ / Π, then

208

Vera Serganova

β = γ1 +α1 ,

ρ (hβ ) = 0,

ρ1 (hβ ) = 1,

rβ (ρ1 ) = ρ1 −β = ρ1 −α1 −γ1 .

Furthermore, since β = α1 + γ1 , rβ (α1 ) = −γ1 ,

rβ (γ1 ) = −α1 ,

therefore rβ rγ1 = r−α1 rβ by (5), which implies that rβ permutes the roots of ∆+ 1 {α1 , γ1 }. Hence rβ eρ1 1 + e−α1 1 + e−γ1 = eρ1 −α1 −γ1 (1 + eα1 ) (1 + eγ1 ) = eρ1 1 + e−α1 1 + e−γ1 , m(α) m(α) rβ 1 + e−α 1 + e−α = . α∈∆+ 1 {α1 ,γ1 }

α∈∆+ 1 {α1 ,γ1 }

This finishes the proof.

Remark 12.9. It is useful to note that D1 is not only W -invariant, but also independent of a choice of Σ. In other words, D1 = eρ0 −ρΣ 1 + e−α α∈∆+ 1 (Σ)

for any Σ obtained from Π by odd reflections.

13 Lie superalgebras sl (1|n)(1) , osp (2|2n)(1) (1)

From now on we assume that g is sl (1|n)

(1)

or osp (2|2n)

.

Lemma 13.1. There exists an invariant symmetric even form on g such that the corresponding form on Q × Q is semi-positive, takes integer values, and (α, α) > 0 for any even real root α. Proof. Just use the form in (7). The positivity conditions follow from those on sl (1|n) and osp (2|2n). Remark 13.2. One can normalize an invariant form on g so that the corresponding form on h∗ satisfies the relations (ε, ε) = −1,

(εi , εj ) = δij ,

(ε, εi ) = (ε, δ) = (εi , δ) = 0.

One can see that (δ, Q) = 0. On the other hand, (λ, δ) = λ (K) for any (1) λ ∈ h∗ . In the case of sl (1|n) , all real even roots have the same length. (1) If g = osp (2|2n) with n ≥ 2, then (β, β) = 2 for a short real even root, 2β and (β, β) = 4 for a long real even root. Since η (hβ ) = (β,β) , given an integrable weight λ, we have (λ + ρ, β) ∈ Z for any real even short root β, and (λ + ρ, β) ∈ 2Z for any real even long root β. Moreover, (ρ, δ) = N − 1 for sl (1|n)(1) , (ρ, δ) = N for osp (2|2n)(1) .


209

For the convenience of the reader, let us recall the properties of weights we use in what follows (we reformulate them here using the invariant form). λ is regular if (λ, β) = 0 for any even real β; λ is typical if (λ + ρ, β) = 0 for any real isotropic β; λ is integrable if L (λ) is an integrable module. Lemma 13.3. Let µ ∈ h∗ be regular and (µ, δ) = 0. Then there is at most one α ∈ ∆+ 1 such that (µ, α) = 0. Proof. Let α, γ ∈ ∆+ 1 and α = γ. By a direct inspection of the list of roots, one can check that either α+γ, or α−γ is an even root β. So if (µ, α) = (µ, γ) = 0, then (µ, β) = 0. Contradiction, since every β is either real or mδ. Lemma 13.4. Let Π be now an arbitrary base, let ν ∈ h∗ be such that ν (hβ ) ∈ Z≥0 for all β ∈ B, Σ be obtained from Π by odd reflections. Then νΣ + ρΣ , hβ ∈ Z≥0 for any β ∈ B ∩ Π, and νΣ + ρΣ , hβ ∈ Z≥−1 for β ∈ B Π. Proof. First, observe that Yβ is locally nilpotent on L (ν) for any β ∈ B ∩ Π. Since LΣ (νΣ ) ∼ = LΠ (ν), then νΣ , hβ ∈ Z≥0 for any β ∈ B ∩ Π. Note also that ρΣ (hβ ) = 1 if β ∈ B ∩ Σ, and ρΣ (hβ ) = 0 if β ∈ B Σ. Hence it suffices to consider the case β = α1 + γ1 ∈ B Π. Let α0 = −γ1 , Π0 = Π, Πk = rαk . . . rα1 (Π). We will prove the statement k for Σ = Πk (for Σ = rγk . . . rγ1 (Π) the proof is similar). Let λ = νΠk + ρΠk , λki = λk , αi . Let k be such that λk−1 (hβ ) ≥ 0, λk (hβ ) < 0; let s > k be the minimal such that αs+1 − αs = αk+1 − αk (it exists since αt+1 − αt is N -periodic). Note that β = α1 − α0 . Lemma 13.5. One has λkk = 0, αk+1 − αk = β, λk = λk+1 , λk (hβ ) = λk+1 (hβ ) = · · · = λs−1 (hβ ) = −1. Proof. By inspection of the formulae for roots, and by (12) and (13), αi (hβ ) = −1 if β = αi+1 − αi ,

αi (hβ ) = 1 if β = αi − αi−1 ,

αi (hβ ) = 0 otherwise; λt = λt−1

if λt−1 = 0, t

λt = λt−1 + αt

if λt−1 = 0. t

Therefore, λkk = 0; αk+1 −αk = β, since αk (hβ ) < 0 (by λk (hβ ) < λk−1 (hβ )); this implies λkk+1 < 0, thus λk+1 = λk . Moreover, λk (hβ ) = λk+1 (hβ ) = · · · = λs−1 (hβ ) = −1, because αt (hβ ) = 0, k + 2 ≤ t < s, and αk (hβ ) = −1. Since αi+N = αi + δ, one has λki+N = λki + M , here M = (ν + ρ, δ). Moreover, (ν, δ) ≥ 0 (since δ is a positive root of g (B)), thus M ≥ (ρ, δ); (1) (1) therefore M ≥ N − 1 if g = sl (1|n) , M ≥ N if g = osp (2|2n) .

210

Vera Serganova

Lemma 13.6. One has λks − λkk ≥ s − k − 2 + (β, β)/2. For k + 2 ≤ i ≤ s − 1, one has (αi , αs ) = −1. Proof. We prove it case-by-case. Let g = sl (1|n)(1) ; hence N = n. In this case αi+1 − αi = αj+1 − αj iff i ≡ j mod N . Therefore, s = k + N . It is enough to note that (αi , αj ) = −1 if i ≡ j mod N , λks − λkk = M ≥ s − k − 1, and (β, β) = 2. (1) Let g = osp (2|2n) ; hence N = 2n. Then αi+1 − αi = αj+1 − αj iff i ≡ j mod N or i + j ≡ 2p mod N ; here p is the smallest 0 ≤ p < n such that αp+1 − αp is a long root. Moreover, (αi , αj ) = −1 if i ≡ j, i + j ≡ 2p + 1 mod N . If p = 0, then (β, β) = 4, n | k, s = k + N , and M ≥ s − k imply what is needed. Now assume p > 0; then k ≡ 0 mod N,

s ≡ 2p mod N,

or k ≡ 2p

mod N,

s≡0

mod N.

Since (β, β) = 2, what remains to prove is λks − λkk ≥ s − k − 1. Note that for t between k and s, the only long root among αt+1 − αt is one with t = def r = (k + s)/2. We already know that λkk+1 − λkk = − (β, β)/2 = −1, and λkt+1 − λkt ≥ 1 for t ≡ k, s mod N . Note that λkr+1 − λkr ≥ 2; indeed, β = αr+1 − αr is a long root, β = β, thus ρΠr (hβ ) = 1, and (ρΠr , β ) = 2. Hence λks − λkk ≥ s − k − 1 indeed. Since λs = λk + αi1 + · · · + αij for some k + 1 < i1 < i2< · · · < ij ≤ s, and s k (αs , αs ) = 0, at most s−k −2 terms αit can ; hence contribute to λ − λ , α s λss −λks ≥ − (s − k − 2). Therefore λss = λss − λks + λks − λkk ≥ (β, β)/2 > 0. Either λs−1 = λs , or λs−1 = λs − αs ; since (αs , αs ) = 0, one can conclude s−1 λs > 0. Therefore λs = λs−1 , hence λs (hβ ) = −1. Lemma 13.7. If λss ≥ 1 + (β, β)/2, then λt = λs for any t > s. Proof. By induction in t, it is enough to show that λst ≥ 1 for any t > s. Since λki+N = λki + M > λki , it is enough to consider t < s + N . Since λt (hβ ) ≥ 0 for β = β, one has λtj+1 − λtj ≥ 0 unless αj+1 − αj = αk+1 − αk ; as shown at the beginning of the proof, the inequalities are strict for t = k, s. Hence, λsj may decrease only after j such that αj+1 − αj = β, and the decrease is (β, β)/2. (1) By inspection, unless g = osp (2|2n) and β is a short root, there are at most one such value of j per an interval of length N . Otherwise, there are two such values, they are not adjacent, and the decrease in each such value is by (β, β)/2 = 1. Now strict increase at other values implies λst ≥ 1. Therefore, if λss ≥ 1 + (β, β)/2, then λt , hβ = −1 for any t ≥ k. On the other hand, if λss = (β, β)/2, then λss+1 = λss +(λs , β) = (1 + λs (hβ )) (β, β)/2 = 0; hence λs+1 = λs + αs+1 . Since αs+1 (hβ ) = 1, this implies λs+1 (hβ ) = 0.


211

In any case, this implies that if λt (hβ ) < 0 for t0 ≤ t ≤ t1 , then λt (hβ ) = −1 (indeed, taking t0 the minimal possible with the given t1 , we may assume t0 = k). This finishes the proof of Lemma 13.4. Lemma 13.8. Let λ, µ ∈ h∗ , λ (hβ ) > 0, µ (hβ ) ≥ 0 for any β ∈ B, and λ − µ = β∈B mβ β for some mβ ∈ Z≥0 . If (λ, λ) = (µ, µ), then λ = µ. Proof. We use the fact that all principal roots have positive square (see Remark 13.2). ⎞ ⎛ mβ β ⎠ = 0. (λ, λ) − (µ, µ) = (λ + µ, λ − µ) = ⎝λ + µ, β∈B

But (λ + µ, β) =

(β,β) 2

λ + µ, hβ is positive. Hence all mβ = 0.

14 On affine character formulae Let g be a regular Kac–Moody superalgebra with a fixed base Π and Cartan matrix A; let λ be an integrable weight. For a symmetrizable matrix A without zeros on diagonal, the character is described by the famous Weyl character formula ch L (λ) = U (λ + ρ) , (22) which was proven by Kac, see [10]. The proof is a straightforward generalization of his proof in Lie algebra case (see, for example, [2]); it is based on existence of the Casimir element corresponding to the invariant form. If g is a finite-dimensional Kac–Moody superalgebra, (22) holds for a typical weight. It was also proven by Kac [6], but the proof is more complicated. One has to use either the Shapovalov form, or a complete description of the center of the universal enveloping algebra U (g). The reason why a simpler proof from [2] does not work is the existence of real roots of nonpositive square of length. In this section we provide a generalization of these results to some infinitedimensional algebras, as well as to the case of atypical weights: we calculate the characters of all simple integrable highest weight modules over sl (1|2n)(1) , and over osp (2|2n)(1) . We use the invariant form, odd reflections and the fact (1) (1) that the defect of sl (1|2n) and osp (2|2n) is 1. Recall that the defect is the maximal number of linearly independent pairwise orthogonal real isotropic roots; defect is 1 iff for any real isotropic α and γ one of α ± γ is a root (compare 13.3). Our proof is an adaptation of the Bernstein–Leites proof of character formula for sl (1|n), see [14]. Since in our case the defect is 1, our formulae are easier than those for general finite-dimensional superalgebras (see [4, 15]).

212

Vera Serganova (1)

(1)

In what follows we assume g = sl (1|n) or osp (2|2n) , and we use the same notations as in the previous section. Let LΣ (λ) be an integrable simple g-module, and let λ + ρΣ be regular. Let v be a highest vector of MΣ (λ). For any principal root β put k (λ, β) = k(λ,β) λ + ρΣ , hβ . For β ∈ Σ, define vβ = Yβ v. If β ∈ / Σ, i.e., β = α1 + γ1 for isotropic α1 , γ1 ∈ Σ, then, by regularity of λ+ρΣ , one can choose α ∈ {α1 , γ1 } k(λ,β) such that (λ, α) = (λ + ρΣ , α) = 0; define vβ as Yβ Yα v. Let VΣ (λ) be the quotient of MΣ (λ) by the submodule generated by vβ , β ∈ B. Lemma 14.1. Let LΣ (λ) be an integrable simple g-module, and λ + ρΣ be regular. Then VΣ (λ) is an integrable g-module, and λ−α1 , λ−γ1 ∈ P (VΣ (λ)). Proof. Let v¯ be the image of v under the natural projection MΣ (λ) → VΣ (λ). k(λ,β) Then Yβ v¯ = 0 for any β ∈ Σ ∩ B. Since v¯ generates VΣ (λ), Yβ is locally k(λ,β)

Yα v¯ = 0. nilpotent on VΣ (λ) for any β ∈ Σ ∩ B. If β ∈ B Σ, we have Yβ But Xα Yα v¯ = λ (hα ) v¯, therefore Yα v¯ generates VΣ (λ), and Yβ is also locally nilpotent if β ∈ B Σ. Thus, VΣ (λ) is integrable. Denote by µβ the weight of vβ . To show that λ − α1 , λ − γ1 ∈ P (VΣ (λ)) it is sufficient to check that λ − α1 , λ − γ1 are not weights of a submodule + generated − by vβ for each β ∈ B. If β ∈ Σ, then vβ is nΣ -invariant, so U (g) vβ = U nΣ vβ , and obviously the inequalities −λ − α 1 ≤Σ µβ , λ − γ1 ≤Σ µβ do not / P U nΣ vβ . If β ∈ B Σ, consider hold. Therefore λ − α1 , λ − γ1 ∈ = −Σ + rα (Σ) and note that vβ is nΣ -invariant. In this case U (g) vβ = U nΣ vβ . Assume without loss of generality that α = α1 . Then the inequality λ−α1 ≤Σ µβ never holds, and λ − γ1 ≤Σ µβ holds only if k (λ, β) = 1. But then µβ = λ − α1 − β = λ − 2α1 − γ1 , and λ − γ1 = µβ + 2α1 ∈ v / P U n− Σ β v . / P U n− since −α1 ∈ Σ is isotropic. Therefore λ − α1 , λ − γ1 ∈ β Σ One can immediately see that VΣ (λ) is the maximal integrable quotient of MΣ (λ). The following lemma can be proven exactly as Lemma 10.2 (b). Therefore we omit the proof. Lemma 14.2. If λ + ρ is regular, Σ is obtained from Π by odd reflections, and λΣ + ρΣ = λ + ρ, then VΣ (λΣ ) ∼ = VΠ (λ). Recall that λΣ + ρΣ = λ + ρ holds when λΣ is obtained from λ by a chain of “typical” reflections. Lemma 14.3. Let LΣ (λ) be integrable, and let λ + ρΣ be regular. If λ is typical, then VΣ (λ) = LΣ (λ). If (λ + ρΣ , α) = 0 for a simple isotropic root α ∈ Σ, then there is a short exact sequence of g-modules 0 → LΣ (λ − α) → VΣ (λ) → LΣ (λ) → 0. Proof. Let [VΣ (λ) : LΣ (µ)] > 0. Since VΣ (λ) is integrable, LΣ (µ) is integrable. Hence µ is an integrable weight. By Lemma 12.3, λ + ρΣ =


213

µ + ρΣ + β∈B mβ β for some mβ ∈ Z≥0 , some Σ obtained from Σ by odd reflections, and µ such that LΣ (µ) ∼ = LΣ (µ ). Since (µ + ρΣ , µ + ρΣ ) = (µ + ρΣ , µ + ρΣ ) = (λ + ρΣ , λ + ρΣ ) , Lemma 13.8 and regularity of λ + ρΣ imply λ + ρΣ = µ + ρΣ . Without loss of generality, one may assume that Σ = rαk . . . rα1 (Σ). Then (23) µ + ρΣ = µ + ρΣ + αi1 + αi2 + · · · + αir , for some 1 ≤ i1 < · · · < ir ≤ k such µ + ρΣ + αi1 + · · · + αij , αij+1 = 0 for all j < r. In particular, r = 0 or (λ + ρΣ , αir ) = 0. If λ is typical, the latter case is impossible; hence µ = λ, therefore VΣ (λ) = LΣ (λ). Assume that λ is atypical. Since λ + ρΣ is regular, by Lemma 13.3, there is only one α such that (λ + ρΣ , α) = 0; hence r = 0, or α = αir . Since α ∈ Σ, and Σ contains only two isotropic simple roots α1 and γ1 , either α = α1 or α = γ1 . Therefore, either λ = µ, or α = αir = α1 and λ = µ + α. Thus [VΣ (λ) : LΣ (µ)] > 0 implies µ ∈ {λ, λ − α}. It remains to show that [VΣ (λ) : LΣ (λ − α)] = 1. Since the multiplicity of λ − α in MΣ (λ) is 1, [VΣ (λ) : LΣ (λ − α)] ≤ 1. On the other hand, notice that Yα v¯ ∈ VΣ (λ) is not zero (by Lemma 14.1) and generates a submodule with the highest weight λ − α (see the proof of Lemma 10.2). Hence [VΣ (λ) : LΣ (λ − α)] = 1. The next step is to check that (22) holds if we replace LΣ (λ) by VΣ (λ). Theorem 14.4. Let λ be integrable, λ + ρΣ be regular. Then ch VΣ (λ) = U (λ + ρΣ ) . Proof. First, we note that, by regularity, there is only one α ∈ ∆+ 1 such that (λ + ρΣ , α) = 0. Without loss of generality, one may assume that α ∈ Σ (if necessary, one can change Σ to Σ by “typical” odd reflections so that VΣ (λ + ρΣ − ρΣ ) ∼ = VΣ (λ), α ∈ Σ , as in Lemma 14.2). By Corollary 11.6, ch VΣ (λ) = aκ ch MΣ (κ) , aκ = 0. κ∈T

Since VΣ (λ) is integrable, ch VΣ (λ) is W -invariant. Therefore bν eν ch VΣ (λ) − U (λ + ρΣ ) = =

κ∈T

ν∈F

aκ ch MΣ (κ) −

w

(−1) ch MΣ (w (λ + ρΣ ) − ρΣ )

w∈W

is W -invariant. We assume that bν = 0 for any ν ∈ F . Assume that the theorem does not hold. Then F is nonempty. Not that F is a W -invariant set. Pick up a maximal ν ∈ F . First, maximality of ν implies that

214

Vera Serganova

ν ∈ T ∪ (W (λ + ρΣ ) − ρΣ ). Furthermore, rβ (ν) ≥ ν for any β ∈ B such that ν (hβ ) < 0; hence ν (hβ ) ∈ Z≥0 . In particular, ν = w (λ + ρΣ ) − ρΣ , because w (λ + ρΣ ) − ρΣ ∈ C only when w = 1 and, obviously, λ ∈ / F (see Remark 12.7). Recall that by Corollary 11.6, there exist ν1 , . . . , νk ∈ T , νk = ν, such that [MΣ (λ) : LΣ (ν1 )] > 0, [MΣ (ν1 ) : LΣ (ν2 )] > 0, . . . [MΣ (νk ) : LΣ (ν)] > 0. We claim that νk = w (λ + ρΣ ) − ρΣ for some w ∈ W . Indeed, ν < νk , therefore νk ∈ / F ; since νk ∈ T and ch VΣ (λ) − U (λ + ρΣ ) does not contain the term MΣ (νk ) with nonzero coefficient, mβ β, νk + ρΣ = w (λ + ρΣ ) = λ + ρΣ − β∈B

for some mβ ∈ Z≥0 (see Remark 12.7). On the other hand, by Lemma 12.3, ν + ρΣ = νk + ρΣ −

mβ β

β∈B

for some base Σ obtained from Σ by odd reflections, ν such that LΣ (ν ) ∼ = LΣ (ν) and some mβ ∈ Z≥0 . As a result ν + ρΣ = λ + ρΣ −

mβ β,

mβ ∈ Z≥0 .

β∈B

By Lemma 13.4, ν + ρΣ , hβ ≥ 0 for all β ∈ B ∩ Σ and ν + ρΣ , hβ ≥ −1 for β ∈ B Σ. Since λ + ρΣ is regular and λ is integrable, (λ + ρΣ + ν + ρΣ , β) > 0 for all β ∈ B ∩Σ, and (λ + ρΣ + ν + ρΣ , β) ≥ 0 for β ∈ B Σ. The condition (λ + ρΣ , λ + ρΣ ) = (ν + ρΣ , ν + ρΣ ) implies λ + ρ Σ + ν + ρΣ , mβ β = 0. β∈B

If (λ + ρΣ + ν + ρΣ , β) > 0 for β ∈ B Σ, then ν + ρΣ = λ + ρΣ . Then, as in the proof of Lemma 14.3, one can show that ν = λ − α, but we already proved that the weight λ − α has multiplicity one in VΣ (λ), as well as in U (λ + ρΣ ). Hence ν = λ − α. Therefore, (λ + ρΣ + ν + ρΣ , β) = 0 for β ∈ B Σ; thus mβ = 0 for β ∈ B ∩ Σ, and λ + ρΣ , hβ = 1, ν + ρΣ , hβ = −1 (hence mβ = 1) for β ∈ B Σ. Therefore, ν + ρΣ = rβ (λ + ρΣ ) = λ + ρΣ − β. Without loss of generality, assume that Σ = rαk . . . rα1 (Σ). Then rβ (λ + ρΣ ) = ν + ρΣ = ν + ρΣ + αi1 + · · · + αir .


215

Assume r = 0; then (rβ (λ + ρΣ ) , αir ) = 0. By our assumption in the beginning of the proof, there are only two possibilities: (λ + ρΣ , α1 ) = 0 or (λ + ρΣ , γ1 ) = 0. The first case is impossible, because then we have rβ (α1 ) = −γ1 , (rβ (λ + ρΣ ) , γ1 ) = 0, and by regularity of rβ (λ + ρΣ ), αir = γ1 . In the second case, we have αir = α1 and ν = λ − β − α1 . Since (ρΣ , β) = 0, ν (hβ ) = λ (hβ ) − β (hβ ) − α1 (hβ ) = 1 − 2 − 2 (α1 , β)/(β, β) = −2, and that contradicts the condition rβ (ν) ≤ ν. Corollary 14.5. For any typical integrable λ the character of L (λ) is given by (22). Corollary 14.6. Let LΣ (λ) be integrable, and λ+ ρΣ be regular. Assume that (λ + ρΣ , α) = 0 for some isotropic α ∈ Σ. Then ch LΣ (λ) + ch LΣ (λ − α) = U (λ + ρΣ ) . Theorem 14.7. Let λ be an atypical integrable weight, Π be such that λ + ρ is regular, and there exists an isotropic α ∈ Π such that (λ + ρ, α) = 0. Then w λ+ρ D1 (−1) w e w∈W 1+e−α ch L (λ) = (24) D0 Proof. Without loss of generality, assume that α = α1 = −γ0 . We start with constructing a sequence (λi , Σi , k (i)), where Σi = rγk(i) . . . rγ1 Π, λi is a weight such that λi + ρΣi , γk(i) = 0, and λi + ρΣi is regular. This sequence is defined uniquely by the following rule: 1. 2. 3. 4.

(λ0 , Σ0 , k (0)) = (λ, Π, 0); λi+1 = λi + γk(i) ; Σi+1 = Σi , k (i + 1) = k (i) if λi+1 + ρΣi is regular; if λi+1 + ρΣi is not regular, set, using Lemma 12.6, Σi+1 = rγk(i+1) . . . rγk(i)+1 (Σi )

such that LΣi (λi+1 ) ∼ = LΣi+1 (λi+1 ), and λi+1 + ρΣi+1 is regular. These rules ensure that λi + ρΣi , γk(i) = 0 and λi + ρΣi is regular. Now define D1 1 w λi +ρΣi . χi = ch LΣi (λi ) , ϕi = (−1) w e D0 1 + eγk(i) w∈W

216

Vera Serganova

(The quotients are taken in R ; obviously, they make sense.) The series converges by the same reasons as for the series for U . We want to show that χi = ϕi ; they are both elements of R. Our next step is to prove the identity ϕi + ϕi+1 = U (λi + ρΣi ) . If Σi+1 = Σi , then the identity is straightforward. Otherwise, by Lemma 12.6 (with Σ = Σi , Σ = Σi+1 , λ = λi+1 , k = s − 1) λi+1 + ρΣi+1 = g λi+1 + ρΣi + (s − 1) γk(i) = g λi + ρΣi + sγk(i) , where s is the positive integer such that λi + ρΣi + tγk(i) is not regular for any g s+1 1 ≤ t < s; and g ∈ W is such that γk(i+1) = g γk(i) and (−1) = (−1) . Thus, we obtain ϕi + ϕi+1 D1 D1 1 w g(λi +ρΣi +sγk(i) ) = (−1) w eλi +ρΣi + e D0 1 + eγk(i) 1 + eg(γk(i) ) w∈W w s+1 λi +ρΣ +sγk(i) D1 λi +ρΣi i + (−1) e γk(i) w∈W (−1) w e 1+e = D0 " s−1 ! s−1 D1 w λi +ρΣi +tγk(i) = = (−1) w e U λi + ρΣi + tγk(i) . D0 t=0 t=0 w∈W

However, λi + ρΣi + tγk(i) is not regular for all 1 ≤ t < s. Hence U λi + ρΣi + tγk(i) = 0, t > 0; we obtain the desired identity. Now recall that by Corollary 14.6 the similar identity holds for χi : χi + χi+1 = U (λi + ρΣi ) . Therefore we can conclude that there exists Φ ∈ R such that i

χi = ϕi + (−1) Φ. We want to show now that Φ = 0. It suffices to prove that for every ν ∈ h∗ the monomial eν appears with nonzero coefficient only in finitely many ϕi and in finitely many χi . We claim that if eν appears in χi (or ϕi ) with non-zero coefficient, then ν ≤Σi λi . For χi it follows from the fact that ν must be a weight of LΣi (λi ). To check this in case of ϕi , note that every term which appears in ϕi appears in the expression for U (λi + ρΣi ), so the same argument as for U works. Recall the Z-grading on Q defined in (15), (16). The condition ν ≤Σi λi can be rewritten in the form (see Lemma 12.1 for Π = Σi )


λi − ν = −mi γk(i) +

miβ β,

if λi − ν ∈ Q+ ,

miβ β,

if λi − ν ∈ Q− ;

217

β∈B

λi − ν = mi γk(i)+1 +

β∈B

miβ

here mi , ∈ Z≥0 . Note that if λi ∈ Qj , then λi+1 ∈ Qj−1 . Hence for sufficiently large i only the second case is possible. Then λi+1 − ν = mi γk(i)+1 + γk(i) + miβ β. β∈B

Now rewrite it in a suitable form λi+1 − ν = (mi + 1) γk(i+1)+1 + miβ β + mi γk(i)+1 − γk(i+1)+1 + γk(i) − γk(i+1)+1 . β∈B

However, γk(i)+1 − γk(i+1)+1 is either zero or a negative real root, γk(i) − γk(i+1)+1 is always a negative even root. Thus β∈B

mi+1 ≤ β

miβ − 1.

β∈B

Hence, for sufficiently large i, β∈B miβ becomes negative, therefore ν ≤Σi λi does not hold. The theorem is proven. Remark 14.8. Under conditions of Theorem 14.7, it is not hard to show that the complex · · · → VΣi+1 (λi+1 ) → VΣi (λi ) → · · · → VΣ0 (λ0 ) → 0 with arrows defined by Corollary 14.6 is a resolution of L (λ). Kac and Wakimoto call a representation with highest weight satisfying the condition of Theorem 14.7 tame. Character formula (24) coincides with (1) the Kac–Wakimoto conjectural formula in [12] in case g = sl (1|n) or osp (2|2n)(1) . On the other hand, using Lemma 12.6, one can construct a chain of odd reflections transforming any atypical integrable weight (except weights of one-dimensional representations) to a weight satisfying the conditions of the last theorem. Therefore, we found character formulae for all (1) (1) integrable simple highest weight modules for sl (1|n) and osp (2|2n) . It seems possible that using the Shapovalov form calculated in [16] as a substitute for a missing invariant symmetric form, one can obtain similar formulae for S (1, 2; b).

218

Vera Serganova

References 1. V. G. Kac, Lie superalgebras, Adv. Math. 26 (1977), 8–96. 2. Victor G. Kac, Infinite-dimensional Lie Algebras, Third edition. Cambridge University Press, Cambridge, 1990. 3. C. Hoyt, Kac-Moody superalgebras of finite growth, Ph.D. thesis, UC Berkeley, Berkeley, 2007. 4. V. Serganova, Characters of simple Lie superalgebras, Proceedings of ICM, 1998, pp. 583–594. 5. V. V. Serganova, Automorphisms of simple Lie superalgebras, Izvestiya Akademii Nauk SSSR. Seriya Matematicheskaya 48 (1984), no. 3, 585–598 (Russian). 6. Victor G. Kac, Representations of classical Lie superalgebras, Differential geometrical methods in mathematical physics, II (Berlin), Lecture Notes in Math., vol. 676, Proc. Conf., Univ. Bonn, Bonn, 1977, Springer, 1978, pp. 597–626. 7. V. G. Kac and J. van de Leur, On classification of superconformal algebras, Strings’88, World Sci. Publ. Teaneck, NJ, 1089, pp. 77–106. 8. J. van de Leur, A classification of contragredient Lie superalgebras of finite growth, Comm. Algebra 17 (1989), 1815–1841. 9. C. Hoyt and V. Serganova, Classification of finite-frowth general Kac-Moody superalgebras, Comm. Algebra 35 (2007), 851–874. 10. V. G. Kac, Infinite-dimensional algebras, Dedekind’s η-function, classical Mobius function and the very strange formula, Advances in Mathematics 30 (1978), no. 2, 85–136. 11. V. G. Kac and M. Wakimoto, Integrable highest weight modules over affine superalgebras and number theory, Progress in Math. 123 (1994), 415–456. 12. V.G. Kac and M. Wakimoto, Integrable highest weight modules over affine superalgebras and Appel’s function, Commun. Math. Phys. 215 (2001), 631–682. 13. D. A. Leites, M. V. Savelev, and V. V. Serganova, Embeddings of osp(n/2) and the associated nonlinear supersymmetric equations, Group theoretical methods in physics (Utrecht), vol. I, Yurmala, 1985, VNU Sci. Press, 1986, pp. 255–297. 14. Joseph N. Bernstein and D. A. Leites, A formula for the characters of the irreducible finite-dimensional representations of Lie superalgebras of series Gl and sl, Doklady Bolgarskoi Akademii Nauk. Comptes Rendus de l’Academie Bulgare des Sciences 33 (1980), no. 8, 1049–1051 (Russian). 15. Jonathan Brundan, Kazhdan-Lusztig polynomials and character formulae for the Lie superalgebra gl(m, n), J. of AMS 16 (2003), no. 1, 185–231. 16. V.G. Kac and M. Wakimoto, Quantum reduction and representation theory of conformal superalgebras, Adv. Math. 185 (2004), 400–458.

Part B

Geometry of Infinite-Dimensional Lie (Transformation) Groups

Jordan Structures and Non-Associative Geometry Wolfgang Bertram Institut Elie Cartan, Université Nancy I, Faculté des Sciences, B.P. 239, 54506 Vandœuvre-lès-Nancy, Cedex, France, [email protected] Summary. We give an overview of constructions of geometries associated to Jordan structures (algebras, triple systems and pairs), featuring analogues of these constructions with the Lie functor on one hand and with the approach of non-commutative geometry on the other. Key words: Jordan pair, Lie triple system, graded Lie algebra, filtered Lie algebra, generalized projective geometry, flag geometries, (non-) associative algebra and geometry. 2000 Mathematics Subject Classifications: 17C37, 17B70, 53C15.

Introduction Let us compare two aspects of the vast mathematical topic “links between geometry and algebra”: on the one hand, the Lie functor establishes a close relation between Lie groups (geometric side) and Lie algebras (algebraic side); this is generalized by a correspondence between symmetric spaces and Lie triple systems (see [Lo69]). On the other hand, the philosophy of noncommutative geometry generalizes the relation between usual, geometric point spaces M (e.g., manifolds) and the commutative and associative algebra Reg(M, K) of “regular” (e.g., bounded, smooth, algebraic, according to the context) K-valued functions on M , where K = R or C, by replacing the algebra Reg(M, K) by more general, possibly non-commutative algebras A. The interaction between these two aspects seems to be rather weak; indeed, in some sense they are “orthogonal” to each other. First, the classical setting of Lie’s third theorem is finite dimensional, whereas the algebras A of commutative or non commutative geometry are typically infinite dimensional; second, and more importantly, taking a commutative and associative algebra as input, we obtain as a Lie bracket [x, y] = xy − yx = 0, and hence (if we forget the associative structure and retain only the Lie bracket) we are left with constructing a Lie group with a zero Lie bracket. But this is not very interesting: K.-H. Neeb and A. Pianzola (eds.), Developments and Trends in Infinite-Dimensional Lie Theory, Progress in Mathematics 288, DOI 10.1007/978-0-8176-4741-4_7, © Springer Basel AG 2011

221

222

Wolfgang Bertram

it does not even capture the specific information encoded in “commutative geometries”.1 This remark suggests that, if one looks for a link between the two aspects just mentioned, it should be interesting to ask for an analogue of the Lie functor for a class of algebras that faithfully contains the class of commutative associative algebras, but rather sacrifices associativity instead of commutativity. At this point let us note that the algebras A of non-commutative geometry are of course always supposed to be associative, so that the term “associative geometry” might be more appropriate than “non-commutative geometry”. Indeed, behind the garden of associative algebras starts the realm of general algebras, generically neither associative nor commutative, and, for the time being, nobody has an idea of what their “geometric interpretation” might be. Fortunately, two offsprings of associative algebras grow not too far behind the garden walls: Lie algebras right on the other side (the branches reach over the wall so abundantly that some people even consider them as still belonging to the garden), and Jordan algebras a bit further. Let us just recall that the former are typically obtained by skew-symmetrizing an associative algebra, [x, y] = xy − yx, whereas the latter are typically obtained by symmetrizing them, x • y = 12 (xy + yx) (the factor 12 being conventional, in order to obtain the same powers xk as in the associative algebra). So let us look at Jordan algebras—their advantage being that the class of commutative associative algebras is faithfully embedded. In this survey paper we will explain their geometric interpretation via certain generalized projective geometries, emphasizing that this interpretation really combines both aspects mentioned above. Of course, it is then important to include the case of infinite-dimensional algebras, and to treat them in essentially the same way as finite-dimensional ones. This is best done in a purely algebraic framework, leaving aside all questions of topologizing our algebras and geometric spaces. Such questions form an interesting topic for the further development of the theory: in a very general setting (topological algebraic structures over general topological fields or rings) basic results are given in [BN05], and many results from the more specific setting of Jordan operator algebras (Banach– Jordan structures over K = R or C; see [HOS84]) admit interesting geometric interpretations in our framework. Beyond the Banach setting, it should be interesting to develop a theory of locally convex topological Jordan structures and their geometries, taking up the associative theory (cf. [Bil04]). The general construction for Jordan algebras (and for other members of the family of Jordan algebraic structures, namely Jordan pairs and Jordan triple systems, which in some sense are easier to understand than Jordan algebras; see Section 1) being explained in the main text (Section 2 and Section 3), here 1

Of course, this does not exclude other ways of associating Lie groups to function algebras from being interesting, for example by looking at their derivations and automorphisms; but our point is precisely that such constructions need more input than the trivial Lie bracket on the algebra of functions.

Jordan Structures and Non-Associative Geometry

223

let us just look at the special case of “commutative geometry”, i.e., the case of the commutative and associative algebra A = Reg(M, K) of “regular K-valued functions” on a geometric space M (here K is a commutative field with unit 1; in the main text we will also allow commutative rings with 1). Looking at A as a Jordan algebra, we associate to it the space X := Reg(M, KP1 ) of “regular functions from M into the projective line KP1 ”. Of course, X is no longer an algebra or a vector space; however, we can recover all these structures if we want: we call two functions f, g ∈ X “transversal” and write f g, if f (p) = g(p) for all points p ∈ M . Now choose three functions f0 , f1 , f∞ ∈ X that are mutually transversal. For each point p ∈ M , the value f∞ (p) singles out an “affine gauge” (by taking f∞ (p) as a point at infinity in the projective line KP1 and looking at the affine line KP1 \ {f∞ (p)}); then f0 (p) singles out a “linear gauge” (origin in the affine line) and f1 (p) a “unit gauge” (unit element in the vector line). Obviously, any such choice of f0 , f1 , f∞ leads to an identification of the set (f∞ ) of all functions that are transversal to f∞ with the usual algebra A of regular functions on M . In other words, A and (X ; f0 , f1 , f∞ ) carry the same information, and thus we see that the encoding via A depends on various choices, from which we are freed by looking instead at the “geometric space” X . It thus becomes evident that A really is a sort of “tangent algebra” of the geometric (“non-flat”) space X at the point f0 , in a similar way as the Lie algebra of a Lie group reflects its tangent structure at the origin. This way of looking at ordinary “commutative geometry” leads to generalizations that are different from non-commutative geometry [Co94] but still have much in common with it. For instance, both theories have concepts of “states”, i.e., a notion that, in the commutative case, amounts to recovering the point space M from the algebra of regular functions. However, in the case of non-commutative geometry, this concept relies heavily on positivity (and hence on the ordered structure of the base field R), and the corresponding concept for generalized projective geometries is purely geometric and closely related to the notion of inner ideals in Jordan theory (Section 4). There are many interesting open problems related to these items, some of which are mentioned in Section 4. Summing up, it seems that the topic of geometrizing Jordan structures, incorporating both ideas from classical Lie theory and basic ideas from associative geometry, is well suited for opening the problem of “general nonassociative geometry”, i.e., the problem of finding geometries corresponding to more general non-associative algebraic structures. Remark (added in may 2010): While this paper was being processed, the joint work [BeKi09] with Michael Kinyon gave a new turn to the whole question. We cannot go into details here, and we strongly suggest the interested reader to have a look at that work. Acknowledgements: This paper is partly based on notes of a series of lectures given at the Université Paul Verlaine (Metz university) in April 2007, and I would like to thank Said Benayadi for inviting me on this occasion.

224

Wolfgang Bertram

Notation: Throughout, K is a commutative base ring with unit 1 and such that 2 and 3 are invertible in K.

1 Jordan pairs and graded Lie algebras 1.1 Z/(2)-graded Lie algebras and Lie triple systems Let (Γ, +) be an abelian group. A Lie algebra g over K is called Γ -graded if it is of the form g= gn , with [gm , gn ] ⊂ gn+m . n∈Γ

Let us consider the example Γ = Z/(2). Here, g = g0 ⊕ g1 , with a subalgebra g0 and a subspace g1 which is stable under g0 and such that [g1 , g1 ] ⊂ g0 . It is then easily seen that the linear map σ : g → g which is 1 on g0 and −1 on g1 is an involution of g (automorphism of order 2). Conversely, every involution gives rise to a Z/(2)-grading of g, and hence such gradings correspond bijectively to symmetric Lie algebras (g, σ), i.e., to Lie algebras g together with an automorphism σ of order 2. Then g1 (the −1-eigenspace of σ) is stable upon taking triple Lie brackets [[x, y], z]. This leads to the notion of a Lie triple system. Definition 1.1. A Lie triple system (LTS) is a K-module q together with a trilinear map q × q × q → q, (X, Y, Z) → [X, Y, Z], or, equivalently, with a bilinear map R : q × q → End(q),

(X, Y ) → R(X, Y ) =: [X, Y, ·],

such that (LT1) R(X, X) = 0 (LT2) R(X, Y )Z + R(Y, Z)X + R(Z, X)Y = 0 (the Jacobi identity) (LT3) the endomorphism D := R(X, Y ) is a derivation of the trilinear product, i.e., D[U, V, W ] = [DU, V, W ] + [U, DV, W ] + [U, V, DW ]. If g is a Z/(2)-graded Lie algebra, then g1 with [X, Y, Z] := [[X, Y ], Z] becomes an LTS, and every LTS arises in this way since one may reconstruct a Lie algebra from an LTS via the standard imbedding (see [Lo69]): let h ⊂ Der(q) be the subalgebra of derivations of the LTS q spanned by all operators R(X, Y ), X, Y ∈ q, and define a bracket on g := q ⊕ h by [(X, D), (Y, E)] := DY − EX, [D, E] − R(X, Y ) . One readily checks that g is a Z/(2)-graded Lie algebra. As a side remark, one may note that this construction (the standard imbedding) is in general not functorial. See [Sm05] for a discussion of this topic.


225

1.2 3-graded Lie algebras and Jordan pairs A Lie algebra is called 2n + 1-graded if it is Z-graded, with gj = 0 if j ∈ / {−n, −n + 1, . . . , n}. The linear map D : g → g with Dx = jx for x ∈ gj is then a derivation of g; if it is an inner derivation, D = ad(E), the element E ∈ g is called an Euler element. To any Z-graded algebra one may associate a Z/(2)-grading g = geven ⊕ godd , by putting together all homogeneous parts with even, resp. odd index. In the sequel we are mainly interested in 3-graded and 5-graded Lie algebras; in both cases, godd = g1 ⊕ g−1 . We then let V ± := g±1 and define trilinear maps by T ± : V ± × V ∓ × V ± → V ±,

(x, y, z) → [[x, y], z].

The maps T± satisfy the identity (LJP2) T ± (u, v, T ± (x, y, z)) = T ± (T ± (u, v, x), y, z) − T ± (x, T ∓ (v, u, y), z) + T ± (x, y, T ± (u, v, z)). Indeed, this is just another version of the identity (LT3), reflecting the fact that ad[u, v] is a derivation of g respecting the grading since [u, v] ∈ g0 . In the 3-graded case, we have moreover, for all x, z ∈ V ± , y ∈ V ∓ : (LJP1) T ± (x, y, z) = T ± (z, y, x). This follows from the fact that g1 and g−1 are abelian, and so by the Jacobi identity [[x, y], z] − [[z, y], x] = [[z, x], y] = 0. Definition 1.2. A pair of K-modules (V + , V − ) together with trilinear maps T ± : V ± × V ∓ × V ± → V ± is called a (linear) Jordan pair if (LJP1) and (LJP2) hold. (The term “linear” refers to the fact that the identities (LJP1) and (LJP2) are linear in each of their variables, whereas the definition given in [Lo75], which is also valid in the case where 2 or 3 are not invertible in K, is based on identities that are quadratic in some of their variables.) Every linear Jordan pair arises by the construction just described: if (V + , V − ) is a linear Jordan pair, let q := V + ⊕ V − and define a trilinear map T˜ : q3 → q via T˜ (x, x ), (y, y ), (z, z ) := T + (x, y , z), T −(x , y, z ) . Then the map T˜ satisfies the same identities as (T + , T − ), with T + and T − both replaced by T˜ (in other words, (q, T˜) is a Jordan triple system, see below). Now we define a trilinear bracket q3 → q by (x, x ), (y, y ), (z, z ) := T˜ ((x, x ), (y, y ), (z, z )) − T˜ ((y, y ), (x, x ), (z, z )).

226

Wolfgang Bertram

By a direct calculation (cf. Lemma 1.4 below), one checks that q with this bracket is an LTS, and its standard imbedding g = q ⊕ h with g±1 = V ± and g0 = h, is a 3-graded Lie algebra. Without loss of generality we may assume that g contains an Euler element E: in fact, the endomorphism E : q → q which is 1 on V + and −1 on V − is a derivation of q commuting with h, and hence we may replace h by h + KE in the construction of the standard imbedding. For 5-graded Lie algebras, the identity (LJP1) has to be replaced by another, more complicated identity, which leads to the notion of a Kantor pair, see [AF99]. As for Jordan pairs, Kantor pairs give rise to LTSs of the form q = V + ⊕ V − , where now the standard imbedding leads back to a 5-graded Lie algebra. 1.3 Involutive Z-graded Lie algebras An involution of a Z-graded Lie algebra is an automorphism θ of order 2 such that θ(gj ) = g−j for j ∈ Z. If g is 3- or 5-graded, we let V := g1 and define T : V × V × V → V,

(X, Y, Z) → [[X, θY ], Z].

Then T satisfies the identity (LJT2) obtained from (LJP2) by omitting the indices ±, and if g is 3-graded, we also have the analogue of (LJP1). Definition 1.3. A K-module V with a trilinear map T : V 3 → V satisfying the identities (JP1) and (JP2) obtained from (LJP1) and (LJP2) by omitting indices is called a (linear) Jordan triple system (JTS). Every linear JTS arises by the construction just described: just let V + := V := V and T ± := T ; then (V + , V − ) is a Jordan pair carrying an involution (= isomorphism τ from (V + , V − ) onto the Jordan pair (V − , V + ), namely τ (x, x ) = (x , x)). Let g be the 3-graded Lie algebra associated to this Jordan pair; then the involution τ of the Jordan pair induces an involution θ of g. By the way, these arguments show that JTSs are nothing but Jordan pairs with involution (cf. [Lo75]). For any JTS (V, T ), the Jordan pair (V + , V − ) = (V, V ) with T ± = T is called the underlying Jordan pair. −

The Lie algebra g now carries two automorphisms of order 2, namely θ and the automorphism σ corresponding to the Z/(2)-grading g = geven ⊕ godd . These two automorphisms commute. In particular, σ restricts to the θ-fixed subalgebra gθ , with the −1-eigenspace being {X + θ(X)| X ∈ g1 }. The LTSstructure of this space is described by the following lemma. Lemma 1.4. (The Jordan–Lie functor.) If T is a JTS on V , then [X, Y, Z] = T (X, Y, Z) − T (Y, X, Z) defines an LTS on V .


227

Proof. The proof of (LT1) is trivial, and (LT2) follows easily from the symmetry of T in the outer variables. In order to prove (LT3), we define the endomorphism R(X, Y ) := T (X, Y, ·) − T (Y, X, ·) of V . Then R(X, Y ) is a derivation of the trilinear map T , as follows from the second defining identity (JT2). But every derivation of T is also a derivation of R, since R is simply defined by skew-symmetrization of T in the first two variables. The lemma defines a functor from the category of JTS to the category of LTS over K, which we call the Jordan–Lie functor. In general, it is neither injective nor surjective; all the more surprising is the fact that, in the real, simple and finite-dimensional case, the Jordan–Lie functor is not too far from setting up a one-to-one correspondence (classification results due to E. Neher, cf. tables given in [Be00]). 1.4 The link with Jordan algebras Whereas all results from the preceding sections are naturally and easily understood from the point of view of Lie theory, the results presented next are “genuinely Jordan theoretic”—by this we mean that proofs by direct calculation in 3-graded Lie algebras are much less straightforward. The reader may try some of these proofs, in order to better appreciate the examples to be given in the next section. In the following, g is a 3-graded Lie algebra, with associated Jordan pair (V + , V − ) = (g1 , g−1 ). For x ∈ V − , we define a K-linear map Q− (x) : V + → V − ,

y → Q− (x)y :=

1 − 1 T (x, y, x) = − [x, [x, y]]. 2 2

In the same way Q+ (y) for y ∈ V + is defined. Then the following hold. (1) The Fundamental Formula. For all x ∈ V − , y ∈ V + , Q− (Q− (x)y) = Q− (x)Q+ (y)Q− (x). (2) Meyberg’s theorem. Fix a ∈ V − . Then Va := V + with product x •a y :=

1 + T (x, a, y) 2

is a Jordan algebra (called the a-homotope algebra). Recall (see, e.g., [McC04]) that a (linear) Jordan algebra is a K-module J with a bilinear and commutative product x • y such that the identity (J2) x • (x2 • y) = x2 • (x • y) holds. (3) Invertible elements and unital Jordan algebras. An element a ∈ V − is called invertible in (V + , V − ) if the operator Q− (a) : V + → V − is invertible.

228

Wolfgang Bertram

The element a ∈ V − is invertible in (V + , V − ) if, and only if, the Jordan algebra Va described in the preceding point admits a unit element; this unit element is then a := (Q(a))−1 (a). Moreover, the Jordan pair (V + , V − ) can then be recovered from the Jordan algebra (V, •) = (Va , •a ) as follows: let V + := V − := V as K-modules and let T : V 3 → V , T (x, y, z) := (x • y) • z − y • (x • z) + x • (y • z). Then (V, T ) is a JTS, and its underlying Jordan pair (V + , V − ) is the one we started with. 1.5 Some examples (1) Jordan pairs of rectangular matrices. Let A and B be two K-modules and W := A ⊕ B. Let I : W → W be the linear map that is 1 on A and −1 on B. Then the Lie algebra g := glK (W ) is 3-graded, with, for i = 1, 0, −1, gj = {X ∈ g| [I, X] = 2jX}. In an obvious matrix notation, this corresponds to describing g by 2 × 2-matrices, g0 as diagonal and g1 as upper and g−1 as lower triangular matrices: g0 V + . V − g0 Therefore,

(V + , V − ) = (HomK (B, A), HomK (A, B))

carries the structure of a linear Jordan pair. The Jordan pair structure is given by T ± (X, Y, Z) = XY Z + ZY X. In fact, this is proved by the following calculation (using matrix notation for elements of g): XY 0 Z 0 XY Z + ZY X 0 0 0 0Z 0X . = , = , , 0 0 0 0 0 −Y X 0 0 Y 0 0 0 The proof of the Fundamental Formula for this Jordan pair is easy since we may calculate in the associative algebra HomK (W, W ), where Q(x) = x ◦ rx is simply composition of left and of right translation by x. Similarly, the proof of Meyberg’s theorem becomes an easy exercise: for any a ∈ HomK (W, W ), the product (x, y) → xay is again an associative product, and by symmetrizing and restricting to V + , if a ∈ V − , we get the Jordan algebra •a . In general, the Jordan pair (V + , V − ) does not contain invertible elements. For instance, if K is a field and A and B are not isomorphic as vector spaces, then it is easily seen that Q− (x) is never bijective. Finally, it is clear that, if A and B are isomorphic as K-modules, and fixing such an isomorphism in order to identify A and B, an element a ∈ V + = A = V −


229

is invertible if it is invertible in the usual sense in Hom(A, A), and then •a has a−1 as a unit element. Note that, in particular, this happens for the Lie algebra gl(2, K), with V + ∼ =V− ∼ = K, where x •a y = xay = yax. (2) Function spaces. As for any algebraic structure defined by identities, spaces of functions with values in such a structure, equipped with the “pointwise product”, again form a structure of the given type. In our case, if M is a set and (V + , V − ) a Jordan pair, then (Fun(M, V + ), Fun(M, V − )) is again a Jordan pair, and similarly for JTSs and Jordan algebras. Moreover, if (V + , V − ) corresponds to a 3-graded Lie algebra g, then the function space corresponds to the 3-graded Lie algebra Fun(M, g). In particular, pairs of scalar functions (Fun(M, K), Fun(M, K)) form a Jordan pair with pointwise product (T (f, g, h))(p) = 2f (p) g(p) h(p). Under suitable assumptions, “regular” functions (continuous, smooth, algebraic, . . .) will form subpairs. (3) The classical examples. Besides rectangular matrices, these include: • full associative algebras A. Here (V + , V − ) = (A, A), and g ⊂ gl(2, A) is the subalgebra generated by the strictly upper triangular and strictly lower triangular matrices; • Hermitian elements of an involutive associative algebra (A, ∗). Here, (V + , V − ) = (Ah , Ah ) is the fixed space of ∗, and g is the symplectic algebra of (A, ∗) (cf. [BN04], Section 8.2). Note that symmetric and Hermitian matrices are a special case; • skew-Hermitian elements of an involutive associative algebra (A, ∗): similar to that above, replacing ∗ by its negative. As above, Jordan pairs of skew-symmetric or skew-Hermitian matrices are a special case; • conformal geometries (or “spin factors”). Here, V + = V − = V is a K-module with a non-degenerate symmetric bilinear form (·|·) : V × V → K and trilinear map T (x, y, z) = (x|z)y − (x|y)z − (z|y)x . Then g is the orthogonal Lie algebra of the quadratic form on K ⊕ V ⊕ K given by β((r, v, s), (r, v, s)) = rs + (v|v) . There are also exceptional Jordan systems, namely octonionic 1 × 2-matrices, resp. the Hermitian octonionic 3 × 3-matrices, corresponding to 3-gradings of the exceptional Lie algebras of type E6 , resp. E7 .

2 The generalized projective geometry of a Jordan pair 2.1 The construction The following construction of a pair (X + , X − ) of homogeneous spaces associated to a Jordan pair (V + , V − ) is due to J.R. Faulkner and O. Loos

230

Wolfgang Bertram

(see [Lo95]). Starting with a (linear) Jordan pair (V + , V − ), let g be the 3-graded Lie algebra g constructed in Section 1.2; as explained there, we may assume that g contains an Euler element E, i.e., an element such that [E, X] = iX for X ∈ gi . Next we define two abelian groups U ± = exp(g±1 ) by observing that, for X ∈ g±1 , the operator ad(X) : g → g is 3-step nilpotent, and that 1 exp(X) := ead(X) = id + ad(X) + ad(X)2 2 is an automorphism of g (for this we need that 2 and 3 are invertible in K). Since g±1 are abelian, U ± are abelian subgroups of Aut(g). The maps exp± : g± → Aut(g) are injective (here we use our Euler operator: exp± (X) = id implies exp± (X)E = E ± X = E, whence X = 0). Thus U ± is isomorphic to (V ± , +). The elementary projective group associated to the Jordan pair (V + , V − ) is the subgroup G := PE(V + , V − ) := U + , U − of Aut(g) generated by U + and U − . Let H be the subgroup of G stabilizing the grading (i.e., commuting with ad(E)); then U + and U − are normalized by H, and the groups P ± generated by H and U ± are semidirect products: P ± := H, U ± ∼ = H V ±. Finally, we define two homogeneous spaces X ± := G/P ∓ with base points o± = e/P ∓ (e being the unit element of G). The pair of “geometric spaces” (X + , X − ) is called the generalized projective geometry associated to (V + , V − ). It is indeed the geometric object associated to the Jordan pair (V + , V − ), in a similar way as a Lie group is the geometric object associated to a Lie algebra. Let us explain this briefly. 2.2 Generalized projective geometries The spaces X ± being defined as above, the direct product X + ×X − is equipped with a transversality relation: call (x, α) ∈ X + × X − transversal, and write xα, if they are conjugate under G to the base point (o+ , o− ). It is easy to describe the set α := {x ∈ X + | xα} : if α = o− is the base point (whose stabilizer is P + ), then this is the P + -orbit P + .o+ which is isomorphic to V+ ∼ = U + .o+ . Note here that the set α carries a canonical structure of an affine space over K, since the vector group U + acts on it simply transitively. By homogeneity, the same statements hold for any α ∈ X − . This observation leads to the following definition. Definition 2.1. A pair geometry is a pair of sets X = (X + , X − ) together with a binary relation ⊂ X + × X − (called transversality, and we write xα for (x, α) ∈ ) such that X ± is covered by subsets of the form


α = {x ∈ X ± | xα},

231

α ∈ X ∓.

A linear pair geometry (over K) is a pair geometry (X + , X − , ) such that, for any transversal pair xα, the set α is equipped with a structure of linear space over K (i.e., a K-module), with origin x, and the same property holds by exchanging the rˆ oles of X + and X − . An affine pair geometry is a linear pair geometry such that the underlying affine structure of the linear space (α , x) does not depend on x, for all transversal pairs (x, α), and dually. In other words, for all α ∈ X − , the set α carries the structure of an affine space over K, and dually. The discussion above shows that, to any Jordan pair (V + , V − ) over K, we can associate an affine pair geometry (X + , X − , ). The link with triple products is given by introducing the structure maps of a linear pair geometry: if xα, yα, zα and r ∈ K, then let rx,α (y) := ry denote the product r · y, and y +x,α z the sum of y and z in the K-module α with zero vector x. In other words, we define maps of three (resp. four) arguments by Pr : (X + × X − × X + ) → X + , (x, α, y) → Pr (x, α, y) := rx,α (y), + S : (X × X − × X + × X + ) → X + , (x, α, y, z) → S(x, α, y, z) := y +x,α z, where the domain of Pr is the “space of transversal triples”, (X + × X − × X + ) = {(x, α, y) ∈ X + × X − × X + | xα, yα}, and the domain of S is the similarly defined space of generic quadruples. Of + course, we should write P+ instead of S; dually, the r instead of Pr and S − − maps Pr and S are then also defined. The structure maps encode all the information of a linear pair geometry: by fixing the pair (x, α), the structure maps describe the linear structure of (α , x), resp. of (x , α). In this way, linear pair geometries can be regarded as objects whose structure is defined by (one or several) “multiplication maps”, just like groups, rings, modules, symmetric spaces, etc. This point of view naturally leads to the question of whether there are more specific “identities” satisfied by the structure maps. If (X + , X − ; ) is the geometry associated to a Jordan pair, then this is indeed the case. There are two such identities, denoted by (PG1) and (PG2) in [Be02]. The first identity (PG1) can be seen as an “integrated version” of the defining identity (LJP2) of a Jordan pair; it implies the existence of a “big” automorphism group, whereas (PG2) is rather a global version of the Fundamental Formula and implies the existence of “structural maps” exchanging the two partners X + and X − (we refer the reader to [Be02] for details). The important point about these identities is that we can also go the other way: by “deriving” them in a suitable way, one can show that any affine pair geometry satisfying (PG1) and (PG2) in all scalar extensions gives rise to a Jordan pair (V + , V − ) as a “tangent geometry” with respect to a fixed base point (o+ , o− ).

232

Wolfgang Bertram

This construction clearly parallels the correspondence between Lie groups and Lie algebras (and even more closely the one between symmetric spaces and Lie triple systems [Lo69]; see also [Be08]). However, in contrast to Lie theory, the constructions are algebraic in nature and therefore work much more generally in arbitrary dimension and over general base fields and rings. They define a bijection (even an equivalence of categories) between Jordan pairs and connected generalized projective geometries with base point and, in the same way, between Jordan triple systems (resp. unital Jordan algebras) and generalized polar (resp. null) geometries, i.e., generalized projective geometries with the additional structure of a certain kind of involution, cf. [Be02], [Be03]. Thus, in principle, all Jordan-theoretic notions can be translated into geometric ones; in general, it is not at all obvious what the correct translation should be, but once it has been found, it often sheds new light on the algebraic notion. In Section 4 we illustrate this with the example of the notion of inner ideals and their complementation relation. 2.3 The geometric Jordan–Lie functor Symmetric spaces are well-known examples of “non-associative geometries” (cf. [Lo69]): they are manifolds M together with a point reflection σx attached to each point x ∈ M such that the multiplication map µ : M × M → M , (x, y) → σx (y) satisfies the properties (M1) µ(x, x) = x, (M2) µ(x, µ(x, y)) = y, (M3) µ(x, µ(y, z)) = µ(µ(x, y), µ(x, z)), (M4) the tangent map Tx (σx ) is the negative of the identity of the tangent space Tx M . Recall from Lemma 1.4 the (algebraic) Jordan–Lie functor, associating an LTS to a JTS. The geometric analogue of this functor associates to each generalized polar geometry (X + , X − ; p) the symmetric space M (p) of its nonisotropic points. The geometric construction is very simple: essentially, the multiplication map µ is obtained from the structure maps Pr by taking the scalar r = −1. More precisely, given a polarity (i.e., a pair of bijections (p+ , p− ) : (X + , X − ) → (X − , X + ) that are inverses of each other, compatible with the structure maps and such that there exists a non-isotropic point x ∈ X + , which means that xp+ (x)), for each non-isotropic point x we define the point-reflection by σx (y) := P−1 x, p+ (x), y . The above-mentioned identity (PG1) then implies in a straightforward way that the set M (p) of non-isotropic points of p becomes a symmetric space (cf. [Be02] for the purely algebraic setting and [BN05] for the smooth setting). Moreover, the LTS of the symmetric space M (p) is precisely the one defined


233

by Lemma 1.4. The well-known construction of finite- and infinite-dimensional bounded symmetric domains (cf. [Up85]) arises as a special case of the geometric Jordan–Lie functor. The remarks following Lemma 1.4 imply that in fact “most” symmetric spaces (in the finite-dimensional real case) are obtained by the preceding construction (essentially, all classical finite-dimensional real ones and about half of the exceptional ones; cf. the tables in [Be00]). 2.4 Examples revisited (1) Rectangular matrices correspond to Grassmannian geometries. Let W = A ⊕ B be as in Example (1) of Section 1.5 and let GrasB A (W ) be the set of all submodules V ⊂ W that are isomorphic to A (“type”) and admit a complement U ⊂ W isomorphic to B (“cotype”). Then A (X + , X − ) = (GrasB A (W ), GrasB (W )),

with transversality of (V, U ) ∈ X + × X − being the usual complementarity of subspaces (W = U ⊕ V ), is an affine pair geometry over K. It is a standard exercise in linear algebra to show that the set U of complements of U carries a natural structure of affine space. Note that by our definition of Grassmannians as complemented Grassmannians this affine space is never empty; if K is a field this condition is automatic. (However, if K is a topological field or ring and W a topological K-module, then one will prefer modified definitions of “restricted Grassmannians” by imposing conditions closedness, boundedness, Fredholm type or other; cf., e.g., [PS86], Chapter 7. All such conditions lead to subgeometries of the algebraically defined geometries considered here.) Fixing the decomposition W = A⊕B as base point (o+ , o− ) in (X + , X − ; ), by elementary linear algebra one may identify the pair of linear spaces ((o− ) , (o+ ) ) with the pair (V + , V − ) = (Hom(B, A), Hom(A, B)). By still elementary (though less standard) linear algebra, one can now give explicit expressions for the structure maps introduced in Section 2.2 and check their fundamental identities, thus describing a direct link between the geometry and its associated Jordan pair (see [Be04]). Namely, elements of X + are realized as images of injective maps f : A → W , modulo equivalence under the general linear group GlK (A) (f ∼ f iff ∃g ∈ GlK (A): f = f ◦ g), and similarly elements of X − are realized as kernels of surjective maps φ : W → A, again modulo equivalence under GlK (A). Two such elements are transversal, [f ][φ], if and only if φ ◦ f : A → A is a bijection, and the structure map Pr is now given by Pr ([f ], [φ], [h]) = (1 − r)f ◦ (φ ◦ f )−1 + rh ◦ (φ ◦ h)−1 . As in ordinary projective geometry, an affinization is given by writing f : A → W = B ⊕A as a “column vector” and normalizing the second component to be 1A , the identity map of A, and similarly for the “row vector” φ : B ⊕ A → A

234

Wolfgang Bertram

(see [Be04]). To get a feeling for the kind of non-linear formulas that appear in such contexts, the reader may rewrite the preceding formula for Pr by replacing f, φ and h by such column, resp. row vectors, and then renormalize the right-hand side, in order to get the formula for the multiplication map in the affine picture. The special case r = 12 (“midpoint map”) is particularly important from a Jordan-theoretic point of view. One may as well also consider the total Grassmannian geometry of all complemented subspaces of W , Gras(W ) := {V ⊂ W | (V submodule), ∃V : W = V ⊕ V (V submodule)}. Then the pair (Gras(W ), Gras(W )) is still a generalized projective geometry, but it is not connected in general (there is a natural notion of connectedness for any affine pair geometry; see [Be02]). In infinite dimension, the geometries (X + , X − ) introduced above need not be connected either, but in finite dimension over a field they are; in fact, they then reduce to the usual Grassmannians Grask (Kn ) ∼ = Gl(n, K)/(Gl(k, K) × Gl(n − k, K)) of k-spaces in Kn . The case n = 2k, or, more generally, A ∼ = B, deserves special attention: in this case, X + and X − really are the same sets; the identity map X + → X − is a canonical antiautomorphism of the geometry. Its properties are similar to the well-known null systems from classical projective geometry. In particular, in the case of the projective line (Gras1 (K2 ), Gras1 (K2 )), this really is the canonical null system coming from the (up to a scalar) unique symplectic form on K2 . (2) Geometries of maps and functions. We fix some generalized projective geometry X = (X + , X − ), and let M be any set. Then the geometry of maps (Fun(M, X + ), Fun(M, X − )) with “pointwise transversality and structure maps” is again a generalized projective geometry. By arguments similar to those above, we see that it corresponds to the Jordan pair of functions introduced in Example (2) of Section 1.5. As explained there, the case of scalar-valued functions corresponds to V + = V − = Fun(M, K) with pointwise triple product (T (f, g, h))(p) = 2f (p) g(p) h(p). On the geometric side, then X + = X − = KP1 is the projective line, and Fun(M, X + ) = Fun(M, X − ) is the space of functions from M into the projective line. Note that these definitions parallel the usual ones for groups (cf. [PS86]); in particular, taking for M the unit circle, we would get “loop geometries”. (3) The classical examples. We briefly characterize the geometries (X + , X − ) corresponding to the examples of item (3) of Section 1.5: • full associative algebras A correspond to the projective line (AP1 , AP1 ) over A (cf. [BN04]);


235

• Hermitian elements of an involutive associative algebra (A, ∗) correspond to the ∗-Hermitian projective line over A (cf. [BN04], Section 8.3), and similarly for skew-Hermitian elements. Lagrangian geometries are a special case of this construction, corresponding to (skew-) Hermitian or (skew-) symmetric matrices or operators; • conformal geometries (or “spin factors”): here, X + ∼ = X − is the projective quadric of K ⊕ V ⊕ K with the same quadratic form as in Section 1.5. As to the exceptional Jordan systems, their geometries are among those constructed by Tits; to our knowledge, their structure of generalized projective geometry has not yet been fully investigated.

3 The universal model We have given the construction of the generalized projective geometry (X + , X − ) associated to a 3-graded Lie algebra g via homogeneous spaces (G/P − , G/P + ). For a more detailed study, a geometric realization of these spaces is useful, which in some sense generalizes the example of the Grassmannian geometries. This “universal model” has been introduced in [BN04] and used in [BN05] to define (under suitable topological assumptions) a manifold structure on (X + , X − ). 3.1 Ordinary flag geometries We fix some K-module g and denote by Fk the set of all flags f = 0 = f0 ⊂ f1 ⊂ · · · ⊂ fk = g of length k in E (i.e., the fi are linear subspaces of E, and all inclusions are supposed to be strict). We say that two flags e, f ∈ F are transversal if they are “crosswise complementary” in the sense that ∀i = 1, . . . , k :

g = fi ⊕ ek−i .

It is a nice exercise in linear algebra to show that e and f are transversal if, and only if, they come from a grading of g of length k. Given a grading g = (g1 , . . . , gk ), i.e., a decomposition g = g1 ⊕ · · · ⊕ gk , we define two flags f+ (g) and f− (g) via f+ i (g) := g1 ⊕ · · · ⊕ gi ,

f− k−i (g) := gi+1 ⊕ · · · ⊕ gk .

Then f+ (g) and f− (g) are obviously transversal, but the converse is also true: every pair of transversal flags can be obtained in this way! From this exercise, one deduces easily that we get a linear pair geometry (X + , X − ; ) by taking X + := X − := X := {f ∈ Fk | ∃e ∈ Fk : ef},

236

Wolfgang Bertram

the space of all flags that admit at least one transversal flag (cf. [BL08]). For k = 2, this is the total Grassmannian geometry, and similarly as in that case, one may for general k single out connected components by taking flags of fixed type and cotype. In general, these geometries will be linear, but not affine pair geometries, and we do not know what kind of “laws” (in a sense generalizing the laws of generalized projective geometries) can be used to describe them. 3.2 Filtrations and gradings of Lie algebras Let us assume now that g is a Lie algebra with Lie bracket denoted by [·, ·], and that all our gradings and filtrations are compatible with the Lie bracket in the sense that [fi , fj ] ⊂ fi+j , [gi , gj ] ⊂ gi+j . We will assume that our Lie algebra is 2k + 1-graded, i.e., as an index set we take Ik = {−k, −k + 1, . . . , k}, and we say that a 2k + 1-grading is inner if it can be defined by an Euler element, i.e., by an element E ∈ g such that [E, X] = iX for all X ∈ gi , i ∈ Ik . We denote by Gk the set of all inner 2k + 1-gradings of g (it is harmless to assume that the center of g is reduced to zero, and so Gk can be identified with the set of all Euler elements). An inner 2k + 1-filtration of g is a flag f = (0 = fk+1 ⊂ fk ⊂ · · · ⊂ f−k = g) such that there is some inner 2k+1-grading with fi = gi ⊕gi+1 ⊕· · ·⊕gk , for all i ∈ Ik ; we write Fk for the set of all such filtrations. Then the following holds ([BN04] for the case of 3-gradings (k = 1), and [Ch09] for the general case). We have to impose some mild restrictions on the characteristic

∞ of K since we want to use, for X ∈ f1 , the operator exp(X) = ead(X) = j=0 j!1 ad(X)j , the sum being finite since [f1 , fi ] ⊂ fi+1 , so that ad(X) is nilpotent. Theorem. Assume that K is a commutative ring such that the integers 2, 3, . . . , 2k + 1 are invertible in K. (1) Two inner 2k + 1-filtrations f and e are transversal if, and only if, there exists a 2k+1-grading g = (g−k , . . . , gk ) such that f = f+ (g) and e = f− (g). (2) The geometry of inner 2k + 1-filtrations (Fk , Fk ; ) is a linear pair geometry. More precisely, for any f ∈ Fk , the nilpotent Lie algebra f1 is in bijection with f via X → exp(X).e, for an arbitrary choice of e ∈ f . Now assume that k = 1, i.e., we are in the 3-graded case. Then [f1 , f1 ] ⊂ f2 = 0 (i.e., f1 is an abelian subalgebra), whence part (2) of the theorem says that we have a simply transitive action of the vector group f1 ∼ = exp(f1 ) on f , and thus we see that the geometry of inner 3-filtrations is an affine pair geometry. In fact, fixing some inner 3-grading as base point in the space of all inner 3-gradings, it now easily follows that the geometry (G/P − , G/P + ) constructed


237

in Section 2 is imbedded into the geometry from the theorem simply by taking the G-orbit of the base point. As mentioned above, this “universal geometric realization” turns out to be useful for giving a precise description of the intersections of the “chart domains” α and β for α, β ∈ X − and for calculating the corresponding “transition functions” (see [BN04]); calculations become similar to the case of Grassmannian geometries (Example (1) in Section 2.4 above), the difference being that the usual fractional linear transformations have to be replaced by certain fractional quadratic transformations (this difference corresponds to the fact that we are now working with true flags instead of single subspaces). Using these purely algebraic results, one can now give necessary and sufficient conditions for defining on (X + , X − ) the structure of a smooth manifold ([BN05]). The theory works nicely over general topological base fields and rings (cf. [Be08] for basic differential geometry and Lie theory in this general framework); complex or real Banach, Fréchet or locally convex manifolds are special cases of it. For higher gradings, similar results hold (see [Ch09] and [Ch10]).

4 The geometry of states Let us come back to the analogies, mentioned in the Introduction, of the preceding constructions with methods of commutative and non-commutative geometry. In the language of classical physics, one wants to recover the pure states of a system (the point space M ) from its observables (the function algebra A = Reg(M, R)). The usual procedure is to look at the (maximal) ideals Ip = {f ∈ A| f (p) = 0} corresponding to points p ∈ M . Left or right ideals in associative algebras are generalized by inner ideals in Jordan theory. Therefore, we shall interpret states in generalized projective geometries as the geometric objects corresponding to inner ideals. 4.1 Intrinsic subspaces Assume (X + , X − ; ) is a linear pair geometry over the commutative ring K. Definition 4.1. A pair (Y + , Y − ) of subsets Y ± ⊂ X ± is called a subspace of (X + , X − ) if for all x ∈ Y ± there exists an element α ∈ Y ∓ such that xα, and if for every such pair (x, α) the set Yα := Y ∩ α is a linear subspace of (α , x) and Yx := Y ∩ x is a linear subspace of (x , α). A subset I ⊂ X + is called a state or an intrinsic subspace (in X + ) if I “appears linearly to all possible observers”: for all α ∈ X − with α ∩ I = ∅ and for all x ∈ α ∩ I, the set (α ∩ I, x) is a linear subspace of (α , x). For instance, if (X + , X − ) = (Gras1 (Kn+1 ), Grasn (Kn+1 )) is an ordinary projective geometry, then all projective subspaces of X + are intrinsic

238

Wolfgang Bertram

subspaces, and every affine subspace of an affinization of X + is obtained in this way. In contrast, for Grassmannian geometries of higher rank the situation becomes more complicated: only rather specific linear subspaces of the affinization M (p, q; K) of Grasp (Kp+q ) are obtained in this way, namely the inner ideals. This observation generalizes to all geometries (X + , X − ) associated to a Jordan pair (V + , V − ): subspaces containing the base point (o+ , o− ) correspond to subpairs of (V + , V − ), and intrinsic subspaces of X + containing o+ to inner ideals I ⊂ V + , i.e., submodules such that T + (I, V − , I) ⊂ I (see [BL08]). The notions of, e.g.s minimal, maximal, principal inner ideals may also be suitably translated into a geometric language. (Minimal intrinsic subspaces will also be called intrinsic lines and maximal ones intrinsic hyperplanes. Readers coming from axiomatic geometry or from Jordan theory will consider the former as more fundamental and call them accordingly pure states, whereas readers coming from algebraic geometry or C ∗ -algebra theory would instead reserve this term for the latter ones.) The collection of states associated to a linear pair geometry can again be turned into a pair geometry by introducing the following transversality relation. Definition 4.2. Let I be an intrinsic subspace in X + and J one in X − . We say that I and J are transversal, and we write again IJ if (1) the pair (I, J ) is a subspace, (2) the linear pair geometry (I, J ) is faithful in the following sense. A linear pair geometry (Y + , Y − ) is called faithful if Y − is faithfully represented by its effect of linearizing Y + , and vice versa: whenever α = β as sets and as linear spaces (with respect to some origin o), then α = β, and the dual property holds. Let us give some motivation for this definition (which does not appear in [BL08]). First of all, in general, a linear pair geometry need not be faithful (take, for instance, a pair of K-modules with trivial structure maps), but ordinary projective geometries over a field are. (In case of geometries corresponding to Jordan pairs, faithfulness corresponds to non-degeneracy in the Jordan-theoretic sense.) Next note that, even if the geometry (X + , X − ) was faithful, the geometry (I, X − ) will in general not be faithful: there is a kernel, i.e., the equivalence relation “α ∼I β iff α and β induce the same linear structure on I” will be non-trivial on X − . The condition IJ means, then, that the space J ⊂ X − is transversal to the fibers of this equivalence relation, in the usual sense. (This is the geometric translation of the notion of complementation of inner ideals introduced in [LoN95]; the fiber of the equivalence relation ∼I corresponds to the kernel of an inner ideal introduced in [LoN95].) For instance, if I is a projective line in an ordinary projective space X + = Gras1 (Kn+1 ), corresponding to a 2-dimensional subspace I of Kn+1 , then J should be a projective subspace of the dual projective space X − = Grasn (Kn+1 ) such that different elements of J define different affine


239

lines in I. If J is the set of hyperplanes lying over some k-dimensional subspace of Kn+1 , then this means that I should contain some complement of J. Vice versa, J then also should contain some complement of I, and hence I and J have to be complements of each other in Kn+1 . Thus (I, J ) is a projective line which is faithfully imbedded in the projective geometry. Moreover, in this example we clearly recover the Grassmannian geometry of 2-spaces as the geometry of intrinsic lines in the 1-Grassmannian. Parts of the observations from the example generalize. Let S ± be the set of all intrinsic subspaces I in X ± that admit a transversal intrinsic subspace J in X ∓ . It is a natural question to ask whether (S + , S − ) is again a linear pair geometry, or under which conditions on (X + , X − ) and on the states we obtain one. For the time being, we have no general answers, but results from [BL08] point into the direction that, indeed, for the most interesting cases we get geometries (S + , S − ) that are again linear pair geometries. More precisely, if (X + , X − ) corresponds to a Jordan pair (V + , V − ) and we consider only states that are associated, via a Peirce-decomposition, to idempotents, we obtain geometries coming from 5-graded Lie algebras ([BL08], Theorem 5.8), and these are indeed linear pair geometries, according to Theorem 3.2. 4.2 Examples (1) Classical states revisited. Let M be a set and A = Fun(M, K) the Jordan algebra of all functions from M to K. It corresponds to the geometry (X , X ) with X = Fun(M, KP1 ). We claim that (if K is a field) the set M can be recovered from the geometry X as the geometry of all pure states (intrinsic lines) running through the “zero function” f0 ≡ o, by associating to a point p ∈ M the intrinsic line Lp = {f : M → KP1 | f (x) = o if x = p} ⊂ Fun(M, KP1 ). Indeed, note first that Lp is indeed a minimal intrinsic subspace since Lp ∼ = KP1 and the geometry (KP1 , KP1 ) itself does not admit any proper intrinsic subspaces that are not points. Now assume f : M → KP1 is a function taking non-zero values at two different points p, q ∈ M . Then a look at the direct product geometry (KP1 , KP1 ) × (KP1 , KP1 ) shows that the intrinsic subspace generated by f contains at least all functions obtained from f by altering the values at p and q in an arbitrary way, and hence contains a homomorphic image of KP1 × KP1 and thus is not minimal, proving our claim. Of course, we may as well work with the maximal intrinsic subspaces Hp = {f : M → KP1 | f (p) = o } ⊂ Fun(M, KP1 ). This is certainly closer to the philosophy of algebraic geometry and to the usual imbedding of “classical” into “non-commative” geometry, when M is a smooth or algebraic manifold over R or C and we look at geometries of

240

Wolfgang Bertram

smooth or algebraic functions f : M → KP1 . Both points of view are dual to each other in the sense that one really should look at the geometry (S + , S − ), where one of the S ± is the collection of minimal, and the other of maximal intrinsic subspaces, each having a complement (in the sense of Definition 4.2) of the other type. (2) Grassmannian geometries X + = GrasB A (W ). Let us fix a flag in W of length two, f : 0 ⊂ f1 ⊂ f2 ⊂ W . Then the set of all subspaces of W that are “squeezed” by this flag, If := {E ∈ X + | f1 ⊂ E ⊂ f2 }, is an intrinsic subspace of X , and if K is a field and W is finite dimensional over K, then all intrinsc subspaces are of this form for a suitable flag f (see [BL08], Theorem 3.11). Moreover, one can show that the intrinsic subspaces If and Ie are transversal in the sense defined above if, and only if, the flags e and f are transversal in the sense defined in Section 3.1 above. Therefore, the geometry of intrinsic subspaces of the Grassmannian geometry is a geometry of flags of length two and hence is a linear pair geometry. The case of ordinary projective geometry (i.e., A ∼ = K) is somewhat degenerate. In this case there is not much choice for the first component f1 of the flag f, and therefore the intrinsic subspace If is already determined by f2 alone; we get back the usual projective subspaces of a projective space. In all other cases, the geometry of intrinsic subspaces is a true flag geometry corresponding to 5-gradings and hence is a linear, but no longer an affine, pair geometry. (3) Quantum states. In the language of quantum mechanics (also widely used in non-commutative geometry), states are defined as positive normalized linear functionals φ : A → C on a C ∗ -algebra (A, ∗). They form a convex set, whose extremal points are interpreted as pure states. These, in turn, admit a Jordan-theoretic interpretation via (primitive) idempotents. See [FK94] for the theory of finite-dimensional symmetric cones and their geometry; there is a vast literature on infinite-dimensional generalizations; cf., e.g., [HOS84], [ER92], [Up85]. Whereas the notion of state just mentioned depends highly on the ordered structure of the base field R (via positivity and convexity), this is not the case for the notion of inner ideal and intrinsic subspace, which hence seems to be more general and more geometric. Accordingly, it has been advocated to use inner ideals as a basic ingredient for an approach to quantum mechanics (see [F80]).

References [AF99] [Be00] [Be02]

Allison, B.N. and J.R. Faulkner, Elementary groups and invertibility for Kantor pairs, Comm. Algebra 27 (1999), no. 2, 519–556. Bertram, W., The geometry of Jordan- and Lie structures, Springer Lecture Notes in Mathematics 1754, Berlin 2000. Bertram, W., Generalized projective geometries: general theory and equivalence with Jordan structures, Adv. Geom. 2 (2002), 329–369.

Jordan Structures and Non-Associative Geometry [Be03] [Be04] [Be08] [BeKi09]

[BL08] [BN04] [BN05] [Bil04] [Ch09] [Ch10] [Co94] [ER92] [FK94] [F80] [HOS84] [Lo69] [Lo75] [Lo95] [LoN95] [McC04] [PS86] [Sm05] [Up85]

241

Bertram, W., The geometry of null systems, Jordan algebras and von Staudt’s Theorem, Ann. Inst. Fourier 53 (2003) fasc. 1, 193–225. Bertram, W., From linear algebra via affine algebra to projective algebra, Linear Algebra and its Applications 378 (2004), 109–134. Bertram, W., Differential geometry, Lie groups and symmetric spaces over general base fields and rings. Mem. AMS 192, 900 (2008), math.DG/ 0502168. Bertram, W. and M. Kinyon, Associative Geometries. I: Torsors, Linear Relations and Grassmannians. II: Involutions, the Classical Torsors, and their Homotopes. To appear in Journal of Lie Theory, arxiv: math.Ra/0903.5441 and math.Ra/0909.4438 Bertram, W. and H. L¨ owe, Inner ideals and intrinsic subspaces, Adv. in Geom. 8 (2008), 53–85. Bertram, W. and K.-H. Neeb, Projective completions of Jordan pairs. I: The generalized projective geometry of a Lie algebra, J. Algebra 277(2) (2004), 193–225; math.RA/0306272. Bertram, W. and K.-H. Neeb, Projective completions of Jordan pairs. II: Manifold structures and symmetric spaces, Geom. Dedicata 112 (2005), 73–113; math.GR/0401236. Biller, H., Continuous Inverse Algebras and Infinite-Dimensional Linear Lie Groups, Habilitationsschrift, Technische Universit˝at Darmstadt, 2004. Chenal, J., Generalized flag geometries and manifolds associated to short Z-graded Lie algebras, C.R.Acad. Sci. Paris, Ser. I 347 (2009) 21–25. Chenal, J.: Géométries liées aux algèbres de Lie graduées, thesis, Nancy 2010. Connes, A.: Non-commutative Geometry. Academic Press, San Diego 1994. Edwards, C.M. and G.T. R¨ uttimann, A characterization of inner ideals in JB ∗ -triples, Proc. Amer. Math. Soc. 116 (1992), no. 4, 1049–1057. Faraut, J. and A. Koranyi, Analysis on Symmetric Cones, Clarendon Press, Oxford, 1994. Faulkner, J.R., Incidence Geometry of Lie Groups in Quantum Theory, Group theoretical methods in physics (Proc. Eigth Internat. Colloq., Kiryat Anavim, 1979) Ann. Israel Phys. Soc., 3, 73–79, Hilger, Bristol, 1980. Hanche-Olsen, H. and E. Størmer, Jordan Operator Algebras, Pitman, Boston, 1984. Loos, O., Symmetric Spaces I, Benjamin, New York, 1969. Loos, O., Jordan Pairs, Springer LNM 460, New York, 1975. Loos, O., Elementary groups and stability for Jordan pairs, K-Theory 9 (1995), no. 1, 77–116. Loos, O. and E. Neher, Complementation of inner ideals in Jordan pairs, J. Alg. 166 (1994), 255–295. Mc. Crimmon, K. A Taste of Jordan Algebras. Springer, New York, 2004. Pressley, A. and G. Segal, Loop Groups, Clarendon Press, Oxford 1986. Smirnov, O.N., Imbedding of Lie triple systems into Lie algebras, preprint, 2005. arxiv: math/0906.1170 Upmeier, H., Symmetric Banach Manifolds and Jordan C ∗ -algebras, North-Holland Mathematics Studies, Amsterdam, 1985.

For electronic versions of some of the above-mentioned papers that are not available on the arXiv, see the Jordan preprint server at the address http://homepage.uibk.ac.at/∼c70202/jordan/index.html

Direct Limits of Infinite-Dimensional Lie Groups Helge Gl¨ ockner Universit¨ at Paderborn, Institut f¨ ur Mathematik, Warburger Str. 100, 33098 Paderborn, Germany, [email protected]

Ë

Summary. Many infinite-dimensional Lie groups G of interest can be expressed as the union G = n∈N Gn of an ascending sequence G1 ⊆ G2 ⊆ · · · of (finite- or infinite-dimensional) Lie groups. In this survey article, we compile general results concerning such ascending unions, describe the main classes of examples and explain what the general theory tells us about them. Key words: Lie group, direct limit, inductive limit, ascending sequence, ascending union, directed union, Silva space, regularity, initial Lie subgroup, homotopy group, small subgroup. 2000 Mathematics Subject Classifications: Primary 22E65. Secondary 26E15, 46A13, 46G20, 46T05, 46T20, 46T25, 54B35, 54D50, 55Q10, 58B05, 58D05.

1 Introduction Many infinite-dimensional Lie groups G can be expressed as the union G = n∈N Gn of a sequence G1 ⊆ G2 ⊆ · · · of (finite- or infinite-dimensional) Lie groups, such that the inclusion maps jn : Gn → G and jm,n : Gn → Gm (for n ≤ m) are smooth homomorphisms. Typically, the steps Gn are Lie groups of a simpler type, and one hopes to (and often does) deduce results concerning G from information available for the Lie groups Gn . The goals of this article are twofold: • •

To survey general results on ascending unions of Lie groups and their properties; To collect concrete classes of examples and explain how the general theory specializes in these cases.

One typical class of examples is given by the groups Diff c (M ) of smooth diffeomorphisms φ : M → M of σ-compact, finite-dimensional smooth manifolds M K.-H. Neeb and A. Pianzola (eds.), Developments and Trends in Infinite-Dimensional Lie Theory, Progress in Mathematics 288, DOI 10.1007/978-0-8176-4741-4_8, © Springer Basel AG 2011

243

244

Helge Gl¨ ockner

which are compactly supported in the sense that the set {x ∈ M : φ(x) = x} has compact closure. The group operation is composition of diffeomorphisms. It is known that Diff c (M ) is a Lie group (see [Mc80] or [Gl02d]). Furthermore, Diff Kn (M ) Diff c (M ) = n∈N

for each exhaustion K1 ⊆ K2 ⊆ · · · of M by compact sets (with Kn in the interior of Kn+1 ), where Diff Kn (M ) is the Lie group of smooth diffeomorphisms of M supported in Kn . The manifold structure of Diff c (M ) is modelled on the space Vc (M ) of compactly supported smooth vector fields, which is an LF-space with a complicated topology. In contrast, Diff Kn (M ) is modelled on the space VKn (M ) of smooth vector fields supported in Kn , which is a Fréchet space. Many specific tools of infinite-dimensional calculus can be applied to VKn (M ), e.g., to clarify differentiability questions for functions on this space. In other typical cases, each Gn is finite dimensional (a particularly well-understood situation) or modelled on a Banach space, whence again special tools are available to deal with the Lie groups Gn (but not a priori for G). Besides diffeomorphism groups, we shall also discuss the following major classes of examples (described in more detail in Section 6): ∞ • The “test function groups” Cc∞ (M, H) = n∈N CK (M, H) of compactly n supported Lie group-valued smooth mappings on a σ-compact smooth manifold M = n∈N Kn ; ∗ n • Weak direct products n∈N Hn := n∈N k=1 Hk of Lie groups Hn ; • Unions A× = n∈N A× n of unit groups of Banach algebras A1 ⊆ A2 ⊆ · · · ; • The groups Germ(K, H) of germs of analytic mappings on open neighbourhoods of a compact subset K of a metrizable complex locally convex space, with values in a complex Banach–Lie group H; 1 • The group H ↓s (K, F ) = t>s H t (K, F ) = n∈N H s+ n (K, F ), where K is a compact smooth manifold, s ≥ dim(K)/2, F a finite-dimensional Lie group and H t (K, F ) ⊆ C(K, F ) the integral subgroup whose Lie algebra is the Sobolev space H t (K, L(F )) of functions with values in the Lie algebra L(F ) of F . For s = dim(K)/2, the Lie group H ↓s (K, F ) is particularly interesting, because a Hilbert–Lie group H s (K, F ) is not available in this case. In some situations, H ↓s (K, F ) may serve as a substitute for the missing group. We shall also discuss the group GermDiff(K, X) of germs of analytic diffeomorphisms γ around a compact set K in a finite-dimensional complex vector space X, such that γ|K = idK . This group is not considered as a union of groups, but as a union of Banach manifolds M1 ⊆ M2 ⊆ · · · . Among others, we shall discuss the following topics in our general setting (and for the preceding examples): •

Direct limit properties of ascending unions;

Direct Limits of Infinite-Dimensional Lie Groups

• • •

245

Homotopy groups of ascending unions; When ascending unions are regular Lie groups in Milnor’s sense; Questions concerning subgroups of ascending unions.

We now describe the main problems and questions in more detail, together with some essential concepts. As a rule, references to the literature, answers (and partial answers) will only be given in later sections. 1.1 Direct limit properties of ascending unions Consider a Lie group G which is an ascending union G = groups, and a map f : G → X. It is natural to ask:

n∈N

Gn of Lie

(a) If X is a smooth manifold (modelled on a locally convex space) and f |Gn is smooth for each n ∈ N, does it follow that f is smooth? (b) If X is a topological space and f |Gn is continuous for each n ∈ N, does it follow that f is continuous? (c) If X is a Lie group, f is a homomorphism of groups and f |Gn is smooth for each n ∈ N, does it follow that f is smooth? (d) If X is a topological group, f is a homomorphism and f |Gn is continuous for each n ∈ N, does it follow that f is continuous? As we shall see, (a) and (b) are frequently not true (unless compactness can be brought into play), while (c) and (d) hold for our typical examples. The preceding questions can be recast in category-theoretic terms: They amount to asking if G is the direct limit lim Gn in the categories of smooth −→ manifolds, topological spaces and Lie groups, resp., topological groups (see 2.8). The relevant concepts from category theory will be recalled in Section 2. Questions (b) and (d) can be asked just as well if G and each Gn are merely a topological group, and each inclusion map is a continuous homomorphism. Essential progress concerning direct limits of topological groups and their relations to direct limits of topological spaces was achieved in the last ten years, notably by N. Tatsuuma, E. Hirai, T. Hirai and N. Shimomura (see [TSH98] and [HST01]) as well as A. Yamasaki [Ya98]. In Section 3, we recall the most relevant results. 1.2 Existence of direct limit charts: an essential hypothesis Meaningful results concerning the preceding topics can only be expected under additional hypotheses. For instance, our general setting includes the situation where each Gn is discrete but G is not (as we only assume that the inclusion maps Gn → G are smooth). In this situation, algebraic properties of the groups Gn (like simplicity or perfectness) pass to G, but we cannot expect to gain information concerning the topological or differentiable structure of G from information on the groups Gn .

246

Helge Gl¨ ockner

A very mild additional hypothesis is the existence of a direct limit chart. Roughly speaking, this is a chart of G juxtaposed from charts of the Lie groups Gn . The formal definition reads as follows (cf. [Gl07b, Definition 2.1]). Definition. A Lie group G = n∈N Gn is said to admit a weak direct limit chart if there exists n0 ∈ N, charts φn : Un → Vn from open identity neighbourhoods Un ⊆ Gn onto open 0-neighbourhoods Vn ⊆ L(Gn ) in the tangent space L(Gn ) := T1 (Gn ) at 1 for n ≥ n0 and a chart φ : U → V from an open identity neighbourhood U ⊆ G onto an open 0-neighbourhood V ⊆ L(G), such that (a) U = n≥n0 Un and Un ⊆ Un+1 for each integer n ≥ n0 ; and (b) φn+1 |Un = L(jn+1,n ) ◦ φn and φ|Un = L(jn ) ◦ φn for each n ≥ n0 . If, furthermore, L(G) = lim L(Gn ) as a locally convex space,1 then −→ G = n∈N Gn is said to admit a direct limit chart. Note that (b) implies that the linear maps L(jn ) and L(jn+1,n ) are injective on some 0-neighbourhood and thus injective. Hence, identifying L(Gn ) with its image under L(jn ) in L(G), we can rewrite (b) as (b) φ|Un = φn and φn+1 |Un = φn , for each n ≥ n0 . Furthermore, we now simply have V = n≥n0 Vn . To assume the existence of a direct limit chart is a natural requirement, which is satisfied by all of our main examples. It provides a link between the topologies (resp., manifold structures) on G and the Lie groups Gn , and will be encountered in connection with most of the topics from above. 1.3 Homotopy groups of ascending unions of Lie groups Given a Lie group G = n∈N Gn , it is natural to ask if its k-th homotopy group can be calculated in terms of the homotopy groups πk (Gn ) in the form πk (G) = lim πk (Gn ) , (1) −→ for each k ∈ N0 . This is quite obvious if G = n∈N Gn is compactly regular in the sense that each compact subset K of G is a compact subset of some Gn (see [Gl10, Proposition 3.3]; cf. [Gl05b, Remark 3.9] and [Ne04c, Lemma A.7] for special cases, as well as works on stable homotopy theory and K-theory). There is another, non-trivial condition: If G = n∈N Gn admits a weak direct limit chart, then(1) holds [Gl10, Theorem 1.2]. A variant of this condition even applies if n∈N Gn is merely dense in G (see Theorem 1.13 in [Gl10]). Moreover, ascending unions can be replaced with directed unions over uncountable families, and Lie groups with manifolds (see Section 9). These results are based on approximation arguments. Analogous results for open subsets of locally convex spaces are classical [Pa66]. 1

Here, we use the bonding maps L(jm,n ) : L(Gn ) → L(Gm ) and the limit maps L(jn ) : L(Gn ) → L(G).


247

We mention that knowledge of π0 (G) = G/G0 , the fundamental group π1 (G) and π2 (G) is essential for the extension theory of G. It is needed to → G → 1 of G with abelian understand the Lie group extensions 1 → A → G kernels, by recent results of K.-H. Neeb (see [Ne02b], [Ne04b] and [Ne07]). 1.4 Regularity in Milnor’s sense Roughly speaking, a Lie group G (modelled on a locally convex space) is called a regular Lie group if all differential equations on G which are of relevance for Lie theory can be solved, and their solutions depend smoothly on parameters. To make this more precise, given g, h ∈ G and v ∈ Th (G), let us write g · v := (Th λg )(v) ∈ Tgh (G), where λg : G → G, x → gx denotes left translation by g. Definition. A Lie group G modelled on a locally convex space is called a regular Lie group (in Milnor’s sense) if for each smooth curve γ : [0, 1] → L(G), there exists a (necessarily unique) smooth curve η = ηγ : [0, 1] → G (a “product integral”) which solves the initial value problem η(0) = 1

(2)

η (t) = η(t) · γ(t) for all t ∈ [0, 1]

(3)

(with 1 ∈ G the identity element), and the “evolution map” evol: C ∞ ([0, 1], L(G)) → G ,

evol(γ) := ηγ (1)

(4)

is smooth (see [Mr84], [GN10] and [Ne06]). Regularity is a useful property which provides a link between G and its Lie algebra. In particular, regularity ensures the existence of a smooth exponential map expG : L(G) → G, i.e., a smooth map such that, for each v ∈ L(G), γv : R → G ,

t → expG (tv)

is a homomorphism of groups with initial velocity γv (0) = v (cf. [Mr84]). The modelling space E of a regular Lie group is necessarily Mackey 1 complete in the sense that the Riemann integral 0 γ(t) dt exists in E for each smooth curve γ : R → E (cf. Lemma A.5 (1) and p. 21 in [NW08]). Lie groups modelled on non-Mackey complete locally convex spaces need not even have an exponential map. For example, this pathology occurs for group G = A× of invertible elements in the normed algebra A ⊆ C[0, 1] of all restrictions to [0, 1] of rational functions without poles in [0, 1] (with the supremum norm). Because A× is an open subset of A, it is a Lie group, and it is not hard to see that a smooth homomorphism γv : R → A× with γ (0) = v exists for v ∈ A = L(A× ) only if v is a constant function [Gl02c, Proposition 6.1]. Further information concerning Mackey completeness can be found in [KM97]. At the time of writing, it is unknown whether non-regular Lie groups modelled on Mackey complete locally convex spaces exist. However, there

248

Helge Gl¨ ockner

is no general method of proof; for each individual class of Lie groups, very specific arguments are required to verify regularity. It is natural to look for conditions ensuring that a union G = n∈N Gn is a regular Lie group if each Gn is one. Already the case of finite-dimensional Lie groups Gn is not easy [Gl05b]. In Section 8, we preview work in progress concerning the general case. We also describe a construction which might lead to non-regular Lie groups (Proposition 8.7). The potential counterexamples are weak direct products of suitable regular Lie groups. 1.5 Subgroups of direct limit groups It is natural to try to use information concerning the subgroups of Lie groups Gn to deduce results concerning the subgroups of a Lie group G = n∈N Gn . Aiming at a typical example, let us recall that a topological group G is said to have no small subgroups if there exists an identity neighbourhood U ⊆ G containing no subgroup of G except for the trivial subgroup. Although finite-dimensional (and Banach–) Lie groups do not have small subgroups, already for Fréchet–Lie groups the situation changes: The additive group of the Fréchet space RN has small subgroups. In fact, every 0-neighbourhood contains ]−r, r[n ×R{n+1,n+2,...} for some n ∈ N and r > 0. It therefore contains the non-trivial subgroup {0} × R{n+1,n+2,...} . It is natural to ask whether a Lie group G = n∈N Gn does not have small subgroups if none of the Lie groups Gn has small subgroups. In Section 10, we describe the available answers to this question and various other results concerning subgroups of direct limit groups. 1.6 Constructions of Lie group structures on ascending unions So far, we assumed that G is already equipped with a Lie group structure. Sometimes, only an ascending sequence G1 ⊆ G2 ⊆ · · · of Lie groups is given such that all inclusion maps Gn → Gn+1 are smooth homomorphisms. It is then natural to ask whether the union G = n∈N Gn can be given a Lie group structure making each inclusion map Gn → G a smooth homomorphism.2 We shall also discuss this complementary problem (in Section 5). If each Gn is finite dimensional, then a Lie group structure on G is always available. 1.7 Properties of locally convex direct limits To enable an understanding of direct limits of Lie groups, an understanding of various properties of locally convex direct limits is essential, i.e., of direct limits in the category of locally convex spaces. For instance, we shall see that if a Lie group G = n∈N Gn admits a direct limit chart, then G = lim Gn as −→

2

Or even making G the direct limit lim Gn in the category of Lie groups. −→


249

a topological space if and only if L(G) = lim L(Gn ) as a topological space. −→ The latter property is frequently easier to prove (or refute) than the first. Also compact regularity of G (as in 1.3 above) can be checked on the level of the modelling spaces (see Lemma 6.1). Another property is useful: Consider a locally convex space E which is a union n∈N En of locally convex spaces, such that all inclusion maps are continuous linear maps. We say that E is regular (or boundedly regular, for added clarity) if every bounded subset of E is a bounded subset of some En . If one wants to prove that a Lie group n Gn is regular in Milnor’s sense, then it helps a lot if one knows that L(G) = n L(Gn ) is compactly or boundedly regular (see Section 8). 1.8 Further comments, and some historical remarks The most typical examples of direct limits of finite-dimensional Lie groups are unions of classical groups like GL (C) = GL (C) and its sub∞ n n∈N groups GL (R) = GL (R), O (R) = O (R) and U (C) = ∞ n ∞ n ∞ n∈N n∈N n∈N Un (C), where A ∈ GLn (C) is identified with the block matrix A 0 0 1 in GLn+1 (C). Thus GL∞ (C) is the group of invertible matrices of countable size which differ from the identity matrix only at finitely many entries. Groups of this form (and related ascending unions of homogeneous spaces) have been considered for a long time in (stable) homotopy theory and K-theory. Furthermore, results concerning their representation theory can be traced back to the 1970s (more details are given below). However, only the group structure or topology was relevant for these studies. Initially, no attempt was made to consider them as Lie groups. The Lie group structure on GL∞ (C) was first described in [Mr82] and that on U∞ (C) and O∞ (R) was mentioned (cf. also p. 1053 in the survey article [Mr84]). The first systematic discussion of direct limits of finite-dimensional Lie groups was given in [NRW91] and [NRW93]. Notably, a Lie group structure on G = lim Gn was constructed there under technical conditions which ensure, −→ in particular, that lim expGn : lim L(Gn ) → lim Gn , x → expGn (x) for all n ∈ N, x ∈ L(Gn ) −→

−→

−→

is a local homeomorphism at 0 (see Section 5 for the sketch of a more general construction from [Gl05b]). Moreover, situations were described in [NRW93] where ascending unions of Lie groups (or the corresponding Lie algebras) can be completed with∞respect to some coarser topology. Also, the Lie group Cc∞ (M, H) = n∈N CK (M, H) is briefly discussed in [NRW93] (for finiten dimensional H), and in [AHK93]. Test function groups with values in possibly infinite-dimensional Lie groups were treated in [Gl02b]. Compare [ACM89]

250

Helge Gl¨ ockner

for an early discussion of gauge groups and automorphism groups of principal bundles over non-compact manifolds in the inequivalent “convenient setting” of infinite-dimensional calculus (cf. also [KM97]). The first construction of the Lie group structure on Diff c (M ) was given in [Mc80], as part of a discussion of manifold structures on spaces of mappings between non-compact manifolds. Groups of germs of complex analytic diffeomorphisms of Cn around 0 were studied in [Pi77]. The real analytic analogue was discussed in [Le94], and groups of germs of more general diffeomorphisms in [KR01]. Further recent works will be described later. It should be stressed that the current article focusses on direct limit groups as such, i.e., on their structure and properties. Representation theory and harmonic analysis on such groups are outside its scope. For completeness, we mention that the study of (irreducible) unitary representations of ascending unions of finite groups started in the 1960s (see [Th64a] and [Th64b]), notably for the symmetric group S∞ := lim Sn . Representations −→ of direct limits of finite-dimensional Lie groups were first investigated in the 1970s (see [Vo76] for representations of U∞ (C), and [KS77] for representations of U∞ (C) and SO∞ (R)). Ol’shanski˘ı [Ol83] studied representations for infinite-dimensional spherical pairs like (GL∞ (R), SO∞ (R)),(GL∞ (C),U∞ (C)) and (U∞ (C), SO∞ (R)). The representation theory of direct limits of both finite groups and finite-dimensional Lie groups remains an active area of research. Representations of O(∞, ∞), U(∞, ∞) and Sp(∞, ∞) were studied in [Dv02] using infinite-dimensional adaptations of Howe duality. Novel results concerning the representation theory of U∞ (C) and S∞ were obtained in [Ol03] and [KOV04], respectively. Versions of the Bott–Borel–Weil theorem for direct limits of finite-dimensional Lie groups were established in [NRW01] and [DPW02] (in a more algebraic setting). J. A. Wolf also investigated principal series representations of suitable direct limit groups [Wo05], as well as the regular representation on some direct limits of compact symmetric spaces [Wo08]. The paper [AK06] discusses representations of an infinite-dimensional subgroup of unipotent matrices in GL∞ (R). Finally, a version of Bochner’s theorem for infinite-dimensional spherical pairs was obtained in [Ra08]. There also is a body of literature devoted to irreducible representations of diffeomorphism groups of (compact or) non-compact manifolds, as well as quasi-invariant measures and harmonic analysis thereon (see, e.g., [VGG75], [Ki81], [Hi93], [Sh01] and [Sh05]). Representations of Cc∞ (M, H) were studied by R. S. Ismagilov [Is76] and in [AHK93]. Such groups, Diff c (M ) and semidirect products thereof, arise naturally in mathematical physics [Go03]. Direct limits of finite-dimensional Lie groups are also encountered as dense subgroups of some interesting Banach–Lie groups (like the group U2 (H) of unitary operators on a complex Hilbert space H which differ from idH by a Hilbert–Schmidt operator) and other groups of operators. This frequently enables the calculation of the homotopy groups of such groups (see [Pa65], also [Ne02a]), exploiting that the homotopy groups of many direct limits of


251

classical groups (like U∞ (C)) can be determined using Bott periodicity. Dense unions of finite-dimensional Lie groups are also useful in representation theory (see [Ne98] and [Ne04c]). We mention a more specialized result: For very particular classes of direct limits G of finite-dimensional Lie groups, a classification is possible which uses the homotopy groups of G (notably π1 (G) and π3 (G)); see [Ku06]. In contrast to direct (or inductive) limits, the dual notion of an inverse (or projective) limit of Lie groups was used much earlier in infinite-dimensional Lie theory. Omori’s theory of ILB-Lie groups (which are inverse l imits of B anach manifolds) gave a strong impetus to the development of the area in the late 1960s and early 1970s (see [Om97] and the references therein). Many important examples of infinite-dimensional Lie groups could be discussed in this approach, e.g., the group C ∞ (K, H) = k∈N0 C k (K, H) of smooth maps on a compact manifold K with values in a finite-dimensional Lie group H, and the group Diff(K) = k∈N Diff k (K) of C ∞ -diffeomorphisms of a compact manifold. The passage from compact to non-compact manifolds naturally leads to the consideration of direct limits of compactly supported objects.

2 Preliminaries, terminology and basic facts General conventions. We write N := {1, 2, . . .}, and N0 := N∪{0}. As usual, R and C denote the fields of real and complex numbers, respectively. If (E, .) is a normed space, x ∈ E and r > 0, we write BrE (x) := {y ∈ E : y − x < r}. Topological spaces, topological groups and locally convex topological vector spaces are not assumed Hausdorff. However, manifolds are assumed Hausdorff, and whenever a locally convex space serves as the domain or range of a differentiable map, or as the modelling space of a Lie group or manifold, it is tacitly assumed Hausdorff. Moreover, all compact and all locally compact topological spaces are assumed Hausdorff. We allow non-Hausdorff topologies because direct limits are much easier to describe if the Hausdorff property is omitted (further explanations will be given at the end of this section). Infinite-dimensional calculus. We are working in the setting of Keller’s Cck theory [Ke74], in a topological formulation that avoids the use of convergence structures (as in [Mc80], [Mr84], [Gl02a], [Ne06] and [GN10]). For more information on analytic maps, see, e.g., [Gl02a], [GN10] and (for K = C) [BS71]. 2.1 Let K ∈ {R, C}, r ∈ N ∪ {∞}, E and F be locally convex K-vector spaces and f : U → F be a map on an open set U ⊆ E. If f is continuous, we say that f is C 0 . We call f a CKr -map if f is continuous, the iterated real (resp., complex) directional derivatives dk f (x, y1 , . . . , yk ) := (Dyk · · · Dy1 f )(x) exist for all k ∈ N such that k ≤ r, x ∈ U and y1 , . . . , yk ∈ E, and the maps dk f : U × E k → F so obtained are continuous. If K is understood, we write

252

Helge Gl¨ ockner

C r instead of CKr . If f is C ∞ , we also say that f is smooth. If K = R, we say → FC on an that f is real analytic (or CRω ) if f extends to a CC∞ -map f : U open neighbourhood U of U in the complexification EC of E. 2.2 We mention that a map f : E ⊇ U → F is CC∞ if and only if it is complex analytic, i.e., f is continuous and for each x ∈ U , there exists a 0-neighbourhood Y ⊆ E with x + Y ⊆ U and continuous homogeneous polynomials pn : E → F of degree n such that ∞

pn (y) . (∀y ∈ Y ) f (x + y) = n=0

Complex analytic maps are also called C-analytic or CCω . 2.3 It is known that compositions of composable CKr -maps are CKr , for each r ∈ N0 ∪ {∞, ω}. Thus a CKr -manifold M modelled on a locally convex K-vector space E can be defined in the usual way, as a Hausdorff topological space, together with a maximal set of homeomorphisms from open subsets of M to open subsets of E, such that the domains cover M and the transition maps are CKr . Given r ∈ {∞, ω}, a CKr -Lie group is a group G, equipped with a structure of CKr -manifold modelled on a locally convex space, such that the group multiplication and group inversion are CKr -maps. Unless the contrary is stated, we consider CK∞ -Lie groups. Throughout the following, the words “manifold” and “Lie group” will refer to manifolds and Lie groups modelled on locally convex spaces. We shall write Tx M for the tangent space of a manifold M at x ∈ M and L(G) := T1 (G) for the (topological) Lie algebra of a Lie group G. Given a CK1 -map f : M → N between CK1 -manifolds, we write Tx f : Tx M → Tf (x) N for the tangent map at x ∈ M . Given a smooth homomorphism f : G → H, we let L(f ) := T1 (f ) : L(G) → L(H). Direct limits. We recall terminology and basic facts concerning direct limits. 2.4 (General definitions). Let (I, ≤) be a directed set, i.e., I is a non-empty set and ≤ a partial order on I such that any two elements have an upper bound. Recall that a direct system (indexed by (I, ≤)) in a category A is a pair S := ((Xi )i∈I , (φji )j≥i ), where each Xi is an object of A and φji : Xi → Xj a morphism such that φii = idXi and φkj ◦ φji = φki , for all elements k ≥ j ≥ i in I. A cone over S is a pair (X, (φi )i∈I ), where X is an object of A and each φi : Xi → X a morphism such that φj ◦ φji = φi whenever j ≥ i. A cone (X, (φi )i∈I ) is a direct limit of S (and we write (X, (φi )i∈I ) = lim S or X = −→ lim Xi ), if for every cone (Y, (ψi )i∈I ) over S, there exists a unique morphism −→ ψ : X → Y such that ψ ◦ φi = ψi for all i ∈ I. If T = ((Yi )i∈I , (ψji )j≥i ) is another direct system over the same index set, (Y, (ψi )i∈I ) a cone over T and (ηi )i∈I a family of morphisms ηi : Xi → Yi which is compatible with the direct systems in the sense that ψji ◦ηi = ηj ◦φji for all j ≥ i, then (Y, (ψi ◦ηi )i∈I ) is a cone over S. We write lim ηi for the induced morphism ψ : X → Y , determined −→ by ψ ◦ φi = ψi ◦ ηi .


253

Direct limits in the categories of sets, groups and topological spaces are particularly easy to understand, and we discuss them now. Direct limits of topological groups (which are a more difficult topic) and direct limits of locally convex spaces will be discussed in Sections 3 and 4, respectively. We concentrate on direct sequences (viz., the case I = N) and actually on ascending sequences, to avoid technical complications. This is the more justified because (except for some counterexamples) hardly anything is known about direct limits of direct systems of Lie groups which do not admit a cofinal subsequence. 2.5 (Ascending unions of sets). If X1 ⊆ X2 ⊆ · · · is an ascending sequence of sets, let φm,n : Xn → Xm be the inclusion map for m, n ∈ N with m ≥ n. Then S := ((Xn )n∈N , (φm,n )m≥n ) is a direct system in the category SET of sets and maps. Define X := n∈N Xn and let φn : Xn → X be the inclusion map. Then (X, (φn )n∈N ) is a cone over S in SET. A sequence of maps ψn : Xn → Y to a set Y gives rise to a cone (Y, (ψn )n∈N ) if and only if ψm |Xn = ψn

for all m, n ∈ N with m ≥ n.

Then ψ : X → Y , ψ(x) := ψn (x) if n ∈ N and x ∈ Xn is a well-defined map, and is uniquely determined by the requirement that ψ ◦ φn = ψ|Xn = ψn for each n ∈ N. Thus (X, (φn )n∈N ) = lim S in SET. −→

2.6 (Direct limits of groups). If each Xn is a group in the situation of 2.5 and each φm,n a homomorphism, then S is a direct system in the category G of groups and homomorphisms. If x, y ∈ X, there exists n ∈ N such that x, y ∈ Xn . We define the product of x and y in X as their product in Xn , i.e., x · y = φn (x) · φn (y) := φn (x · y). Since each φm,n is a homomorphism, x · y is independent of the choice of n, and it is clear that the product so defined makes X a group and each φn : Xn → X a homomorphism. If (Y, (ψn )n∈N ) is a cone over S in G, let ψ : X → Y be the unique map such that ψ ◦ φn = ψn for each n ∈ N, as in 2.5. Given x, y ∈ X, say x, y ∈ Xn , we then have ψ(xy) = ψ(φn (xy)) = ψn (xy) = ψn (x)ψn (y) = ψ(x)ψ(y), whence ψ is a homomorphism. Thus (X, (φn )n∈N ) = lim S in G. −→

If S = ((Xn )n∈N , (φm,n )) is a direct sequence of groups (with φm,n : Xn → Xm not necessarily injective), then Kn := m≥n ker(φm,n ) is a normal subgroup of Xn . Consider the quotient groups Gn := Xn /Kn , the canonical quotient maps qn : Xn → Gn and the homomorphisms ψm,n : Gn → Gm determined by ψm,n ◦ qn = qm ◦ φm,n . Then each ψm,n is injective, and it is clear that the direct limit (G, (ψn )n∈N ) of the “injective quotient system” ((Gn )n∈N , (ψm,n )) yields a direct limit (G, (ψn ◦ qn )n∈N ) of S (cf. [NRW91, §3]). 2.7 (Direct limits of topological spaces). If each Xn is a topological space in the situation of 2.5 and each φm,n : Xn → Xm a continuous map, we equip X = n∈N Xn with the finest topology ODL making each inclusion

254

Helge Gl¨ ockner

map φn : Xn → X continuous (called the direct limit topology). Thus U ⊆ X is open (resp., closed) if and only if φ−1 n (U ) = U ∩ Xn is open (resp., closed) in Xn for each n ∈ N. Then (X, (φn )n∈N ) = lim S in the category TOP of −→ topological spaces and continuous maps. To see this, let (Y, (ψn )n∈N ) be a cone over S in TOP. Let ψ : X → Y be the unique map with ψ ◦ φn = ψn for each n. If U ⊆ Y is open, then ψ −1 (U ) ∩ Xn = (ψ|Xn )−1 (U ) = (ψn )−1 (U ) is open in Xn , for each n. Hence ψ −1 (U ) is open in X and thus ψ is continuous. The direct system S is called strict if each φn is a topological embedding (i.e., Xn+1 induces the topology of Xn ). Then also the inclusion map φn : Xn → X is a topological embedding for each n [NRW93, Lemma A.5]. It is also known that X has the separation property T1 if each Xn is T1 (see, e.g., [Gl05b, Lemma 1.7 (a)]). And in the case of a direct sequence X1 ⊆ X2 ⊆ · · · of locally compact spaces Xn , the direct limit topology on n∈N Xn is Hausdorff (as observed in [Gl05b, Lemma 1.7 (c)], the strictness hypotheses in [Ha71, Proposition 4.1 (ii)] and [Gl03a, Lemma 3.1] is unnecessary). First remarks on ascending unions of Lie groups and direct limits. Consider an ascending sequence G1 ⊆ G2 ⊆ · · · of CK∞ -Lie groups, such that the inclusion maps jm,n : Gn → Gm are C ∞ -homomorphisms for all m, n ∈ N with m ≥ n. Then S := ((Gn )n∈N , (jm,n )m≥n ) is a direct system in the category LIEK of CK∞ -Lie groups and CK∞ -homomorphisms. One would not expect that S always has a direct limit in the category of CK∞ -Lie groups (although no counterexamples are known at the time of writing). Also, there is no general construction principle for a Lie group structure on n∈N Gn such that all inclusion maps jn : Gn → G are CK∞ -homomorphisms (unless restrictive conditions are imposed, as in Section 5). 2.8 However, in many concrete cases we are given such a Lie group structure on G := n∈N Gn . Then (G, (jn )n∈N ) is a cone over S in LIEK , and it is natural to ask if G = lim Gn as a Lie group. A sequence of CK∞ -homomorphisms −→ fn : Gn → H to a CK∞ -Lie group H is a cone over S if and only if fm |Gn = fn

for all m, n ∈ N with m ≥ n.

Then f : G → H, f (x) := fn (x) if n ∈ N and x ∈ Gn is a well-defined homomorphism. This map is uniquely determined by the requirement that f ◦ jn = f |Gn = fn for each n ∈ N. Therefore, (G, (jn )n∈N ) = lim S holds in −→ LIEK if and only if each f of the preceding form is CK∞ . A similar argument applies if H is a topological group, smooth manifold or topological space. Thus questions (a)–(d) posed in 1.1 amount to asking: Does G = lim Gn −→ hold : (a) in the category of smooth manifolds (modelled on locally convex spaces) and smooth maps between them? (b) in the category of topological spaces and continuous maps?


255

(c) in the category LIEK of Lie groups? (d) in the category of topological groups and continuous homomorphisms? The Hausdorff property. We allow non-Hausdorff topologies because direct limits are much easier to describe if the Hausdorff property is omitted. In fact, we have already seen that it is always possible to topologize a union X = n∈N Xn of topological spaces in such a way that it becomes the direct limit lim Xn in the category of topological spaces (see 2.7), and likewise a −→ union of topological groups (resp., locally convex spaces) can always be made the direct limit in the category of topological groups, resp., locally convex spaces (see Sections 3 and 4). A mere union X = n∈N Xn is a very concrete object, and easy to work with. In contrast, if each Xn is Hausdorff, then the direct limit lim Xn in the −→ category of Hausdorff topological spaces (resp., Hausdorff topological groups, resp.,Hausdorff locally convex spaces) can only be realized as a quotient of X = n∈N Xn in general, and is a much more elusive object in this case. Luckily, in all situations of relevance for this article, X from above injects continuously into a Lie group and thus X is Hausdorff. Then automatically X also is the direct limit in the category of Hausdorff topological spaces (resp., Hausdorff topological groups, resp., Hausdorff locally convex spaces).

3 Direct limits of topological groups As an intermediate step towards the study of Lie groups, let us consider a sequence G1 ⊆ G2 ⊆ · · · of topological groups, such that all inclusion maps Gn → Gn+1 are continuous homomorphisms. We make G = n∈N Gn the direct limit group (as in 2.6) and give it the finest group topology ODLG making each inclusion map Gn → G continuous. Then G = lim Gn in the −→ category of (not necessarily Hausdorff) topological groups. Moreover, if each Gn is Hausdorff, then the factor group of G modulo the closure {1} ⊆ G is the direct limit in the category of Hausdorff topological groups.3 Unfortunately, the preceding description of the topology ODLG on the direct limit topological group is not at all concrete. Various questions are natural (and also relevant for our studies of Lie groups): Does ODLG coincide with the direct limit topology ODL (as in 2.7)? Can ODLG be described more explicitly? Given a group topology on G = n∈N Gn , how can we prove that it agrees with ODLG ? We now give some answers to the first and last questions. An answer to the second question, namely the description of ODLG as a “bamboo-shoot” topology, can be found in [TSH98] and [HST01] (under suitable hypotheses).

3

If G is Hausdorff, then no passage to the quotient is necessary.

256

Helge Gl¨ ockner

Comparison of ODL and ODLG . It is clear from the definition that the direct limit topology ODL is finer than ODLG . Moreover, ODL may be properly finer than ODLG , as emphasized by Tatsuuma et al. [TSH98].4 To understand this difficulty, let ηn : Gn → Gn , x → x−1 and η : G → G be the inversion maps and µn : Gn × Gn → Gn , (x, y) → xy as well as µ : G × G → G be the respective group multiplication. Then η = lim ηn : lim Gn , ODL → lim Gn , ODL −→

−→

−→

is always continuous. However, it may happen that µ is discontinuous (with respect to the product topology on G × G), in which case (G, ODL ) is not a topological group and hence ODL = ODLG . We recall a simple example for this pathology from [TSH98]. Example 3.1. Let Gn := Q × Rn−1 with the addition and topology induced by Rn . Identifying Rn−1 with the vector subspace Rn−1× {0} of Rn , we obtain a strict direct sequence G1 ⊆ G2 ⊆ · · · of metrizable topological groups. It can be shown by direct calculation that the direct limit topology ODL does not make the group multiplication on G := n∈N Gn continuous (see [TSH98, Example 1.2]). To understand the difficulties concerning the group multiplication (in contrast to the group inversion) on G = n∈N Gn , note that we always have a continuous map lim µn : lim (Gn × Gn ), ODL → lim Gn , ODL . −→

−→

−→

Thus µ is continuous as a map from (G×G, ODL ) to (G, ODL ); i.e., it becomes continuous if, instead of the product topology, the topology ODL is used on G × G, which makes it the direct limit topological space lim (Gn × Gn ). This −→ topology is finer than the product topology and, in general, properly finer. If the direct limit topology on G × G happens to coincide with the product topology, then (G, ODL ) is a topological group and thus ODL = ODLG (cf. [HST01] and [Gl03a, §3]). The following proposition describes a situation where the two topologies coincide. We recall that a topological space X is said to be a kω -space if it is the direct limit topological space of an ascending sequence K1 ⊆ K2 ⊆ · · · of compact topological spaces (see, e.g., [GGH10] and the references therein).5 Such spaces are always Hausdorff (see 2.7). For example, every σ-compact, locally compact space is a kω -space. A topological space X is called locally kω if every point x ∈ X has an open neighbourhood in X which is a kω -space in the induced topology [GGH10, Definition 4.1]. For example, every locally compact topological space is locally kω . The topological space underlying a topological group G is locally kω if and only if G has an open subgroup which is a kω -space [GGH10, Proposition 5.3]. See [GGH10, 4 5

In part of the older literature, there was some confusion concerning this point. These spaces can also be characterized as the hemicompact k-spaces.


257

Proposition 4.7] for the following fact. The special case where each Xn and Yn is locally compact was first proved in [HST01, Theorem 4.1] (cf. also [Gl03a, Proposition 3.3] for the strict case). Proposition 3.2. Let X1 ⊆ X2 ⊆ · · · and Y1 ⊆ X2 ⊆ · · · be topological spaces with continuous inclusion maps Xn → Xn+1 and Yn → Yn+1 . If each Xn and each Yn is locally kω , then lim (Xn × Yn ) = lim Xn × lim Yn −→

−→

−→

as a topological space. Using the fact that direct limits of ascending sequences of locally kω -spaces are locally kω by [GGH10, Proposition 4.5] (and thus Hausdorff), the preceding discussion immediately entails the following conclusion from [GGH10] (cf. [TSH98, Theorem 2.7] for locally compact Gn , as well as [Gl03a, Corollary 3.4] (in the case of a strict direct system)). Corollary 3.3. Consider a sequence G1 ⊆ G2 ⊆ · · · of topological groups such that each inclusion map Gn → Gn+1 is a continuous homomorphism. If the topological space underlying Gn is locally kω for each n ∈ N (for example, if each Gn islocally compact), then the direct limit topology is Hausdorff and makes G = n∈N Gn the direct limit topological group. 3.4 Given topological groups G1 ⊆ G2 ⊆ · · · such that all inclusion maps Gn → Gn+1 are continuous homomorphisms, consider the conditions: (a) Gn is an open subgroup of Gn+1 (with the induced topology) for all sufficiently large n. (b) For each sufficiently large n, the topological group Gn has an identity neighbourhood U whose closure in Gm is compact for some m ≥ n. Then ODL = ODLG holds if (a) or (b) is satisfied [Ya98, Theorems 2 and 3]. By a most remarkable theorem of Yamasaki [Ya98, Theorem 4], the validity of (a) or (b) is also necessary in order that ODL = ODLG , provided that each Gn is metrizable and the inclusion maps Gn → Gn+1 are topological embeddings. Criteria ensuring that a given group topology coincides with ODLG . Frequently, a given topological group G is a union G = n∈N Gn of topological groups, such that all inclusion maps Gn → Gn+1 and Gn → G are continuous homomorphisms. In many cases, a criterion from [Gl07b] helps to show that the given topology on G coincides with ODLG (cf. [Gl07b, Proposition 11.8]). The criterion uses the weak direct product ∗n∈N Gn as a tool. The latter can be formed for any sequence (Gn )n∈N of topological groups. It is defined as the subgroup of all (gn )n∈N ∈ n∈N Gn such that gn = 1 for all but finitely many n. The weak direct product is a topological group; a basis for ∗its topology (the “box topology”) is given by sets of the form n∈N Un ∩ n∈N Gn (the “boxes”), where Un ⊆ Gn is open for each n and 1 ∈ Un for almost all n.

258

Helge Gl¨ ockner

Returning to the case where G1 ⊆ G2 ⊆ · · · and G = n∈N Gn , we can consider the “product map” ∗ π : n∈N Gn → G , (gn )n∈N → g1 g2 · · · gN , where N ∈ N is so large that gn = 1 for all n > N . ∗ Proposition 3.5. If the product map π : n∈N Gn → G is open at 1, then the given topology on G coincides with ODLG and thus G = lim Gn as a −→ topological ∗group. The openness of π at 1 is guaranteed if there exists a map σ : Ω → n∈N Gn on an identity neighbourhood Ω ⊆ G such that π ◦ σ = idΩ , σ(1) = 1 and σ is continuous at 1. Remark 3.6. Such a section σ to π might be called a fragmentation map, in analogy to concepts in the theory of diffeomorphism groups (cf. [Ba97, §2.1]). Example 3.7. It canbe shown that the Lie groups Diff c (M ) = n∈N Diff Kn (M ) r and Ccr (M, H) = n∈N CK (M, H) (as defined in the Introduction) always n admit fragmentation maps (even smooth ones); cf. [Gl07b, Lemmas 5.5 and r 7.7]. Hence Diff c (M ) = lim Diff Kn (M ) and Ccr (M, H) = lim CK (M, H) as n −→ −→ topological groups.

4 Non-linear mappings on locally convex direct limits Consider a sequence E1 ⊆ E2 ⊆ · · · of locally convex spaces, such that each inclusion map En → En+1 is continuous and linear. Then there is a finest locally convex vector topology Olcx on E := n∈N En , making each inclusion map En → E continuous, called the locally convex direct limit topology.6 4.1 Some basic properties of locally convex direct limits are frequently used. (a) If the direct sequence E1 ⊆ E2 ⊆ · · · is strict, then (E, Olcx ) induces the given topology on En , for each n ∈ N (see Proposition 9 (i) in [Bo87, Chapter II, §4, no. 6]). (b) If the direct sequence E1 ⊆ E2 ⊆ · · · is strict and each En is Hausdorff, then (E, Olcx ) is also Hausdorff (see Proposition 9 (i) in [Bo87, Chapter II, §4, no. 6]). (c) If the direct sequence E1 ⊆ E2 ⊆ · · · isstrict and each En complete, then the locally convex direct limit E = n∈N En is boundedly regular (cf. Proposition 6 in [Bo87, Chapter III, §1, no. 4]) and hence also compactly regular, in view of (a).

6

This topology can be described in various ways. We mention: (1) A convex set U ⊆ E is open if and only if U ∩En is open in En , for each n ∈ N. (2) A seminorm q : E → [0, ∞[ is continuous if and only if q|En is continuous, for each n ∈ N.


259

(d) If the direct sequence E1 ⊆ E2 ⊆ · · · is strict and each En complete, then also the locally convex direct limit E is complete (see Proposition 9 (iii) in [Bo87, Chapter II, §4, no. 6]). (e) If also F1 ⊆ F2 ⊆ · · · is an ascending sequence of locally convex spaces, with locally convex directlimit F = n∈N Fn , then the locally convex direct limit topology on n∈N (En × Fn ) and the product topology on E × F coincide [HST01, Theorem 3.4] (because finite direct products coincide with finite direct sums in the category of locally convex spaces). The reader may find [Fl80] and [Bi88] convenient points of entry to the research literature on locally convex direct limits. We mention that few general results ensuring the Hausdorff property for locally convex direct limits E = lim En are known (besides 4.1 (b) just encoun−→ tered and Proposition 4.4 below). Some Hausdorff criteria for direct limits of Banach spaces (and normed spaces) can be found in [Fl79] (see also [Fl80, p. 214]). In concrete examples, a very simple argument frequently works: If one can find an injective continuous linear map from E to some Hausdorff locally convex space, then E is Hausdorff. However, the fact remains that non-Hausdorff locally convex direct limits do exist: See [Mk63] for examples where each En is a Banach space, [Fl80, p. 207] for a simple example (due to L. Waelbroeck) with each En a normed space and [Fl80, p. 227, Corollary 2] for an example where each En is a nuclear Fréchet space (cf. also [Sv69]). It is well known that O and O coincide on E = lcx DLG n∈N En , because

n on E = E , both O and O coincide with the box n k lcx DLG n∈N n∈N k=1 topology and n∈N En (with either topology) can be considered as a quotient of the direct sum (see [Gl07b, Lemma 2.7]; cf. [HST01, Proposition 3.1] for a different argument. See also [Ko69] and [Bo87, Chapter II, Exercise 14 to §4]). It is also known that Olcx need not coincide with ODL (see [Sr59] or Exercise 16 (a) to §4 in [Bo87, Chapter II]; cf. also [Du64, p. 506]). For example, Olcx is properly coarser than ODL if each En is an infinite-dimensional Fréchet space and En is a proper vector subspace of En+1 with the induced topology, for each n ∈ N ([KM97, Proposition 4.26 (ii)]; cf. Yamasaki’s theorem recalled in Section 3). The following concrete example shows that not even smoothness or analyticity of f |En ensures that a map f : E → F on a locally convex direct limit E = n∈N En is continuous (let alone smooth or analytic). Example 4.2. Consider the map g : Cc∞ (R, C) → Cc∞ (R × R, C) ,

g(γ) := γ ⊗ γ

between spaces of compactly supported smooth functions, where (γ ⊗ γ)(x, y) := γ(x)γ(y) for x, y ∈ R. It can be shown that g is discontinuous, although ∞ ∞ ∞ g|C[−n,n] (R,C) : C[−n,n] (R, C) → Cc (R × R, C) is a continuous homogeneous polynomial (and hence complex analytic), for each n ∈ N (see Remark 7.9 in [Gl07b], based on [HST01, Theorem 2.4]).

260

Helge Gl¨ ockner

Remark 4.3. Consider the locally convex direct limit E = n∈N En of Hausdorff locally convex spaces E1 ⊆ E2 ⊆ · · · over K ∈ {R, C}. Let U1 ⊆ U2 ⊆ · · · be an ascending sequence of open sets Un ⊆ En , and U := n∈N Un . Let r ∈ N ∪ {∞}, F be a Hausdorff locally convex space and f : U → F be a map such that f |Un : En ⊇ Un → F is CKr for each n ∈ N. Assume that E is Hausdorff and U ⊆ E is open.7 Then the iterated directional derivatives dk f (x, y1 , . . . , yk ) = (Dyk · · · Dy1 f )(x) exist for all k ∈ N with k ≤ r and all x ∈ U and y1 , . . . , yk ∈ E, because x ∈ Un and y1 , . . . , yk ∈ En for some n ∈ N and then (Dyk · · · Dy1 f )(x) = dk (f |Un )(x, y1 , . . . , yk ). Hence only continuity of the maps dk f , which satisfy dk f |Un ×(En )k = dk (f |Un ) for all n ∈ N,

(5)

may be missing for some k, and may prevent f from being a CKr -map. We mention that locally convex direct limits of ascending sequences of Banach spaces (resp., Fréchet spaces) are called (LB)-spaces (resp., (LF)-spaces). If the sequence is strict, we speak of LB-spaces (resp., LF-spaces).8 A locally convex space E is called a Silva space if it is the locally convex direct limit of an ascending sequence E1 ⊆ E2 ⊆ · · · of Banach spaces, such that all inclusion maps En → En+1 are compact operators (cf. [Se55] and [Fl71]).9 Silva spaces are very well-behaved direct limits. We recall the following from [Fl71]. Proposition 4.4. If E = n∈N En is a Silva space, then the following hold: (a) E is Hausdorff and complete; (b) E = n∈N En is boundedly regular and hence also compactly regular; 10 (c) The locally convex direct limit topology on E coincides with the direct limit topology ODL ; (d) If F = F n∈N n is also a Silva space, with Fn → Fn+1 compact, then E × F = n∈N (En × Fn ) is a Silva space.11 Some interesting infinite-dimensional Lie groups are modelled on Silva spaces, e.g., the group Diff ω (K) of real analytic diffeomorphisms of a compact real analytic manifold K (see [Ls82]; cf. [KM97, Theorem 43.4]). More examples will be encountered below.

7 8 9

10 11

For example, we might start with an open set U ⊆ E and set Un := U ∩ En . These conventions are local. The meanings of LF and (LF) vary in the literature. A locally convex space is a Silva space if and only if it is isomorphic to the dual of a Fréchet–Schwartz space [Fl71]; therefore, Silva spaces are also called (DFS)-spaces. Using the fact that the inclusion maps En → En+1 are compact operators. The inclusions En × Fn → En+1 × Fn+1 are compact operators, and 4.1 (e) holds.


261

Mappings on Silva spaces or unions of kω -spaces. In good cases, the pathology described in Remark 4.3 cannot occur (see [Gl07b, Lemma 9.7] and [GGH10, Proposition 8.13]). Proposition 4.5. Consider the locally convex direct limit E = n∈N En of Hausdorff locally convex spaces E1 ⊆ E2 ⊆ · · · over K ∈ {R, C}. Let U1 ⊆ U2 ⊆ · · · be an ascending sequence of open sets Un ⊆ En , and U := n∈N Un . Let r ∈ N0 ∪ {∞}, F be a Hausdorff locally convex space and f : U → F be a map such that f |Un is CKr for each n ∈ N. Assume that (a) Each En is a kω -space, or (b) En is a Banach space and the inclusion map En → En+1 is a compact operator, for each n ∈ N (in which case E is a Silva space). Then E is Hausdorff and the locally convex direct limit topology on E coincides with ODL . Moreover, U is open in E and f : U → F is CKr . In the Silva case, the hypotheses of Proposition 4.5 can be relaxed (cf. [Ls85, Proposition 2.8]). Real analyticity is more elusive. For example, there exists a real-valued map f on the Silva space R(N) := lim Rn which is not real analytic −→ although f |Rn is real analytic for each n ∈ N (cf. [KM97, Example 10.8]). Complex analytic maps on (LB)-spaces. A very useful result from [Da10] frequently helps us to check complex analyticity beyond Silva spaces. Theorem 4.6 (Dahmen’s theorem). Let E1 ⊆ E2 ⊆ · · · be an ascending sequence of normed spaces (En , .n ) over C such that, for each n ∈ N, the inclusion map En → En+1 is continuous and complex linear, of operator norm at most 1. Let r ∈ ]0, ∞[, Un := {x ∈ En : xn < r} for n ∈ N and F be a complex locally convex space. Assume that the locally convex direct limit E = n∈N En is Hausdorff. Then U := n∈N Un is open in E, and if f: U → F is a map such that f |Un : En ⊇ Un → F is complex analytic and bounded for each n ∈ N, then f is complex analytic. Mappings between direct sums. If (En )n∈N is a sequence of locally

convex spaces, we equip n∈N En with the box topology (as introduced before Proposition 3.5). See [Gl03b, Proposition 7.1] for the following result. Proposition 4.7. Let (En )n∈N and (Fn )n∈N be sequences of Hausdorff locally convex spaces, r ∈ N0 ∪ {∞}, Un ⊆ En be open and fn : Un → Fn be C r . Assume that 0 ∈ Un and fn (0) = 0 for all but finitely many n ∈ N. Then U := ( E ) ∩ n∈N n∈N n∈N Un is open in n∈N En and the map

n

n r ⊕n∈N fn : U → F , (x ) → (f (x )) n n∈N n n n∈N is C . n∈N n n∈N n Non-linear maps between spaces of test functions. Let r, s ∈ N0 ∪ {∞}, M be a σ-compact, finite-dimensional C r -manifold, N be a σ-compact,

262

Helge Gl¨ ockner

finite-dimensional C s -manifold, E, F be Hausdorff locally convex spaces, Ω ⊆ Ccr (M, E) be open and f : Ω → Ccs (N, F ) be a map. We say that f is almost local if there exist locally finite covers (Un )n∈N and (Vn )n∈N of M (resp., N ) by relatively compact, open sets Un ⊆ M (resp., Vn ⊆ N ) such that f (γ)|Vn only depends on γ|Un , i.e., (∀n ∈ N) (∀γ, η ∈ Ω)

γ|Un = η|Un ⇒ f (γ)|Vn = f (η)|Vn .

For example, f is almost local if M = N and f is local in the sense that f (γ)(x) only depends on the germ of γ at x ∈ M . As shown in [Gl02d] (see also [Gl04b, Theorem 10.4]), almost locality prevents pathologies as in Example 4.2. Proposition 4.8. Let r, s, t ∈ N0 ∪ {∞} and f : Ccr (M, E) ⊇ Ω → Ccs (N, F ) r be an almost local map. Assume that the restriction of f to Ω ∩ CK (M, E) is t t C , for each compact set K ⊆ M . Then f is C . An analogous result is available for mappings between open subsets of spaces of compactly supported sections in vector bundles. Almost local maps between subsets of the space of compactly supported smooth vector fields occur in the construction of the Lie group structure on Diff c (M ) (see [Gl02d]; cf. [Gl05a] and [Gl04b]). Proof of Proposition 4.8. The proof exploits the fact that the map σ : Ccs (N, F ) → C s (Vn , F ) , γ → (γ|Vn )n∈N n∈N

is a linear topological embedding with closed image [Gl04b, Proposition 8.13], for each locally finite cover (Vn )n∈N of N by relatively compact, open sets Vn . It hence suffices to show that σ◦f is C t . Let us assume that Ω = Ccr (M, E) for n )n∈N of M by relatively compact, simplicity. There is a locally finite cover (U n → R be a open sets such that Un contains the closure of Un . Let hn : U compactly supported smooth map such that hn |Un = 1. Then the following map is C t : n , E) → C s (Vn , F ) , fn (γ) := f (hn · γ)|Vn . fn : C r (U

n , E), γ → (γ| )n . Then σ◦f = (⊕n∈N fn )◦ρ, Set ρ : Ccr (M, E) → n∈N C r (U Un t where ⊕n∈N fn is C by Proposition 4.7. Hence σ ◦ f and thus f is C t .

5 Lie group structures on directed unions of Lie groups In some situations, it is possible to construct Lie group structures on ascending unions of Lie groups. Unions of finite-dimensional Lie groups. In the case of finite-dimensional manifolds, an extension lemma for charts is available [Gl05b, Lemma 2.1].


263

If M and N are finite-dimensional C ∞ -manifolds such that M ⊆ N and the inclusion map M → N is an immersion, then each chart φ : U → V of M which is defined on a relatively compact, (smoothly) contractible subset U ⊆ M extends to a chart of N on a domain with analogous properties. Now consider a sequence G1 ⊆ G2 ⊆ · · · of finite-dimensional Lie groupssuch that the inclusion maps are smooth homomorphisms. Let x ∈ G := n∈N Gn , say x ∈ Gn0 . We then pick a chart φn0 of Gn0 around x whose domain is relatively compact and contractible, and use the extension lemma to obtain charts φn of Gn for n > n0 which are defined on relatively compact, contractible open sets, and such that φn extends φn−1 . One then easily verifies (using Proposition 4.5) that the homeomorphisms φ := lim φn so obtained define a C ∞ -atlas on G (equipped with the −→ direct limit topology), which makes the lattera Lie group modelled on lim L(Gn ) (see [Gl05b]).12 By construction, G = n∈N Gn admits direct limit −→ charts. Moreover, it is clear from the construction that G = lim Gn as a topo−→ logical space and as a topological group. Using Proposition 4.5, one easily infers that G = lim Gn also as a smooth manifold and as a Lie group (see −→ [Gl05b, Theorem 4.3]). Remark 5.1. The preceding construction applies just as well to ascending unions of finite-dimensional smooth manifolds Mn , such that all inclusion maps are immersions.13 This enables G/H to be turned into the direct limit C ∞ -manifold lim Gn /(H ∩ Gn ), for each closed subgroup H ⊆ G (see [Gl05b, −→ Proposition 7.5]). Then the quotient map G → G/H makes G a principal H-bundle over G/H, using a suitable extension lemma for sections in nested principal bundles [Gl05b, Lemma 6.1]. We mention that an equivariant version of the above extension lemma (namely [Wk07, Lemma 1.13]) can be used to turn the gauge group Gau(P ) into a Lie group, for each smooth principal bundle P → K over a compact smooth manifold K whose structure group is a direct limit G = lim Gn of −→ finite-dimensional Lie groups (see [Wk07, Lemma 1.14 (e) and Theorem 1.11]). Remark 5.2. Direct limits G = lim Gn of finite-dimensional Lie groups are −→ regular Lie groups in Milnor’s sense [Gl05b, Theorem 8.1], but they can be quite pathological in other ways. For example, the exponential map expG = lim expGn need not be injective on any 0-neighbourhood, and the exponential −→ image need not be an identity neighbourhood in G. Both pathologies occur for G := C(N) × R = lim Cn × R , −→

via t.(zk )k∈N := (eikt zk )k∈N . This can be checked where t ∈ R acts on C quite easily, using the fact that the exponential map of G is given explicitly ikt by expG ((zk )k∈N , t) = e ikt−1 zk k∈N , t (see [Gl03a, Example 5.5]). (N)

12 13

Ë

See [NRW91], [KM97, Theorem 47.9] and [Gl03a] for earlier, less general results. Compare [Ha71] for n∈N Mn as a topological manifold.

264

Helge Gl¨ ockner

The preceding general construction implies thatevery countably dimensional locally finite Lie algebra g (i.e., each union g = n∈N gn of finite-dimensional Lie algebras g1 ⊆ g2 ⊆ · · · ), when endowed with the finest locally convex vector topology, arises as the Lie algebra of some regular Lie group.14 Such locally finite Lie algebras have been much studied in recent years, e.g., by Yu. Bahturin, A. A. Baranov, G. Benkart, I. Dimitrov, K.-H. Neeb, I. Penkov, H. Strade, N. Stumme, A. E. Zalesski˘ı and others (see [BBZ04], [BB04], [DP04], [Ne00], [NS01], [PS03], [St99] and the references therein). Unions of Banach–Lie groups. These are Lie groups under additional hypotheses (which, e.g., exclude the pathologies described in Remark 5.2). Theorem 5.3. Let G1 ⊆ G2 ⊆ · · · be Banach–Lie groups over K ∈ {R, C}, ∞ such that all inclusion maps λn : Gn → Gn+1 are CK -homomorphisms. Set G := n∈N Gn . Assume that (a)–(c) are satisfied. (a) For each n ∈ N, there exists a norm .n on L(Gn ) defining its topology, such that [x, y]n ≤ xn yn for all x, y ∈ L(Gn ) and the continuous linear map L(λn ) : L(Gn ) → L(Gn+1 ) has operator norm at most 1. (b) The locally convex direct limit topology on g := n∈N L(Gn ) is Hausdorff. (c) expG := lim expGn : g → G is injective on some 0-neighbourhood. −→

Then there exists a K-analytic Lie group structure on G which makes expG a K-analytic local diffeomorphism at 0. If, furthermore, g = n∈N L(Gn ) is compactly regular, then G is a regular Lie group in Milnor’s sense. Sketch of proof. The Lie group structure is constructed in [Da10], along the following lines: Applying Dahmen’s theorem 4.6 (to the complexification gC , if K = R), one finds that the Baker–Campbell–Hausdorff (BCH-) series con L(G ) L(G ) verges to a K-analytic map n∈N Br n (0) × Br n (0) → g for some r > 0. Because expG is locally injective, it induces an isomorphism φ of local groups L(G ) from some 0-neighbourhood U ⊆ n∈N Br n (0) onto some subset V of G. We give V the CKω -manifold structure making φ a CKω -diffeomorphism. Now standard arguments can be used to make G a Lie group with V as an open submanifold. The proof of regularity will be sketched in Section 8. The author does not know whether the Lie groups G in Theorem 5.3 are always the direct limit lim Gn in the category of Lie groups (unless additional −→ hypotheses are satisfied). Another construction principle. There is another construction principle for a Lie group structureon a union G = n∈N Gn of Lie groups (or a group which is a union G = n∈N Mn of manifolds), which produces Lie groups modelled on Silva spaces or ascending unions of kω -spaces. A direct limit Lie group structure can be constructed on G if (1) there are compatible charts φn 14

One chooses a simply connected Lie group Gn with Lie algebra gn and forms the direct limit group G = lim Gn (see [Gl05b, Theorem 5.1] for the details). −→


265

of the Lie groups Gn (resp., the manifolds Mn ) around each point in G; and (2) suitable hypotheses are satisfied which ensure that the transition maps between charts of the form lim φn are CK∞ , because they are mappings of the −→ form discussed in Proposition 4.5 (see [Gl07b, Lemma 14.5]).

6 Examples of directed unions of Lie groups The main examples of ascending unions of infinite-dimensional Lie groups were already briefly described in the Introduction. We now provide more details. Notably, we discuss the existence of direct limit charts and compact regularity. As already mentioned, the latter gives information on the homotopy groups (see (1)) and can help to verify regularity in Milnor’s sense (see Theorem 5.3 and Section 8). A special case of [Gl10, Corollary 3.6] is useful. Lemma 6.1. If the Lie group G = n∈N Gn admits a weak direct limit chart, then G = n∈N Gn is compactly regular if and only if L(G) = n∈N L(Gn ) is compactly regular. In the case of an (LF)-space E = n∈N En , there is a quite concrete characterization of compact regularity only in terms of properties of the steps En (see [We03, Theorem 6.4 and its corollary]). Theorem 6.2. Let E1 ⊆ E2⊆ · · · be Fréchet spaces, with continuous linear inclusion maps. Give E = n∈N En the locally convex direct limit topology. Then E = n∈N En is compactly regular if and only if for each n ∈ N, there exists m ≥ n such that for all k ≥ m, there is a 0-neighbourhood U in En on which Ek and Em induce the same topology. In this case, E is also boundedly regular and complete. We mention that a Hausdorff (LF)-space is boundedly regular if and only if it is Mackey complete [Fl80, 1.4 (f), p. 209]. Groups ofcompactly supported diffeomorphisms. The Lie group Diff c (M ) = n∈N Diff Kn (M ) (discussed in the Introduction) admits a direct limit chart (cf. [Gl07b, §5.1]). Moreover, the LF-space Vc (M ) = n∈N VKn (M ) is compactly regular (see 4.1 (c)) and hence also Diff c (M ) (by Lemma 6.1). To avoid exceptional cases in our later discussions of direct limit properties, we assume henceforth that M is non-compact and of positive dimension. Test function groups. Let M and an exhaustion K1 ⊆ K2 ⊆ · · · of M be as in the definition of Diff c (M ), H be a Lie group modelled on a locally convex space and r ∈ N0 ∪ {∞}. We consider the “test function group” Ccr (M, H) of C r -maps γ : M → H such that the closure of {x ∈ M : γ(x) = 1} (the r support of γ) is compact. Let CK (M, H) be the subgroup of functions supn r r ported in Kn . Then CKn (M, H) is a Lie group modelled on CK (M, L(H)), n r and Cc (M, H) is a Lie group modelled on the locally convex direct limit

266

Helge Gl¨ ockner

r Ccr (M, L(H)) = lim CK (M, L(H))([Gl02b]; cf. 1.8 for special cases). Also, n −→

Ccr (M, H) =

n∈N

r CK (M, H) n

admits a direct limit chart (cf. [Gl07b, §7.1]). Furthermore, Ccr (M, L(H)) = r n CKn (M, L(H)) is compactly regular as a consequence of 4.1 (c). We now assume that H is non-discrete and M non-compact, of positive dimension. Weak direct products of Lie groups. Given a sequence (Hn )n∈N of Lie groups, its weak direct product G := ∗n∈N Hn (as introduced before Proposition 3.5) has a natural §7], modelled on the locally

Lie group structure [Gl03b, convex direct sum L(H ). Then G = n n∈N Gn , identifying the n n∈N partial product Gn := k=1 Hk with a subgroup of G. By construction, G = n∈N Gn

has a direct limit chart. Furthermore, L(G) = n∈N L(Hn ) = lim L(Gn ) is −→ compactly regular, as locally convex direct sums are boundedly regular [Bo87, Ch. 3, §1, no. 4, Proposition 5] and induce the given topology on each finite partial product (cf. Propositions 7 or 8 (i) in [Bo87, Ch. 2, §4, no. 5]). To avoid exceptional cases, we assume henceforth that each Hn is non-discrete. Unit groups of unions of Banach algebras. Let A1 ⊆ A2 ⊆ · · · be unital complex Banach algebras (such that all inclusion maps are continuous homo morphisms of unital algebras). Give A := n∈N An the locally convex direct limit topology. Then A× is open in A and if A is Hausdorff (which we assume now), then A× is a complex Lie group [Gl07b, Proposition 12.1]. Moreover, A× = n∈N A× n , and the identity map idA× is a direct limit chart (cf. [DW97] and [Ed99] for related results). If each inclusion map An → An+1 is a topological embedding or each inclusion map a compact operator, then A = A and hence also A× = n∈N n × An are compactly regular. However, for particular choices of the steps, n∈N A = n∈N An is not compactly regular (see [Gl10, Example 7.8], based on [BMS82, Remark 1.5]). Lie groups of germs of analytic mappings. Let H be a complex Banach– Lie group, . be a norm on L(H) defining its topology, X be a complex metrizable locally convex space and K ⊆ X be a non-empty compact set. Let W1 ⊇ W2 ⊇ · · · be a fundamental sequence of open neighbourhoods of K in X such that each connected component of Wn meets K. Then the set Germ(K, H) of germs around K of H-valued complex analytic functions on open neighbourhoods of K can be made a Lie group modelled on the locally convex direct limit Germ(K, L(H)) = lim Holb (Wn , L(H)) −→

of the Banach spaces gn := Holb (Wn , L(H)) of bounded L(H)-valued complex analytic functions on Wn , equipped with the supremum norm (see [Gl04a]). The group operation arises from pointwise multiplication of representatives of germs. The identity component Germ(K, H)0 is the union


Germ(K, H)0 =

267

Gn

n∈N

of the Banach–Lie groups Gn := [expH ◦ γ] : γ ∈ gn , and Germ(K, H)0 = n∈N Gn admits a direct limit chart [Gl07b, §10.4]. Theorem 6.2 implies that Germ(K, L(H)) = n∈N gn is compactly regular (see [DG10]), and thus Germ(K, H)0 = n∈N Gn is compactly regular (see [Ch85, Theorems 21.15 and 21.23] for the bounded regularity and completeness of Germ(K, L(H)) if X is a normed space; cf. [Mj79]). In the most relevant case where X and H are finite-dimensional, we can choose Wn+1 relatively compact in Wn . Then the restriction maps Holb (Wn , L(H)) → Holb (Wn+1 , L(H)) are compact operators [Gl07b, §10.5] and thus Germ(K, L(H)) is a Silva space. Lie groups of germs of analytic diffeomorphisms. If X is a complex Banach space and K ⊆ X a non-empty compact set, let GermDiff(K, X) be the set of germs around K of C-analytic diffeomorphisms γ : U → V between open neighbourhoods U and V of K (which may depend on γ), such that γ|K = idK . Then GermDiff(K, X) is a Lie group modelled on the locally convex direct limit Germ(K, X)K := lim Holb (Wn , X)K , −→

where Wn and Holb (Wn , X) are as in the last example and Holb (Wn , X)K := {ζ ∈ Holb (Wn , X) : ζ|K = 0} (see [Gl07b, §15] for the case dim(X) < ∞, and [Da10] for the general result). The group operation arises from composition of representatives of germs. Now the set Mn of all elements of GermDiff(K, X) having a representative in Holb (Wn , X)K is a Banach manifold, and Mn GermDiff(K, X) = n∈N

has a direct limitchart (see [Da10]; cf. [Gl07b, Lemma 14.5 and §15]). GermDiff(K, X) = n Mn is compactly regular by Theorem 6.2 and Lemma 6.1 (see [Da10]); if X is finite dimensional, then Germ(K, X)K is a Silva space. Unions of Lie groups1 modelled on Sobolev spaces. The Lie groups H ↓s (K, F ) = n∈N H s+ n (K, F ) (as in the Introduction) are studied in the work in progress [DG10]. By construction, they admit a direct limit chart, and 1 they are modelled on the Silva space H ↓s (K, L(F )) = n∈N H s+ n (K, L(F )) (and hence compactly regular). We mention that the Lie group structure on H ↓s (K, F ) can be obtained via Theorem 5.3; therefore, H ↓s (K, F ) is a regular Lie group in Milnor’s sense. Compare [Pk08] (in this volume) for analysis and probability theory on variants of the Lie groups H s (K, F ) (with s > dim(K)/2), and limit processes as s ↓ dim(K)/2.

268

Helge Gl¨ ockner

7 Direct limit properties of ascending unions We now discuss the direct limit properties of ascending unions of infinitedimensional Lie groups in the categories of Lie groups, topological groups, smooth manifolds and topological spaces. Tools to prove or disprove direct limit properties. Such tools were provided in [Gl07b]. Recall that a real locally convex space E is said to be smoothly regular (or to admit smooth bump functions) if the topology on E is initial with respect to C ∞ (E, R). Remark 7.1. If U ⊆ E is a 0-neighbourhood and the topology is initial n with respect to C ∞ (E, R), then j=1 fj−1 (]−ε, ε[) ⊆ U for suitable ε > 0 and f1 , . . . , fn ∈ C ∞ (E, R) such that f1 (0) = · · · = fn (0) = 0. Then f −1 (]−δ, δ[) ⊆ U with f := f12 + · · · + fn2 and δ := ε2 . Let g : R → R be a smooth function such that g(R) ⊆ [0, 1], g(0) = 1 and g(x) = 0 if |x| ≥ δ/2. Then h := g ◦ f : E → R is a smooth function such that h(0) = 1 and supp(h) ⊆ U (a “smooth bump function” supported in U ). This explains the terminology. Example 7.2. Every Hilbert space H admits smooth bump functions (because H → R, x → x2 is smooth). As a consequence, every locally convex space which admits a linear topological embedding into a direct product of Hilbert spaces (for example, every nuclear locally convex space) admits smooth bump functions (cf. also [KM97, Chapter III]). Proposition 7.3. Consider a Lie group G = n∈N Gn , where G1 ⊆ G2 ⊆ · · · are Lie groups and all inclusion maps Gn → Gn+1 and Gn → G are smooth homomorphisms. Assume that G = n∈N Gn admits a direct limit chart. Then the following hold. (a) If G = lim Gn as a topological group, then G = lim Gn as a Lie group. −→ −→ (b) G = lim Gn as a topological space if and only if L(G) = lim L(Gn ) as a −→ −→ topological space. (c) If L(G) admits smooth bump functions, then G = lim Gn as a −→ CR∞ -manifold if and only if L(G) = lim L(Gn ) as a CR∞ -manifold. −→

Direct limit properties of the main examples. Using Proposition 7.3, Proposition 3.5 (to recognize direct limits of topological groups) and a counterpart of Proposition 4.5 for analogous ascending unions of manifolds [Gl07b, Proposition 9.8], one obtains the following information concerning the direct limit properties of the examples from Section 6 (see [Gl07b]; the properties of H ↓s (K, F ) follow from [Gl07b, Proposition 9.8]). The entries in Table 7.1 indicate whether G = lim Gn holds in the category −→ shown on the left, for the Lie group described at the top. The abbreviation “dep” is used if the answer depends on special properties of the group(s) involved. We abbreviate “category” by “cat,” “group” by “gp,” “space” by “sp,” “topological” by “top” and “smooth manifold” by “mfd.”


269

Table 7.1: Direct limit properties in various categories cat\gp Diffc (M ) Cc∞ (M,H) Lie gps top gps mfds top sps

yes yes no no

yes yes no no

ÉH ∗ n

Germ(K,H)0 GermDiff H ↓s (K,F ) (K,X) yes yes yes — yes yes yes yes — yes dep* dep** yes† yes†† yes dep* dep** yes† yes†† yes n

A×

* “yes” if each Hn is finite dimensional or modelled on a kω -space; “no” if each Hn is modelled on an infinite dimensional Fréchet space (which we assume nuclear when dealing with the category of smooth manifolds). Other cases unclear. ** “yes” if each An is finite dimensional or each inclusion map λn : An → An+1 a compact operator; “no” (when dealing with the category of topological spaces), if An is infinite dimensional, An ⊂ An+1 and λn a topological embedding for each n. Other cases unclear. † “yes” if X and H are finite dimensional; general case unknown. †† “yes” if X is finite dimensional; general case unknown.

8 Regularity in Milnor’s sense Experience tellsus that if one tries to prove regularity in Milnor’s sense for a Lie group G = n∈N Gn , then regularity of the Lie groups Gn does not suffice to carry out the desired arguments. But strengthened regularity properties increase the chances for success. Definition 8.1. Given k ∈ N0 , we say that a Lie group G is C k -regular if it is a regular Lie group in Milnor’s sense and evolG : C ∞ ([0, 1], L(G)) → G is smooth with respect to the C k -topology on C ∞ ([0, 1], L(G)) (induced by C k ([0, 1], L(G))). If each γ ∈ C k ([0, 1], L(G)) has a product integral ηγ and the map evolG : C k ([0, 1], L(G)) → G , γ → ηγ (1) is smooth, then we say that the Lie group G is strongly C k -regular. For example, every Banach–Lie group is strongly C 0 -regular [GN10]. Although much of the following remains valid for C k -regular Lie groups, we shall presume strong C k -regularity, as this simplifies the presentation. We also suppress possible variants involving bounded regularity instead of compact regularity. All results presented in this section are taken from [DG10]. In the regularity proofs for our main classes of direct limit groups, we always use an isomorphism C k ([0, 1], lim En ) ∼ = lim C k ([0, 1], En ) at a pivotal −→ −→ point. Let us begin with the elementary case of locally convex direct sums.

270

Helge Gl¨ ockner

Lemma 8.2.

If (En )n∈N

is a sequence of Hausdorff locally convex spaces, then C k ([0, 1], n∈N En ) = n∈N C k ([0, 1], En ), for all k ∈ N0 .

Sketch of proof. The locally convex direct sum n∈N En = n∈N (E1 ×· · ·×En ) is compactly regular, because it is boundedly regular by [Bo87, Chapter 3, §1, no. 4, Proposition 5] and induces the given topology on each finite partial product (cf. or 8 (i) in [Bo87, Chapter 2, §4, no. 5]). Therefore,

Propositions 7 k C k ([0, 1], n∈N En ) and n∈N C ([0, 1], En ) coincide as sets. Comparing 0-neighbourhoods, we see that both vector topologies coincide (using the fact that boxes are typical 0-neighbourhoods in a countable direct sum). ∞ (N) ∞ n Remark 8.3. Although C ([0, 1], R ) = n∈N C ([0, 1], R ) as a set, the topology on the left-hand side is properly coarser than the locally convex direct limit topology Olcx on the right-hand side, because n γ = (γn )n∈N ∈ C ∞ ([0, 1], R(N) ) : (∀n ∈ N) ddxγnn (0) ∈ ]−1, 1[ ∈ Olcx is not a 0-neighbourhood in C ∞ ([0, 1], R(N) ) = lim

←− m∈N0

C m ([0, 1], R(N) ). Thus

Lemma 8.2 becomes false for k = ∞, explaining the need for C k -regularity with finite k. Weak direct products of Lie groups.If k ∈ N0 and (Hn )n∈N is a sequence ∗ of strongly C k -regular Lie groups, then n∈N Hn is strongly C k -regular (and hence regular) since its evolution map can be obtained as the composition ∼

⊕n∈N evolHn ∗ = k C k [0, 1], n L(Hn ) −→ → n Hn n C ([0, 1], L(Hn )) −−−−−−−− (cf. Proposition 4.7 for the definition and smoothness of ⊕n evolHn ). Test function groups. Given a σ-compact, finite-dimensional smooth manifold M and a C k -regular Lie group H, pick a locally finite family (Mn )n∈N of compact submanifolds with boundary of M , the interiors of which cover M . Then standard arguments (based on suitable exponential laws for function r k spaces) show that H n := C (Mn , H) is C -regular, for each n ∈ N. The map r r σ : Cc (M, L(H)) → n C (Mn , L(H)), γ → (γ|Mn )n∈N is continuous linear k and hence also the map τ := C ([0, 1], σ) from C k ([0, 1], Ccr (M, L(H))) to

k r C ([0, 1], n C (Mn , L(H))) ∼ = n C k ([0, 1], C r (Mn , L(H))). Furthermore, ρ : G := Ccr (M, H) → ∗n∈N C r (Mn , H) , γ → (γ|Mn )n∈N is an isomorphism of Lie groups onto a closed Lie subgroup (and embedded submanifold) of the weak direct product P . Using point evaluations, one finds that the composition evolP ◦ τ = ⊕n evolHn ◦ τ : Ccr (M, H) → ∗n∈N C r (Mn , H) (which is smooth by the preceding example) takes its image in the image of ρ. Then f := ρ−1 ◦ evolP ◦ τ : C k ([0, 1], Ccr (M, L(H))) → Ccr (M, H) is a smooth


271

map, and one verifies using point evaluations that f = evolG . A similar (but more complicated) argument shows that Diff c (M ) is regular. Ascending unions of Banach–Lie groups. As a preliminary, observe that regularity in Milnor’s sense (and strong C k -regularity) can be defined just as well for local Lie groups G; in this case, one requires that a smooth evolution evolG exists on some open 0-neighbourhood in C ∞ ([0, 1], L(G)) (resp., C k ([0, 1], L(G))). In the case of global Lie groups, the local notions of regularity are equivalent to the corresponding global ones (see [DG10] and [GN10]; cf. [KM97, lemma on p. 409]). See [Sm78] for the next theorem (and [Mj83] or [Sm83, Theorem I.7.2] for a variant beyond compact regularity). Theorem 8.4. Consider a Hausdorff locally convex space E which is the locally convex direct limit of Hausdorff locally convex spaces E1 ⊆ E2 ⊆ · · · . If E = n∈N En is compactly regular, then the natural continuous linear map lim C([0, 1], En ) → C([0, 1], E) −→

is an isomorphism of topological vector spaces. Remark 8.5. If E is a locally convex space which is integral complete,15 then C k ([0, 1], E) ∼ = E k × C([0, 1], E)

(6)

naturally via γ → (γ(0), . . . , γ (k−1) (0), γ (k) ), for each k ∈ N. If E = lim En is a −→ locally convex direct limit of integral complete locally convex spaces and E = n∈N En is compactly regular, then E is also integral complete. Moreover, (6), 4.1 (e) and Theorem 8.4 imply that lim C k ([0, 1], En ) ∼ = C k ([0, 1], lim En ) . −→

−→

(7)

Alternatively, (7) follows from Theorem 8.4 because C k ([0, 1], E) ∼ = C([0, 1], E) naturally if E is integral complete [DG10], as follows from (6) and the fact that C([0, 1], E) ∼ = E m × C([0, 1], E) naturally for each m ∈ N, by a suitable (elementary) variant of Miljutin’s theorem [Mi66] provided in [DG10]. 8.6 Assume that g = n∈N L(Gn ) is compactly regular in Theorem 5.3. Following [DG10], we now explain in the essential case where K = C that G is regular in Milnor’s sense.16 Let r > 0 be as in the earlier parts of the L(G ) proof and Un := Br n (0). Because the BCH-series has the same shape for each n ∈ N, one finds s > 0 such that an evolution evolUn exists as a map from L(G ) C([0, 1], Bs n (0)) to Un , for each n ∈ N. Since Un is bounded, Theorem 4.6 L(G ) shows that evolU := lim evolUn : n∈N C([0, 1], Bs n (0)) → U := n∈N Un −→ is C-analytic and hence also expG ◦ evolU , which is a local group version of the evolution map for G. Hence G is regular and in fact strongly C 0 -regular. 15 16

That is, every continuous curve γ : [0, 1] → E has a Riemann integral in E [LS00]. The real case follows easily via complexification on the level of local groups.

272

Helge Gl¨ ockner

Using Theorem 5.3, one readily deduces H) and H ↓s (K, F ) are that ×Germ(K, 0 × strongly C -regular, and also A = n∈N An if A = n∈N An is compactly regular (see [DG10]). The proof of regularity for GermDiff(K, X) is more involved, but eventually also reduces to Theorem 4.6 (see [Da10]). An idea which might lead to non-regular Lie groups. An observation from [DG10] might be a source of Lie groups which are not regular in Milnor’s sense, although they are modelled on Mackey complete locally convex spaces. Proposition 8.7. Suppose that, for each n ∈ N, there exists a Lie group Hn modelled on a Mackey complete locally convex space which is regular but not C n -regular because evolHn : C ∞ ([0, 1], L(H ∗ n )) → Hn is discontinuous with respect to the C n -topology. Then G := n∈N Hn is a Lie group modelled on the Mackey complete locally convex space L(G) = n∈N L(Hn ). It has an evolution evolG : C ∞ ([0, 1], L(G)) → G, but evolG fails to be continuous and thus G is not a regular Lie group in Milnor’s sense.

9 Homotopy groups of ascending unions of Lie groups We have seen that all main examples of ascending unions G = Lie groups admit a direct limit chart, and thus πk (G) = lim πk (Gn ) −→

for all k ∈ N0

n∈N

Gn of (8)

(see 1.3). Alternatively, many (but not all) of them are compactly regular. In this case, (8) holds by an elementary argument, but one has to pay the price that the proof of compact regularity may require specialized functionalanalytic tools (like Wengenroth’s theorem recalled above). It is an interesting feature that the approach via (weak)direct limit charts even extends to Lie groups G in which an ascending union n∈N Gn is merely dense (and to similar, more general situations). Weak direct limit charts (as defined in 1.2) then have to be replaced by certain “well-filled charts.” The precise setting will be described now. Besides smooth manifolds, it applies to topological manifolds and more general topological spaces (like manifolds with boundary or corners). Given a subset A of a real vector space V , let us write conv2 (A) := {tx + (1 − t)y : x, y ∈ A, t ∈ [0, 1]}. Definition 9.1. Let M be a topological space and (Mα )α∈A be a directed family of topological spaces such that M∞ := α∈A Mα is dense in M and all inclusion maps Mα → M and Mα → Mβ (for α ≤ β) are continuous. We say that a homeomorphism φ : U → V ⊆ E from an open subset U ⊆ M onto an arbitrary subset V of a topological vector space E is a well-filled chart of M if there exist α0 ∈ A and homeomorphisms φα : Uα → Vα ⊆ Eα from open subsets Uα ⊆ Mα onto subsets Vα of certain topological vector spaces Eα for α ≥ α0 such that the following conditions are satisfied.


273

(a) Eα ⊆ E, Eα ⊆ Eβ if α ≤ β and the inclusion maps Eα → E and Eα → Eβ are continuous and linear. (b) For all α ≥ α0 , we have Uα ⊆ U and φ|Uα = φα . (c) For all β ≥ α ≥ α0 , we have Uα ⊆ Uβ and φβ |Uα = φα . (d) U∞ := α≥α0 Uα = U ∩ M∞ . (e) There exists a non-empty (relatively) open set V (2) ⊆ V such that (2) conv2 (V (2) ) ⊆ V and conv2 (V∞ ) ⊆ V∞ , where V∞ := α≥α0 Vα and (2)

V∞ := V (2) ∩ V∞ . (2) (f) For each α ≥ α0 and compact set K ⊆ Vα := V (2) ∩ Vα , there exists β ≥ α such that conv2 (K) ⊆ Vβ . Then U (2) := φ−1 (V (2) ) is an open subset of U , called a core of φ. If cores of well-filled charts cover M , then M is said to admit well-filled charts. On a first reading, the reader may find the notion of a well-filled chart somewhat elusive. Special cases of particular interest (which are more concrete and easier to understand) are described in [Gl10, Examples 1.11 and 1.12]. See [Gl10, Theorem 1.13] for the following result. Theorem 9.2.Let M be a Hausdorff topological space containing a directed union M∞ := α∈A Mα of Hausdorff topological spaces Mα as a dense subset, such that all inclusion maps Mα → Mβ (for α ≤ β) and Mα → M are continuous. If M admits well-filled charts, then πk (M, p) = lim πk (Mα , p) −→

for all k ∈ N0 and p ∈ M∞ .

For a typical application, let H be a Lie group, m ∈ N, S(Rm , L(H)) be the Schwartz space of rapidly decreasing L(H)-valued smooth functions on Rm and S(Rm , H) be the correspondingLie group, as in [BCR81] (for spe∞ m cial H) and [Wa10]. Then Cc∞ (Rm , H) = n∈N C[−n,n] , H) is dense in m (R m m S(R , H), and S(R , H) admits well-filled charts [Gl10, Example 8.4]. Using Theorem 9.2 and approximation results from [Ne04a], it is then easy to see that ∞ m πk (S(Rm , H)) ∼ , H)) ∼ = lim πk (C[−n,n] = πk (Cc∞ (Rm , H)) m (R −→

∼ = πk (C(Sm , H)∗ ) ∼ = πk+m (H) = πk (C0 (Rm , H)) ∼ (see [Gl10, Remark 8.6]). This had been conjectured in [BCR81] and was open since 1981.

10 Subgroups of ascending unions and related topics We now discuss various results concerning subgroups of ascending unions of Lie groups (notably for direct limits of finite-dimensional Lie groups).

274

Helge Gl¨ ockner

Non-existence of small subgroups. It is an open problem whether infinitedimensional Lie groups may contain small torsion subgroups [Ne06, p. 293]. For direct limits of finite-dimensional Lie groups, the pathology could be ruled out by proving that they do not contain small subgroups [Gl07a, Theorem A]. Theorem 10.1. If G1 ⊆ G2 ⊆ · · · is a direct sequence of finite-dimensional Lie groups, then the Lie group G = lim Gn does not have small subgroups. −→

Idea of proof. Given a compact identity neighbourhood C1 ⊆ G1 which does not contain non-trivial subgroups of G1 , there exists a compact identity neighbourhood C2 ⊆ G2 with C1 in its interior relative G2 , which does not contain non-trivial subgroups of G2 (see [Gl07a, Lemma 2.1]). Proceeding in this way, we find a sequence (Cn )n∈N of compact identity neighbourhoods Cn ⊆ Gn 0 not containing non-trivial subgroups, such that Cn ⊆ Cn+1 for each n. Then C := n∈N Cn is an identity neighbourhood in G and we may hope that C does not contain non-trivial subgroups this is not true of G. Unfortunately, in general, as the example R(N) = n∈N Rn = n∈N Cn with Cn := [−n, n]n shows. However, if the sets Cn are chosen carefully (which requires much work), then indeed C will not contain non-trivial subgroups [Gl07a]. We mention that an analogous result is available for certain ascending unions of infinite-dimensional Lie groups G1 ⊆ G2 ⊆ · · · (see [Gl07a, Theorem B]). To enable compactness arguments, each Gn has to be locally kω or each Gn a Banach–Lie group and the tangent map L(λn ) : L(Gn ) → L(Gn+1 ) of the inclusion map λn : Gn → Gn+1 a compact operator.17 Initial Lie subgroups. If G is a Lie group and H ⊆ G a subgroup, then H is called an initial Lie subgroup 18 if it admits a Lie group structure making the inclusion map ι : H → G a smooth map, such that L(ι) is injective and mappings from C k -manifolds M to H are C k if and only if they are C k as mappings to G, for each k ∈ N ∪ {∞}. Answering an open problem from [Ne06] in the negative, it was shown in [Gl08] that subgroups of infinite-dimensional Lie groups need not be initial Lie subgroups. In fact, one can take G = RN (with the product topology) and H = ∞ (see [Gl08, Theorem 1.3]). For direct limits of finite-dimensional Lie groups, G = n∈N Gn , it was already shown in [Gl05b] that every subgroup H ⊆ G admits a natural Lie group structure. By [Gl08, Theorem 2.1], this Lie group structure makes H an initial Lie subgroup of G, and thus the preceding pathology does not occur for such direct limit Lie groups G. Continuous one-parameter groups and the topology on L(G). If G = n∈N Gn is a direct limit of finite-dimensional Lie groups, then every 17 18

Further technical hypotheses need to be imposed, which we suppress here. Some readers may prefer to omit the second condition, or allow M to be a manifold with C k -boundary, with corners or (more generally) a C k -manifold with rough boundary (as introduced in [GN10]). The following results carry over to these varied situations (see [Gl08]).


275

continuous homomorphism (R, +) → G (i.e., each continuous one-parameter subgroup) is a continuous homomorphism to some Gn (by compact regularity) and hence smooth. It easily follows from this that the natural map θ : L(G) → Homcts (R, G) ,

x → (t → expG (tx))

is a bijection onto the set Homcts (R, G) of continuous one-parameter subgroups of G. It was asked in [Ne06, Problem VII.2] whether G is a topological group with Lie algebra in the sense of [HM07, Definition 2.11]. This holds if θ is a homeomorphism onto Homcts (R, G), equipped with the compact-open topology (which is not obvious because expG need not be a local homeomorphism at 0). As shown in [Gl08, Theorem 3.4], the latter property is always satisfied. Thus L(G) is determined by the topological group structure of G. For example, this implies that every continuous homomorphism from a locally exponential Lie group to G is smooth [Gl08, Proposition 3.7] (where a Lie group is called locally exponential if it has an exponential function and the latter is a local diffeomorphism at 0). It is an open problem whether continuous homomorphisms between arbitrary Lie groups are automatically smooth. Acknowledgements: The author is grateful to K.-D. Bierstedt (†) and S. A. Wegner (Paderborn) for discussions related to regularity properties of (LF)spaces, and thanks J. Bonet (Valencia) for comments which entered into Remark 8.5.

References [ACM89] M. C. Abbati, R. Cirelli, A. Mania and P. W. Michor, The Lie group of automorphisms of a principal bundle, J. Geom. Phys. 6:2 (1989), 215–235. [AHK93] S. A. Albeverio, R. J. Høegh-Krohn, J. A. Marion, D. H. Testard and B. S. Torrésani, Noncommutative Distributions, Marcel Dekker, New York, 1993. [AK06] S. A. Albeverio and A. Kosyak, Quasiregular representations of the infinite-dimensional nilpotent group, J. Funct. Anal. 236:2 (2006), 634–681. [BBZ04] Y. A. Bahturin, A. A. Baranov and A. E. Zalesski, Simple Lie subalgebras of locally finite associative algebras, J. Algebra 281:1 (2004), 225–246. [BB04] Y. Bahturin and G. Benkart, Some constructions in the theory of locally finite simple Lie algebras, J. Lie Theory 14:1 (2004), 243–270. [Ba97] A. Banyaga, The Structure of Classical Diffeomorphism Groups, Kluwer, Dordrecht, 1997. [Bi88] K.-D. Bierstedt, An introduction to locally convex inductive limits, pp. 35–133 in: H. Hogbe-Nlend (ed.), Functional Analysis and its Applications, World Scientific, Singapore, 1988. [BMS82] K.-D. Bierstedt, R. Meise and W. H. Summers, A projective description of weighted inductive limits, Trans. Amer. Math. Soc. 272:1 (1982), 107–160.

276 [BS71]

Helge Gl¨ ockner

J. Bochnak and J. Siciak, Analytic functions in topological vector spaces. Studia Math. 39 (1971), 77–112. [BCR81] H. Boseck, G. Czichowski and K.-P. Rudolph, Analysis on Topological Groups – General Lie Theory, Teubner, Leipzig, 1981. [Bo87] N. Bourbaki, Topological Vector Spaces, Chapters 1–5, Springer, Berlin, 1987. [Ch85] S. B. Chae, Holomorphy and Calculus in Normed Spaces, Marcel Dekker, New York, 1985. [Da10] R. Dahmen, Complex analytic mappings on (LB)-spaces and applications in infinite-dimensional Lie theory, Math. Z. 266:1 (2010), 115–140. [DG10] R. Dahmen and H. Gl¨ ockner, Regularity in Milnor’s sense for direct limits of infinite-dimensional Lie groups, in preparation. [DW97] S. Dierolf and J. Wengenroth, Inductive limits of topological algebras, Linear Topol. Spaces Complex Anal. 3 (1997), 45–49. [DP04] I. Dimitrov and I. Penkov, Borel subalgebras of gl(∞), Resenhas 6:2–3 (2004), 153–163. [DPW02] I. Dimitrov, I. Penkov and J. A. Wolf, A Bott-Borel-Weil theory for direct limits of algebraic groups, Amer. J. Math. 124:5 (2002), 955–998. [Du64] R. M. Dudley, On sequential convergence, Trans. Amer. Math. Soc. 112 (1964), 483–507. [Dv02] A. Dvorsky, Direct limits of low-rank representations of classical groups, Comm. Algebra 30:12 (2002), 6011–6022. [Ed99] T. Edamatsu, On the bamboo-shoot topology of certain inductive limits of topological groups, J. Math. Kyoto Univ. 39:4 (1999), 715–724. [Fl71] K. Floret, Lokalkonvexe Sequenzen mit kompakten Abbildungen, J. Reine Angew. Math. 247 (1971), 155–195. [Fl79] K. Floret, On bounded sets in inductive limits of normed spaces, Proc. Amer. Math. Soc. 75:2 (1979), 221–225. [Fl80] K. Floret, Some aspects of the theory of locally convex inductive limits, pp. 205–237 in: K.-D. Bierstedt and B. Fuchssteiner (eds.), Functional Analysis: Surveys and Recent Results II, North-Holland Math. Stud. 38, North-Holland, Amsterdam, 1980. [Gl02a] H. Gl¨ ockner, Infinite-dimensional Lie groups without completeness restrictions, pp. 43–59 in: A. Strasburger et al. (eds.), Geometry and Analysis on Finite- and Infinite-Dimensional Lie Groups, Banach Center Publ. 55, Warsaw, 2002. [Gl02b] H. Gl¨ ockner, Lie group structures on quotient groups and universal complexifications for infinite-dimensional Lie groups, J. Funct. Anal. 194:2 (2002), 347–409. [Gl02c] H. Gl¨ ockner, Algebras whose groups of units are Lie groups, Studia Math. 153:2 (2002), 147–177. [Gl02d] H. Gl¨ ockner, Patched locally convex spaces, almost local mappings and diffeomorphism groups of non-compact manifolds, manuscript, 2002. [Gl03a] H. Gl¨ ockner, Direct limit Lie groups and manifolds, J. Math. Kyoto Univ. 43:1 (2003), 1–26. [Gl03b] H. Gl¨ ockner, Lie groups of measurable mappings, Canad. J. Math. 55:5 (2003), 969–999. [Gl04a] H. Gl¨ ockner, Lie groups of germs of analytic mappings, pp. 1–16 in: T. Wurzbacher (ed.), Infinite-Dimensional Groups and Manifolds, IRMA Lecture Notes in Math. and Theor. Physics, de Gruyter, 2004.

Direct Limits of Infinite-Dimensional Lie Groups [Gl04b]

277

H. Gl¨ ockner, Lie groups over non-discrete topological fields, preprint, arXiv:math/0408008. [Gl05a] H. Gl¨ ockner, Diff(Rn ) as a Milnor-Lie group, Math. Nachr. 278:9 (2005), 1025–1032. [Gl05b] H. Gl¨ ockner, Fundamentals of direct limit Lie theory, Compos. Math. 141:6 (2005), 1551–1577. [Gl07a] H. Gl¨ ockner, Direct limit groups do not have small subgroups, Topol. Appl. 154:6 (2007), 1126–1133. [Gl07b] H. Gl¨ ockner, Direct limits of infinite-dimensional Lie groups compared to direct limits in related categories, J. Funct. Anal. 245:1 (2007), 19–61. [Gl08] H. Gl¨ ockner, Solutions to questions in Neeb’s recent survey on infinitedimensional Lie groups, Geom. Dedicata 135 (2008), 71–86. [Gl10] H. Gl¨ ockner, Homotopy groups of ascending unions of infinitedimensional manifolds, to appear in Ann. Inst. Fourier (cf. arXiv:0812.4713v1) [GGH10] H. Gl¨ ockner, R. Gramlich and T. Hartnick, Final group topologies, KacMoody groups and Pontryagin duality, Israel J. Math. 177 (2010), 49–102. [GN10] H. Gl¨ ockner and K.-H. Neeb, Infinite-Dimensional Lie Groups, Vol 1, book in preparation. [Go03] G. A. Goldin, Lectures on diffeomorphism groups in quantum physics, pp. 3–93 in: Contemporary Problems in Mathematical Physics, Proceedings of the 3rd Intern. Workshop (Cotonou, 2003), World Scientific, 2004 [Ha71] V. L. Hansen, Some theorems on direct limits of expanding systems of manifolds, Math. Scand. 29 (1971), 5–36. [Hi93] T. Hirai, Irreducible unitary representations of the group of diffeomorphisms of a noncompact manifold, J. Math. Kyoto Univ. 33:3 (1993), 827–864. [HST01] T. Hirai, H. Shimomura, N. Tatsuuma and E. Hirai, Inductive limits of topologies, their direct products, and problems related to algebraic structures, J. Math. Kyoto Univ. 41:3 (2001), 475–505. [HM07] K. H. Hofmann and S. A. Morris, The Lie Theory of Connected Pro-Lie Groups, EMS Tracts in Mathematics 2, European Mathematical Society, Zurich, 2007. [Is76] R. S. Ismagilov, Unitary representations of the group C0∞ (X, G), G = SU2 , Mat. Sb. (N.S.) 100(142):1 (1976), 117–131, 166. [KR01] N. Kamran and T. Robart, A manifold structure for analytic isotropy Lie pseudogroups of infinite type, J. Lie Theory 11:1 (2001), 57–80. [Ke74] H. H. Keller, Differential Calculus in Locally Convex Spaces, Springer, Berlin, 1974. [KOV04] S. Kerov, G. I. Ol’shanski˘ı and A. Vershik, Harmonic analysis on the infinite symmetric group, Invent. Math. 158:3 (2004), 551–642. [Ki81] A. A. Kirillov, Unitary representations of the group of diffeomorphisms and of some of its subgroups, Sel. Math. Sov. 1 (1981), 351–372. [Ko69] J. K¨ ohn, Induktive Limiten nicht lokal-konvexer topologischer Vektorr¨ aume, Math. Ann. 181 (1969), 269–278. [KS77] V. I. Kolomycev and Y. S. Samo˘ılenko, Irreducible representations of inductive limits of groups, Ukr. Mat. J. 29 (1977), 526–531, 565 [KM97] A. Kriegl and P. W. Michor, The Convenient Setting of Global Analysis, Amer. Math. Soc., Providence, RI, 1997.

278 [Ku06]

Helge Gl¨ ockner

K. K¨ uhn, Direct limits of diagonal chains of type O, U, and Sp, and their homotopy groups, Comm. Algebra 34:1 (2006), 75–87. [Le94] F. Leitenberger, Unitary representations and coadjoint orbits for a group of germs of real analytic diffeomorphisms, Math. Nachr. 169 (1994), 185–205. [LS00] A. K. Leonov and S. A. Shkarin, Integral completeness of locally convex spaces, Russ. J. Math. Phys. 7:4 (2000), 402–412. [Ls82] J. Leslie, On the group of real analytic diffeomorphisms of a compact real analytic manifold, Trans. Amer. Math. Soc. 274:2 (1982), 651–669. [Ls85] J. Leslie, Some finite-codimensional Lie subgroups of Diff ω (M ), pp. 359– 372 in: G. M. Rassias and T. M. Rassias (eds.), Differential Geometry, Calculus of Variations, and their Applications, Lecture Notes in Pure and Appl. Math. 100, Marcel Dekker, New York, 1985. [Mk63] B. M. Makarov, Some pathological properties of inductive limits of B-spaces, Uspehi Mat. Nauk 18:3 (1963), 171–178. [Mc80] P. W. Michor, Manifolds of Differentiable Mappings, Shiva Publishing, Nantwich, 1980. [Mr82] J. Milnor, On infinite dimensional Lie groups, Preprint, Institute of Advanced Study, Princeton, 1982. [Mr84] J. Milnor, Remarks on infinite dimensional Lie groups, pp. 1007–1057 in: B. DeWitt and R. Stora (eds.), Relativity, Groups and Topology II, North-Holland, Amsterdam, 1984. [Mi66] A. A. Miljutin, Isomorphism of the spaces of continuous functions over compact sets of the cardinality of the continuum, Teor. Funkci˘ı Funkcional. Anal. i Priloˇzen. Vyp. 2 (1966), 150–156. [Mj79] J. Mujica, Spaces of germs of holomorphic functions, pp. 1–41 in: G.-C. Rota (ed.), Studies in Analysis, Adv. in Math. Suppl. Stud. 4, Academic Press, New York, 1979. [Mj83] J. Mujica, Spaces of continuous functions with values in an inductive limit, pp. 359–367 in: G. I. Zapata (ed.), Functional Analysis, Holomorphy, and Approximation Theory, Lect. Notes Pure Appl. Math. 83, Marcel Dekker, New York, 1983. [NRW91] L. Natarajan, E. Rodr´ıguez-Carrington and J. A. Wolf, Differentiable structure for direct limit groups, Letters in Math. Phys. 23:2, (1991) 99–109. [NRW93] L. Natarajan, E. Rodr´ıguez-Carrington and J. A. Wolf, Locally convex Lie groups, Nova J. Alg. Geom. 2:1 (1993), 59–87. [NRW01] L. Natarajan, E. Rodr´ıguez-Carrington and J. A. Wolf, The Bott-BorelWeil theorem for direct limit groups, Trans. Amer. Math. Soc. 353:11 (2001), 4583–4622. [Ne98] K.-H. Neeb, Holomorphic highest weight representations of infinitedimensional complex classical groups, J. Reine Angew. Math. 497 (1998), 171–222. [Ne00] K.-H. Neeb, Integrable roots in split graded Lie algebras, J. Algebra 225:2 (2000), 534–580. [Ne02a] K.-H. Neeb, Classical Hilbert-Lie groups, their extensions and their homotopy groups, pp. 87–151 in: A. Strasburger et al. (eds.) Geometry and Analysis on Finite- and Infinite-Dimensional Lie Groups, Banach Center Publ. 55, Warsaw, 2002.

Direct Limits of Infinite-Dimensional Lie Groups [Ne02b] [Ne04a]

[Ne04b] [Ne04c] [Ne06] [Ne07] [NS01] [NW08] [Ol83] [Ol03] [Om97] [Pa65] [Pa66] [PS03] [Pk08] [Pi77] [Ra08] [Sm78] [Sm83] [Se55] [Sh01]

279

K.-H. Neeb, Central extensions of infinite-dimensional Lie groups, Ann. Inst. Fourier (Grenoble) 52:5 (2002), 1365–1442. K.-H. Neeb, Current groups for non-compact manifolds and their central extensions, pp. 109–183 in: T. Wurzbacher (ed.) Infinite-Dimensional Groups and Manifolds, IRMA Lecture Notes in Math. and Theor. Physics, de Gruyter, 2004. K.-H. Neeb, Abelian extensions of infinite-dimensional Lie groups, Travaux Math. 15 (2004), 69–194. K.-H. Neeb, Infinite-dimensional groups and their representations, pp. 213–328 in: J. C. Jantzen and K.-H. Neeb (eds.), Lie Theory, Birkh¨ auser, Boston, MA, 2004. K.-H. Neeb, Towards a Lie theory of locally convex groups, Jpn. J. Math. 1:2 (2006), 291–468. K.-H. Neeb, Non-abelian extensions of infinite-dimensional Lie groups, Ann. Inst. Fourier (Grenoble) 57:1 (2007), 209–271. K.-H. Neeb and N. Stumme, The classification of locally finite split simple Lie algebras, J. Reine Angew. Math. 533 (2001), 25–53. K.-H. Neeb and F. Wagemann, Lie group structures on groups of smooth and holomorphic maps on non-compact manifolds, Geom. Dedicata 134 (2008), 17–60. G. I. Ol’shanski˘ı, Unitary representations of infinite-dimensional pairs (G, K) and the formalism of R. Howe, Dokl. Akad. Nauk SSSR 269:1 (1983), 33–36. G. I. Ol’shanski˘ı, The problem of harmonic analysis on the infinitedimensional unitary group, J. Funct. Anal. 205:2 (2003), 464–524. H. Omori, Infinite-Dimensional Lie Groups, Transl. Math. Monographs 158, Amer. Math. Soc., Providence, RI, 1997. R. S. Palais, On the homotopy type of certain groups of operators, Topology 3 (1965), 271–279. R. S. Palais, Homotopy theory of infinite-dimensional manifolds, Topology 5 (1966), 1–16. I. Penkov and H. Strade, Locally finite Lie algebras with root decomposition, Arch. Math. (Basel) 80:5 (2003), 478–485. D. Pickrell Heat kernel measures and critical limits, Karl-Hermann Neeb and Arturo Pianzola (eds.), Developments and Trends in InfiniteDimensional Lie Theory, Birkh¨ auser, Boston, MA, 2010. D. Pisanelli, An example of an infinite Lie group, Proc. Amer. Math. Soc. 62:1 (1977), 156–160. M. Rabaoui, A Bochner type theorem for inductive limits of Gelfand pairs, Ann. Inst. Fourier 58:5 (2008), 1551–1573. J. Schmets, Spaces of vector-valued continuous functions, pp. 368–377 in: R. M. Aron and S. Dineen (eds.), Vector Space Measures and Applications II, Lecture Notes in Math. 644, Springer, Berlin, 1978. J. Schmets, Spaces of Vector-Valued Continuous Functions, Lecture Notes in Math. 1003, Springer, Berlin, 1983. J. Sebasti˜ ao e Silva, Su certe classi di spazi localmente convessi importanti per le applicazioni, Rend. Mat. Appl. 14 (1955), 388–410. H. Shimomura, Quasi-invariant measures on the group of diffeomorphisms and smooth vectors of unitary representations, J. Funct. Anal. 187:2 (2001), 406–441.

280 [Sh05]

Helge Gl¨ ockner

H. Shimomura, Irreducible decompositions of unitary representations of infinite-dimensional groups, Math. Z. 251:3 (2005), 575–587. [Sr59] T. Shirai, Sur les topologies des espaces de L. Schwartz, Proc. Japan Acad. 35 (1959), 31–36. [Sv69] O. G. Smolyanov, Almost closed linear subspaces of strict inductive limits of sequences of Fréchet spaces, Mat. Sb. (N.S.) 80 (122) (1969), 513–520. [St99] N. Stumme, The structure of locally finite split Lie algebras, J. Algebra 220:2 (1999), 664–693. [TSH98] 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. 38:3 (1998), 551–578. ¨ [Th64a] E. Thoma, Uber unit¨ are Darstellungen abz¨ ahlbarer, diskreter Gruppen, Math. Ann. 153 (1964), 111–138. [Th64b] E. Thoma, Die unzerlegbaren, positiv-definiten Klassenfunktionen der abz¨ ahlbar unendlichen, symmetrischen Gruppe, Math. Z. 85 (1964), 40–61. [VGG75] A. M. Verˇsik, I. M. Gel’fand and M. I. Graev, Representations of the group of diffeomorphisms, Uspehi Mat. Nauk 30:6 (1975), 1–50. [Vo76] D. Voiculescu, Représentations factorielles de type II1 de U(∞), J. Math. Pures Appl. 55:1 (1976), 1–20. [Wa10] B. Walter, Weighted diffeomorphism groups of Banach spaces and weighted mapping groups, preprint, arXiv:1006.5580v1. [We03] J. Wengenroth, Derived Functors in Functional Analysis, Lecture Notes in Math. 1810, Springer, Berlin, 2003. [Wk07] C. Wockel, Lie group structures on symmetry groups of principal bundles, J. Funct. Anal. 251:1 (2007), 254–288. [Wo05] J. A. Wolf, Principal series representations of direct limit groups, Compos. Math. 141:6 (2005), 1504–1530. [Wo08] J. A. Wolf, Infinite dimensional multiplicity free spaces I: limits of compact commutative spaces, Karl-Hermann Neeb and Arturo Pianzola (eds.), Developments and Trends in Infinite-Dimensional Lie Theory, Birkhäuser, Boston, MA, 2010. [Ya98] A. Yamasaki, Inductive limits of general linear groups, J. Math. Kyoto Univ. 38:4 (1998), 769–779.

Lie Groups of Bundle Automorphisms and Their Extensions Karl-Hermann Neeb Department of Mathematics, Friedrich-Alexander-Universit¨ at Erlangen-N¨ urnberg, Bismarckstrasse 1 1/2, 91054 Erlangen, Germany, [email protected] Summary. We describe natural abelian extensions of the Lie algebra aut(P ) of infinitesimal automorphisms of a principal bundle over a compact manifold M and discuss their integrability to corresponding Lie group extensions. The case of a trivial bundle P = M × K is already quite interesting. In this case, we show that essentially all central extensions of the gauge algebra C ∞ (M, k) can be obtained from three fundamental types of cocycles with values in one of the spaces z := C ∞ (M, V ), Ω 1 (M, V ) and Ω 1 (M, V )/dC ∞ (M, V ). These cocycles extend to aut(P ), and, under the assumption that T M is trivial, we also describe the space H 2 (V(M ), z) classifying the twists of these extensions. We then show that all fundamental types have natural generalizations to non-trivial bundles and explain under which conditions they extend to aut(P ) and integrate to global Lie group extensions. Key words: gauge group, automorphism group, infinite-dimensional Lie group, central extension, abelian extension, affine connection. 2000 Mathematics Subject Classifications: Primary 22E65. Secondary 22E67, 17B66.

Introduction Two of the most important classes of infinite-dimensional Lie groups are groups of smooth maps, such as the Lie group C ∞ (M, K) of smooth maps of a compact smooth manifold M with values in a Lie group K, and groups of diffeomorphisms, such as the group Diff(M ) of diffeomorphisms of a compact smooth manifold M . A case of particular importance arises for the circle M = S1 , where LK := C ∞ (S1 , K) is called the loop group of K. If K is a compact sim by the circle ple Lie group, then LK has a universal central extension LK 1 group T. Furthermore, the group Tr ⊆ Diff(S ) of rigid rotations of S1 acts smoothly by automorphisms on LK, this action lifts to the central extension


281

282

Karl-Hermann Neeb

and we thus obtain the affine Kac–Moody Lie groups LK Tr .1 The LK twisted affine Kac–Moody Lie groups can be realized in this picture as the fixed point groups for an automorphism σ acting trivially on Tr , and inducing on LK an automorphism of the form σ(f )(t) = ϕ(f (ζt)), where ϕ ∈ Aut(K) is an automorphism of finite order m and ζ ∈ Tr satisfies ζ m = 1. On each affine Kac–Moody group, the group Diff + (S1 ) of orientation-preserving diffeomor Tr embeds into LK phisms of S1 acts by automorphism, so that LK 1 Diff + (S ). Actually, the latter group permits an interesting twist, where the subgroup T×Diff + (S1 ) is replaced by the Virasoro group, a non-trivial central extension of Diff + (S1 ) by T. The purpose of this article is to discuss several infinite-dimensional Lie groups that are constructed in a similar fashion from higher dimensional compact manifolds M . Since the one-dimensional manifold M = S1 is a rather simple object, the general theory leaves much more room for different Lie group constructions, extensions and twistings thereof. Here we shall focus on recent progress in several branches of this area, in particular relating the Lie algebra picture to global objects. We shall also explain how this relates to other structures, such as multiloop algebras, which are currently under active investigation from the algebraic point of view ([ABP06], [ABFP08]). We also present open problems and describe several areas where the present knowledge is far from satisfactory. The analogue of the loop algebra C ∞ (S1 , k) of a Lie algebra k is the Lie algebra of smooth maps g = C ∞ (M, k) on the compact manifold M . Here the compactness assumption is convenient if we want to deal with Lie groups, because it implies that g is the Lie algebra of any group G = C ∞ (M, K), where K is a Lie group with Lie algebra k. This group can also be identified with the gauge group of the trivial K-principal bundle P = M × K, so that gauge groups of principal bundles are natural generalizations of mapping groups. Twisted loop groups are gauge groups of certain non-trivial principal bundles over S1 (see Section 4). For a K-principal bundle q : P → M over the compact manifold M , we write Aut(P ) := Diff(P )K for the group of all diffeomorphisms of P commuting with the K-action, i.e., the group of bundle automorphisms. The gauge group Gau(P ) is the normal subgroup of all bundle isomorphisms inducing the identity on M . Writing Diff(M )P for the set of all diffeomorphisms of M that can be lifted to bundle automorphisms, which is an open subgroup of the Lie group Diff(M ), we obtain a short exact sequence of Lie groups 1 → Gau(P ) → Aut(P ) → Diff(M )P → 1

(1)

(cf. [ACM89], [Wo07]). On the Lie algebra level, we have a corresponding short exact sequence of Lie algebras of vector fields 0 → gau(P ) → aut(P ) := V(P )K → V(M ) → 1. 1

(2)

Strictly speaking, these are the “unitary forms” of the affine Kac–Moody groups. Starting with a complex simple Lie group K instead, we obtain complex versions.

Lie Groups of Bundle Automorphisms and Their Extensions

283

It is an important problem to understand the central extensions of gauge groups by an abelian Lie group Z on which Diff(M )P acts naturally and the extent to which they can be enlarged to abelian extensions of the full group Aut(P ), or at least its identity component. Whenever such an enlargement exists, one has to understand the set of all enlargements, the twistings, which leads to the problem of classifying all abelian extensions of Diff(M )P by Z. Below we shall discuss various partial results concerning these questions. We also describe several tools to address them and to explain what still remains to be done. The content of the paper is as follows. After discussing old and new results on Lie group structures on gauge groups, mapping groups and diffeomorphism groups in Section 1, we turn in Section 2 to central extensions of mapping groups G := C ∞ (M, K). Here we first exhibit three fundamental classes of cocycles over which all other non-trivial cocycles can be factored, up to cocycles vanishing on the commutator algebra. The fundamental cocycles are defined by an invariant symmetric bilinear form κ : k × k → V and an alternating map η : k × k → V as follows: 1

ωκ (ξ1 , ξ2 ) := [κ(ξ1 , dξ2 )] ∈ Ω (M, V ) := Ω 1 (M, V )/dC ∞ (M, V ), ωη (ξ1 , ξ2 ) := η(ξ1 , ξ2 ) ∈ C ∞ (M, V ) for dk η = 0, and if dk η(x, y, z) = κ([x, y], z), we also have ωκ,η (ξ1 , ξ2 ) := κ(ξ1 , dξ2 ) − κ(ξ2 , dξ1 ) − d(η(ξ1 , ξ2 )) ∈ Ω 1 (M, V ). We then discuss the action of G0 Diff(M ) on the corresponding central Lie algebra extension and explain under which conditions it integrates to an extension of Lie groups. Since we are also interested in the corresponding abelian extensions of the full automorphism group Aut(M ×K) ∼ = GDiff(M ) and its Lie algebra, we turn in Section 3 to abelian extensions of V(M ), resp., Diff(M ) by the three types of target spaces z from above. If (the tangent bundle) T M is trivial and V = R, then a full description of H 2 (V(M ), z) in all cases has been obtained recently in [BiNe08] (Theorem 3.8). As explained in some detail in Appendix B (Corollary 8.6), the space H 2 (V(M ), z) classifies the twists of the abelian extensions of the semidirect sum g V(M ) which are of the form g V(M ), where g is a central extension of g by z. See [FRST90] for a discussion of central extensions of Lie algebras of the form C ∞ (M, k) V(M, µ), where µ is a volume form on M . In Section 4 we then turn to gauge and automorphism groups of general principal bundles, where the first step consists in the construction of analogs of the three fundamental types of cocycles. A crucial difficulty arises from the fact that one has to consider principal bundles with non-connected structure groups to realize natural classes of Lie algebras such as twisted affine algebras as gauge algebras. We therefore have to consider target spaces V on which the quotient group π0 (K) = K/K0 acts non-trivially and consider K-invariant forms κ : k×k → V . A typical example is the target space V (k) of the universal

284

Karl-Hermann Neeb

invariant symmetric bilinear form of a semisimple Lie algebra k on which K = Aut(k) acts non-trivially. Replacing the exterior differential by a covariant derivative, we thus obtain cocycles 1

ωκ (ξ1 , ξ2 ) := [κ(ξ1 , dξ2 )] ∈ Ω (M, V) := Ω 1 (M, V)/dΓ V, where V → M is the flat vector bundle associated to P via the K-module structure on V . For the cocycles ωκ there are natural criteria for integrability whenever π0 (K) is finite ([NeWo09]), but it is not clear how to extend ωκ to aut(P ). However, the situation simplifies considerably for the cocycle d ◦ ωκ with values in Ω 2 (M, V), which extends naturally to a cocycle on aut(P ) that integrates to a smooth group cocycle on Aut(P ). The analogs of C ∞ (M, V )-valued cocycles are determined by a central extension k of k by V to which the adjoint action of K lifts. Here the central Lie algebra extension g au(P ) is a space of sections of an associated Lie algebra bundle with fiber k. We then show that for any central extension K ∼ of K by V there exists a K-bundle P with P /V = P , so that Aut(P ) is an abelian extension of Aut(P ) by C ∞ (M, V ) containing the central extension Gau(P ) of Gau(P ) with Lie algebra g au(P ). There are also analogs of the Ω 1 (M, V )-valued cocycles which exist whenever the 3-cocycle κ([x, y], z) is a coboundary. For non-trivial bundles the fact that the extension aut(P ) of V(M ) generally does not split makes the analysis of the abelian extensions of aut(P ) considerably more difficult than in the trivial case, where we have a semidirect product Lie algebra. As a consequence, the results on group actions on the Lie algebra extensions and on integrability are much less complete than for trivial bundles. To connect our geometric setting with the algebraic setup, we describe in Section 5 the connection between multiloop algebras and gauge algebras of flat bundles over tori which can be trivialized by finite coverings. The paper concludes with some remarks on bundles with infinite-dimensional structure groups and appendices on integrability of abelian Lie algebra extensions, extensions of semidirect products (such as C ∞ (M, k) V(M )) and the triviality of the action of a connected Lie group on the corresponding continuous Lie algebra cohomology. Throughout we only consider Lie algebras and groups of smooth maps endowed with the compact open C ∞ -topology. To simplify matters, we focus on compact manifolds to work in a convenient Lie-theoretic setup for the corresponding groups. For non-compact manifolds, the natural setting is provided by spaces of compactly supported smooth maps with the direct limit LF -topology (cf. [KM97], [ACM89], [Mi80], [Gl02]). However, there are some classes of non-compact manifolds on which we still have (a) Lie group structure on the full group of smooth maps (Theorem 1.4). For groups of sections of group bundles one can also develop a theory for C k - and Sobolev sections, but this has the disadvantage that smooth vector fields do not act as derivations ([Sch04]).


285

In [Pi93] D. Pickrell studies central extensions of loop groups C ∞ (S1 , K), K a compact connected Lie group, which are invariant under the group Diff(S1 )0 of orientation-preserving diffeomorphisms of the circle. Such extensions turn out to define modular functors, in the sense of three-dimensional topological quantum field theory, if they have a certain reciprocity property. Some generalizations of these results to non-connected compact groups K can be found in [Pi00]. Unitary representations of central extensions of loop groups and the existence of intertwining actions of Diff(S1 ) are discussed in [TL99]. Smooth mapping groups on surfaces and realizations of their central extensions via complex geometry have been studied in [EF94] and [FK96]. Connections between central extensions of mapping groups and algebraic geometry and rational conformal field theory are exhibited in [LMNS96]. The automorphism group Aut(P ) also shows up naturally as a symmetry group in the symplectic approach to classical field theory ([GIMM04]). Notation and basic concepts A Lie group G is a group equipped with a smooth manifold structure modeled on a locally convex space for which the group multiplication and the inversion are smooth maps. We write 1 ∈ G for the identity element and λg (x) = gx, resp., ρg (x) = xg for the left, resp., right multiplication on G. Then each x ∈ T1 (G) corresponds to a unique left invariant vector field xl with xl (g) := dλg (1).x, g ∈ G. The space of left invariant vector fields is closed under the Lie bracket of vector fields, and hence it inherits a Lie algebra structure. In this sense we obtain on T1 (G) a continuous Lie bracket which is uniquely determined by [x, y]l = [xl , yl ] for x, y ∈ T1 (G). We write L(G) = g for the so-obtained locally convex Lie algebra and note that for morphisms ϕ : G → H of Lie groups we obtain with L(ϕ) := T1 (ϕ) a functor from the category of Lie 0 → G0 groups to the category of locally convex Lie algebras. We write qG : G for the universal covering map of the identity component G0 of G and identify 0 with π1 (G) ∼ the discrete central subgroup ker qG of G = π1 (G0 ). In the following we always write I = [0, 1] for the unit interval in R. A Lie group G is called regular if for each ξ ∈ C ∞ (I, g), the initial value problem γ(0) = 1,

γ (t) = γ(t).ξ(t) = T (λγ(t) )ξ(t)

has a solution γξ ∈ C ∞ (I, G), and the evolution map evolG : C ∞ (I, g) → G,

ξ → γξ (1)

is smooth (cf. [Mil84]). For a locally convex space E, the regularity of the Lie group (E, +) is equivalent to the Mackey completeness of E, i.e., to the existence of integrals of smooth curves γ : I → E. We also recall that for each regular Lie group G, its Lie algebra g is Mackey complete and that all Banach–Lie groups are regular ([GN10]). For a smooth map f : M → G we define the left, resp., right logarithmic derivative in Ω 1 (M, g) by

286

Karl-Hermann Neeb

δ l (f )v := f (m)−1 ·Tm (f )v and δ r (f )v := Tm (f )v·f (m)−1 , where · refers to the two-sided action of G on its tangent bundle T G. A smooth map expG : L(G) → G is said to be an exponential function if for each x ∈ L(G), the curve γx (t) := expG (tx) is a homomorphism R → G with γx (0) = x. Currently, all known Lie groups modeled on complete locally convex spaces are regular, and hence possess an exponential function. For Banach–Lie groups, the existence of an exponential function follows from the theory of ordinary differential equations in Banach spaces. A Lie group G is called locally exponential if it has an exponential function mapping an open 0-neighborhood in L(G) diffeomorphically onto an open neighborhood of 1 in G. For more details, we refer to Milnor’s lecture notes [Mil84], the survey [Ne06a] or the monograph [GN10]. If q : E → B is a smooth fiber bundle, then we write Γ E := {s ∈ C ∞ (B, E) : q ◦ s = idB } for its space of smooth sections. If g is a topological Lie algebra and V a topological g-module, we write (C • (g, V ), dg ) for the corresponding Lie algebra complex ([ChE48]).

1 Lie group structures on mapping groups and automorphism groups of bundles Before we turn to extensions, in this short section we collect some results on Lie group structures on gauge groups and automorphism groups of bundles and introduce the notation and conventions used below. Some of the material is classical, but there are also some new interesting results. 1.1 Automorphism groups of bundles Theorem 1.1. Let M be a compact manifold and K a Lie group with Lie algebra k. Then the following assertions hold. (1) Diff(M ) carries the structure of a Fréchet–Lie group whose Lie algebra is the space V(M ) of smooth vector fields on M . (2) G := C ∞ (M, K) carries a natural Lie group structure with Lie algebra g := C ∞ (M, k), endowed with the pointwise bracket. If (ϕ, U ) is a k-chart of K, then (ϕG , UG ) with UG := {f ∈ G : f (M ) ⊆ U } and ϕG (f ) = ϕK ◦f is a g-chart of G. For (1) we refer to [Les67], [Omo70] and [Ha82], and for (2) we refer to [Mil84] and [Mi80] (see also [GN10] for both assertions). Theorem 1.2. If q : P → M is a smooth principal bundle over a compact manifold M with locally exponential structure group K, then its gauge group Gau(P ) and its automorphism group Aut(P ) carry Lie group structures turning (1) into a Lie group extension, where the group Diff(M )P is open in Diff(M ). In particular, the conjugation action of Aut(P ) on Gau(P ) is smooth.


287

For the case where K is finite dimensional, this can be found in [KYMO85], [ACM89] and [KM97]. For general locally exponential Lie groups K, this has recently been proved by Wockel [Wo07]. Remark 1.3. (a) If P = M ×K is a trivial bundle, the group extension Aut(P ) splits and we have Aut(P ) ∼ = Gau(P ) Diff(M )

where

Gau(P ) ∼ = C ∞ (M, K).

Here Diff(M ) acts on P by ϕ.(m, k) := (ϕ(m), k) and C ∞ (M, K) acts by f.(m, k) = (m, f (m)k). (b) For any principal K-bundle P , each gauge transformation ϕ ∈ Gau(P ) determines a unique smooth function f : P → K by ϕ(p) = p.f (p) for p ∈ P . Then f satisfies f (p.k) = k −1 f (p)k, and any such function corresponds to a gauge transformation ϕf , so that Gau(P ) ∼ = {f ∈ C ∞ (P, K) : (∀p ∈ P )(∀k ∈ K) f (p.k) = k −1 f (p)k}. Accordingly, we have on the level of vector fields gau(P ) ∼ = {ξ ∈ C ∞ (P, k) : (∀p ∈ P )(∀k ∈ K) ξ(p.k) = Ad(k)−1 ξ(p)}. (c) For any connected manifold M , the universal covering map → M defines on M the structure of a π1 (M )-principal bundle, where qM : M π1 (M ) denotes the group of deck transformations of this bundle, considered as a discrete group. Then ) = Z(π1 (M )) Gau(M

and

) = C Aut(M (π (M )). Diff(M) 1

) carries a natural Lie group structure, turning If M is compact, then Aut(M it into a covering of the open subgroup Diff(M )[M] of Diff(M ) with kernel ) carries a Lie group Z(π1 (M )). The normalizer Diff(M ) of π1 (M ) in Diff(M structure that leads to a short exact sequence 1 → π1 (M ) → Diff(M ) → Diff(M ) → 1. This is due to the fact that each diffeomorphism ϕ of M lifts to some dif normalizing π1 (M ), but in general it does not centralize feomorphism of M π1 (M ). (d) Let ρ : π1 (M ) → K be a homomorphism and ×ρ K := (M × K)/π1 (M ) Pρ := M via ρ. be the corresponding flat K-bundle associated to the π1 (M )-bundle M × K for the right action (m, Then Pρ is the set of π1 (M )-orbits in M k).γ := (m.γ, ρ(γ)−1 k). We write [(m, k)] for the orbit of the pair (m, k). Then → P, s(m) s: M := [(m, 1)] is a smooth function with q ◦ s = qM , satisfying

288

Karl-Hermann Neeb

s(m.γ) = s(m).ρ(γ)

for

, γ ∈ π1 (M ). m ∈M

).K, any bundle automorphism ϕ In view of P = s(M is determined by its values on the image of s. If ϕ ∈ Diff(M ) is the corresponding diffeomorphism , then ϕ of M and ϕ is a lift of ϕ to a diffeomorphism of M can be written ϕ(s( m)) = s(ϕ( m)).f (m)

(3)

→ K, satisfying for some smooth function f : M f (m)ρ(γ) = ρ(ϕγ ϕ −1 )f (m.γ)

for

, γ ∈ π1 (M ). m ∈M

(4)

Conversely, if (4) is satisfied, then (3) defines an automorphism of P . We thus obtain a gauge transformation if and only if ϕ = idM , which leads to , K) : (∀γ ∈ π1 (M )) f ◦ γ = c−1 ◦ f }, Gau(P ) ∼ = {f ∈ C ∞ (M ρ(γ) ) defined under (c) acts naturally where ck (h) := khk −1 . The group Aut(M ), via ϕ(s( on P , preserving s(M m)) := s(ϕ( m)). An element of this group induces the identity on M if and only if it comes from some γ ∈ Z(π1 (M )), → K with the value ρ(γ). and this corresponds to the constant function M We thus obtain an open subgroup of Aut(P ) as the quotient ))/Z(π1 (M )). (Gau(P ) Aut(M of vector fields on M to vector Accordingly, the horizontal lift X → X fields on P is a homomorphism because the bundle is flat. This yields an isomorphism aut(P ) ∼ = gau(P ) V(M ), , k) : (∀γ ∈ π1 (M )) f ◦ γ = Ad(ρ(γ))−1 ◦ f }. where gau(P ) ∼ = {f ∈ C ∞ (M 1.2 Mapping groups on non-compact manifolds For a non-compact smooth manifold M , the preceding construction of the atlas on C ∞ (M, K) no longer works because the sets UG are not open. However, it turns out that in some interesting cases there are other ways to obtain charts. To formulate the results, we say that a Lie group structure on G = C ∞ (M, K) is compatible with evaluations if g := C ∞ (M, k) is its Lie algebra and all K k evaluation maps evK m : G → K, m ∈ M , are smooth with L(evm ) = evm . We then have the following (cf. [NeWa08b]). Theorem 1.4. Let K be a connected regular real Lie group and M a real finite-dimensional connected manifold (with boundary). Then the group C ∞ (M, K) carries a Lie group structure compatible with evaluations if one of the following conditions holds.


289

of K is diffeomorphic to a locally convex (1) The universal covering group K space, which is the case if K is finite-dimensional solvable. If, in addition, π1 (M ) is finitely generated, the Lie group structure is compatible with the smooth compact open topology. (2) dim M = 1. (3) M ∼ = Rk × C, where C is compact. For complex groups we have the following ([NeWa08b], Thms. III.12, IV.3). Theorem 1.5. Let K be a connected regular complex Lie group and M a finite-dimensional connected complex manifold without boundary. Then the group O(M, K), endowed with the compact open topology, carries a Lie group structure with Lie algebra O(M, k) compatible with evaluations if is diffeomorphic to a locally convex space. If, in addition, π1 (M ) is (1) K finitely generated, the Lie group structure is compatible with the compact open topology. (2) dimC M = 1, π1 (M ) is finitely generated and K is a Banach–Lie group Problem 1.6. (1) Do the results of Section 3 on central extensions of mapping group C ∞ (M, K) extend to corresponding groups of smooth and holomorphic maps on non-compact manifolds whenever Theorem 1.4 or Theorem 1.5 provides a suitable Lie group structure? For corresponding results on groups of compactly supported maps we refer to [Ne04b]. (2) Find an example of a connected non-compact smooth manifold M and a finite-dimensional Lie group K for which C ∞ (M, K) does not carry a Lie group structure compatible with evaluations. The first candidate causing problems is the compact group K = SU2 (C), and M should be a manifold that is not a product of some Rn with a compact manifold (see Theorem 1.4(3)), which excludes all Lie groups and all Riemannian symmetric spaces. (3) Do the preceding theorems generalize in a natural way to gauge groups? Which restrictions does one have to impose on the bundles?

2 Central extensions of mapping groups In this section we discuss several issues concerning central extensions of mapping groups. We start with a description of the fundamental types of 2-cocycles on the Lie algebra g := C ∞ (M, k), where M is a finite-dimensional smooth manifold and k a finite-dimensional Lie algebra over K ∈ {R, C}. These 1 cocycles have values in spaces like C ∞ (M, V ), Ω 1 (M, V ) and Ω (M, V ) = Ω 1 (M, V )/dC ∞ (M, V ). In particular, we describe what is known about their integrability to corresponding central Lie group extensions. This is best understood for the identity component C ∞ (M, K)0 , but the extendibility to all of C ∞ (M, K) is not clear (see Problem 7.5 in Appendix A). In Section 4 we outline some generalizations of the results in the present section to gauge groups of non-trivial bundles.

290

Karl-Hermann Neeb

2.1 Central extensions of C ∞ (M, k) 1

We write [α] ∈ Ω (M, V ) for the image of a 1-form α ∈ Ω 1 (M, V ) in this space. The subspace dC ∞ (M, V ) of exact V -valued 1-forms is characterized by the vanishing of all integrals over loops in M , hence a closed subspace, and 1 therefore Ω (M, V ) carries a natural Hausdorff locally convex topology. For a trivial k-module V , we write Sym2 (k, V )k for the space of V -valued symmetric invariant bilinear forms, and recall the Cartan map Γ : Sym2 (k, V )k → Z 3 (k, V ),

Γ (κ)(x, y, z) := κ([x, y], z).

We call κ exact if Γ (κ) is a coboundary. Theorem 2.1. Each continuous 2-cocycle ω with values in the trivial g-module K can be written as ω = β1 ◦ ω1 + β2 ◦ ω2 + β3 ◦ ω3 + ω4 , 1

where β1 , β2 , resp., β3 is a continuous linear map on a space Ω (M, V1 ), C ∞ (M, V2 ), resp., Ω 1 (M, V3 ), and ωi , i = 1, . . . , 4, is a cocycle of type i described as follows: 1

(I) Cocycles ωκ ∈ Z 2 (g, Ω (M, V )) of the form ωκ (ξ1 , ξ2 ) = [κ(ξ1 , dξ2 )],

κ ∈ Sym2 (k, V )k ,

where κ(ξ1 , dξ2 ) is considered as a V-valued 1-form on M ; (II) Cocycles ωη ∈ Z 2 (g, C ∞ (M, V )) of the form ωη (ξ1 , ξ2 ) = η(ξ1 , ξ2 ) := η ◦ (ξ1 , ξ2 ),

η ∈ Z 2 (k, V );

(III) Cocycles ωκ,η ∈ Z 2 (g, Ω 1 (M, V )) of the form ωκ,η (ξ1 , ξ2 ) = κ(ξ1 , dξ2 ) − κ(ξ2 , dξ1 ) − d(η(ξ1 , ξ2 )), where η ∈ C 2 (k, V ) and κ ∈ Sym2 (k, V )k are related by Γ (κ) = dk η; (IV) Cocycles vanishing on g × g, where g = C ∞ (M, k ) is the commutator algebra, i.e., pull-backs of cocycles of an abelian quotient of g. Proof. First we apply [NeWa08a] (Lemma 1,1, Thm. 3.1, Cor. 3.5 and Section 6) to g = C ∞ (M, K) ⊗ k to show that, after subtracting a cocycle of type (IV), each continuous cocycle ω ∈ Z 2 (g, K) can be written as ω(f ⊗ x, g ⊗ y) = βa (f dg − gdf )(x, y) − βs (f g)(x, y), where βa : Ω 1 (M, K) → Sym2 (k, K)k

and

βs : C ∞ (M, K) → C 2 (k, K)

(5)


291

are continuous linear maps coupled by the condition Γ (βa (df )) = dk (βs (f ))

for

f ∈ C ∞ (M, K).

(6)

Conversely, any pair (βa , βs ) satisfying (6) defines a cocycle via (5): ω(f f ⊗ [x, x ], f ⊗ x ) (dg ω)(f ⊗ x, f ⊗ x , f ⊗ x ) = − =−

cyc.

βa (f f df − f d(f f ))([x, x ], x ) +

cyc.

βs (f f f )([x, x ], x )

cyc.

= Γ (βa (d(f f f )))(x, x , x ) − dk (βs (f f f ))(x, x , x ). Let V (k) := S 2 (k)/k.S 2 (k) be the target space of the universal invariant symmetric bilinear form κu (x, y) := [x ∨ y] and observe that Sym2 (k, K)k ∼ = V (k)∗ . We write Sym2 (k, K)kex for the subspace of Sym2 (k, K)k consisting of exact forms, and V (k)0 ⊆ V (k) for its annihilator. Then there exists a linear map χ : Sym2 (k, K)kex → C 2 (k, K) with dk (χ(β)) = Γ (β). As βa (dC ∞ (M, K)) ⊆ Sym2 (k, K)kex , we may use the Hahn–Banach extension theorem to extend βa from dC ∞ (M, K) to a continuous linear map βa : Ω 1 (M, K) → Sym2 (k, K)kex and put βs (f ) := χ(βa (df )), so that ω (f ⊗ x, f ⊗ x ) := βa (f df − f df )(x, x ) − βs (f f )(x, x ) is a 2-cocycle. Further, βa := βa − βa vanishes on dC ∞ (M, K), so that the values of βs := βs − βs are cocycles. Hence both pairs (βa , 0), resp., (0, −βs ) define cocycles, ω2 , resp., ω3 , satisfying ω=ω + ω2 − ω3 . We now show that each of these summands factors through a cocycle of the form (I)–(III). For ω2 we put V := V (k), so that we may write ω2 (f ⊗ x, f ⊗ x ) = βa (f df − f df )(x, x ) = β a (f df − f df ⊗ κu (x, x )), where β a : Ω 1 (M, V ) ∼ = Ω 1 (M, K) ⊗ V → K is a continuous linear map van∞ ishing on dC (M, V ). This shows that ω2 factors through a cocycle of type 1 (I) with values in Ω (M, V ). For ω3 we put V := Z 2 (k, K)∗ , and write ηu ∈ Z 2 (k, V ) for the 2-cocycle defined by ηu (x, x)(f ) = f (x, x ) for f ∈ Z 2 (k, K). Then ω3 (f ⊗ x, f ⊗ x ) = βs (f f )(x, x ) = ηu (x, x )(βs (f f )) = β s (f f ⊗ ηu (x, x )), where β s : C ∞ (M, V ) ∼ = C ∞ (M, K) ⊗ V → K is a continuous linear map. Hence ω3 factors through a cocycle of type (II) with η = ηu .

292

Karl-Hermann Neeb

Finally, we turn to ω . Now we put V := V (k)/V (k)0 , so that V ∗ ∼ = 2 k Sym (k, K)ex and write κ : k × k → V for the symmetric bilinear map obtained from κu . We note that χ : V ∗ → C 2 (k, K) defines a map χ ∈ C 2 (k, V ) by χ(λ)(x, y) = λ(χ (x, y))

for

x, y ∈ k, λ ∈ V ∗ .

For each λ ∈ V ∗ we then identify λ with the corresponding bilinear form λ ◦ κ and obtain λ ◦ dk (χ ) = dk (λ ◦ χ ) = dk (χ(λ)) = Γ (λ) = Γ (λ ◦ κ) = λ ◦ Γ (κ), showing that dk (χ ) = Γ (κ). Let β a : Ω 1 (M, V ) ∼ = Ω 1 (M, K) ⊗ V → K be the continuous linear functional defined by β a (α ⊗ v) = βa (α)(v). Then β a κ(f ⊗ x, df ⊗ x ) − κ(f ⊗ x , df ⊗ x) − d(χ (f ⊗ x, f ⊗ x )) = β a (f df − f df ) ⊗ κ(x, x ) − d(f f ⊗ χ (x, x )) = βa (f df − f df )(x, x ) − βa (d(f f ))(χ (x, x )) = βa (f df − f df )(x, x ) − χ(βa (d(f f )))(x, x ) = βa (f df − f df )(x, x ) − βs (f f )(x, x ) = ω (f ⊗ x, f ⊗ x ). This completes the proof.

Remark 2.2. (a) The central extension g := C ∞ (M, V ) ⊕ωη g defined by a cocycle ωη of type (II) can also be described more directly: If k = V ⊕η k is the central extension defined by η, then g∼ = C ∞ (M, k) ∼ = C ∞ (M, K) ⊗ k, i.e., ∞ the C (M, K)-Lie algebra g is obtained from k by extension of scalars from K to the ring C ∞ (M, K). (b) For any η ∈ C ∞ (M, Z 2 (k, V )), ωη(ξ1 , ξ2 )(m) := η(m)(ξ1 (m), ξ2 (m)) also defines a continuous C ∞ (M, V )-valued 2-cocycle on g. To see how these cocycles fit into the scheme of the preceding theorem, we observe that Z 2 (k, V ) is finite dimensional (if k and V are finite dimensional), so that n η = i=1 αi ·ηi for some αi ∈ C ∞ (M, R) and ηi ∈ Z 2 (k, V ). This leads to ωη = factors through the cocycle ω(η1,...,ηn ) with values i αi · ωηi , showing that ωη in C ∞ (M, V )n ∼ = C ∞ (M, V n ). All the cocycles ωη are C ∞ (M, R)-bilinear and if such a cocycle is a coboundary, there exists a continuous linear map β : g → C ∞ (M, V ) with ωη (ξ1 , ξ2 ) = β([ξ2 , ξ1 ]) for ξ1 , ξ2 ∈ g. This relation easily implies that the restriction of β to the commutator algebra g = C ∞ (M, k ) is C ∞ (M, R)bilinear so that we may w.l.o.g. assume that β itself has this property. Then there exists a smooth map β : M → Hom(k, V ) with β(ξ)(m) = β(m)(ξ(m)) and we obtain


= η(m) dk β(m)

for

293

m ∈ M.

Thus ωη is a coboundary if and only if η(M ) consists of coboundaries. In particular, [ωη ] = 0 is equivalent to [η] = 0 if η = η is constant, which corresponds to cocycles of type (II). (c) Each cocycle ωκ,η of type (III) may be composed with the quotient 1 map q : Ω 1 (M, V ) → Ω (M, V ), which leads to the cocycle q ◦ ωκ,η = 2ωκ of type (I). Here the exactness of κ ensures the existence of a lift of ωκ to an Ω 1 (M, V )-valued cocycle. If η ∈ C 2 (k, V ) also satisfies dk η = Γ (κ), then η − η ∈ Z 2 (k, V ) and ωκ,η = ωκ,η + d ◦ ωη−η (7) is another lift of 2ωκ . (d) If k is a semisimple Lie algebra and κ : k × k → V (k) is the universal invariant symmetric bilinear form, then the corresponding cocycle ωκ with 1 values in Ω (M, V (k)) is universal; i.e., up to coboundaries each 2-cocycle with values in a trivial module can be written as f ◦ ωκ for a continuous linear 1 map f : Ω (M, V (k)) → V (cf. [Ma02], [PS86]). (e) Cocycles of type (III) exist if and only if k carries an exact invariant symmetric bilinear form κ with Γ (κ) = 0. Typical examples of Lie algebras with this property are cotangent bundles k = T ∗ (h) := h∗ h with κ((α, x), (α , x )) := α(x ) + α (x) and η((α, x), (α , x )) := α (x) − α (x) (cf. [NeWa08a], Example 5.3). (f) Cocycles of the types (I)–(III) can also be defined if k is an infinite-dimensional locally convex Lie algebra. In this case we require κ and η to be continuous. The only point where we have used the finite dimension of k in the proof of Theorem 2.1 is to show that every K-valued cocycle is a sum of cocycles factoring through a cocycle of type (I)–(IV); the corresponding arguments do not carry over to the infinite-dimensional case. Remark 2.3. Many results concerning cocycles of type (I)–(III) can be reduced from general manifolds to simpler ones as follows.

1 For cocycles of type (I), the integration maps γ : Ω (M, V ) → V, γ ∈ C ∞ (S1 , M ), separate the points. Accordingly, we have pull-back homomorphisms of Lie groups γ ∗ : C ∞ (M, K) → C ∞ (S1 , K), f → f ◦ γ satisfying 1 ◦ωκM = ◦ωκS ◦ (L(γ ∗ ) × L(γ ∗ )), (8) γ

S1

which can be used to reduce many things to M = S1 (cf. [MN03]). For cocycles of type (II), the evaluation maps evVm : C ∞ (M, V ) → V,

f → f (m)

separate the points, and the corresponding evaluation homomorphism of Lie ∞ groups evK m : C (M, K) → K, f → f (m), satisfies

294

Karl-Hermann Neeb K k k evVm ◦ ωη = η ◦ (L(evK m ) × L(evm )) = η ◦ (evm × evm ).

(9)

Finally, we observe that for cocycles of type (III), the integration maps 1 ∞ : γ Ω (M, V ) → V, γ ∈ C (I, M ), separate the points and that the ∗ pull-back homomorphism γ : C ∞ (M, K) → C ∞ (I, K) satisfies M I ◦ωκ,η = ◦ωκ,η ◦ (L(γ ∗ ) × L(γ ∗ )). (10)

γ

I

2.2 Covariance of the Lie algebra cocycles In this subsection we discuss the covariance of the cocycles of types (I)–(III) under Diff(M ) and C ∞ (M, K), i.e., the existence of actions of these groups on the centrally extended Lie algebras g = z ⊕ω g, compatible with the actions on g and z. For Diff(M ), the situation is simple because the cocycles are invariant, but for C ∞ (M, K) several problems arise. For the identity component C ∞ (M, K)0 , the existence of a corresponding action on g is equivalent to the vanishing of the flux homomorphism (see Proposition 7.1 of Appendix A), but for the existence of an extension to the full group there is no such simple criterion. Remark 2.4. (Covariance under Diff(M )) (a) The Lie algebra V(M ) acts in a p natural way on all spaces Ω p (M, V ) and Ω (M, V ) := Ω p (M, V )/dΩ p−1 (M, V ) and it acts on the Lie algebra g = C ∞ (M, k) by derivations. With respect to this action, all cocycles ωκ , ωη and ωκ,η are V(M )-invariant, and hence extend by ω ((f, X), (f , X )) := ω(f, f ) to cocycles of the semidirect sum g V(M ) (cf. [Ne06a], Def. I.4.2). (b) These cocycles are invariant under the full diffeomorphism group Diff(M ), which implies that the diagonal action of this group on the corresponding central extension defined by ϕ.(z, f ) := ((ϕ−1 )∗ z, ϕ.f ) is a smooth action by Lie algebra automorphisms ([MN03], Thm. VI.3). 1

Remark 2.5. (Covariance under C ∞ (M, K)) (a) For ωκ ∈ Z 2 (g, Ω (M, V )) of type (I) we also have a natural action of the full group C ∞ (M, K) on the 1 extended Lie algebra g = Ω (M, V ) ⊕ωκ g ([MN03], Thm. VI.3). If δ(f ) := l δ (f ) denotes the left logarithmic derivative of f , then θ(f )(x) := [κ(δ(f ), x)] 1 defines a linear map g → Ω (M, V ), satisfying Ad(f )∗ ωκ − ωκ = dg (θ(f )), so that

Adg (f ).(z, ξ) := (z − θ(f )(ξ), Ad(f )ξ)

defines an automorphism of g (Lemma 7.6), and we thus obtain a smooth action of the Lie group C ∞ (M, K) on g.


295

(b) Let K be a connected Lie group with Lie algebra k, V a trivial k-module and η ∈ Z 2 (k, V ) a 2-cocycle. For the cocycle ωη ∈ Z 2 (g, C ∞ (M, V )) of type (II), the situation is slightly more complicated. We recall from Proposition 7.1 of Appendix A that the action of K on k lifts to an action on k = V ⊕η k if and only if the flux Fη : π1 (K) → HomLie (k, V ) = H 1 (k, V ) vanishes. Let m0 ∈ M be a base point and C∗∞ (M, K) C ∞ (M, K) be the normal subgroup of all maps vanishing in m0 . Then this also is a Lie group, and C ∞ (M, K) ∼ = C∗∞ (M, K) K → K denotes the universal covering group, then a as Lie groups. If qK : K if and only smooth map f : M → K vanishing in m0 lifts to a map M → K if the induced homomorphism π1 (f ) : π1 (M ) → π1 (K) vanishes. We therefore have an exact sequence of groups → C ∞ (M, K) → Hom(π1 (M ), π1 (K)), 1 → C∗∞ (M, K) ∗

(11)

and since π1 (M ) is finitely generated because M is compact, we may consider as an open subgroup of C ∞ (M, K). From the canonical action Ad C∗∞ (M, K) k on of K on k we immediately get a smooth action of C ∞ (M, K) g∼ = C ∞ (M, k) by (Adg (f ).ξ)(m) := Adk (f (m)).ξ(m). This shows that the flux Fωη : π1 (C ∞ (M, K)) ∼ = π1 (C∗∞ (M, K)) × π1 (K) → H 1 (g, C ∞ (M, V )) vanishes on the subgroup π1 (C∗∞ (M, K)), and hence factors through the flux Fη of η. For any f ∈ C ∞ (M, K) and m ∈ M , the connectedness of K implies that Ad(f (m))∗ η − η is a 2-coboundary (Theorem 9.1 of Appendix C), and from that we derive the existence of a smooth map θ: M → C 1 (k, V ) with Ad(f (m)−1 )∗ η − η = dk (θ(m)) for m ∈ M . Now the automorphism Ad(f ) lifts to an automorphism of g (Lemma 7.6). (c) To determine the covariance of the cocycles ωκ,η of type (III) under the group C ∞ (M, K), we first observe that for f ∈ C ∞ (M, K) and ξ1 , ξ2 ∈ g we obtain similarly as in [MN03], Prop. III.3: Ad(f )∗ ωκ,η (ξ1 , ξ2 ) = κ(Ad(f ).ξ1 , d(Ad(f ).ξ2 )) − κ(Ad(f ).ξ2 , d(Ad(f ).ξ1 )) − d (Ad(f )∗ η)(ξ1 , ξ2 ) = κ(ξ1 , dξ2 ) − κ(ξ2 , dξ1 ) + 2κ(δ l (f ), [ξ2 , ξ1 ]) − d (Ad(f )∗ η)(ξ1 , ξ2 ) = ωκ,η (ξ1 , ξ2 ) + 2κ(δ l (f ), [ξ2 , ξ1 ]) − d (Ad(f )∗ η − η)(ξ1 , ξ2 ) .

296

Karl-Hermann Neeb

It is clear that the term κ(δ l (f ), [ξ2 , ξ1 ]) is a Lie algebra coboundary. To see that the second term also has this property, we note that Γ (κ) = dk η implies that for each x ∈ k the cochain ix (dk η)(y, z) = κ([x, y], z) = κ(x, [y, z]) is a coboundary. Hence Lx η = ix dk η + dk (ix η) also is a coboundary and since K is connected, for each m ∈ M the cocycle Ad(f (m))∗ η −η is a coboundary. With Lemma 7.6 we now see that, for each smooth function f : M → K, the adjoint action on g lifts to an automorphism of the central extension g, defined by ωκ,η . If f is not in C ∞ (M, K)0 , i.e., homotopic to a constant map, this does not follow from Theorem 9.1. Problem 2.6. For which groups K and which connected smooth manifolds M does the short exact sequence 1 → C ∞ (M, K)0 → C ∞ (M, K) → π0 (C ∞ (M, K)) ∼ = [M, K] → 1 split? If K ∼ = k/ΓK is abelian, then C ∞ (M, K)0 is a quotient of C ∞ (M, k), hence divisible, and therefore the sequence splits. If K is non-abelian, the situation is more involved and even the case M = S1 is non-trivial. Then [M, K] ∼ = π1 (K) and if K is abelian, then we obtain a splitting from the isomorphism π1 (K) ∼ = Hom(T, K). If K is solvable and T ⊆ K is a maximal torus, then K ∼ = Rm × T as a manifold ([Ho65]), so that [M, K] ∼ = [M, T ] and we obtain a splitting from the splitting [M, T ] → C ∞ (M, T ) ⊆ C ∞ (M, K). The case of compact (semisimple) groups is the crucial one. If M is a product of d spheres (for instance a d-dimensional torus), then [M, K] is nilpotent of length ≤ d which is not always abelian ([Whi78], Thm. X.3.6). Problem 2.7. Can the automorphisms of the central extensions g = z ⊕ω g of types (I)–(III) corresponding to elements of C ∞ (M, K) be chosen in such a way as to define a group action? If k and hence g is perfect, this follows trivially from the uniqueness of lifts to g. In general, lifts are only unique up to an element of the group H 1 (g, z), so that only an abelian extension acts. Are the corresponding cohomology classes in H 2 (C ∞ (M, K), z) that arise this way always trivial? Since the action of g on g is determined by the cocycle, we obtain an action of the identity component C ∞ (M, K)0 whenever the flux vanishes, and then the restriction of the cocycle to the identity component is trivial. Here even the case M = S1 of loop groups is of particular interest. Then π0 (C ∞ (S1 , K)) ∼ = π1 (K), so that we have a natural inflation map H 2 (π1 (K), z) → H 2 (C ∞ (S1 , K), z), and it would be nice if all classes under consideration could be determined this way. If we have a semidirect decomposition C ∞ (M, K) ∼ = C ∞ (M, K)0 [M, K], then the present problem simplifies significantly because then semidirect product techniques similar to those discussed in Appendix B for Lie algebras apply.


297

Problem 2.8. Generalize the description of H 2 (A ⊗ k, K) obtained in [NeWa08a] under the assumption that A is unital to non-unital commutative associative algebras. This would be of particular interest for algebras of the form C∗∞ (M, R) (functions vanishing in one point) or the algebra Cc∞ (M, R) of compactly supported functions. If A+ := A ⊕ K is the algebra with unit 1 = (0, 1), then we have A+ ⊗ k ∼ = (A ⊗ k) k, so that we may use the semidirect product techniques in Appendix B. 2.3 Corresponding Lie group extensions A natural question is: To which extent do the cocycles of the form ωκ , ωη and ωκ,η on g = C ∞ (M, k) actually define central extensions of the corresponding group C ∞ (M, K)? Cocycles of the form ωκ for a vector-valued κ : k × k → V are treated in [MN03], where it is shown that the corresponding period homomorphism 1

perωκ : π2 (C ∞ (M, K)) → Ω (M, V ) 1 (M, V ), and the image con(see Appendix A) has values in the subspace HdR sists of cohomology classes whose integrals over circles take values in the image of the homomorphism perκ : π3 (K) → V, [σ] → Γ (κ)eq , σ

where Γ (κ)eq ∈ Ω 3 (K, V ) is the left invariant closed 3-form whose value in 1 is Γ (κ). The flux homomorphism 1

Fω : π1 (C ∞ (M, K)) → H 1 (g, Ω (M, V )) vanishes because the action of g on the central extension g defined by ωκ integrates to a smooth action of the group C ∞ (M, K) on g (Remark 2.5(a), [MN03], Prop. III.3). This leads to the following theorem ([MN03], Thms. I.6, II.9, Cor. III.7). Theorem 2.9. For the cocycles ωκ and the connected group G := C ∞ (M, K)0 , the following are equivalent. (1) ωκ integrates for each compact manifold M to a Lie group extension of G. (2) ωκ integrates for M = S1 to a Lie group extension of G. 1 1 (3) The image of perωκ in HdR (M, V ) ⊆ Ω (M, V ) is discrete. (4) The image of perκ in V is discrete. These conditions are satisfied if κ is the universal invariant bilinear form with values in V (k).

298

Karl-Hermann Neeb

Remark 2.10. Composing a cocycle ωκ with the exterior derivative, we obtain the Ω 2 (M, V )-valued cocycle (d ◦ ωκ )(ξ1 , ξ2 ) = κ(dξ1 , dξ2 ). In view of [MN03], Thm. III.9, we have a smooth 2-cocycle c(f1 , f2 ) := δ l (f1 ) ∧κ δ r (f2 ) = δ r (f1 ) ∧κ Ad(f1 ).δ r (f2 ) on the full group C ∞ (M, K) which defines a central extension by Ω 2 (M, V ) whose corresponding Lie algebra cocycle is 2d ◦ ωκ . Here ∧κ denotes the natural product Ω 1 (M, k) × Ω 1 (M, k) → Ω 2 (M, V ) defined by κ : k × k → V . In Subsection 4.3, this construction is generalized to gauge groups of non-trivial bundles. For the cocycles of type (II), the situation is much simpler (see Theorem 7.2). Theorem 2.11. Let K be a connected finite-dimensional Lie group with Lie algebra k, η ∈ Z 2 (k, V ), and k = V ⊕η k. Then the following are equivalent. (a) The central Lie algebra extension g = C ∞ (M, V ) ⊕ωη g of g = C ∞ (M, k) integrates to a central Lie group extension of C ∞ (M, K)0 . → K. (b) k → k integrates to a Lie group extension V → K 1 (c) The flux homomorphism Fη : π1 (K) → H (k, V ) vanishes. If this is the case, then → C ∞ (M, K) → 1 1 → C ∞ (M, Z) → C ∞ (M, K) g. defines a central Lie group extension of the full group C ∞ (M, K) integrating ∗ k ∗ Proof. For m ∈ M , the relation evVm ◦ ωη = L(evK m ) η = (evm ) η (see (9) in V K Remark 2.3) leads to evm ◦ perωη = perη ◦π2 (evm ) (see (17) in Remark 7.4). Since π2 (K) vanishes ([CaE36]), the period map perη is trivial; therefore, all maps evVm ◦ perωη vanish, which implies that perωη = 0. A direct calculation further shows that the flux homomorphisms Fη and Fωη : π1 (G) → H 1 (g, C ∞ (M, V )) satisfy

Fωη ([γ])(ξ)(m) = Fη ([evK m ◦ γ])(ξ(m)),

ξ ∈ g, [γ] ∈ π1 (G)

(see (18) in Remark 7.4). Therefore, Fη vanishes if and only if Fωη does. Now Theorem 7.2 of Appendix A shows that (a) is equivalent to (b), which in turn is equivalent to (c) because all period maps vanish. If these conditions are satisfied, we apply Theorem 7.2 with V = A to → K. This extension is a principal V obtain a central extension V → K bundle, hence trivial as such, so that there exists a smooth section σ : K → K. is a central Lie group extension of C ∞ (M, K) with a Therefore, C ∞ (M, K) g (cf. [GN10]).

smooth global section. Its Lie algebra is C ∞ (M, k) ∼ =


299

→ Remark 2.12. If, in the context of Theorem 2.11, qK : K → K is a connected Lie group extension of K by Z = V /ΓZ , ΓZ ⊆ V a discrete subgroup and generally does not have a smooth global section, = k = V ⊕η k, then K L(K) → C ∞ (M, K) only is an open and the image of the canonical map C ∞ (M, K) subgroup. In our context, where K is finite dimensional, the obstruction for a map can be made quite explicit. The existence of a smooth f : M → K to lift to K of f is equivalent to the triviality of the smooth Z-bundle lift f: M → K → M , and the equivalence classes of these bundles are parametrized by f ∗K ˇ H 2 (M, ΓZ ) (cf. [Bry93]). Describing bundles in terms of Cech cocycles, it is easy to see that we thus obtain a group homomorphism C ∞ (M, K) → H 2 (M, ΓZ ),

f → [f ∗ K]

which factors through an injective homomorphism π0 (C ∞ (M, K)) ∼ = [M, K] → H 2 (M, ΓZ ),

[f ] → [f ∗ K]

of discrete groups. Since π2 (K) and Fη vanish, Remark 6.12 in [Ne04a] implies that the ∈ H 2 (K, ΓZ ) vanishes on H2 (K). Therefore, the cohomology class [K] universal coefficient theorem shows that it is defined by an element of as Ext(H1 (K), ΓZ ) = Ext(π1 (K), ΓZ ). Accordingly, we may interpret [f ∗ K] an element of the group Ext(H1 (M ), ΓZ ). The long exact homotopy sequence contains the short exact sequence of the Z-bundle K → π1 (K) → 1, 1 → ΓZ = π1 (Z) → π1 (K) as a central extension of π1 (K) corresponding to the which exhibits π1 (K) class in Ext(π1 (K), ΓZ ). We now see that the homomorphism from above yields an exact sequence of groups → C ∞ (M, K) → Ext(H1 (M ), ΓZ ). C ∞ (M, K) Here the rightmost map can be calculated by first assigning to f ∈ C ∞ (M, K) ∈ the homomorphism H1 (f ) : H1 (M ) → H1 (K) ∼ = π1 (K) and then H1 (f )∗ [K] Ext(H1 (M ), ΓZ ), which can be evaluated easily in concrete cases. Now we turn to cocycles of type (III). We start with a key example. Example 2.13. Let k be a locally convex real Lie algebra and consider the locally convex Lie algebra g := C ∞ (I, k), where I := [0, 1] is the unit interval. Let κ : k × k → V be a continuous invariant symmetric bilinear form with values in the Mackey complete space V . Consider V as a trivial g-module and let η ∈ C 2 (k, V ) with Γ (κ) = dk η. Then the cocycle ωκ,η ∈ Z 2 (g, Ω 1 (I, V )) vanishes on the subalgebra k ⊆ g of constant maps.

300

Karl-Hermann Neeb

Let K be a connected Lie group with Lie algebra k and G := C ∞ (I, K). The map H : I×G → G, H(t, f )(s) := f (ts) is smooth because the correspond : I × G × I → K, H(t, f, s) := f (ts) is smooth (cf. [NeWa08b], ing map H Lemma A.2). Since H1 = idG and H0 (f ) = f (0), H is a smooth retraction of G to the subgroup K of constant maps. Therefore, the inclusion j : K → G induces an isomorphism π2 (j) : π2 (K) → π2 (G). The period map perωκ,η : π2 (G) → V satisfies perωκ,η ◦π2 (j) = perj ∗ ωκ,η = 0 because the cocycle j ∗ ωκ,η = ωκ,η |k×k vanishes, so that the period group of ωκ,η is trivial. Theorem 2.14. Let M be a smooth compact manifold, k a locally convex Lie algebra, κ : k × k → V a continuous invariant symmetric bilinear form and let η ∈ C 2 (k, V ) with Γ (κ) = dk η. For a connected Lie group K with Lie algebra k and G := C ∞ (M, K), the following assertions hold: (a) The period map perωκ,η : π2 (G) → Ω 1 (M, V ) vanishes. (b) If K is simply connected, then the flux Fωκ,η vanishes. (c) If K is finite dimensional, then there exists η ∈ C 2 (k, V ) with dk η = Γ (κ) for which the flux Fωκ,η vanishes. (d) If Fωκ,η = 0, then ωκ,η integrates to a central extension of G0 by Ω 1 (M, V ). Proof. (a) Combining (10) in Remark 2.3 with (17) in Remark 7.4, we obtain for each γ ∈ C ∞ (I, M ): ◦ωκ,η ◦ωκ,η ◦ perωκ,η ◦π2 (γ ∗ ) = ◦ perωκ,η ◦π2 (γ ∗ ), M = per M = per I I γ

γ

I

I

and Example 2.13 implies that perωκ,η vanishes. Since γ was arbitrary, all I M periods of ωκ,η vanish. (b) Next we assume that K is 1-connected. In Example 2.13 we have seen I that the group C ∞ (I, K) is also 1-connected, so that the flux of ωκ,η vanishes. Now we apply (18) in Remark 7.4 with (10) in Remark 2.3 to see that the flux Fωκ,η also vanishes. (c) Assume that K is finite dimensional and pick a maximal compact subgroup C ⊆ K. Since κ is K-invariant, K acts on the affine space {η ∈ C 2 (k, V ) : dk η = Γ (κ)}, and the restriction of this action to the compact subgroup C has a fixed point η. Then ωκ,η is C-invariant, and therefore C acts diagonally on g = Ω 1 (M, V ) ⊕ωκ,η g. It follows, in particular, that the flux vanishes on the image of π1 (C) ∼ = π1 (K) in π1 (G) (Proposition 7.1). For M = I = [0, 1] we immediately derive that the flux vanishes, and in the general case we argue as in (b) by reduction via pull-back maps γ ∗ : C ∞ (M, K) → C ∞ (I, K). (d) This assertion follows from Theorem 7.2.

Remark 2.15. We have seen in the preceding theorem that if K is simply connected, the period homomorphism and the flux of ω := ωκ,η vanish. Since G := C ∞ (I, K) is also 1-connected, one may therefore expect an explicit formula for a corresponding group cocycle integrating ωκ,η because


301

eq 2 the cohomology class [ωκ,η ] ∈ HdR (C ∞ (I, K), C ∞ (I, R)) vanishes (cf. [Ne02], Thm. 8.8). Such a formula can be obtained by a method due to Cartan, combined with fiber integration and the homotopy H : I × G → G to K. Since ω vanishes on k, the fiber integral θ :==I H ∗ ω eq ∈ Ω 1 (G, V ) satisfies dθ = ω eq , so that we may use Prop. 8.2 in [Ne04a] to construct an explicit cocycle. For a more general method, based on path groups, we refer to [Vi08].

Remark 2.16. In view of Remarks 2.4 and 2.5, for cocycles ω of types (I)–(III), the diffeomorphism group Diff(M ) acts diagonally on g. If, in addition, K is 1-connected, then the flux Fω vanishes, so that the identity component G0 := C ∞ (M, K)0 also acts naturally on g (Proposition 7.1), and this action is uniquely determined by the adjoint action of g on g (cf. [GN10]). From that it easily follows that these two actions combine to a smooth action of C ∞ (M, K)0 Diff(M ) on g and the Lifting Theorem 7.7 provides a smooth action of this group by automorphisms on any corresponding central of the simply connected covering group G 0 . extension G Problem 2.17. Do the central extensions of the connected group C ∞ (M, K)0 constructed above extend to the full group? The answer is positive for M = S1 , K compact simple and ω = ωκ if κ is universal ([PS86], Prop. 4.6.9). See Problem 7.5 for the general framework in which this problem can be investigated.

3 Twists and the cohomology of vector fields We have seen above how to obtain central extensions of the Lie algebra C ∞ (M, k) and that the three types of cocycles ωκ , ωη and ωκ,η are V(M )-invariant, which leads to central extensions of the semidirect sum C ∞ (M, k) V(M ) by spaces of the type z := C ∞ (M, V ), Ω 1 (M, V ) and 1 Ω (M, V ). The bracket on such a central extension has the form [(z1 , (ξ1 , X1 )), (z2 , (ξ2 , X2 ))] = X1 .z2 − X2 .z1 + ω(ξ1 , ξ2 ), (X1 .ξ2 − X2 .ξ1 + [ξ1 , ξ2 ], [X1 , X2 ]) , and there are natural twists of these central extensions that correspond to replacing the cocycle ω ((ξ1 , X1 ), (ξ2 , X2 )) := ω(ξ1 , ξ2 ) by ω + η, where η((ξ1 , X1 ), (ξ2 , X2 )) := η(X1 , X2 )

for some

η ∈ Z 2 (V(M ), z)

(see Corollary 8.6 in Appendix B). To understand the different types of twists, one has to determine the cohomology groups Hc2 (V(M ), R),

Hc2 (V(M ), Ω 1 (M, R))

and

1

Hc2 (V(M ), Ω (M, R)).

For a parallelizable manifold M , these spaces have been determined in [BiNe08], and in the following subsections we describe the different types of cocycles that appear.

302

Karl-Hermann Neeb

We also discuss the integrability of these twists to abelian extensions of Diff(M ), or at least of the simply connected covering Diff(M )0 of its identity component Diff(M )0 . 3.1 Some cohomology of the Lie algebra of vector fields The most obvious source of 2-cocycles of V(M ) with values in differential forms is described in the following lemma, applied to p = 2 ([Ne06c], Prop. 6). Lemma 3.1. For each closed (p + q)-form ω ∈ Ω p+q (M, V ), the prescription q

ω [p] (X1 , . . . , Xp ) := [iXp . . . iX1 ω] ∈ Ω (M, V ) q

defines a continuous p-cocycle in Z p (V(M ), Ω (M, V )). The preceding lemma is of particular interest for p = 0. In this case it associates Lie algebra cohomology classes to closed differential forms. For the • following theorem, we recall that, for a subset E ⊆ HdR (M, R), the ideal of • all classes subordinated to E (cf. [Th54]) is the minimal ideal of HdR (M, R) ∼ = • ∗ Hsing (M, R) containing E, which is the kernel of a pull-back map ϕ for some continuous map ϕ : N → M , N a topological space. Theorem 3.2. Each smooth p-form ω ∈ Ω p (M, R) defines a p-linear alternating continuous map V(M )p → C ∞ (M, R), which leads to an inclusion of chain complexes (Ω • (M, R), d) → (C • (V(M ), C ∞ (M, R)), dV(M) ) inducing an algebra homomorphism • Ψ : HdR (M, R) → H • (V(M ), C ∞ (M, R))

whose kernel is the ideal of all classes suborbinated to the Pontrjagin classes of M , i.e., the characteristic classes of the tangent bundle. Proof. The main point is the determination of the kernel of Ψ , for which we refer to [ST77, p. 224]. That ker Ψ contains the Pontrjagin classes follows from Cor. 3.2 in [Lec85], so that the representability of H • (V(M ), C ∞ (M, R)) as the cohomology of a bundle space over M ([Tsu81]) further implies that ker Ψ contains all classes subordinated to the Pontrjagin classes.

Remark 3.3. There are several classes of compact manifolds for which the Pontrjagin classes all vanish but the tangent bundle is non-trivial. According to [GHV72], II, Proposition 9.8.VI, this is in particular the case for all Riemannian manifolds with constant curvature, and hence in particular for all spheres. A second source of cocycles is the following ([Kosz74]).


303

Lemma 3.4. Any affine connection ∇ on M defines a 1-cocycle ζ : V(M ) → Ω 1 (M, End(T M )),

X → LX ∇,

where (LX ∇)(Y )(Z) = [X, ∇Y Z] − ∇[X,Y ] Z − ∇Y [X, Z]. For any other affine connection ∇ , the corresponding cocycle ζ has the same cohomology class. Definition 3.5. We use the cocycle ζ, associated to an affine connection ∇ to define k-cocycles Ψk ∈ Zck (V(M ), Ω k (M, R)). Note that A → Tr(Ak ) defines a homogeneous polynomial of degree k on gld (R), invariant under conjugation. The corresponding invariant symmetric k-linear map is given by β(A1 , . . . , Ak ) = Tr(Aσ(1) · · · Aσ(k) ), σ∈Sk

and we consider it as a linear GLd (R)-equivariant map gld (R)⊗k → R, where GLd (R) acts trivially on R. This map leads to a vector bundle map βM : End(T M )⊗k → M × R, where M × R denotes the trivial vector bundle with fiber R. On the level of bundle-valued differential forms, this in turn yields an alternating k-linear Diff(M )-equivariant map 1 βM : Ω 1 (M, End(T M ))k → Ω k (M, R).

The Diff(M )-equivariance implies the invariance of this map under the natural 1 action of V(M ), so that we can use βM to multiply Lie algebra cocycles (cf. [Fu86] or App. F in [Ne04a]). In particular, we obtain for each k ∈ N a Lie algebra cocycle Ψk ∈ Zck (V(M ), Ω k (M, R)), defined by Ψk (X1 , . . . , Xk ) := (−1)k

1 sgn(σ)βM (ζ(Xσ(1) ), . . . , ζ(Xσ(k) )).

σ∈Sk

For any other affine connection ∇ , the corresponding cocycle ζ and the associated cocycles Ψk , the difference ζ −ζ is a coboundary, and since products of cocycles and coboundaries are coboundaries, Ψk −Ψk is a coboundary. Hence its cohomology class in Hck (V(M ), Ω k (M, R)) does not depend on ∇. Remark 3.6. The Lie algebra 1-cocycle ζ is the differential of the group cocycle ζD : Diff(M ) → Ω 1 (M, End(T M )),

ϕ → ϕ.∇ − ∇,

where (ϕ.∇)X Y := ϕ.(∇ϕ−1 .X ϕ−1 .Y ) for ϕ.X = T (ϕ) ◦ X ◦ ϕ−1 . Now the k-fold cup product of ζD with itself defines an Ω k (M, R)-valued smooth k-cocycle on Diff(M ) whose corresponding Lie algebra cocycle is Ψk (cf. [Ne04a], Lemma F.3).

304

Karl-Hermann Neeb

Remark 3.7. Assume that M is a parallelizable d-dimensional manifold and that the 1-form τ ∈ Ω 1 (M, Rd ) implements this trivialization in the sense that each map τm is a linear isomorphism Tm (M ) → Rd . Then, for each X ∈ V(M ), LX τ can be written as LX τ = −θ(X)·τ for some smooth function θ(X) ∈ C ∞ (M, gld (R)) and θ is a crossed homomorphism, i.e., θ([X, Y ]) = X.θ(Y ) − Y.θ(X) + [θ(X), θ(Y )] ([BiNe08], Example II.3). Moreover, ∇X Y := τ −1 (X.τ (Y )) defines an affine connection on M for which the parallel vector fields correspond to constant functions, and for the corresponding cocycle ζ, the map −1 τ : Γ (End(T M )) → C ∞ (M, gld (R)), τ(ϕ)(m) = τm ◦ ϕm ◦ τm satisfies τ ◦ ζ(X) = −d θ(X) ∈ Ω 1 (M, gld (R)). This leads to sgn(σ) Tr dθ(Xσ(1) ) ∧ . . . ∧ dθ(Xσ(k) ) . Ψk (X1 , . . . , Xk ) =

(12)

σ∈Sk

If T M is trivial, we further obtain the cocycles Φk ∈ Zc2k−1 (V(M ), C ∞ (M, R)), Φk (X1 , . . . , X2k−1 ) = sgn(σ) Tr θ(Xσ(1) ) · · · θ(Xσ(2k−1) ) σ∈S2k−1 k−1

and Ψ k ∈ Zck (V(M ), Ω (M, R)), defined by Ψ k (X1 , . . . , Xk ) = sgn(σ)[Tr θ(Xσ(1) )dθ(Xσ(2) ) ∧ . . . ∧ dθ(Xσ(k) ) ] σ∈Sk

([BiNe08], Def. II.7). For each k ≥ 1 we have d ◦ Ψ k = Ψk . For k = 1 we have in particular 0

Ψ 1 (X) = Tr(θ(X)) ∈ C ∞ (M, R) = Ω (M, R),

Ψ1 (X) = Tr(dθ(X)),

and 1 Ψ 2 (X1 , X2 ) = [Tr θ(X1 )dθ(X2 ) − θ(X2 )dθ(X1 ) ] ∈ Ω (M, R). Collecting the results from [BiNe08], Section IV, we now obtain the following classification result which, for a parallelizable manifold M , provides in particular all the information on the twists of the fundamental central extensions of the mapping Lie algebras defined by cocycles of types (I)–(III). Theorem 3.8. Let M be a compact d-dimensional manifold with trivial tangent bundle. Then the following assertions hold. (1) For p ≥ 2 and d ≥ 2 the map p

p+2 HdR (M, R) → H 2 (V(M ), Ω (M, R)),

is a linear isomorphism.

[ω] → [ω [2] ]


305

(2) For d ≥ 2 we have 3 H 2 (V(M ), Ω (M, R)) ∼ (M, R) ⊕ R[Ψ 1 ∧ Ψ1 ] ⊕ R[Ψ 2 ], = HdR 1

3 where HdR (M, R) embeds via [ω] → [ω [2] ]. (3) For d ≥ 2 the map 2 1 HdR (M, R)⊕HdR (M, R) → H 2 (V(M ), C ∞ (M, R)), ([α], [β]) → [α+β∧Ψ 1 ]

is a linear isomorphism. (4) For M = S1 , dim H 2 (V(M ), C ∞ (M, R)) = 2 and dim H 2 (V(M ), 1 Ω (M, R)) = dim H 2 (V(M ), R) = 1. In addition, for any compact connected manifold M , H 2 (V(M ), Ω 2 (M, R)) = R[Ψ2 ] ⊕ R[Ψ1 ∧ Ψ1 ] is 2-dimensional and 1 H 2 (V(M ), Ω 1 (M, R)) = R[Ψ 1 ∧ Ψ1 ] ⊕ {[α ∧ Ψ1 ] : [α] ∈ HdR (M, R)},

is of dimension 1 + b1 (M ). Remark 3.9. The main point of the preceding theorem is the precise description of the cohomology spaces, which is obtained under the assumption that the tangent bundle is trivial. The cocycles Ψk exist for all manifolds, whereas the construction of the cocycles Φk and Ψ k requires at least the existence of a flat affine connection on M : If ∇ is a flat affine connection on the d-dimensional manifold M , then we have a holonomy homomorphism ρ : π1 (M ) → GLd (R), for which the tangent bundle T M is equivalent to the ×ρ Rd . We then obtain a trivializing 1-form τ ∈ associated bundle M , Rd ) satisfying γ ∗ τ = ρ(γ)−1 ◦ τ for γ ∈ π1 (M ) and a corresponding Ω 1 (M crossed homomorphism ) → C ∞ (M , gld (R)). θ: V(M ∈ V(M )π1 (M) denote the canonical lift of X to M . Then For X ∈ V(M ), let X X) X) ◦ ρ(γ) = ρ(γ)−1 ◦ θ( γ ∗ θ(

for

γ ∈ π1 (M ),

which means that the cocycles Ψk , Φk and Ψ k lead on vector fields of the form to π1 (M )-invariant differential forms, and hence that they actually define X k−1 (M, R). cocycles on V(M ) with values in Ω k (M, R), C ∞ (M, R) and Ω 1

Problem 3.10. Compute the second cohomology of V(M ) in Ω (M, R) for all (compact) connected smooth manifolds. The cohomology with values in Ω p (M, R) can be reduced with results of Tsujishita ([Tsu81], Thm. 5.1.6; [BiNe08], Thm. III.1) to the description of H • (V(M ), C ∞ (M, R)), which depends very much on the topology of M ([Hae76]).

306

Karl-Hermann Neeb

3.2 Abelian extensions of diffeomorphism groups We now address the integrability of the cocycles described in the preceding subsection. Here a key method consists in a systematic use of crossed homomorphisms which are used to pull back central extensions of mapping groups (cf. [Bi03]). Definition 3.11. Let G and N be Lie groups and S : G → Aut(N ) a smooth action of G on N . A smooth map θ : G → N is called a crossed homomorphism if θ(g1 g2 ) = θ(g1 ) · g1 .θ(g2 ), i.e., if θ := (θ, idG ) : G → N S G is a morphism of Lie groups. Crossed homomorphisms are non-abelian generalizations of 1-cocycles. They have the interesting application that, for any smooth G-module V (considered as a module of N S G on which N acts trivially), we have a natural pull-back map θ∗ = (θ, idG )∗ : H 2 (N S G, V ) → H 2 (G, V ).

(13)

Two crossed homomorphisms θ1 , θ2 : G → N are said to be equivalent if there exists an n ∈ N with θ2 = cn ◦ θ1 , cn (x) = nxn−1 . Since N acts trivially by conjugation on the cohomology group H 2 (N S G, V ) (cf. [Ne04a], Prop. D.6), equivalent crossed homomorphisms define the same pull-back map on cohomology. Remark 3.12. (a) If θ : G → N is a crossed homomorphism, then θ = (θ, idG ) yields a new splitting on the semidirect product group N S G, which leads to an isomorphism N S G ∼ = N S G, where S (g) = cθ(g) ◦ S(g). As we shall see, it may very well happen that a Lie group extension of N S G is trivial on {1} × G but not on θ(G), which leads to a non-trivial extension of G (Example 3.16). (b) If θ : G → N is a crossed homomorphism of Lie groups, L(θ) := T1 (θ) : L(G) → L(N ) is a crossed homomorphism of Lie algebras because : L(G) → L(N ) L(G) is a morphism of Lie algebras. (L(θ), idL(G) ) = L(θ) (c) If the central extension n = V ⊕ω n is defined by a g-invariant cocycle ω and ω is the corresponding cocycle of the abelian extension ng, corresponding to the diagonal action of g on n, then the pull-back of this cocycle with respect to a crossed homomorphism θ : g → n has the simple form (θ∗ ω )(x, y) = ω ((θ(x), x), (θ(y), y)) = ω(θ(x), θ(y)) = (θ∗ ω)(x, y). As we shall see here, crossed homomorphisms can be used to prove the integrability of many interesting cocycles of the Lie algebra of vector fields to the corresponding group of diffeomorphisms.


307

Example 3.13. (a) Let q : V → M be a natural vector bundle over M with typical fiber V . Then the action of Diff(M ) on M lifts to an action on V by bundle automorphisms. Any trivialization of V yields an isomorphism Aut(V) ∼ = Gau(V) Diff(M ) → Aut(M × V ) ∼ = C ∞ (M, GL(V )) Diff(M ), and the lift ρ : Diff(M ) → Aut(V) can now be written as ρ = (θ, idDiff(M) ) for some crossed homomorphism θ : Diff(M ) → C ∞ (M, GL(V )). (b) Let τ ∈ Ω 1 (M, Rd ) be such that each τm is invertible in each m ∈ M , so that τ defines a trivialization of the tangent bundle of M . In [BiNe08], Ex. II.3, it is shown that LX τ = −θ(X) · τ defines a crossed homomorphism θ : V(M ) → C ∞ (M, gld (R)) (cf. Remark 3.7). A corresponding crossed homomorphism on the group level is defined by Θ : Diff(M ) → C ∞ (M, GLd (R)),

ϕ.τ = (ϕ−1 )∗ τ = Θ(ϕ)−1 · τ.

It satisfies L(Θ) = θ, taking into account that V(M ) is the Lie algebra of the opposite group Diff(M )op , which causes a minus sign. We shall use this crossed homomorphism to write the cocycles Ψ 2 and Ψ 1 ∧ Ψ1 as pull-backs (cf. [Bi03] for the case where M is a torus). For the Lie algebra k := gld (R), the space Sym2 (k, R)k is two dimensional, spanned by the two forms κ1 (x, y) := tr(x) tr(y)

and

κ2 (x, y) := tr(xy).

Then the relations [tr(θ(X1 )) tr(dθ(X2 )) − tr(θ(X2 )) tr(dθ(X1 ))] = (Ψ 1 ∧ Ψ1 )(X1 , X2 ), [tr(θ(X1 )dθ(X2 ) − θ(X2 )dθ(X1 ))] = Ψ 2 (X1 , X2 ) imply that

2θ∗ ωκ1 = Ψ 1 ∧ Ψ1

and

2θ∗ ωκ2 = Ψ 2 .

With these preparations, we are now ready to prove the following. Proposition 3.14. If the tangent bundle T M is trivial, then the cocycles Ψ 1 ∧ Ψ1

and

Ψ2

in

Z 2 (V(M ), Ω 1 (M, R))

integrate to abelian Lie group extensions of some covering group of Diff(M )0 . Proof. Fix i ∈ {1, 2}. Since the invariant form κi is one of the two components of the universal invariant symmetric bilinear form of gld (R), Theorem 2.9 implies that the cocycles 2ωκi of C ∞ (M, gld (R)) are integrable to a central ex of the simply connected covering group G of G := C ∞ (M, GLd (R))0 tension G

308

Karl-Hermann Neeb 1

by Z := Ω (M, V )/ im(perωκi ) on which Diff(M ) acts smoothly by automorphisms (see Remark 2.4 and the Lifting Theorem 7.7). We thus obtain a := G Diff(M semidirect product group H )0 which is an abelian extension Diff(M )0 by Z. of H := G The crossed homomorphism Θ : Diff(M )0 → G has a unique lift to a and the pull-back extension Ξ ∗H crossed homomorphism Ξ : Diff(M )0 → G ∗ integrates 2θ ωκi . In view of the preceding example, the assertion follows.

Problem 3.15. It would be nice to know if we really need to pass to a covering group in the preceding proposition. As the argument in the proof shows, this is not necessary if the crossed homomorphism Θ lifts to a crossed homomorphism The obstruction for that is the homomorphism into G. π1 (Θ) : π1 (Diff(M )) → π1 (G), but this homomorphism does not seem to be very accessible. In all cases where we have a compact subgroup C ⊆ Diff(M ) for which π1 (C) → π1 (Diff(M )) is surjective (this is in particular the case for dim M ≤ 2; [EE69]), we may choose the trivialization τ in a C-invariant fashion, so that Θ(C) is trivial. This clearly implies that π1 (Θ) vanishes. We now take a closer look at some examples of extensions of Diff(M )0 defined by the two cocycles discussed above. Example 3.16. (cf. [Bi03]) (The Virasoro group) For M = S1 , the preceding constructions can be used in particular to obtain the Virasoro group, resp., a corresponding global smooth 2-cocycle. The simply connected cover of C ∞ (S1 , GL1 (R))0 is the group N := ∞ 1 C (S , R) on which we have a smooth action of G := Diff(S1 )op 0 by ϕ.f := f ◦ ϕ. In the notation of the preceding example, we have κ := κ1 = κ2 , and, identifying V(S1 ) with C ∞ (S1 , R), we obtain the 1-cocycle θ(X) = X . The corresponding 1-cocycle on G with values in N is given by Θ(ϕ) = log(ϕ ). 1 The cocycle ω := 2ωκ on V(S1 ) with values in Ω (S1 , R) ∼ = R is 1 ω(ξ1 , ξ2 ) = ξ1 ξ2 − ξ2 ξ1 dt = 2 ξ1 dξ2 , 0

S1

and 12 ω = ωκ is a corresponding group cocycle, defining the central extension := R ×ω N on which G acts by ϕ.(z, f ) := (z, ϕ.f ), so that the semidirect N G is a central extension of N G (cf. [Ne02]). product N : G → N G. Since we have in We now pull this extension back with Θ G the relation N

(0, Θ(ϕ), ϕ)(0, Θ(ψ), ψ) = ωκ (Θ(ϕ), ϕ.Θ(ψ)), Θ(ϕ)ϕ.Θ(ψ), ϕψ , this leads to the 2-cocycle


309

Ω(ϕ, ψ) = ωκ (Θ(ϕ), ϕ.Θ(ψ)) = ωκ (Θ(ϕ), Θ(ϕψ) − Θ(ϕ)) log(ψ ◦ ϕ) d(log ϕ ). = ωκ (Θ(ϕ), Θ(ϕψ)) = −ωκ (log(ψ ◦ ϕ) , log ϕ ) = − S1

This is the famous Bott–Thurston cocycle for Diff(S1 )op 0 , and the corresponding central extension is the Virasoro group. On the Lie algebra level, the pull-back cocycle is (θ∗ ω)(ξ1 , ξ2 ) = ω(ξ1 , ξ2 ) =

1

ξ1 ξ2 − ξ2 ξ1 dt

0

(cf. [FF01]). Example 3.17. ([Bi03]) The preceding example easily generalizes to tori M = Td . Then we can trivialize T M with the Maurer–Cartan form τ of the group Td , which leads after identification of V(M ) with C ∞ (M, Rd ) to θ(X) = −X (the Jacobian matrix of X). A corresponding crossed homomorphism is Θ : Diff(Td ) → C ∞ (Td , GLd (R)),

Θ(ϕ) := (ϕ−1 ) ,

where the derivative ϕ of an element of Diff(Td ) ⊆ C ∞ (Td , Td ) is considered as a GLd (R)-valued smooth function on Td . The main difference from the one-dimensional case is that here we do not have an explicit formula for the action of Diff(Td ) on the central extension of the mapping group C ∞ (M, GLd (R))0 . As a consequence, we do not get explicit cocycles integrating Ψ 2 and Ψ 1 ∧ Ψ1 . On the integration of cocycles of the form ω [2] Now we turn to the other types of cocycles listed in Theorem 3.8(2), namely those of the form ω [2] , where ω is a closed V -valued 3-form on M . Let G be a Lie group and σ : M × G → M be a smooth right action on the (possibly infinite-dimensional) smooth manifold M . Let ω ∈ Ω p+q (M, V ) be a closed (p + q)-form with values in a Mackey complete space V . If L(σ) : g → V(M ) is the corresponding homomorphism of Lie algebras, then q

ωg := L(σ)∗ ω [p] ∈ Z p (g, Ω (M, V )) is a Lie algebra cocycle, and we have a well-defined period homomorphism q

perωg : πp (G) → Ω (M, V )G . If M is smoothly paracompact, which is in particular the case if M is finite q dimensional, then any class [α] ∈ Ω (M, V ) is determined by its integrals over smooth singular q-cycles S in M , which permits us to determine perωg geometrically.

310

Karl-Hermann Neeb

In the following we write for an oriented compact manifold F : = : Ω p+q (M × F, V ) → Ω q (M, V ) F

for the fiber integral (cf. [GHV72], Ch. VII). Theorem 3.18. (Period Formula) For any smooth singular q-cycle S on M and any smooth map γ : Sp → G, we have the following integral formula: perωg ([γ]) = (−1)pq ω, (14) S

S•γ

where S • γ ∈ Hp+q (M ) denotes the singular cycle obtained from the natural map Hq (M ) ⊗ Hp (G) → Hp+q (M ), [α] ⊗ [β] → [σ ◦ (α × β)] induced by the action map M × G → M . Proof. First, let γ : Sp → G be a smooth map and α : ∆q → M be a smooth singular simplex. We then obtain a smooth map α • γ := σ ◦ (α × γ) : ∆q × Sp → M,

(x, y) → α(x).γ(y).

Integration of ω yields with 7.12/14 in [GHV72] (on pull-backs and Fubini’s theorem): ω= (α • γ)∗ ω = (α × γ)∗ σ ∗ ω α•γ

= ∆q ×Sp

∆q ×Sp

∆q ×Sp

(α × idSp )∗ (idM ×γ)∗ σ ∗ ω =

∆q

= (α × idSp )∗ (idM ×γ)∗ σ ∗ ω Sp

α∗ = (idM ×γ)∗ σ ∗ ω = = (idM ×γ)∗ σ ∗ ω . = p ∆q S α Sp

∈Ω q (M,V )

It therefore remains to show that

perωg ([γ]) = (−1)pq = ((idM ×γ)∗ σ ∗ ω) .

(15)

Sp

So let Y1 , . . . , Yq ∈ V(M ) be smooth vector fields on M and write F jm : F → M × F, x → (m, x) for the inclusion map. We then have Sp ∗ = ((idM ×γ)∗ σ ∗ ω)(Y1 , . . . , Yq )m = (jm ) (iYq · · · iY1 ((idM ×γ)∗ σ ∗ ω)) Sp Sp Sp ∗ ∗ ∗ G ∗ = (jm ) (idM ×γ) (iYq · · · iY1 (σ ω)) = (jm ) (iYq · · · iY1 (σ ∗ ω)). Sp

γ

We further observe that for x1 , . . . , xp ∈ g = T1 (G) we have


311

(iYq · · · iY1 (σ ∗ ω))(m,g) (g.x1 , . . . , g.xp ) = (σ ∗ ω)(m,g) (Y1 (m), . . . , Yq (m), g.x1 , . . . , g.xp ) = ω(m.g) (Y1 (m).g, . . . , Yq (m).g, m.(g.x1 ), . . . , m.(g.xp )) = ω(m.g) (Y1 (m).g, . . . , Yq (m).g, L(σ)(x1 )(m.g), · · · , L(σ)(xp )(m.g)) = (−1)pq (σg∗ (iL(σ)(xp ) · · · iL(σ)(x1 ) ω))(Y1 (m), . . . , Yq (m)) = (−1)pq (g.ωg (x1 , . . . , xp ))(Y1 , . . . , Yq )m = (−1)pq ωgeq (g.x1 , . . . , g.xp )(Y1 , . . . , Yq )m . We thus obtain in Ω q (M, V ) the identity ∗ ∗ pq = (idM ×γ) σ ω = (−1) ωgeq = (−1)pq perωg ([γ]), Sp

γ

which in turn implies (15) and hence completes the proof.

p

Example 3.19. For q = 0 and a closed p-form ω ∈ Ω (M, V ), the preceding theorem implies in particular that perωg : πp (G) → C ∞ (M, V ) satisfies perωg ([γ])(m) := ω with (m • γ)(x) = m.γ(x). m•γ

As an important consequence of the period formula in the preceding theorem, we derive that the 2-cocycles ω [2] of integral (p + 2)-forms integrate to abelian extensions of a covering group of Diff(M )0 . This applies in particular to p = 1, which leads to the integrability of the cocycles occurring in Theorem 3.8(2). Corollary 3.20. Let ω ∈ Ω p+2 (M, V ) be closed with discrete period group Γω . Then there exists an abelian extension of the simply connected covering p group Diff(M )0 by the group Z := Ω (M, V )/ΓZ , where p ΓZ := [α] ∈ HdR (M, V ) : α ⊆ Γω ∼ = Hom(Hp (M ), Γω ), Zp (M)

p

whose Lie algebra cocycle is ω [2] ∈ Z 2 (V(M ), Ω (M, V )). Proof. Applying the period formula in Theorem 3.18 to ωg = ω [2] and G = Diff(M )op , which acts from the right on M , we see that the image of the period map is contained in the group ΓZ , which is discrete. Now Theorem 7.2 of Appendix A applies.

It is a clear disadvantage of the preceding theorem that it only says something about the simply connected covering group, and since π1 (Diff(M )) is not very accessible, it would be nicer if we could say more about the flux homomorphism of ω [2] . Unfortunately, it does not vanish in general, so that we cannot expect an extension of the group Diff(M )0 itself. However, the following result is quite useful to evaluate the flux.

312

Karl-Hermann Neeb

Theorem 3.21. (Flux Formula) Let ω ∈ Ω p+2 (M, V ) be a closed (p+2)-form

1 and γ : S1 → Diff(M ) a smooth loop. Then ηγ := 0 γ(t)∗ iδr (γ)t ω dt is a closed (p + 1)-form and the flux of ω [2] can be calculated as p

Fω[2] ([γ]) = [ηγ[1] ] ∈ H 1 (V(M ), Ω (M, V )). Proof. We recall from Appendix A that Fω[2] ([γ]) = [Iγ ] = −[Iγ −1 ] for 1 Iγ (X) = γ(t).ω [2] (Ad(γ(t))−1 .X, δ l (γ)t ) dt. 0 p

Hence Iγ −1 (X) ∈ Ω (M, V ) is represented by the p-form Iγ −1 (X), defined on vector fields Y1 , . . . , Yp by Iγ −1 (X)(Y1 , . . . , Yp ) 1 γ(t)∗ ω(Ad(γ(t)).X, δ l (γ −1 )t , Ad(γ(t)).Y1 , . . . , Ad(γ(t)).Yp ) dt = 0

1

γ(t)∗

= 0

=− 0

1

iδr (γ)t ω (Ad(γ(t)).X, Ad(γ(t)).Y1 , . . . , Ad(γ(t)).Yp ) dt

γ(t)∗ iδr (γ)t ω (X, Y1 , . . . , Yp ) dt

= −ηγ (X, Y1 , . . . , Yp ) = −ηγ[1] (X)(Y1 , . . . , Yp .) That ηγ is closed follows directly from the fact that, for any smooth path γ : I → Diff(M ), we have 1 1 ∗ ∗ γ(1) ω − ω = γ(t) Lδr (γ)t ω dt = γ(t)∗ diδr (γ)t ω dt 0

=d 0

0

1

γ(t)∗ iδr (γ)t ω dt = dηγ .

In general it is not that easy to get good hold of the flux homomorphism, but there are cases where it factors through an evaluation homomorphism π1 (evm ) : π1 (Diff(M )) → π1 (M, m). According to Lemma 11.1 in [Ne04a], this is the case if ω is a volume form on M . At this point we have discussed the integrability of all the 2-cocycles of V(M ) that are relevant for understanding the twists of the fundamental central 1 extensions of the mapping algebras with values in Ω (M, V ). The twists relevant for cocycles with values in Ω 1 (M, V ) and C ∞ (M, V ) are easier to handle. First we recall that a product of two integrable Lie algebra 1-cocycles is integrable to the cup product of the corresponding group cocycles ([Ne04a], Lemma F.3), so that, combined with the integrability of the Ψk (Remark 3.6), the relevant information is contained in the following remark.


313

Remark 3.22. (a) If α ∈ Ω 1 (M, R) is a closed 1-form, then Example 3.19, applied to p = 1, provides a condition for the integrability of the 1-cocycle → M αg on V(M ) to a 1-cocycle on Diff(M ). Passing to a covering q : M ∗ of M for which q α is exact shows that there exists a covering of the whole diffeomorphism group to which α integrates as a 1-cocycle with values in C ∞ (M, R) (cf. Remark 1.3(c)). (b) If ω ∈ Ω 2 (M, R) is a closed 2-form, then Example 3.19, applied with p = 2, shows that if ω has a discrete period group Γω , then ωg has discrete periods, and hence integrates to a group cocycle on some covering group of Diff(M )0 . This can be made more explicit by the observation that in this case Z := R/Γω is a Lie group and there exists a Z-bundle q : P → M with a connection 1-form θ satisfying q ∗ ω = dθ. Then the group Aut(P ) is an abelian extension of the open subgroup Diff(M )P of all diffeomorphisms lifting to bundle automorphisms of P by C ∞ (M, Z) and ω is a corresponding Lie algebra cocycle (cf. [Ko70], [NV03]). For these cocycles the flux can also be made quite explicit (cf. [Ne04a], Prop. 9.11): For α ∈ C ∞ (S1 , M ) and γ ∈ C ∞ (S1 , Diff(M )0 ), we have Fω ([γ]) = ω, where (γ −1 • α)(t, s) := γ(t)−1 (α(s)). α

γ −1 •α

(c) If µ is a volume form on M , then the cocycle Ψ 1 satisfies LX µ = −Ψ 1 (X)µ ([BiNe08], Lemma III.3), so that Ψ 1 (X) = − div X. Therefore, ζ(ϕ), defined by ϕ.µ = ζ(ϕ)−1 µ, defines a C ∞ (M, R)-valued group cocycle integrating Ψ 1 . → M such If M is not orientable, then there is a 2-sheeted covering q : M that M is orientable. Accordingly, we have a 2-fold central extension Diff(M ) of Diff(M ) by Z/2 acting on M and the preceding construction provides a , R). If σ : M →M is the non-trivial deck transcocycle ζ : Diff(M ) → C ∞ (M and there exists a volume formation, then σ reverses the orientation of M ∗ ), all functions form µ on M with σ µ = −µ. Since σ commutes with Diff(M ∞ σ ∼ ∞ ζ(ϕ) are σ-invariant, and hence are in C (M , R) = C (M, R). We also note that σ ∗ µ = −µ implies that σ is not contained in the identity component of Diff(M ), so that Diff(M )0 ∼ = Diff(M )0 and Ψ 1 also integrates to a cocycle on Diff(M )0 . Toroidal groups and their generalizations 1

For the spaces z = Ω (M, V ), C ∞ (M, V ) and Ω 1 (M, V ), the cocycles ω of types (I)–(III) on g = C ∞ (M, k) are V(M )-invariant, and hence extend trivially to cocycles on h := g V(M ), and we have also seen how to classify the z-cocycles τ on V(M ) that occur as twists (cf. Theorems 3.8 and 8.5). We write ωτ for the corresponding twisted cocycle on g V(M ). Let G := C ∞ (M, K)0 . If ω is integrable to a Lie group extension, then for a corresponding extension of G by Z ∼ we write G = z/ΓZ . This extension

314

Karl-Hermann Neeb

is uniquely determined by [ω], which has the consequence that the natural which leads to a action of Diff(M ) on g integrates to a smooth action on G, Diff(M ). semidirect product G If p : Diff(M ) → Diff(M )0 is an abelian extension of Diff(M )0 by Z cor Diff(M ) is an extension of H := G Diff(M )0 by responding to τ , then G Z × Z, and the anti-diagonal ∆Z in Z is a normal split Lie subgroup, so that := (G Diff(M H ))/∆Z ∼ carries a natural Lie group structure, and it is now easy to verify that L(H) = z ⊕ωτ h. For the special case where M = Td is a torus, k is a simple complex algebra, κ is the Cartan–Killing form, V = C, ω = ωκ and τ is a linear combination whose Lie of the two cocycles Ψ 2 and Ψ 1 ∧ Ψ1 , we thus obtain Lie groups H algebras are Fréchet completions of the toroidal Lie algebras. These algebras (with the twists) have been introduced by Rao and Moody in [ERMo94], and since then their representation theory has been an active research area (cf. [Bi98], [Lar99], [Lar00], [ERC04], [FJ07]).

4 Central extensions of gauge groups In this section we shall see how, and under which circumstances, the fundamental types of cocycles on C ∞ (M, k) generalize to gauge algebras gau(P ). Therefore, it becomes a natural issue to understand the corresponding central extensions of the gauge group Gau(P ), resp., its identity component. Depending on the complexity of the bundle P , this leads to much deeper questions, most of which are still open. 4.1 Central extensions of gau(P ) In this section we describe natural analogs of the cocycles of types (I)–(III) for gauge Lie algebras of non-trivial bundles. Let q : P → M be a K-principal bundle and gau(P ) its gauge algebra, which we realize as gau(P ) ∼ = {f ∈ C ∞ (P, k) : (∀p ∈ P )(∀k ∈ K) f (p.k) = Ad(k)−1 .f (p)}, which is the space of smooth sections of the associated Lie algebra bundle := P/K0 denote the Ad(P ) = P ×Ad k (see Remark 1.3(b)). Further, let M corresponding squeezed bundle, which is a, not necessarily connected, covering of M and a π0 (K)-principal bundle. The squeezed bundle is associated to the → M by the homomorphism δ1 : π1 (M ) → π0 (K) from π1 (M )-bundle qM : M the long exact homotopy sequence of P . Sometimes it is convenient to reduce


315

matters to connected structure groups, which amounts to considering P as a of M . K0 -principal bundle over the covering space M Let ρ : K → GL(V ) be a homomorphism with K0 ⊆ ker ρ, so that V is a π0 (K)-module. Then the associated bundle V := P ×ρ V is a flat vector bundle. It is associated to the squeezed bundle P/K0 via the representation ρ : π0 (K) → GL(V ). We have a natural exterior derivative on 1 the space Ω • (M, V) of V-valued differential forms and we define Ω (M, V) := 1 Ω (M, V)/d(Γ V). If V is finite dimensional, then d(Γ V) is a closed subspace of the Fréchet space Ω 1 (M, V), so that the quotient inherits a natural Hausdorff locally convex topology (cf. [NeWo09]). The Lie algebra aut(P ) ⊆ V(P ) acts in a natural way on all spaces of sections of vector bundles associated to P , via their realization as smooth functions on P , and hence in particular on the spaces Ω p (M, V), which can be realized as V -valued differential forms on P . Since V is flat, the action on Ω p (M, V) factors through an action of the Lie algebra V(M ) ∼ = aut(P )/gau(P ). := P/K0 , the squeezed Remark 4.1. The passage from M to the covering M bundle associated to P , provides a simplification of our setting to the case where the structure group under consideration is connected if we consider P . One has to pay for this reduction by changing M . In as a K0 -bundle over M is non-compact. particular, if π0 (K) is infinite, the manifold M We now take a closer look at analogs of the cocycles of types (I)–(III) for gauge algebras g := gau(P ) of non-trivial bundles. Here the natural target 1 spaces are Γ V, Ω 1 (M, V) and Ω (M, V), replacing C ∞ (M, V ), Ω 1 (M, V ) and 1 Ω (M, V ). (I) First we choose a connection 1-form θ ∈ Ω 1 (P, k) and write ∇ for the corresponding covariant derivative, which induces covariant derivatives on all associated vector bundles such as Ad(P ). For the flat bundle V the corresponding covariant derivative coincides with the Lie derivative. Let κ : k × k → V be a K-invariant symmetric bilinear map. For ξ ∈ gau(P ), the 1-form ∇ξ ∈ Ω 1 (P, k) is K-equivariant and basic; hence it defines a bundle-valued 1-form in Ω 1 (M, Ad(P )). Now κ defines a 2-cocycle 1

ωκ ∈ Z 2 (gau(P ), Ω (M, V)), 1

ωκ (ξ1 , ξ2 ) := [κ(ξ1 , ∇ξ2 )]

with values in the trivial module Ω (M, V) (cf. [NeWo09]; and [LMNS98] for bundles with connected K). (II) Let K be a connected Lie group with Lie algebra k and η ∈ Z 2 (k, V ) be a 2-cocycle, defining a central Lie algebra extension qk : k → k of k by V of K by some quotient which integrates to a central Lie group extension K Z := V /ΓZ of V by a discrete subgroup. Then K acts trivially on V , so that the bundle V is trivial.

316

Karl-Hermann Neeb

) with The K action on k defines an associated Lie algebra bundle Ad(P fiber k which is a central extension of the Lie algebra bundle Ad(P ) by the ) of trivial bundle M × V (cf. [Ma05]). Now the Lie algebra g au(P ) := Γ Ad(P smooth sections of this bundle is a central extension of gau(P ) by the space C ∞ (M, V ) ∼ = Γ V. Any splitting of the short exact sequence of vector bundles ) → Ad(P ) → 0 0 → M × V → Ad(P leads to a description of the central extension g au(P ) by a 2-cocycle ω ∈ Z 2 (gau(P ), C ∞ (M, V )) which is C ∞ (M, R)-bilinear, and hence can be identified with a section of the associated bundle with fiber Z 2 (k, V ). (III) To see the natural analogs of cocycles of type (III), we first note that the natural analogs of the cocycles ω0,η = d ◦ ωη are cocycles of the form d◦ω ∈ Z 2 (gau(P ), Ω 1 (M, V )), where ω ∈ Z 2 (gau(P ), C ∞ (M, V )) corresponds ) of the Lie algebra bundle Ad(P ) by the trivial to a central extension Ad(P bundle M × V . Let κ be an invariant V -valued symmetric bilinear form κ and θ a principal connection 1-form on P with associated covariant derivative ∇. We are looking for a C ∞ (M, R)-bilinear alternating map η ∈ C 2 (gau(P ), C ∞ (M, V )) for which ωκ,∇,η(ξ1 , ξ2 ) := κ(ξ1 , ∇ξ2 ) − κ(ξ2 , ∇ξ1 ) − d( η (ξ1 , ξ2 )) defines a Lie algebra cocycle with values in Ω 1 (M, V ). For ω κ (ξ1 , ξ2 ) := κ(ξ1 , ∇ξ2 ) − κ(ξ2 , ∇ξ1 ) we obtain as in Section 3: dg ( ωκ )(ξ1 , ξ2 , ξ3 ) = d(Γ (κ)(ξ1 , ξ2 , ξ3 )). Here we use the fact that, in the realization of g = gau(P ) in C ∞ (P, k), we have ∇ξ = dξ + [θ, ξ], so that ω κ (ξ1 , ξ2 ) = κ(ξ1 , dξ2 ) − κ(ξ2 , dξ1 ) + 2κ(θ, [ξ2 , ξ1 ]). Therefore, dg ωκ,∇,η = 0 is equivalent to Γ (κ) − dg η having values in constant functions, and since this map is C ∞ (M, R)-trilinear, we find the condition Γ (κ) = dg η

in

Z 3 (gau(P ), C ∞ (M, V )).

This means that, for each m ∈ M , we have fiberwise in local trivializations Γ (κ) = dk η(m). In particular, κ has to be exact. Lemma 4.2. If Γ (κ) is a coboundary, then there exists a C ∞ (M, R)-bilinear alternating map η ∈ C 2 (gau(P ), C ∞ (M, V )) for which ωκ,∇,η is a 2-cocycle.


317

Proof. The set C 2 (k, V )κ := {η ∈ C 2 (k, V ) : dk η = Γ (κ)} is an affine space on which K acts by affine map. Now the associated bundle A := P ×K C 2 (k, V )η has a smooth section η, and any such section defines a C ∞ (M, R)-bilinear element of C 2 (g, C ∞ (M, V )) with dg (d ◦ η) = d ◦ Γ (κ) = dg ( ωκ ).

In view of the preceding lemma, analogs of the type (III) cocycles on the Lie algebra gau(P ) with values in Ω 1 (M, V ) always exist if κ is exact. Problem 4.3. (Universal central extensions) Suppose that k is a semisimple Lie algebra and q : P → M a principal K-bundle. Find a universal central extension of gau(P ). If P is trivial, then we have a universal central extension 1 by Ω (M, V (k)), given by the universal invariant symmetric bilinear form κu of k (see Remark 2.2(d)). The construction described above yields a central 1 extension of gau(P ) by Ω (M, V) for the associated bundle with fiber V (k), but it is not clear that this extension is universal. A class of gauge algebras closest to those of trivial bundles are those of flat bundles which can be trivialized by a finite covering of M . It seems quite probable that the above central extension is universal. The analogous result for multiloop algebras (see Section 5) has recently been obtained by E. Neher ([Neh07], Thm. 2.13; cf. also [PPS07], p. 147). 4.2 Covariance of the Lie algebra cocycles Remark 4.4. (Covariance for type (I)) (a) Realizing g = gau(P ) in C ∞ (P, k), we have ∇ξ = dξ + [θ, ξ],so that ωκ (ξ1 , ξ2 ) = [κ(ξ1 , ∇ξ2 )] = [κ(ξ1 , dξ2 ) + κ(θ, [ξ2 , ξ1 ])]. If X = Xξ ∈ aut(P ) corresponds to ξ ∈ gau(P ) in the sense that θ(Xξ ) = −ξ, then LXξ ξ = −[θ(Xξ ), ξ ] = [ξ, ξ ], and the fact that the curvature dθ + 12 [θ, θ] is a basic 2-form leads to 1 LXξ θ = d(iXξ θ) + iXξ dθ = −dξ − iXξ [θ, θ] = −dξ − [θ(Xξ ), θ] 2 = −dξ + [ξ, θ] = −∇ξ, so that ωκ can also be written as ωκ (ξ1 , ξ2 ) = −[κ(ξ1 , LXξ2 θ)]. The group Aut(P ) acts on the affine space A(P ) ⊆ Ω 1 (P, k) of principal connection 1-forms by ϕ.θ := (ϕ−1 )∗ θ and on g = gau(P ) by ϕ.ξ = ξ ◦ ϕ−1 . We then have ϕ.∇ξ = ϕ.(dξ + [θ, ξ]) = d(ϕ.ξ) + [ϕ.θ, ϕ.ξ]) = ∇(ϕ.ξ) + [ϕ.θ − θ, ϕ.ξ].

318

Karl-Hermann Neeb

This leads to (ϕ.ωκ )(ξ1 , ξ2 ) = ϕ.ωκ (ϕ−1 .ξ1 , ϕ−1 .ξ2 ) = ωκ (ξ1 , ξ2 ) + [κ(ϕ.θ − θ, [ξ2 , ξ1 ])]. Note that ζ : Aut(P ) → Ω 1 (M, Ad(P )),

ϕ → ϕ.θ − θ

is a smooth 1-cocycle, so that 1

Ψ : Aut(P ) → Hom(g, Ω (M, V)),

Ψ (ϕ)(ξ) := [κ(ϕ.θ − θ, ξ)]

is a 1-cocycle with dg (Ψ (ϕ)) = ϕ.ωκ − ωκ . Therefore, Lemma 7.6 implies that ϕ.([α], ξ) := ([ϕ.α] + [κ(ϕ.θ − θ, ϕ.ξ)], ϕ.ξ) defines an automorphism of the central extension g, and it is easy to see that we thus obtain a smooth action of Aut(P ) on g (cf. [NeWo09]). If ϕf ∈ Gau(P ) is a gauge transformation corresponding to the smooth function f : P → K, then ϕ∗f θ = δ l (f ) + Ad(f )−1 .θ implies ϕf .θ = δ l (f −1 ) + Ad(f ).θ = −δ r (f −1 ) + Ad(f ).θ. (b) If P = M × K is the trivial bundle and pM : P → M and pK : P → K are the two projection maps, then the identification of gauge and mapping groups can be written as C ∞ (M, K) → Gau(P ) ⊆ C ∞ (P, K),

∗ f → f := p−1 K · pM (f ) · pK ,

and the canonical connection 1-form is θ = δ l (pK ) ∈ Ω 1 (P, k). We further write α := Ad(pK )−1 .p∗M α for the realization of α ∈ Ω 1 (M, k) as an element 1 of Ω (M, Ad(P )) ⊆ Ω 1 (P, k). If ϕ ∈ Aut(P ) is the lift of ϕ ∈ Diff(M ) defined by ϕ(m, k) := (ϕ(m), k), then pK ◦ ϕ = pK implies that ϕ.θ = θ, so that ζ(ϕ) = 0. To determine ζ on the gauge group, we calculate ϕf.θ = −δ r (f) + Ad(f).θ = −δ r (f) − Ad(f).δ r (p−1 K )

r −1 ∗ r −1 −1 r ∗ .δ (pM (f )) = −δ r (f · p−1 K ) = −δ (pK · pM (f )) = −δ (pK ) − Ad(pK )

= δ l (pK ) − Ad(pK )−1 .(p∗M δ r (f )) = θ − δ r (f ). In this sense, the restriction of the cocycle ζ to Gau(P ) is ζ(ϕf) = −δ r (f ), so that it corresponds to the 1-cocycles defined by the right logarithmic derivative δ r : C ∞ (M, K) → Ω 1 (M, k). We therefore recover the formulas for the actions of C ∞ (M, K) Diff(M ) on g from Remarks 2.4 and 2.5. For later reference, we also note that the action of the gauge group on 1-forms is given in these terms by


319

∗ ϕf. α = (ϕ−1 K ).(pM α) )∗ α = (ϕ−1 )∗ Ad(p−1 f

f

∗ ∗ )∗ α = Ad(p−1 = Ad(pK · f−1 )−1 .(pM ◦ ϕ−1 K · pM f )pM α f ∗ α. = Ad(p−1 K ).pM (Ad(f ).α) = Ad(f ).α = Ad(f ).

Remark 4.5. (Covariance for type (II)) In the type (II) situation we have con ), and since Ad(P ) structed the Lie algebra extension g = g au(P ) as Γ Ad(P is a Lie algebra bundle associated to P via an action of K on k by automorphisms, the group Aut(P ) and its Lie algebra aut(P ) act naturally on g au(P ) by automorphisms, resp., derivations, lifting the action on gau(P ). Problem 4.6. (Covariance for type (III)) It would be interesting to see if the type (III) cocycles on gau(P ) with values in Ω 1 (M, V ) are aut(P )-covariant in the sense that aut(P ) acts on g au(P ). Problem 4.7. (Corresponding crossed modules) For cocycles of type (I) and (II) we have seen that the action of aut(P ) on gau(P ) lifts to an action by derivations on the central extension g au(P ), which defines a crossed module µ : g au(P ) → aut(P ) of Lie algebras. For type (I), the characteristic class of this crossed module is an element 1 of H 3 (V(M ), Ω (M, V )), and for type (II), in H 3 (V(M ), C ∞ (M, V )). The vanishing of these characteristic classes is equivalent to the embeddability of the central extension g au(P ) to an abelian extension of aut(P ) (cf. [Ne06b], Thm. III.5). For type (II) it has been shown in [Ne06b], Lemma 6.2, that the characteristic class can be represented by a closed 3-form, so that it comes from an ele3 ment of HdR (M, V ). It seems that a proper understanding of the corresponding situation for type (I) requires a description of the third cohomology space 1 H 3 (V(M ), Ω (M, R)), so that we need a generalization of Theorem 3.8 for cohomology in degree 3. For type (III) cocycles, one similarly needs the cohomology spaces H 3 (V(M ), Ω 1 (M, R)), but they can be described with Tsujishita’s work ([Tsu81], Thm. 5.1.6; [BiNe08], Thm. III.1), which asserts that the space H • (V(M ), Ω • (M, R)) is a free module of the algebra H • (V(M ), C ∞ (M, R)), generated by the non-zero products of the cohomology classes [Ψk ]. This leads to H 3 (V(M ), Ω 1 (M, R)) ∼ = H 2 (V(M ), C ∞ (M, R)) · [Ψ1 ]. If T M is trivial, we obtain with Theorem 3.8 1 2 (M, R) · [Ψ1 ] ⊕ HdR (M, R) · [Ψ 1 ∧ Ψ1 ]. H 3 (V(M ), Ω 1 (M, R)) = HdR

320

Karl-Hermann Neeb

4.3 Corresponding Lie group extensions 1

(I) For cocycles of the form ωκ ∈ Z 2 (gau(P ), Ω (M, V)), it seems quite difficult to decide their integrability. However, if π0 (K) is finite, we have the following generalization of Theorem 2.9 ([NeWo09]). Theorem 4.8. If π0 (K) is finite and G := Gau(P )0 , then the following are equivalent. (1) ωκ integrates for each principal K-bundle P over a compact manifold M to a Lie group extension of G. (2) ωκ integrates for the trivial K-bundle P = S1 × K over M = S1 to a Lie group extension of G. (3) The image of perκ : π3 (K) → V is discrete. These conditions are satisfied if κ is the universal invariant bilinear form with values in V (k). As we have already seen for trivial bundles, the cocycle d ◦ ωκ with values in Ω 2 (M, V) can be integrated quite explicitly. We now explain the geometric background for that in the general case. We start with a general group-theoretic remark. Remark 4.9. Let E and F be vector spaces and ω : E × E → F an alternating bilinear form. (a) We write H := H(E, F, ω) for the corresponding Heisenberg group, which is the product set F × E, endowed with the product (f1 , e1 )(f2 , e2 ) = (f1 + f2 + ω(e1 , e2 ), e1 + e2 ). The invariance group of ω is Sp(E, F, ω) := {(ϕ, ψ) ∈ GL(E) × GL(F ) : ϕ∗ ω = ψ ◦ ω}, and the natural diagonal action of this group on F × E preserves ω, so that we may form the semidirect product group HSp(E, F, ω) := H(E, F, ω) Sp(E, F, ω) with the product (f1 , e1 , ϕ1 , ψ1 )(f2 , e2 , ϕ2 , ψ2 ) = (f1 +ψ1 .f2 +ω(e1 , ϕ1 .e2 ), e1 +ϕ1 .e2 ,ϕ1 ϕ2 ,ψ1 ψ2 ). This group is an abelian extension of the group E Sp(E, F, ω) ⊆ Aff(E) × GL(F ) by F , and the corresponding cocycle is


321

Ω((e1 , ϕ1 , ψ1 ), (e2 , ϕ2 , ψ2 )) = ω(e1 , ϕ1 .e2 ). (b) Now let G be a group, and let ρE : G → Aff(E) and ρF : G → GL(F ) be group homomorphisms such that ρ := (ρE , ρF ) maps G into E Sp(E, F, ω). := ρ∗ HSp(E, F, ω) is an abelian extension of G by V . To Then the pull-back G make the cocycle ωG of this extension explicit, we write ρE (g) = (µ(g), ϕ(g)) and obtain ωG (g1 , g2 ) = ω(µ(g1 ), ρ(g1 ).µ(g2 )). This 2-cocycle ωG ∈ Z 2 (G, F ) is the cup product ωG = µ ∪ω µ of the 1-cocycle µ ∈ Z 1 (G, E) with itself with respect to the equivariant bilinear map ω (cf. [Ne04a], App. F). If we think of E as an abstract affine space without specifying an origin, then any point o ∈ E leads to a 1-cocycle µ(g) := g.o − o with values in the translation group (E, +) and, accordingly, we get ωG (g1 , g2 ) = ω(g1 .o − o, ρ(g1 )(g2 .o − o)). Remark 4.10. On the translation space Ω 1 (M, Ad(P )) of the affine space A(P ) of principal connection 1-forms, any V -valued K-invariant symmetric bilinear map κ on k defines an alternating C ∞ (M, R)-bilinear map Ω 1 (M, Ad(P )) × Ω 1 (M, Ad(P )) → Ω 2 (M, V),

(α, β) → α ∧κ β,

so that A(P ) is an affine space with a translation-invariant Ω 2 (M, V)-valued 2-form that is invariant under the affine action of Aut(P ) by pull-backs. Proposition 4.11. The smooth 2-cocycle ζκ ∈ Z 2 (Aut(P ), Ω 2 (M, V)),

ζκ (ϕ1 , ϕ2 ) = (ϕ1 .θ − θ) ∧κ ϕ1 .(ϕ2 .θ − θ)

defines an abelian extension of Aut(P ) by Ω 2 (M, V) whose corresponding Lie algebra cocycle is given by (X1 , X2 ) → (LX1 θ) ∧κ (LX2 θ) − (LX2 θ) ∧κ (LX1 θ) = 2(LX1 θ) ∧κ (LX2 θ). Remark 4.12. If P = M × K is a trivial bundle, then we obtain on Gau(P ) ∼ = C ∞ (M, K) with Remark 4.4(b): g ) = δ l (f) ∧κ δ r ( g ) = δ l (f ) ∧κ δ r (g) . ζκ (ϕf, ϕg ) = δ r (f) ∧κ Ad(f).δ r ( Problem 4.13. From the existence of the cocycle ζκ , it follows that d ◦ c = 0 holds for the obstruction class c of the corresponding crossed module (see Problem 4.7). Therefore, the long exact cohomology sequence associated to the short exact sequence 1

1 2 0 → HdR (M, V) → Ω (M, V) → BdR (M, V) → 0

322

Karl-Hermann Neeb

1 implies that c lies in the image of the space H 3 (V(M ), HdR (M, V)). As the 1 action of V(M ) on HdR (M, V) is trivial, we have to consider H 3 (V(M ), R). This space vanishes for dim M ≥ 3 (cf. [BiNe08], Cor. D.6, resp., [Hae76], Thm. 3/4), so that in this case the central extension g au(P ) corresponding to ωκ embeds into an abelian extension of aut(P ). It is an interesting problem to describe these extensions more explicitly. Are there any two-dimensional manifolds for which the obstruction class in c is non-zero?

(II) Now we turn to cocycles of type (II). As before, we fix a central of K by Z ∼ Lie group extension K = V /ΓZ . Then there is an obstruction 2 ˇ class χK ∈ H (M, Z) with values in the sheaf of germs of smooth Z-valued functions that vanishes if and only if there exists a K-principal bundle P with ∼ P /Z = P , as a K-bundle ([GP78]). Using the isomorphisms 3 ˇ 2 (M, Z) ∼ ˇ 3 (M, ΓZ ) ∼ H (M, ΓZ ), =H = Hsing 3 (M, V ), and it has been shown we see that χ maps to some class χ ∈ HdR in [LW08] that χ yields (up to sign) the characteristic class χ(µ) of the Lie algebra crossed module µ : g au(P ) → aut(P ). (see Theorem 3.2). On the level of de Rham classes, the following theorem can also be found in [LW08]; the present version is a refinement.

→ K a central Theorem 4.14. Let K be a connected Lie group and qK : K Lie group extension by Z with Z0 ∼ and q : P → M a principal KV /Γ = Z bundle. Then the following assertions hold. (a) If Z is connected and π2 (K) vanishes, which is the case if dim K < ∞, then there is a K-bundle P with P /Z ∼ = P and we obtain an abelian Lie group extension 1 → C ∞ (M, Z) → Aut(P) → Aut(P ) → 1 containing the central extension 1 → C ∞ (M, Z) → Gau(P) → Gau(P ) → 1 integrating the Lie algebra g au(P ). 3 (b) If dim K < ∞, then the obstruction class in HdR (M, V ) vanishes. (c) If K is 1-connected and K is connected, then Z is connected. ´ Cartan Proof. (a) That π2 (K) vanishes if K is finite dimensional is due to E. ([CaE36]). If π2 (K) vanishes and Z is connected, then Theorem 7.12 of [Ne02] of K by Z ∼ implies that the central extension K = V /ΓZ can be lifted to a ∼ K /Γ central extension K of K by V with K = Z . This implies that the which obstruction class to lifting the structure group of the bundle P to K, 2 ˇ is an element of the sheaf cohomology group H (M, Z), is the image of an ˇ 2 (M, V ) = {0}, and hence trivial. Here we have used the fact that element of H


323

ˇ p (M, V ), p > 0, vanish because the sheaf of the sheaf cohomology groups H germs of smooth maps with values in the vector space V is soft (cf. [God73]). (b) Since Z0 is divisible, we have a direct product decomposition Z ∼ = is a Baer sum of an Z0 × π0 (Z) and, accordingly, the central extension K extension by Z0 and an extension by π0 (Z). According to (a), the extension by Z0 does not contribute to the obstruction class, so that it is an element of H 2 (M, π0 (Z)), and therefore the corresponding de Rham class in H 3 (M, V ) vanishes. 0 is a connected covering group of K, and hence trivial, (c) The group K/Z and therefore Z = Z0 .

is a central T-extension of some infinite-dimensional group Remark 4.15. If K 3 for which the corresponding de Rham obstruction class in HdR (M, R) is non-zero, Theorem 3.2 implies that the corresponding class in H 3 (V(M ), C ∞ (M, R)) is also non-zero because the degrees on the Pontrjagin classes are multiples of 4, so that the ideal they generate vanishes in degrees ≤ 3. This means that the class of the corresponding crossed module µ : g au(P ) → aut(P ) is non-trivial, and hence the central extension g au(P ) does not embed equivariantly into an abelian extension of aut(P ). In a similar fashion as the group H 2 (M, Z) classifies T-bundles over M , the group H 3 (M, Z) classifies certain geometric objects called bundle gerbes ([Bry93], [Mu96]). As the preceding discussion has shown, non-trivial bundle gerbes can only be associated to the central T-extensions of infinitedimensional Lie groups. For more on bundle gerbes and compact groups, we refer to [SW07] in this volume.

5 Multiloop algebras In this section we briefly discuss the connection between gauge algebras of flat bundles and the algebraic concept of a (multi-)loop algebra. These are infinite-dimensional Lie algebras which are presently under active investigation from the algebraic point of view ([ABP06], [ABFP08]). To simplify matters, we discuss only the case where k is a complex Lie algebra. 5.1 The algebraic picture Let k be a complex Lie algebra, m1 , . . . , mr ∈ N, ζi ∈ T with ζimi = 1 and σ1 , . . . , σr ∈ Aut(k) with σimi = idk . Then we obtain an action of the group ∆ := Z/m1 × · · · × Z/mr on the Lie algebra −1 k ⊗ C[t1 , t−1 1 , . . . , tr , tr ]

by letting the i-th generator σ i of ∆ act by σ i (x ⊗ tα ) := σi (x) ⊗ ζiαi tα .

324

Karl-Hermann Neeb

The algebra −1 ∆ M (k, σ1 , . . . , σr ) := (k ⊗ C[t1 , t−1 1 , . . . , tr , tr ])

of fixed points of this action is called the corresponding multiloop algebra. For r = 1 we also write L(k, σ) := M (k, σ) and call it a loop algebra. The loop algebra construction can be iterated: If we start with σ1 ∈ Aut(k) with σ1m1 = id and pick σ2 ∈ Aut(L(k, σ1 )) with σ2m2 = id, we put L(k, σ1 , σ2 ) := L(L(k, σ1 ), σ2 ). Repeating this process, we obtain the iterated loop algebras L(k, σ1 , . . . , σr ) := L(L(k, σ1 , . . . , σr−1 ), σr ), m

where σj ∈ Aut(L(k, σ1 , . . . , σj−1 )) is assumed to satisfy σj j = id. Clearly, every multiloop algebra can also be described as an iterated loop algebra, but the converse is not true (see Example 5.3). 5.2 Geometric realization of multiloop algebras There is a natural analytic variant of the loop algebra construction. Here the main idea is that the analytic counterpart of the algebra C[t, t−1 ] of Laurent polynomials is the Fréchet algebra C ∞ (S1 , C) of smooth complex-valued functions on the circle. Its dense subalgebra of trigonometric polynomials is isomorphic to the algebra of Laurent polynomials, so that it is a completion of this algebra. A drawback of this completion process is that the algebra of smooth functions has zero divisors, but this in turn facilitates localization arguments. −1 The analytic analogue of the algebra k ⊗ C[t1 , t−1 1 , . . . , tr , tr ] is the ∞ r Fréchet–Lie algebra C (T , k) of smooth k-valued functions on the r-dimensional torus Tr , endowed with the pointwise Lie bracket. To describe the analytic analogs of the multiloop algebras, let ∆ := Z/m1 × · · · × Z/mr and fix a homomorphism ρ : ∆ → Aut(k), which is determined by a choice of automorphisms σi of k with σimi = idk . Let m := (m1 , . . . , mr ) and consider the corresponding torus Trm := (R/m1 Z) × · · · × (R/mr Z), on which the group ∆ ∼ = Zr /(m1 Z ⊕ . . . ⊕ mr Z) acts in the natural fashion from the right by factorization of the translation action of Zr on Rr . Now we obtain an action of ∆ on the Lie algebra C ∞ (Trm , k) by (γ.f )(x) := γ.f (x.γ) Then we have a dense embedding M (k, σ1 , . . . , σr ) → C ∞ (Trm , k)∆ . To realize this Lie algebra geometrically, we note that the quotient map q : Trm → Tr is a regular covering for which ∆ acts as the group of deck transformations on


325

Trm . Then the homomorphism ρ : ∆ → Aut(k) from above defines a flat Lie algebra bundle K := (Trm × k)/∆, [t, x] → q(t), over Tr , where ∆ acts by γ.(t, x) := (t.γ −1 , γ.x), and the space Γ K of smooth sections of this Lie algebra bundle can be realized as {f ∈ C ∞ (Trm , k) : (∀γ ∈ ∆)(∀x ∈ Trm ) f (x.γ) = γ −1 .f (x)} = C ∞ (Trm , k)∆ . In this sense, the Lie algebra Γ K is a natural geometric analogue of a multiloop algebra. 5.3 A generalization of multiloop algebras Our geometric realization of multiloop algebras suggests the following geo → M be a regular metric generalization. Let ∆ be a discrete group and q : M as the group of covering, where ∆ acts from the right via (t, γ) → t.γ on M deck transformations. Let k be a locally convex Lie algebra and ρ : ∆ → Aut(k) an action of ∆ by topological automorphisms of k. Let K0 be a 1-connected (regular or locally exponential) Lie group with Lie algebra k. Then the action of ∆ integrates to an action ρK : ∆ → Aut(K0 ) by Lie group automorphisms ([GN10]), so that we may form the semidirect product Lie group K := K0 ∆ with π0 (K) = ∆. Next we form the flat Lie algebra bundle × k)/∆, K := (M

qK : K → M, [t, x] := ∆.(t, x) → q(t),

over M , where ∆ acts on M × K by γ.(t, x) := (t.γ −1 , γ.x). Each section of → k, K can be written as s(q(t)) = [t, fs (t)] with a smooth function f : M −1 satisfying f (t.γ) = γ .f (t) for t ∈ M , γ ∈ ∆. We thus obtain a realization of the space Γ K of smooth sections of this Lie algebra bundle as , k) : (∀γ ∈ ∆) f (t.γ) = γ −1 .f (t)} = C ∞ (M , k)∆ , {f ∈ C ∞ (M

(16)

, k) is defined by (γ.f )(t) := γ.f (t.γ). where the Γ -action on C ∞ (M × K)/∆ is the flat K-bundle associated to the Proposition 5.1. If P := (M inclusion ∆ → K, then gau(P ) ∼ = Γ K. Hence the Lie algebra of sections of a flat Lie algebra bundle can always be realized as a gauge algebra of a flat principal bundle; the converse is trivial. Proof. Our assumptions on K0 imply that Aut(K0 ) ∼ = Aut(k) ([GN10]), so that the action of ∆ on k integrates to an action ρK on K0 by automorphisms. × K0 is a trivial K0 -bundle, but it is not trivial as a Clearly, P ∼ = M × K0 by K-bundle, because the group ∆ acts on M (m, k).γ = (m.γ, ρK (γ)−1 (k)).

326

Karl-Hermann Neeb

Let Ad(P ) = (P × k)/K denote the corresponding gauge bundle with Γ Ad(P ) ∼ = gau(P ). Then gau(P ) ∼ = {f ∈ C ∞ (P, k) : (∀p ∈ P )(∀k ∈ K) f (p.k) = Ad(k)−1 .f (p)} ∼ , k) : (∀m ∈ M )(∀γ ∈ ∆) f (m.γ) = Ad(γ)−1 .f (m)} = {f ∈ C ∞ (M , k) : (∀m ∈ M )(∀γ ∈ ∆) f (m.γ) = ρ(γ)−1 .f (m)}. = {f ∈ C ∞ (M

5.4 Connections to forms of Lie algebras over rings In this subsection we assume that ∆ is finite and briefly describe the connection to forms of Lie algebras over rings. For more details on this topic we refer to [PPS07]. The Lie algebra Γ K is a module of the algebra R := C ∞ (M, C), and the Lie bracket is R-bilinear. The pull-back bundle × K : q(t) = qK (x)} q ∗ K = {(t, x) ∈ M × k → q ∗ K, (t, y) → (t, [t, y]) is an isomorphism is trivial, because the map M of Lie algebra bundles. Accordingly, the space Γ (q ∗ K), which is a module of , C), is isomorphic to Γ (M × k) ∼ , k). On the level of S := C ∞ (M = C ∞ (M sections, we have an R-linear embedding q ∗ : Γ K → Γ (q ∗ K),

(q ∗ s)(t) := (t, s(q(t))),

which corresponds to the realization of Γ K as the subspace (16) of ∆-fixed , k). Using the local triviality of the bundle and the finiteness points in C ∞ (M of ∆, one easily verifies that the canonical map ∼ , k), Γ K ⊗R S → Γ K = C ∞ (M

s ⊗ f → (q ∗ s) · f

is a linear isomorphism (cf. Thm 7.1 in [ABP06]). In this sense, the R-Lie , k). algebra Γ K is an R-form of the S-Lie algebra C ∞ (M denote the completion of the algebraic tensor product Problem 5.2. Let ⊗ as a locally convex space. Determine to which extent the map , K) → Γ K ∼ , k), C ∞ (M,K) C ∞ (M Γ K⊗ = C ∞ (M

s ⊗ f → f · (q ∗ s)

is an isomorphism of Lie algebras if ∆ is not necessarily finite. In the proof of Theorem 7.1 in [ABP06], the finiteness of ∆ is used to conclude that its cohomology with values in any rational representation space is trivial. Example 5.3. (cf. Example 9.7 in [ABP06]) We consider k = sln (C) for n ≥ 2 and the iterated loop algebra L(k, σ1 , σ2 ) defined as follows. As σ1 ∈ Aut(k) we choose an involution not contained in the identity component of Aut(k),


327

so that L(k, σ1 ) is a non-trivial loop algebra corresponding to the affine (2) Kac–Moody algebra of type An−1 (cf. [Ka90], Ch. 8). As σ2 , we take the involution on L(k, σ1 ) = {f ∈ C ∞ (T, k) : f (−t) = σ1 (f (t))}, where T is realized as the unit circle in C× , defined by σ2 (f )(t1 ) := f (t−1 1 ) and put L := L(k, σ1 , σ2 ) = {f ∈ C ∞ (T, L(k, σ1 )) : : f (−t2 ) = σ2 (f (t2 ))}. We may thus identify elements of L with smooth functions f : T2 → k satisfying f (t1 , −t2 ) = f (t−1 1 , t2 )

and

f (−t1 , t2 ) = σ1 (f (t1 , t2 )).

For τ1 (t1 , t2 ) := (−t1 , t2 ) and τ2 (t1 , t2 ) := (t−1 1 , −t2 ) we thus obtain a fixed point free action of ∆ := τ1 , τ2 on T2 and a representation ρ : ∆ → Aut(k), defined by ρ(τ1 ) := σ1 and ρ(τ2 ) := id. Then L∼ = {f ∈ C ∞ (T2 , k) : (∀γ ∈ ∆)(∀t ∈ T2 ) f (t.γ) = γ −1 .f (t)}. This means that L ∼ = Γ K for the Lie algebra bundle K over M := T2 /∆, where M is the Klein bottle (τ2 is orientation reversing). This implies that the centroid of L is C ∞ (M, C) ([Lec80]), and hence not isomorphic to C ∞ (T2 , C), so that L cannot be a multiloop algebra.

6 Concluding remarks From the structure-theoretic perspective on infinite-dimensional Lie groups, it is natural to ask for a classification of gauge groups of principal bundles and associated structural data, such as their central extensions, invariant forms, etc. One readily observes that two equivalent K-principal bundles Pi → M have isomorphic gauge groups, so that a classification of gauge groups or automorphism groups of K-principal bundles can be achieved, in principle, by a classification of all K-principal bundles and then an identification of those whose gauge groups are isomorphic. A key observation is that if a bundle P is twisted by a Z(K)-bundle PZ , then the corresponding gauge groups are isomorphic Lie groups, where the isomorphism on the Lie algebra level is C ∞ (M, R)-linear. Taking this refined structure into account leads to the concept of (transitive) Lie algebroids and the perspective on Gau(P ) as the group of sections of a Lie group bundle (cf. [Ma89] for results in this direction

328

Karl-Hermann Neeb

and [Ma05] for more on Lie groupoids). Currently the theory of Lie groupoids is only available for finite-dimensional structure groups K. It would be natural to develop it also for well-behaved infinite-dimensional groups over finitedimensional manifolds. If K is finite dimensional with finitely many connected components, then the classification of K-principal bundles can be reduced to the corresponding problem for a maximal compact subgroup, so that the well-developed theory of bundles with compact structure group applies ([Hu94]). If K is an infinite-dimensional Lie group, it might not possess maximal compact subgroups,2 so that a similar reduction does not work. However, recent results of M¨ uller and Wockel show that each topological K-principal bundle over a finite-dimensional paracompact smooth manifold M is equivalent to a smooth one, and that two smooth K-principal bundles which are topologically equivalent are also smoothly equivalent ([MW09]). Therefore, the set Bun(M, K) of equivalence classes of smooth K-principal bundles over M can be identified with the set Bun(M, K)top of topological K-bundles over M , and the latter set can be identified with the set [M, BK] of homotopy classes of continuous maps of M into a classifying space BK of the topological group K ([Hu94], Thms. 9.9, 12.2, 12.4). On the topological level, each continuous map σ : M → BK defines an algebra homomorphism • H • (σ, R) : H • (BK, R) → H • (M, R) ∼ (M, R), = HdR

whose range is the characteristic classes of the corresponding bundle P . The advantage of these characteristic classes is that they can be represented by closed differential forms on M ; hence, they can be evaluated by integrals over compact submanifolds or other piecewise smooth singular cycles. These characteristic classes are quite well understood for finite-dimensional connected Lie groups K, because in this case essentially everything can be reduced to compact Lie groups, for which a powerful theory exists. However, for infinite-dimensional structure groups, the corresponding theory of characteristic classes has only been explored in very special cases. We refer to Morita’s excellent textbook for more details and references (cf. [Mo01]). Central extensions of gauge groups of bundles with infinite-dimensional fiber over the circle have recently been studied in the context of integrable systems in [OR07].

7 Appendix A. Abelian extensions of Lie groups In this appendix we result some facts on the integration of Lie algebra 2-cocycles from [Ne04a]. They provide a general set of tools to integrate abelian extensions of Lie algebras to extensions of connected Lie groups. 2

The representation theory of compact groups implies that the unitary group of an infinite-dimensional Hilbert space is such an example.


329

Let G be a connected Lie group and V a Mackey complete G-module. Further, let ω ∈ Z k (g, V ) be a k-cocycle and ω eq ∈ Ω k (G, V ) be the corresponding left equivariant V -valued k-form with ω1eq = ω. Then each continuous map Sk → G is homotopic to a smooth map, and ω eq = σ ∗ ω eq perω : πk (G) → V G , [σ] → σ

Sk

defines the period homomorphism whose values lie in the G-fixed part of V ([Ne02], Lemma 5.7). For k = 2 we define the flux homomorphism Fω : π1 (G) → H 1 (g, V ),

[γ] → [Iγ ],

where we define for each piecewise smooth loop γ : S1 → G the 1-cocycle 1 eq Iγ : g → V, Iγ (x) := ixr ω = γ(t).ω(Ad(γ(t))−1 .x, δ l (γ)t ) dt, γ

0

where xr is the right invariant vector field on x with xr (1) = x. If V is a trivial module, then dg V = {0}, so that H 1 (g, V ) consists of linear maps g → V , and we may think of the flux as a map Fω : π1 (G) × g → V. Proposition 7.1. ([Ne02], Prop. 7.6) If V is a trivial G-module, then the adjoint action of g on the central extension g := V ⊕ω g integrates to a smooth action Adg of G if and only if Fω = 0. Theorem 7.2. Let G be a connected Lie group, A a smooth G-module of the form A ∼ = a/ΓA , where ΓA ⊆ a is a discrete subgroup of the Mackey complete space a and qA : a → A the quotient map. Then the Lie algebra extension g := a ⊕ω g of g defined by the cocycle ω integrates to an abelian Lie group extension of G by A if and only if qA ◦ perω : π2 (G) → A

and

Fω : π1 (G) → H 1 (g, a)

vanish. Remark 7.3. If only qA ◦ perω vanishes, then the preceding theorem applies of G, so that we only obtain an to the simply connected covering group G extension of G by a non-connected group A , which itself is a central extension of the discrete group π1 (G) by A. For a more detailed discussion of these aspects, we refer to Section 7 of [Ne04a]. Remark 7.4. To calculate period and flux homomorphisms, it is often convenient to use related cocycles on different groups. So, let us consider a morphism ϕ : G1 → G2 of Lie groups and ωi ∈ Z 2 (gi , V ), V a trivial Gi -module, satisfying L(ϕ)∗ ω2 = ω1 . Then a straightforward argument shows that perω2 ◦π2 (ϕ) = perω1 : π2 (G1 ) → V.

(17)

330

Karl-Hermann Neeb

For the flux we likewise obtain Fω2 ◦ (π1 (ϕ) × L(ϕ)) = Fω1 : π1 (G1 ) × g1 → V,

(18)

if we consider the flux as a bihomomorphism π1 (G) × g → V . Problem 7.5. It is a crucial assumption in the preceding theorem that the group G is connected, and it would be very desirable to have a suitable generalization to non-connected groups G. The generalization to non-connected Lie groups G means deriving accessible criteria for the extendibility of a 2-cocycle, resp., abelian extensions, from the identity component G0 to the whole group G. From the short exact sequence G0 → G → → π0 (G), we obtain maps I

R

H 2 (π0 (G), A)−−→H 2 (G, A)−−→H 2 (G0 , A)G = H 2 (G0 , A)π0 (G) , but it seems to be difficult to describe the image of the restriction map R. If the identity component G0 of G splits, i.e., if G ∼ = G0 π0 (G), then it is easy to see that R is surjective, because the invariance of the cohomology 0 permits us to lift the conjugation action of π0 (G) to an action on class of G 0 π0 (G) is an abelian extension of G G0 , and then the semidirect product G 0 as a subgroup. by A, containing G If A is a trivial module, one possible approach is to introduce additional 0 of G0 by A, so that the map q : G 0 → G structures on a central extension G describes a crossed module of groups (cf. [Ne07]), which requires an exten 0 , resp., its Lie algebra sion of the natural G0 -action of G on G g, to an action of G. If we have such an action lifting the conjugation action of G 0 ) of this crossed module is an eleon G0 , then the characteristic class χ(G 3 ment of the cohomology group H (π0 (G), A), which vanishes if there exists an of G into which G 0 embeds in a G-equivariant fashion A-extension G (cf. [Ne07], Thm. III.8). The following lemma is easy to verify ([MN03], Lemma V.1; see also [Ne06b], Lemma II.5 for a generalization to general Lie algebra extensions). Lemma 7.6. Let V ⊕ω g be a central extension of the Lie algebra g and γ = (γV , γg ) ∈ GL(V ) × Aut(g). For θ ∈ C 1 (g, V ) the formula γ (z, x) := (γV (z) + θ(γg (x)), γg (x)),

x ∈ g, z ∈ V,

defines a continuous Lie algebra automorphism of g if and only if γ.ω − ω = dg θ,

resp.,

dg θ = ω − γ −1 .ω

for

θ := θ ◦ γg .

The following theorem can be found in [MN03], Thm. V.9. → G be a central Lie group Theorem 7.7. (Lifting Theorem) Let q : G extension of the 1-connected Lie group G by the Lie group Z ∼ = z/ΓZ . Let


331

σG : H × G → G, resp., σZ : H × Z → Z be smooth automorphic actions of the Lie group H on G, resp., Z and σg a smooth action of H on g compatible with →G the actions on z and g. Then there is a unique smooth action σG : H × G by automorphisms compatible with the actions on Z and G, for which the corresponding action on the Lie algebra g is σg .

8 Appendix B. Abelian extensions of semidirect sums We consider a semidirect product of topological Lie algebras h = n S g and a topological h-module V . We are interested in a description of the cohomology space H 2 (h, V ) in terms of H 2 (n, V ) and H 2 (g, V ). To this end, we have to study the inflation, resp., the restriction map I : H 2 (g, V n ) → H 2 (h, V ),

resp., Rn : H 2 (h, V ) → H 2 (n, V )g ,

satisfying Rn I = 0. Here we use the fact that h acts trivially on H 2 (h, V ) to see that the image of Rn is contained in H 2 (n, V )h = H 2 (n, V )g . We further have a restriction map Rg : H 2 (h, V ) → H 2 (g, V ). The composition Rg ◦I : H 2 (g, V n ) → H 2 (g, V ) is the natural map induced by the inclusion V n → V of g-modules. If we interpret the elements of H 2 (h, V ) as abelian extensions of h by V , then the inflation map leads to twistings of these extensions by extensions of g by V n . We now construct an exact sequence describing the kernel and cokernel of the map (Rn , Rg ) which provides a quite accessible description of H 2 (h, V ). Definition 8.1. (a) We write C p (n, V ) for the space of continuous Lie algebra p-cochains with values in V and C p (g, C q (n, V ))c ⊆ C p (g, C q (n, V )) for the subspace consisting of those cochains defining a continuous (p + q)-linear map gp × nq → V . Accordingly, we define Z p (g, C q (n, V ))c , H p (g, C q (n, V ))c , etc. (b) We further write H 2 (n, V )[g] ⊆ H 2 (n, V )g for the subspace of those cohomology classes [f ] for which there exists a θ ∈ C 1 (g, C 1 (n, V ))c with dn (θ(x)) = x.f

for x ∈ g.

(19)

Because of the continuity requirement for θ, this is stronger than the g-invariance of the cohomology class [f ] ∈ H 2 (n, V ).

332

Karl-Hermann Neeb

Remark 8.2. Any continuous g-module action by derivations on a central extension n = V ⊕f n, compatible with the actions on V and n is of the form x.(v, n) = (x.v + θ(x)(n), x.n), (20) where θ ∈ Z 1 (g, C 1 (n, V ))c satisfies (19). In particular, the existence of such an action implies that [f ] ∈ H 2 (n, V )[g] . Lemma 8.3. For [fn ] ∈ H 2 (n, V )[g] we choose θ ∈ C 1 (g, C 1 (n, V ))c as in (19). We thus obtain a well-defined linear map, γ : H 2 (n, V )[g] → H 2 (g, Z 1 (n, V ))c ,

[fn ] → [dg θ].

Proof. Using θ, we obtain a linear map ψ : g → der( n),

ψ(x)(v, n) = (x.v + θ(x)(n), x.n)

which is a linear lift of the given Lie algebra homomorphism ψ = (S, ρV ) : g → (der(n) × gl(V )). In view of [Ne06b], Prop. A.7, the corresponding Lie algebra cocycle on g is given by fg = dg θ ∈ Z 2 (g, Z 1 (n, V ))c . The corresponding extension g = Z 1 (n, V ) ×fg g → g → 0 0 → Z 1 (n, V ) → acts naturally on n by (α, x).(v, n) = (α(n) + x.v + θ(x)(n), x.n). Next we observe that γ is well defined. Indeed, if θ ∈ C 1 (g, C 1 (n, V ))c also satisfies (19), then θ − θ ∈ C 1 (g, Z 1 (n, V ))c , so that [dg (θ − θ)] = 0 in H 2 (g, Z 1 (n, V ))c . Moreover, if fn = fn + dn β for some β ∈ C 1 (n, V ), then x.fn = x.fn + dn (x.β) = dn (θ(x) + x.β), so that θ(x) := θ(x) + x.β leads to dn (θ(x)) = x.fn . Then dg θ = dg θ + d2g β = dg θ. This shows that γ is well defined. The linearity is clear.

Lemma 8.4. For each θ ∈ Z 1 (g, Z 1 (n, V ))c , x.(v, n) := (x.v + θ(x)(n), x.n) defines a continuous action of g on the semidirect sum n := V n, so that hθ := n g is an extension of h by V . We thus obtain a well-defined map, ϕ

ϕ : H 1 (g, Z 1 (n, V ))−−→H 2 (h, V ),

[θ] → [ hθ ].

Proof. A 2-cocycle defining the extension hθ is given by ωθ ((n1 , x1 ), (n2 , x2 )) = [(0, (n1 , x1 )), (0, (n2 , x2 ))] − (0, [n1 , n2 ] + x1 .n2 − x2 .n1 , [x1 , x2 ]) = (0, θ(x1 )(n2 ) − θ(x2 )(n1 )).


333

If θ is a coboundary, i.e., there exists a β ∈ Z 1 (n, V ), with θ(x) = x.β, x ∈ g, then θ(x)(n) = x.β(n) − β(x.n) leads to ωθ ((n1 , x1 ), (n2 , x2 )) = θ(x1 )(n2 ) − θ(x2 )(n1 ) = x1 .β(n2 ) − x2 .β(n1 ) + β(x1 .n2 − x2 .n1 ). If β : h → V denotes the continuous linear map extending β and vanishing on g, then ωθ is a coboundary: (dh β)((n 1 , x1 ), (n2 , x2 )) = x1 .β(n2 ) − x2 .β(n1 ) + dn (β)(n1 , n2 ) − β(x1 .n2 − x2 .n1 ) = x1 .β(n2 ) − x2 .β(n1 ) − β(x1 .n2 − x2 .n1 ) = ωθ ((n1 , x1 ), (n2 , x2 )).

Theorem 8.5. With the linear map η = (dn )∗ : H 2 (g, V ) → H 2 (g, Z 1 (n, V ))c ,

[fg ] → [−dn ◦ fg ],

the following sequence is exact: (Rn ,Rg )

ϕ

H 1 (g, Z 1 (n, V ))−−→H 2 (h, V )−−−−−−→H 2 (n, V )[g] ⊕ H 2 (g, V ) γ−η

−−−−→H 2 (g, Z 1 (n, V ))c . Proof. To see that the sequence is exact in H 2 (h, V ), we simply note that a V -extension corresponds to a class [f ] ∈ ker(Rn , Rg ) if and only if its restrictions to both subalgebras n and g are trivial. This is equivalent to h being of the form h = (V n)g, and hence equivalent to hθ for some θ ∈ Z 1 (g, Z 1 (n, V ))c . It remains to verify that im(Rn , Rg ) = ker(γ − η). If [fn ] ∈ im(Rn ), then there exists a topologically split extension h of h by the topological h-module V . Then h contains n as an ideal, so that the adjoint action defines an action of h on n. As V acts non-trivially on n, this action generally does not factor through an action of h. Let p : h → h denote the quotient map. Then n = p−1 (n), and we put g := p−1 (g). As we have seen in Remark 8.2, the action of g is described by some θ ∈ Z 1 ( g, C 1 (n, V ))c satisfying = x.f dn (θ(x))

for x ∈ g.

(21)

x) satisfies (19), so that Writing g as V ⊕fg g, it follows that θ(x) := θ(0, 2 [g] [fn ] ∈ H (n, V ) . For x, y ∈ g we then obtain the relation y) − y.θ(0, x) = θ([(0, x.θ(y) − y.θ(x) = x.θ(0, x), (0, y)]) = −dn (fg (x, y)) + θ([x, y]), i.e., dg θ = −dn ◦ fg . Hence, any element ([fn ], [fg ]) ∈ im(Rn , Rg ) is contained in the kernel of γ − η : H 2 (n, V )[g] ⊕ H 2 (g, V ) → H 2 (g, Z 1 (n, V ))c .

334

Karl-Hermann Neeb

If, conversely, [fn ] ∈ ker(γ − η), then there exists θ ∈ C 1 (g, C 1 (n, V ))c x) := θ(x) − dn v defines a representation with dg θ = −dn ◦ fg , and then θ(v, g := V ⊕fg g → der n, compatible with the actions of g on n and V . Then the semidirect product n g is an extension of n g = h by the space V × V . Since both V -factors act in the same way on n, the anti-diagonal ∆V acts trivially, and hence commutes with n. Since g, resp., g, acts diagonally on V × V , it also preserves the anti-diagonal. We thus obtain the V -extension h := ( n g)/∆V of h. This extension contains n and g as subalgebras. This completes the proof.

Corollary 8.6. If n is topologically perfect and V = V n , then (Rn , Rg ) : H 2 (h, V ) → H 2 (n, V )[g] ⊕ H 2 (g, V ) is a linear bijection. Proof. In this case Z 1 (n, V ) vanishes, and the assertion follows from Theorem 8.5.

9 Appendix C. Triviality of the group action on Lie algebra cohomology Theorem 9.1. Let G be a connected Lie group with Lie algebra g and V a Mackey complete smooth G-module. Then the natural action of G on the Lie algebra cohomology H • (g, V ) is trivial. Proof. Since the space H • (g, V ) carries no natural locally convex topology for which the action of G on this space is smooth, we cannot simply argue that the triviality of the G-action follows from the triviality of the g-action. The assertion is clear for p = 0, because H 0 (g, V ) = V g = V G follows from the triviality of the g-action on the closed G-invariant subspace V g of V ([Ne06a], Rem. II.3.7). We may therefore assume p > 0. Let ω ∈ Z p (g, V ) be a continuous p-cocycle and g ∈ G. We have to find an η ∈ C p−1 (g, V ) with g.ω − ω = dg η. Let γ : [0, 1] → G be a smooth curve with γ(0) = 1 and γ(1) = g. We write δ l (γ)t := γ(t)−1 .γ (t) for its left logarithmic derivative and recall the Cartan formula Lx = ix ◦ dg + dg ◦ ix for the action of g on C • (g, V ). We thus obtain in the pointwise sense on each p-tuple of elements of g: 1 1 d g.ω − ω = γ(t).ω dt = γ(t). Lδl (γ)t ω dt 0 dt 0 1 1 γ(t). iδl (γ)t dg ω + dg iδl (γ)t ω dt = dg γ(t).iδl (γ)t ω dt = 0

= dg 0

0

1

γ(t).iδl (γ)t ω dt.


For p = 2 we obtain in particular g.ω − ω = dg ηg for 1 ηg (x) := γ(t).ω δ l (γ)t , Ad(γ(t))−1 .x dt.

335

(22)

0

References [ABFP08] [ABP06] [ACM89] [Bi98] [Bi03] [BiNe08] [Bry93] [CaE36] [ChE48] [EE69] [ERMo94] [ERC04] [EF94] [Fa84] [FF01] [FK96] [FJ07] [FRST90]

Allison, B. N., Berman, S., Faulkner, J. R., and A. Pianzola, Realization of graded-simple algebras as loop algebras, Forum Math. 20:3 (2008), 395–432 Allison, B. N., S. Berman, and A. Pianzola, Iterated loop algebras, Pacific J. Math. 227:1 (2006), 1–41 Abbati, M. C., R. Cirelli, A. Mania, and P. Michor, The Lie group of automorphisms of a principal bundle, JGP 6:2 (1989), 215–235 Billig, Y., Principal vertex operator representations for toroidal Lie algebras, J. Math. Phys. 39:7 (1998), 3844–3864 Billig, Y., Abelian extensions of the group of diffeomorphisms of a torus, Lett. Math. Phys. 64 (2003), 155–169 Billig, Y., and K.-H. Neeb, On the cohomology of vector fields on parallelizable manifolds, Ann. Inst. Fourier 58 (2008), 1937–1982 Brylinski, J.-L., “Loop Spaces, Characteristic Classes and Geometric Quantization,” Progr. in Math. 107, Birkh¨ auser Verlag, 1993 ´ La topologie des groupes de Lie. (Exp. de géom. No. 8.), Cartan, E., Actual. Sci. Industr. 358 (1936), 28 p Chevalley, C., and S. Eilenberg, Cohomology theory of Lie groups and Lie algebras, Transactions of the Amer. Math. Soc. 63 (1948), 85–124 Eearle, C. J., and J. Eells, A fiber bundle description of Teichm¨ uller theory, J. Diff. Geom. 3 (1969), 19–43 Eswara Rao, S., and R. V. Moody, Vertex representations for n-toroidal Lie algebras and a generalization of the Virasoro algbra, Commun. Math. Phys. 159 (1994), 239–264 Eswara Rao, S., and C. Jiang, Classification of irreducible integrable representations for the full toroidal Lie algebras, J. Pure Appl. Algebra 200:1-2 (2005), 71–85 Etingof, P. I., and I. B. Frenkel, Central extensions of current groups in two dimensions, Commun. Math. Phys. 165 (1994), 429–444 Faddeev, L. D., Operator anomaly for the Gauss Law, Physics Letters 145B:1,2 (1984), 81–84 Feigin, B. L. and D. B. Fuchs, Cohomologies of Lie Groups and Lie Algebras, in “Lie Groups and Lie Algebras II”, A. L. Onishchik and E. B. Vinberg (Eds.), Encyclop. Math. Sci. 21, Springer-Verlag, 2001 Frenkel, I. B., and B. Khesin, Four dimensional realization of two dimensional current groups, Commun. Math. Phys. 178 (1996), 541–562 Fu, J., and C. Jiang, Integrable representations for the twisted full toroidal Lie algebras, J. Algebra 307:2 (2007), 769–794 Frappat, L., Raggoucy, E., Sorba, P., and F. Thuillier, Generalized Kac–Moody algebras and the diffeomorphism group of a closed surface, Nuclear Physics B334 (1990), 250–264

336

Karl-Hermann Neeb

[Fu86] [GIMM04] [Gl02] [GN10] [God73] [GHV72]

[GP78] [Hae76] [Ha82] [Ho65] [Hu94] [Ka90] [KYMO85] [Ko70] [Kosz74]

[KM97] [Lar99] [Lar00] [LW08] [Lec80]

Fuks, D. B. “Cohomology of Infinite-Dimensional Lie Algebras,” Consultants Bureau, New York, London, 1986 Gotay, M. J., J. Isenberg, J. E. Marsden, and R. Montgomery, Moment Maps and Classical Fields. Part I: Covariant Field Theory, arXiv:physics/9801019v2, August 2004 Gl¨ ockner, H., Patched locally convex spaces, almost local mappings, and diffeomorphism groups of non-compact manifolds, Manuscript, TU Darmstadt, 26.6.02 Gl¨ ockner, H., and K.-H. Neeb, “Infinite-dimensional Lie groups, Vol. I, Basic Theory and Main Examples,” book in preparation Godement, R., “Théorie des faisceaux”, Hermann, Paris, 1973 Greub, W., S. Halperin, and R. Vanstone, “Connections, Curvature, and Cohomology. Vol. I: De Rham Cohomology of Manifolds and Vector Bundles”, Pure and Applied Mathematics 47, Academic Press, New York-London, 1972 Greub, W., and H.-R. Petry, On the lifting of structure groups, in “Differential geometric methods in math. physics, II”, Lecture Notes Math. 676 (1978), 217–246 Haefliger, A., Sur la cohomologie de l’algèbre de Lie des champs de vecteurs, Ann. Sci. Ec. Norm. Sup. 4e série 9 (1976), 503–532 Hamilton, R., The inverse function theorem of Nash and Moser, Bull. Amer. Math. Soc. 7 (1982), 65–222 Hochschild, G., “The Structure of Lie Groups,” Holden Day, San Francisco, CA, 1965. Husemoller, D., “Fibre Bundles”, Graduate Texts in Math. 20, Springer-Verlag, New York, Berlin, 1994 Kac, V. G., “Infinite-dimensional Lie Algebras,” Cambridge University Press, 3rd printing, Cambridge, UK, 1990 Kobayashi, O., A. Yoshioka, Y. Maeda, and H. Omori, The theory of infinite-dimensional Lie groups and its applications, Acta Appl. Math. 3:1 (1985), 71–106 Kostant, B., Quantization and unitary representations, Lectures in modern analysis and applications III, Lecture Notes Math. 170 (1970), 87–208 Koszul, J.-L., Homologie des complexes de formes différentielles d’ordre supérieur, in “Collection of articles dedicated to Henri ´ Cartan on the occasion of his 70th birthday, I”, Ann. Sci. Ecole Norm. Sup. (4) 7 (1974), 139–153 Kriegl, A., and P. Michor, “The Convenient Setting of Global Analysis,” Math. Surveys and Monographs 53, Amer. Math. Soc., 1997 Larsson, T. A., Lowest-energy representations of non-centrally extended diffeomorphism algebras, Commun. Math. Phys. 201 (1999), 461–470 Larsson, T. A., Extended diffeomorphism algebras and trajectories in jet space, Commun. Math. Phys. 214:2 (2000), 469–491 Laurent-Gengoux, C., and F. Wagemann, Obstruction classes of crossed modules of Lie algebroids linked to existence of principal bundles, Ann. Global Anal. Geom. 34:1 (2008), 21–37 Lecomte, P., Sur l’algèbre de Lie des sections d’un fibré en algèbre de Lie, Ann. Inst. Fourier 30 (1980), 35–50

Lie Groups of Bundle Automorphisms and Their Extensions [Lec85] [Les67] [LMNS96] [LMNS98] [Ma89] [Ma05] [Ma02]

[MN03] [Mi80] [Mil84] [Mo01] [Mu96] [MW09] [Ne02] [Ne04a] [Ne04b]

[Ne06a] [Ne06b] [Ne06c] [Ne07]

337

Lecomte, P., Sur la suite exacte canonique associée ` a un fibré principal, Bull Soc. Math. Fr. 13 (1985), 259–271 Leslie, J. A., On a differential structure for the group of diffeomorphisms, Topology 6 (1967), 263–271 Losev, A., G. Moore, N. Nekrasov, and S. Shatashvili, Fourdimensional avatars of two-dimensional RCFT, in “Strings 95” (Los Angeles, CA, 1995), World Sci. Publ., NJ, 1996, 336–362 Losev, A., G. Moore, N. Nekrasov, and S. Shatashvili, Central extensions of gauge groups revisited, Sel. Math., New series 4 (1998), 117–123 Mackenzie, K., Classification of principal bundles and Lie groupoids with prescribed gauge group bundle, J. Pure Appl. Algebra 58:2 (1989), 181–208 Mackenzie, K. C. H., “General Theory of Lie Groupoids and Lie Algebroids,” London Mathematical Society Lecture Note Series 213, Cambridge University Press, Cambridge, UK, 2005 Maier, P., Central extensions of topological current algebras, in “Geometry and Analysis on Finite- and Infinite-Dimensional Lie Groups,” A. Strasburger et al. (eds.), Banach Center Publications 55, Warszawa 2002; 61–76 Maier, P., and K.-H. Neeb, Central extensions of current groups, Math. Annalen 326:2 (2003), 367–415 Michor, P. W., “Manifolds of Differentiable Mappings,” Shiva Publishing, Orpington, Kent (UK), 1980 Milnor, J., Remarks on infinite-dimensional Lie groups, in DeWitt, B., Stora, R. (eds.), “Relativité, groupes et topologie II” (Les Houches, 1983), North Holland, Amsterdam, 1984; 1007–1057 Morita, S., “Geometry of Characteristic Classes,” Transl. Math. Monographs 199, 1999, American Math. Soc., 2001 Murray, M. K., Bundle gerbes, London Math. Soc. (2) 54 (1996), 403–416 M¨ uller, Chr., and Chr. Wockel, Equivalences of smooth and continuous principal bundles with infinite-dimensional structure group, Adv. Geom. 9:4 (2009), 605–626 Neeb, K.-H., Central extensions of infinite-dimensional Lie groups, Annales de l’Inst. Fourier 52:5 (2002), 1365–1442 Neeb, K.-H., Abelian extensions of infinite-dimensional Lie groups, Travaux mathématiques 15 (2004), 69–194 Neeb, K.-H., Current groups for non-compact manifolds and their central extensions, in “Infinite Dimensional Groups and Manifolds”. T. Wurzbacher (ed.). IRMA Lectures in Mathematics and Theoretical Physics 5, de Gruyter Verlag, Berlin, 2004; 109–183 Neeb, K.-H., Towards a Lie theory of locally convex groups, Jap. J. Math. 3rd series 1:2 (2006), 291–468. Neeb, K.-H., Nonabelian extensions of topological Lie algebras, Communications in Alg. 34 (2006), 991–1041 Neeb, K.-H., Lie algebra extensions and higher order cocycles, J. Geom. Sym. Phys. 5 (2006), 48–74 Neeb, K.-H., Nonabelian extensions of infinite-dimensional Lie groups, Ann. Inst. Fourier 56 (2007), 209–271

338

Karl-Hermann Neeb

[NV03] [NeWa08a] [NeWa08b] [NeWo09] [Neh07] [Omo70] [OR07] [PPS07] [Pi93] [Pi00] [PS86] [Sch04] [SW07]

[ST77] [Th54] [TL99] [Tsu81] [Vi08] [Whi78] [Wo07]

Neeb, K.-H. and C. Vizman, Flux homomorphisms and principal bundles over infinite-dimensional manifolds, Monatshefte Math. 139 (2003), 309–333 Neeb, K.-H., and F. Wagemann, The second cohomology of current algebras of general Lie algebras, Canadian J. Math. 60:4 (2008), 892–922 —, Lie group structures on groups of maps on non-compact manifolds, Geom. Dedicata 134 (2008), 17–60 Neeb, K.-H., and Chr. Wockel, Central extensions of groups of section, Annals of Global Analysis and Geometry 36:4 (2009), 381–418 Neher, E., Lectures on root graded and extended affine Lie algebras, Version of May 11, 2007 Omori, H., On the group of diffeomorphisms on a compact manifold, in “Global Analysis” (Proc. Sympos. Pure Math., Vol. XV, Berkeley, CA, 1968), Amer. Math. Soc., Providence, RI, 1970, 167–183 Ovsienko, V., and C. Roger, Looped cotangent Virasoro algebra and non-linear integrable systems in dimension 2 + 1, Comm. Math. Phys. 273:2 (2007), 357–378 Pianzola, A., D. Prelat, and J. Sun, Descent constructions for central extensions of infinite dimensional Lie algebras, Manuscripta Math. 122:2 (2007), 137–148 Pickrell, D., Extensions of loop groups, in “The Mathematical Legacy of Hanno Rund,” Hadronic Press, Palm Harbor, FL, 1993; pp. 87–134 Pickrell, D., Extensions of Loop Groups, H 4 (BK, Z), and Reciprocity, Unpublished preprint, 2000 Pressley, A., and G. Segal, “Loop Groups,” Oxford University Press, Oxford, 1986 Schmid, R., Infinite dimensional Lie groups with applications to mathematical physics, J. Geom. Symm. Phys. 1 (2004), 54–120 Schweigert, Chr., and K. Waldorf, Gerbes and Lie groups, Karl-Hermann Neeb and Arturo Pianzola (eds.), “Developments and Trends in Infinite-Dimensional Lie Theory,” Birkh¨ auser, Boston, MA, 2010 Shiga, K., and T. Tsujishita, Differential representations of vector fields, Kodai Math. Sem. Rep. 28 (1977), 214–225 Thom, R., Opérations en cohomologie r´ eelle, Séminaire Henri Cartan, tome 7, no. 2, (1954–1955), exp. no.17, 1–10 Toledano Laredo, V., Positive energy representations of the loop groups of non-simply connected Lie groups, Commun. Math. Phys. 207 (1999), 307–339 Tsujishita, T., “Continuous Cohomology of the Lie Algebra of Vector Fields,” Memoirs of the Amer. Math. Soc. 34:253, 1981 Vizman, C., The path group construction of Lie group extensions, J. Geom. Phys. 58:7 (2008), 860–873 Whitehead, G. W., “Elements of Homotopy Theory,” Graduate Texts in Mathematics 61, Springer-Verlag, New York, Berlin, 1978 Wockel, Chr., Lie group structures on symmetry groups of principal bundles, J. Funct. Anal. 251:1 (2007), 254–288

Gerbes and Lie Groups Christoph Schweigert and Konrad Waldorf Organisationseinheit Mathematik, Universit¨ at Hamburg, Bundesstraße 55, D-20146, Germany, [email protected], [email protected]

Summary. Compact Lie groups do not only carry the structure of a Riemannian manifold, but also canonical families of bundle gerbes. We discuss the construction of these bundle gerbes and their relation to loop groups. We present several algebraic structures for bundle gerbes with connection, such as Jandl structures, gerbe modules and gerbe bimodules, and indicate their applications to Wess–Zumino terms in two-dimensional field theories. Key words: bundle gerbe, loop group, Jandl structure, gerbe module and bimodule. 2000 Mathematics Subject Classifications: 22E67, 55R65, 81T40.

Introduction Compact Lie groups come not only with a canonical metric (the Killing form), but also with a canonical family of bundle gerbes. These bundle gerbes are geometric objects made of finite-dimensional manifolds and maps between them, and they provide a way of understanding structure over the infinitedimensional loop group. As a motivation, consider a central extension of the loop group of a compact connected and simply connected Lie group G, 0 −→ C× −→ LG −→ LG −→ 0.

(1)

Such extensions are classified by H2 (LG, Z). By transgression, this cohomology group is in turn isomorphic to the cohomology of G, H3 (G, Z) ∼ = H2 (LG, Z).

(2)

While the cohomology group H2 (LG, Z) classifies line bundles over LG by their Chern class, H3 (G, Z) classifies bundle gerbes over G. In this way, every


339

340

Christoph Schweigert and Konrad Waldorf

bundle gerbe over the finite-dimensional manifold G gives rise to a line bundle over the infinite-dimensional manifold LG. This contribution is organized as follows. In Section 1 we describe bundle gerbes on general manifolds and their classification. In Section 2 we explain how bundle gerbes can be equipped with a connection which allows to define surface holonomies: from this point of view, bundle gerbes generalize the holonomy of principal bundles around curves. In case that the base manifold is a compact Lie group G, we construct examples of bundle gerbes over G in Section 3. Then we explain in Section 4 how a bundle gerbe gives rise to a line bundle over the loop space. In Section 5 we discuss additional structures for bundle gerbes like bundle gerbe modules, bimodules and Jandl structures. Finally, in Section 6 we outline the applications of bundle gerbes on Lie groups to two-dimensional conformal field theory and string theory, which are closely related to the surface holonomy from Section 2. In these theories one also recovers the loop space as the space of configurations.

1 Bundle gerbes Let M be a smooth manifold. We shall briefly review the classification of complex line bundles over M . For this purpose, let us choose a good open cover V = {Vi }i∈I of M ; i.e., every finite intersection of open sets Vi is contractible. In particular, every line bundle L admits local non-zero sections which determine smooth transition functions gij : Vi ∩ Vj → C× .

(3)

On three-fold intersections Vi ∩ Vj ∩ Vk , these transition functions satisfy the cocycle condition gik = gij · gjk . (4) It is fair to call this equality a cocycle condition, since it means that δg = 0 ˇ for the element g := (gij ) ∈ Cˇ 1 (V, C× M ) in the Cech cohomology of the sheaf × of smooth C -valued functions on M with respect to the cover V. Since we have chosen the cover V to be good, there is a canonical isomorphism 2 ˇ 1 (V, C× ) ∼ H M = H (M, Z) using the exponential sequence. The image of the 2 class [g] in H (M, Z) is independent of the choice of the sections, and is called the (first) Chern class c1 (L) of the line bundle L. This defines an isomorphism c1 : Pic(M ) → H2 (M, Z)

(5)

from the group Pic(M ) of isomorphism classes of line bundles to the cohomology group H2 (M, Z), providing a geometric realization of this group. A bundle gerbe is a geometric object which realizes the cohomology group H3 (M, Z) in a similar way. To prepare its definition, we fix the

Gerbes and Lie Groups

341

following notation. For a surjective submersion π : Y → M we denote the k-fold fibre product of Y over M by Y [k] = {(y1 , . . . , yk ) ∈ Y k |π(y1 ) = · · · = π(yk )}.

(6)

This is again a smooth manifold, having canonical projections πi1 ,...,i : Y [k] → Y [] on the respective factors. Definition 1.1 ([Mur96]). A bundle gerbe G over a manifold M is a triple (π, L, µ) consisting of a surjective submersion π : Y → M , a line bundle L over Y [2] and an isomorphism ∗ ∗ ∗ µ : π12 L ⊗ π23 L → π13 L

(7)

of line bundles over Y [3] , such that µ is associative in the sense that the diagram ∗ ∗ ∗ π12 L ⊗ π23 L ⊗ π34 L

∗ π123 µ⊗id

∗ ∗ / π13 L ⊗ π34 L

∗ id⊗π234 µ

(8)

∗ π134 µ

∗ ∗ π12 L ⊗ π24 L

∗ π124 µ

∗ / π14 L

of isomorphisms of line bundles over Y [4] is commutative. Let us now describe how bundle gerbes realize the cohomology group H3 (M, Z). Let us again choose a good open cover V = {Vi }i∈I of M which admits sections si : Vi → Y into the manifold Y of the surjective submersion of a bundle gerbe G = (π, L, µ). If we denote by MV the disjoint union of all the open sets Vi , the sections si patch together to a smooth map s : MV → Y sending a point x ∈ Vi to si (x) ∈ Y . Note that there are induced maps [k] MV → Y [k] on fibre products (all denoted by s in order to simplify the nota[k] tion), where MV is the disjoint union of all k-fold intersections of open sets [2] Vi . Now we pull back the line bundle L along s to a line bundle over MV , [3] and the isomorphism µ to an isomorphism of line bundles over MV . For a ∗ choice σij : Vi ∩ Vj → s L of local sections into the pull-back line bundle, we obtain smooth functions

by

gijk : Vi ∩ Vj ∩ Vk → C×

(9)

s∗ µ(σij ⊗ σjk ) = gijk · σik.

(10)

The associativity condition (8) leads to the equality gijk · gik = gij · gjk

(11)

342


of functions on four-fold intersections Vi ∩ Vj ∩ Vk ∩ V . In other words, the ˇ element g = (gijk ) ∈ Cˇ 2 (V, C× M ) is a cocycle and defines a class in the Cech × 2 3 ˇ cohomology group H (V, CM ). Its image in the cohomology group H (M, Z) is called the Dixmier–Douady class dd(G) of the bundle gerbe G; it is analogous to the Chern class of a line bundle. To obtain a classification result for isomorphism classes of bundle gerbes, we first have to define morphisms between bundle gerbes. To simplify the notation, we work with the convention that we do not write pull-backs along canonical maps, as in (12) and (13) below. Definition 1.2. A morphism A : G1 → G2 between two bundle gerbes G1 = (π1 , L1 , µ1 ) and G2 = (π2 , L2 , µ2 ) is a pair A = (A, α) consisting of a vector bundle A over the fibre product Z := Y1 ×M Y2 (whose surjective submersion to M is denoted by ζ) and an isomorphism α : L1 ⊗ ζ2∗ A → ζ1∗ A ⊗ L2

(12)

of vector bundles over Z [2] such that the diagram ∗ ∗ ζ12 L1 ⊗ ζ23 L1 ⊗ ζ3∗ A

µ1 ⊗id

∗ / ζ13 L1 ⊗ ζ3∗ A

(13)

∗ id⊗ζ23 α

∗ ∗ ζ12 L1 ⊗ ζ2∗ A ⊗ ζ23 L2

∗ ζ13 α

∗ α⊗id ζ12

∗ ∗ ζ1∗ A ⊗ ζ12 L2 ⊗ ζ23 L2

id⊗µ2

∗ / ζ1∗ A ⊗ ζ13 L2

of isomorphisms of vector bundles over Z [3] is commutative. The definition of the composition of two morphisms A : G1 → G2 and A : G2 → G3 is quite involved, and we omit its discussion for the purposes of this article; see [Ste00, Wal07] for more details. Bundle gerbes and their morphisms fit into the structure of a 2-category rather than that of a category. This becomes obvious when comparing two morphisms A and A between the same bundle gerbes: since A and A themselves consist of vector bundles, the natural way to compare them is a morphism of vector bundles. Definition 1.3. Let A = (A, α) and A = (A , α ) be two morphisms from G1 = (π1 , L1 , µ1 ) to G2 = (π2 , L2 , µ2 ). A 2-morphism β : A ⇒ A

(14)

is an isomorphism β : A → A of vector bundles over Z, which is compatible with the isomorphisms α and α in the sense that the diagram


L1 ⊗ ζ2∗ A 1⊗ζ2∗ β

L1 ⊗ ζ2∗ A

α

α

/ ζ1∗ A ⊗ L2

343

(15)

ζ1∗ β⊗1

/ ζ1∗ A ⊗ L2

of isomorphisms of vector bundles over Z [2] is commutative. The 2-categorical setup is also the appropriate context in which to address the question of which of the morphisms between two bundle gerbes are invertible. Proposition 1.4 ([Wal07]). A morphism A = (A, α) is invertible if and only if the vector bundle A is of rank 1. So, the invertible morphisms in the 2-category are the stable isomorphisms from [MS00], Section 3. Let us now return to the relation between bundle gerbes and the cohomology group H3 (M, Z). Let g and g be the cocycles extracted from two bundle gerbes G and G over M with respect to the same cover V of M and sections si : Vi → Y and si : Vi → Y . Now let A : G → G be an isomorphism, A = (A, α). Let t : MV → Z := Y ×M Y be the map sending a point x ∈ Vi to the pair (si (x), si (x)). Using t we pull back the line bundle A and choose non-zero sections σi : Vi → A. Then we obtain smooth functions hij : Vi ∩ Vj → C× by α(σi ⊗ ζ2∗ σij ) = hij · (ζ1∗ σij ⊗ σi ).

(16)

The compatibility condition (13) between α and the isomorphisms µ and µ of the bundle gerbes leads to the equation hij hjk , hik gijk = gijk

(17)

ˇ equivalently, in terms of the Cech coboundary operator, g = g ·δh. This means that the Dixmier–Douady classes [g] and [g ] of the isomorphic bundle gerbes G and G are equal. This is the main ingredient for the following classification result. Theorem 1.5 ([MS00]). The Dixmier–Douady class defines a bijection between the set of isomorphism classes of bundle gerbes and the cohomology group H3 (M, Z). In particular, consider M = G a compact, simple, connected and simply connected Lie group. There is a canonical identification H3 (G, Z) = Z, so that we have a canonical sequence of bundle gerbes associated to G. In Section 3, we give a geometric construction of these bundle gerbes. The 2-category of bundle gerbes admits several additional structures such as pull-backs, tensor products and duals. All these structures are compatible with the Dixmier–Douady class

344


dd(G1 ⊗ G2 ) = dd(G1 ) + dd(G2 ) ,

dd(G ∗ ) = −dd(G)

(18)

and, for a smooth map f : X → M , dd(f ∗ G) = f ∗ dd(G).

(19)

To close, let us construct a bundle gerbe whose Dixmier–Douady class vanishes, representing the neutral element in H3 (M, Z). For this purpose, consider the identity idM : M → M as the surjective submersion and the trivial line bundle M × C over M . Now, the isomorphism µ can be chosen to be the identity idC , so that I := (idM , M × C, idC ) is a bundle gerbe. It is easy to verify that its Dixmier–Douady class vanishes, dd(I) = 0.

2 Connections on bundle gerbes and holonomy Before we discuss more examples of bundle gerbes in Section 3, we introduce several additional structures on bundle gerbes and the appropriate cohomology theory for their classification. Again, we first review similar additional structures for complex line bundles. One can equip every line bundle with a hermitian metric and a (unitary) connection ∇. In addition to the transition function gij : Vi ∩ Vj → C× , which now can be determined such that it takes values in U (1), we obtain local connection 1-forms Ai ∈ Ω(Vi ) by writing the covariant derivatives of the sections si : Vi → L as ∇si = Ai ⊗ si . The Leibniz rule implies the equality Aj − Ai − dlog(gij ) = 0.

(20)

As we shall see next, the local data (g, A) define a cocycle in the Deligne k complex DV (n) for n = 1. The first cochain groups of this complex are 0 (1) = Cˇ 0 (V, U (1)M ) DV 1 1 (1) = Cˇ 1 (V, U (1)M ) ⊕ Cˇ 0 (V, ΩM ) DV 2 2 1 1 ˇ ˇ DV (1) = C (V, U (1)M ) ⊕ C (V, ΩM ),

(21) (22) (23)

and its differential is given by 1 2 (1) → DV (1) : (g, A) → (δg, δA − dlog(g)). D : DV

(24)

k ˇ (n) is the total complex of the Cech– In general, the Deligne complex DV ˇ Deligne double complex, whose rows are Cech cochain groups of the sheaf complex

0

/ U (1)M

dlog

/ Ω1 M

d

/ ...

d

n / ΩM ,

(25)

ˇ truncated in degree n, and the columns are the usual Cech complexes assok ciated to the sheaves U (1)M and ΩM . The truncation is necessary to obtain objects that are not flat.


345

Thus we can regard the local data (g, A) of a hermitian line bundle with 1 connection as an element of DV (1). By equations (4) and (20) it satisfies D(g, A) = 0, and thus defines a class in the first cohomology group of the Deligne complex, denoted by H1 (M, D(1)). One can show that this defines a bijection Pic∇ (M ) ∼ (26) = H1 (M, D(1)) between the group Pic∇ (M ) of isomorphism classes of hermitian line bundles with connection and Deligne cohomology [Bry93]. Now we discuss bundle gerbes with similar additional structures. Definition 2.1. Let G = (π, L, µ) be a bundle gerbe over M. It can be equipped successively with the following additional structures. (a) A hermitian structure on G is a hermitian metric on the line bundle L, such that the isomorphism µ is an isometry. (b) A connection on the hermitian bundle gerbe G is a connection ∇ on the hermitian line bundle L, such that the isomorphism µ respects the induced connections. (c) A curving of a connection ∇ on the hermitian bundle gerbe G is a 2-form C ∈ Ω 2 (Y ), such that π2∗ C − π1∗ C = curv(∇),

(27)

where curv(∇) ∈ Ω 2 (Y [2] ) is the curvature of the connection ∇ on L. Every bundle gerbe admits all of these additional structures [Mur96]. Because the applications of bundle gerbes that we discuss later require all this additional structure, we work from now on only with hermitian bundle gerbes with connection and curving. An important feature of those gerbes is that they provide a notion of curvature. To see this, consider the derivative of equation (27): since the curvature of the connection ∇ is a closed form, we obtain π1∗ dC = π2∗ dC, which means that dC is the pull-back of a 3-form on M , dC = π ∗ H.

(28)

This 3-form H is uniquely determined and closed; it is called the curvature of the curving of the connection ∇ on the hermitian bundle gerbe G, and denoted by curv(C) := H. To provide a simple example of such additional structures, the bundle gerbe I = (idM , M × C, idC ) with vanishing Dixmier–Douady class becomes a hermitian bundle gerbe with connection by taking the canonical hermitian metric and the trivial flat connection ∇ := d on the trivial line bundle M × C. Note that now any 2-form C ∈ Ω 2 (M ) satisfies the condition (27) for a curving on I, because ∇ is flat and π1 = π2 = idM . So we have a canonical hermitian bundle gerbe IC := (idM , M × C, idC , d, C) with connection and curving for every 2-form C ∈ Ω 2 (M ). The curvature of its curving C is curv(C) = dC.

346


Now we extend the cohomological classification from bundle gerbes to hermitian bundle gerbes G = (π, L, µ, ∇, C) with connection and curving, using a good open cover V of M . Recall that we have extracted a transition function gijk : Vi ∩ Vj ∩ Vk → C× using a choice of sections si : Vi → Y and σi : Vi → s∗ L, defining a representative g ∈ Cˇ 2 (V, C× ) of the Dixmier– Douady class of the bundle gerbe. Now that L is a hermitian line bundle, we choose the sections σi such that gijk is U (1)-valued. Furthermore, by using the connection s∗ ∇ on s∗ L, we obtain local connection 1-forms Aij ∈ Ω 1 (Vi ∩ Vj ). The condition that µ preserves connections implies Ajk − Aik + Aij + dlog(gijk ) = 0.

(29)

Finally, the curving C gives rise to local 2-forms Bi := s∗i C ∈ Ω 2 (Vi ), and the compatibility (27) with the curvature of ∇ implies Bj − Bi − dAij = 0.

(30)

Note that the curvature curv(C) of the curving can be computed from the local data by H|Vi = dBi . By (30), this indeed gives a globally defined 3-form. ˇ In terms of Cech cohomology, we have extracted an element (g, A, B) in the second cochain group of the Deligne complex in degree 2, 2 DV (2) = Cˇ 2 (V, U (1)) ⊕ Cˇ 1 (V, Ω 1 ) ⊕ Cˇ 0 (V, Ω 2 ).

(31)

Here the differential is 2 3 (2) → DV (2) : (g, A, B) → (δg, δA + dlog(g), δB − dA), D : DV

(32)

so that the cocycle condition (27) and equations (29) and (30) imply the cocycle condition D(g, A, B) = 0. This way, a hermitian bundle gerbe with connection and curving defines a class in the cohomology of the Deligne complex in degree 2, H2 (M, D(2)). As an exercise, the reader may compute the Deligne class of the canonical hermitian bundle gerbe IC with connection and curving associated to any 2-form C ∈ Ω 2 (M ). 1 2 Note that both the Deligne cochain groups DV (1) from (22) and DV (2) from (31) have projections on the first summand, which commute with the ˇ Deligne differential and the Cech coboundary operator, so that we get induced (surjective) group homomorphisms [Bry93] Hk (M, D(k)) → Hk+1 (M, Z).

(33)

This way we obtain the Chern class and the Dixmier–Douady class of a hermitian line bundle with connection and of a hermitian bundle gerbe with connection and curving, respectively, from their Deligne classes. The surjectivity of (33) means that Deligne cohomology refines the ordinary singular cohomology with Z coefficients. To achieve a classification result for hermitian bundle gerbes with connection and curving similar to the result (26) for hermitian line bundles with


347

connection, we have to adapt the definition of a morphism between bundle gerbes to morphisms between hermitian bundle gerbes with connection and curving. Definition 2.2. A morphism A : G → G between two hermitian bundle gerbes G = (π, L, µ, ∇, C) and G = (π , L , µ , ∇ , C ) with connection and curving is a morphism A = (A, α) as in Definition 1.2, together with a connection on the vector bundle A, such that 1. the isomorphism α respects connections, 2. the curvature of is related to the curvings by 1 tr(curv()) = C − C. n

(34)

As in Proposition 1.4, a morphism is invertible precisely if the vector bundle is of rank 1. It is again an easy exercise to check that an isomorphism A : G → G of hermitian bundle gerbes with connection and curving with Deligne cocycles (g, A, B) and (g , A , B ), respectively, gives rise to a Deligne 1 cochain (h, W ) ∈ DV (2) which satisfies (g , A , B ) = (g, A, B) + D(h, W ).

(35)

This shows that isomorphism classes of hermitian bundle gerbes with connection and curving have well-defined Deligne classes. Theorem 2.3 ([MS00]). Isomorphism classes of hermitian bundle gerbes with connection and curving correspond bijectively to the Deligne cohomology group H2 (M, D(2)). Particular examples of morphisms are trivializations: a trivialization of a hermitian bundle gerbe G with connection and curving is an isomorphism T : G → Iρ

(36)

for some 2-form ρ ∈ Ω 2 (M ). In terms of local data, this isomorphism corre1 sponds to a Deligne cochain (h, W ) ∈ DV (2) with (1, 0, ρ) = (g, A, B) + D(h, W ),

(37)

if (g, A, B) is local data of the bundle gerbe G. In particular, the existence of a trivialization implies that the Dixmier–Douady class of G vanishes. Many assertions about bundle gerbes and their isomorphisms can be proven either in a geometrical way or by computations in Deligne cohomology. As an illustration, we shall prove the following important lemma. Lemma 2.4. Two trivializations T1 : G → Iρ1

and

T2 : G → Iρ2

(38)

of the same hermitian bundle gerbe G over M with connection and curving determine a hermitian line bundle over M with connection of curvature ρ2 −ρ1 .

348


Proof 1 (2-categorical). Using the features of the 2-category of bundle gerbes, we can give a quite conceptual proof: by taking the inverse and composition (which we have not explained in this article, but can be found in [Wal07]), we obtain an isomorphism T2 ◦ T1−1 : Iρ1 → Iρ2

(39)

of trivial bundle gerbes. From the definitions of isomorphisms and trivial bundle gerbes it follows immediately that T2 ◦ T1−1 is a line bundle with curvature ρ2 − ρ1 . Proof 2 (geometrical). The two isomorphisms Ti = (Ti , τi , i ) consist of hermitian line bundles Ti over Z := Y ×M M ∼ = Y , connections i of curvature curv(i ) = π ∗ ρi − C and isomorphisms τi : L ⊗ π2∗ Ti → π1∗ Ti of hermitian line bundles respecting the connections. They can be composed to an isomorphism τ2−1 ⊗ τ1∗ : π1∗ (T2 ⊗ T1∗ ) → π2∗ (T2 ⊗ T1∗ )

(40)

of hermitian line bundles with connection over Y [2] , which satisfies the obvious cocycle condition over Y [3] , due to the commutative diagram (13) for the τi . This is the condition for the hermitian line bundle T2 ⊗ T1∗ with connection 2 − 1 to descend from Y to M . The descent line bundle has the claimed curvature. Proof 3 (cohomological). If the isomorphisms T1 and T2 have local data (h1 , W1 ) and (h2 , W2 ), respectively, both satisfying equation (37), their difference satisfies D(h2 · h−1 (41) 1 , W2 − W1 ) = 0, which is the cocycle condition for a hermitian line bundle with connection of curvature d(W2 − W1 ) = curv(2 ) − curv(1 ) = ρ2 − ρ1 . One of the most important aspects of the theory of hermitian bundle gerbes with connection and curving is that they provide a notion of holonomy around surfaces. Definition 2.5 ([CJM02]). Let G be a hermitian bundle gerbe with connection and curving over M . For a closed oriented surface Σ and a smooth map φ : Σ → M , let T : φ∗ G → Iρ (42) be a trivialization of the pullback of G along φ. Then we define ρ holG (φ) := exp Σ

to be the holonomy of G around φ : Σ → M .

(43)


349

Note that the Dixmier–Douady class of φ∗ G vanishes by dimensional reasons, so that the existence of the trivialization T is guaranteed. However, it is not unique, and different trivializations may have different 2-forms ρ. Now, by Lemma 2.4 we know that the difference ρ2 − ρ1 between two such 2-forms is the curvature of some line bundle over M , in particular: it is a closed form with integral class. Then, the calculation exp (44) ρ2 = exp ρ2 − ρ1 · exp ρ1 = exp ρ1 Σ

Σ

Σ

Σ

shows that the definition holG (φ) is independent of the choice of the trivialization. There also exist expressions for the holonomy holG (φ) in terms of local data of G. They generalize the local formulae for the holonomy of hermitian line bundles with connection. In Section 6 we describe the applications of holonomy of bundle gerbes in conformal field theory.

3 Bundle gerbes over compact Lie groups Now that we have introduced bundle gerbes as geometric objects over arbitrary manifolds, we specialize to manifolds which are Lie groups. We describe in this section how the Lie group structure allows constructions of examples of bundle gerbes. First constructions of gerbes over different types of compact Lie groups (in realizations different from bundle gerbes) can be found in [Cha98, Bry]. Bundle gerbes (with connection and curving) have been constructed in [GR02, Mei02, GR03]. As already mentioned, for a compact, simple, connected and simply connected Lie group G, we have H3 (G, Z) = Z. The (up to isomorphism unique) bundle gerbe over G whose Dixmier–Douady class corresponds to 1 ∈ Z is called the basic bundle gerbe, and is denoted by G0 . The bundle gerbe with Dixmier–Douady class k ∈ Z can then be obtained from G0 or G0∗ by a k-fold tensor product. For the purposes of this article, we will restrict ourselves to the construction given in [GR02], by which we obtain the basic bundle gerbe over the special unitary groups SU(n) and the symplectic groups Sp(n). First we consider a general compact, simple and simply connected Lie group G with Lie algebra g. We choose a maximal torus T with Lie algebra t of rank r. We further choose a set of simple roots α1 , ..., αr and denote the associated positive Weyl chamber by C. Let α0 be the highest root and let A := {ξ ∈ C | α0 (ξ) ≤ 1}

(45)

be the fundamental alcove. It is bounded by the hyperplanes Hi perpendicular to the roots αi , and the additional hyperplane H0 consisting of elements

350


ξ ∈ t with α0 (ξ) = 1. Thus, it is an r-dimensional simplex in t with vertices µ0 , . . . , µr determined by the condition that µi ∈ Hj for all j = i. For simple and simply connected groups, the fundamental alcove parameterizes conjugacy classes of G in the sense that each conjugacy class contains a unique point exp ξ with ξ ∈ A. This defines a continuous map q : G → A.

(46)

Let Ai be the open complement of the face opposite to the vertex µi in A, and consider the open sets Ui := q −1 (Ai ). More generally, for any subset I ⊂ r = {0, ..., r} denote by UI the intersection of all Ui with i ∈ I, and similarly by AI the intersection of all Ai with i ∈ I. Of course, UI = q −1 (AI ). We use the open sets Ui to construct the surjective submersion Ui and π : Y → G : (i, x) → x. (47) Y := i∈r

Note that the k-fold fibre products are disjoint unions of intersections UI . Y [k] =

(48)

|I|=k

The surjective submersion π : Y → M will serve as the first ingredient of the bundle gerbe we want to construct. To construct the line bundle L over Y [2] , we show next that Y [2] projects onto a union of coadjoint orbits. For any I ⊂ r, all group elements exp ξ with ξ in the open face spanned by the vertices µi with i ∈ I have the same centralizer GI . For any inclusion I ⊂ J it follows that GJ ⊂ GI ; for I = r we obtain Gr = T . Let SI be the orbit of exp AI ⊂ T under the conjugation with GI . Consider the set G×GI SI consisting of equivalence classes of pairs (g, s) ∈ G × SI under the equivalence relation (g, s) ∼ (gh, h−1 sh) for h ∈ GI . We have the canonical projection ρI : G ×GI SI → G/GI and a smooth map u I : G × G I SI → U I

(49)

which sends a representative (g, s) to gsg −1 ∈ G. This is well defined on equivalence classes, and for h ∈ GI and ξ ∈ AI with s := h exp ξ h−1 we find q(gsg −1 ) = ξ and hence gsg −1 ∈ UI . The map uI is even a diffeomorphism: for g ∈ UI let ξ := q(g) ∈ AI and h ∈ G such that g = h exp ξ h−1 . Then, the inverse sends g to the equivalence class of (h, exp ξ). Since Gij fixes the difference µij := µj − µi , the quotient G/Gij projects on the coadjoint orbit Oij of µij in g∗ . Now we specialize our construction to the case that µij is a weight. This is the case for SU(n) and Sp(n), where the vertices of the fundamental alcove are contained in the weight lattice. For µij being a weight there is a canonical line bundle Lij over the coadjoint orbit Oij . Let us recall the construction of the pull-back of this line bundle to G/Gij . Let χij : Gij → U (1) be the character associated to the weight µij .


351

Then the line bundle is the bundle associated to the principal Gij -bundle G over G/Gij , namely, Lij := G ×Gij C. (50) The line bundles Lij over G/Gij can be pulled back along ∼

Ui ∩ Uj

/ G ×Gij Sij

/ G/Gij

(51)

to line bundles over Ui ∩ Uj , and their disjoint union gives a line bundle L over Y [2] . To close the definition of a bundle gerbe over G, it remains to construct the isomorphism µ of line bundles over Y [3] ; i.e., we need an isomorphism ∗ ∗ ∗ L ⊗ π23 L → π13 L µ : π12

(52)

of line bundles over Y [3] . Over each connected component Ui ∩ Uj ∩ Uk , this can be chosen as the pull-back of the canonical identification Lij ⊗ Ljk ∼ = Lik

(53)

of line bundles over G/Gijk , which comes from the coincidence µik = µij +µjk of the weights from which the line bundles are constructed. This identification is obviously associative. In order to calculate the Dixmier–Douady class of the bundle gerbe we have just constructed, we choose a connection and a curving. This procedure is analogous to the calculation of Chern classes of complex vector bundles using a choice of a hermitian metric and a connection; see [MS76]. Recall that this is based on the fact that the de Rham cohomology class of the curvature of a hermitian line bundle with connection equals the image of its Chern class under the induced map ι∗ : H2 (M, Z) → H2 (M, R). The same holds for a bundle gerbe G = (π, L, µ, ∇, C): [curv(C)] = ι∗ dd(G),

(54)

ˇ which can be proven by a zigzag-argument in the Cech–Deligne double complex [Bry93]. Note that—in a certain normalization of the Killing form [PS86]—the bi-invariant closed 3-form H :=

1 θ ∧ [θ ∧ θ] ∈ Ω 3 (G) 6

(55)

is a generator of the cohomology group H3 (G, Z). Here θ is the left invariant Maurer–Cartan form, a g-valued 1-form on G; at the unit element, the 3-form H takes the value H1 (X, Y, Z) = X, [Y, Z] on elements X, Y, Z ∈ g. Our goal is now to define a connection and a curving with curvature H. First note that the line bundle Lij from (50) inherits a hermitian metric from the standard metric on C, and that the isomorphism µ is an isometry.

352


The line bundle Lij can also be equipped with a connection: we consider the 1-form Aij := µij , θ on the total space of the principal Gij -bundle G. It induces a connection on the associated line bundle Lij because µij is preserved under the action of Gij . In this way the bundle gerbe is a hermitian bundle gerbe with connection. To define a curving for this bundle gerbe, i.e., a 2-form C ∈ Ω 2 (Y ), we use the fact that the linear retraction of Ai to the vertex µi lifts to a smooth retraction of Ui to the conjugacy class Cµi [Mei02], ri : Ui × [0, 1] → Ui .

(56)

On the conjugacy class Cµi , the 3-form H becomes exact, ι∗i H = dωµi where ιi : Cµi → G is the inclusion, and Ad−1 + idg ∗ ιi θ ∈ Ω 2 (Cµi ) ωµi := ι∗i θ ∧ Ad−1 − idg

(57)

is an invariant 2-form on the conjugacy class [BRS01]. Here the notation is to be understood as follows: at some element h ∈ Cµi , the 2-form ωµi is obtained by considering the Maurer–Cartan forms in h and taking the inverse of the adjoint action (Adh )−1 : g → g. The endomorphism (Adh )−1 − idg becomes invertible when restricted to the image of ι∗i θ so that the fraction makes sense. By pull-back along ri and fibre integration, one obtains a 2-form Ci ∈ Ω 2 (Ui ) with H|Ui = dCi . (58) One can now show that Cj − Ci = µij , dθ on Ui ∩ Uj , which is the condition on the curving. So, our construction realizes, for G = SU(n) and G = Sp(n), a hermitian bundle gerbe with connection and curving with Dixmier–Douady class 1 ∈ Z, the basic bundle gerbe. For the other compact, simple, connected and simply connected Lie groups, one can find an integer k0 for which the vertices of k0 A are weights. Using the weights k0 µij in the construction of the line bundles Lij , and the 2-forms k0 Ci in the definition of the curving, we obtain bundle gerbes with Dixmier-Douady class k0 ∈ Z [Mei02]. The smallest such integer k0 is tabulated in [Bou68]: G SU(n) Spin(n) Sp(2n) E6 E7 E8 F4 G2 k0 1 2 1 3 12 60 6 2 The construction of the basic bundle gerbes on the groups with k0 > 1 requires more advanced techniques [Mei02, GR03]. Here it becomes in particular important that the definition of a bundle gerbes admits π : Y → G to be a surjective submersion, which is more general than just an open cover of G. Starting from a bundle gerbe G on a simply connected Lie group G, we now describe a method to obtain bundle gerbes on the non simply connected Lie groups G/Z which are quotients of G by a subgroup Z of the center of G. More generally, let Γ be a finite group acting on a manifold M by diffeomorphisms.


353

Definition 3.1. A Γ -equivariant structure (Aγ , ϕγ1 ,γ2 ) on a bundle gerbe G over M consists of isomorphisms Aγ : G → γ ∗ G

(59)

for each γ ∈ Γ and of 2-isomorphisms ϕγ1 ,γ2 : γ1∗ Aγ2 ◦ Aγ1 =⇒ Aγ2 γ1

(60)

for each pair γ1 , γ2 ∈ Z, such that the diagram γ1∗ γ2∗ Aγ3 ◦ γ1∗ Aγ2 ◦ Aγ1

id◦ϕγ1 ,γ2

(61)

ϕγ2 γ1 ,γ3

γ1∗ ϕγ2 ,γ3 ◦id

γ1∗ Aγ3 γ2 ◦ Aγ1

+3 γ1∗ γ2∗ Aγ3 ◦ Aγ2 γ1

ϕγ1 ,γ3 γ2

+3 Aγ3 γ2 γ1

of 2-isomorphisms is commutative. Not every bundle gerbe G over M admits a Γ -equivariant structure, and if it does, it may not be unique. To obtain obstructions and classifications we use the cohomological language. For this purpose, we impose the structure n of a Γ -module on the Deligne cochain groups DV (k) [GSW]. We assume the existence of a good open cover V = {Vi }i∈I of M which is compatible with the group action in the sense that there is an induced action of Γ on the index set I such that γ(Vi ) = Vγi . For example, the open sets Ui that we have used in the construction of the basic bundle gerbe satisfy this condition. Then Γ n acts by pull-back on the cochain groups DV (k). For each Γ -module W , one can build the usual group cohomology complex, consisting of cochain groups CΓp (W ) := Map(Γ p+1 , W ) and the usual coboundary operator d : CΓk−1 (W ) → CΓk (W ). (62) A simple key observation is that the coboundary operator d and the Deligne differential D commute so that we have a double complex with cochain groups k CΓp (DV (n)) in degree (p, k). We denote the total cohomology of this double complex by HqΓ (M, D(n)). In particular, note that we have a natural group homomorphism eq : H2 (M, D(2)) → H2Γ (M, D(2)) : ξ → (γ ∗ ξ − ξ, 0, 0)

(63)

that includes an ordinary Deligne cohomology class into the cohomology of the total complex, that we just have defined. Now let G be a hermitian bundle gerbe with connection and curving with Deligne class ξ, and let (Az , ϕz1 ,z2 ) be a Γ -equivariant structure on G. For 1 local data aγ ∈ CΓ1 (DV (2)) of the isomorphism Aγ , i.e.,

354


Daγ = γ ∗ ξ − ξ,

(64)

0 and local data bγ1 ,γ2 ∈ CΓ2 (DV (2)) of the 2-isomorphisms ϕγ1 ,γ2 , i.e.,

Dbγ1 ,γ2 = (da)γ1 ,γ2 ,

(65)

the commutativity of diagram (61) imposes the condition (db)γ1 ,γ2 ,γ3 = 0. This means that, for a bundle gerbe with Deligne class ξ, the class eq(ξ) ∈ H2Γ (M, D(2)) is the obstruction class for G to admit Γ -equivariant structures. Furthermore, the cohomology group H1Γ (M, D(2)) classifies the inequivalent choices. In the case of bundle gerbes over a compact, simple, connected and simply connected Lie group G, we consider the action of a subgroup Z of the center of G by multiplication. In this case the relevant cohomology groups reduce to the usual group cohomology of the finite group Z, so that there is an obstruction class in H3Grp (Z, U (1)), and the possible Z-equivariant structures are classified by H2Grp (Z, U (1)). For the bundle gerbes we have constructed above, all obstruction classes against Z-equivariant structures for all subgroups Z of the center of G can be calculated [GR02, GR03]. Let us now describe how a choice (Az , ϕz1 ,z2 ) of a Γ -equivariant structure on a given bundle gerbe G = (π, L, µ, ∇, C) with connection and curving over M defines a quotient bundle gerbe G = (π, L, µ, ∇, C) over M := M/Γ . Following [GR02], we set Y := Y and π := p ◦ π : Y → M , where p : M → M is the projection to the quotient. Note that the fibre products are [2] [3] Y = Z γ and Y = Z γ1 ,γ2 (66) γ∈Γ

γ1 ,γ2 ∈Γ

for the smooth manifolds Z γ := Yγ ×M Y

and

Z γ1 ,γ2 = Yγ1 γ2 ×M Z γ1 ,

(67)

where Yγ := Y as manifolds but with projection γ −1 ◦ π instead of π. The manifolds Y

[2]

and Y

[3]

again have projections π i := πi and π ij := πij to

[2]

Y and to Y , respectively. The curving of the quotient bundle gerbe will be [2] C := C ∈ Ω 2 (Y ), and the line bundle L → Y will be L|Z γ := Aγ . Condition (27) for the curvature of L is curv(L)|Z γ = curv(Aγ ) = π2∗ C − π1∗ C = π ∗2 C − π ∗1 C, and hence is satisfied. The isomorphism µ over Y µ|Z γ1 ,γ2 := ϕZ γ1 ,γ2 ,

[3]

(68)

is defined by (69)

and its associativity by means of diagram (8) is nothing but the condition on the isomorphisms ϕγ1 ,γ2 from Definition 3.1. So we have defined a bundle gerbe G over M with connection and curving.


355

Applying this procedure to the bundle gerbes over the simply connected Lie groups and their equivariant structures, one obtains examples of bundle gerbes over all compact, simple, connected and simply connected Lie groups.

4 Structure on loop spaces from bundle gerbes We describe a construction of a line bundle over the loop space LM of a manifold M from a given hermitian bundle gerbe over M with connection and curving. This construction is adapted from the one in [Bry93]. For preparation, we have to describe the set of isomorphisms between two fixed hermitian bundle gerbes G1 and G2 with connection and curving. We denote the set of isomorphism classes of morphisms A : G1 → G2 by Iso(G1 , G2 ). The following lemma can easily be shown using the cohomological description. Lemma 4.1 ([SSW07]). The group Pic∇ 0 (M ) of isomorphism classes of flat line bundles with connection over M acts freely and transitively on the set Iso(G1 , G2 ). Now let LM := C ∞ (S 1 , M ) be the free loop space of M , equipped with a smooth manifold structure as described in [Bry93]. Let G = (π, L, µ, ∇, C) be a hermitian bundle gerbe on M with connection and curving. The total space of the line bundle we are going to define is, as a set, L := Iso(γ ∗ G, I0 ). (70) γ∈LM

It comes with the evident projection to LM , and by Lemma 4.1 every fibre is a torsor over the group 1 ∼ 1 ∼ Pic∇ 0 (S ) = Hom(π1 (S ), U (1)) = U (1).

(71)

Note that the canonical projection p : L → LM admits local sections: for a contractible open subset U ⊂ M we have an isomorphism T : G|U → Iρ which provides a section σ : LU → L : γ → (γ, γ ∗ T ). Proposition 4.2. There is a unique differentiable structure on L, such that the projection p and the sections σ are smooth, and L becomes a principal U (1)-bundle over LM . Proof. Since every gerbe over S 1 is trivializable, none of the fibres p−1 (γ) is empty. Hence each fibre is a U (1)-torsor. For the same reason, the image γ(S 1 ) of any loop γ has an open neighbourhood U ⊂ M (cf. the proof of Proposition 6.2.1 in [Bry93]), such that G|U admits a trivialization T : G|U → Iρ . The corresponding section σ identifies p−1 (LU ) with LU × U (1), and thus defines a topology and a differentiable structure on each preimage p−1 (LU ). Let a

356


topology on L be generated by all open subsets of all the fibres p−1 (LU ). Now, for two intersecting subsets U1 and U2 and trivializations T1 and T2 , respectively, let N be the line bundle over U1 ∩ U2 from Lemma 2.4. The transition map LU1 × U (1) → LU2 × U (1) is then given by (γ, z) → (γ, z · holγ ∗ N (S 1 )−1 ), and is hence differentiable with respect to the loop γ : S 1 → U1 ∩ U2 . Instead of a principal U (1)-bundle, we will often and equivalently consider L as a hermitian line bundle. The construction just discussed also applies to the case when the bundle gerbe G is hermitian and has a connection. A curving defines a connection ∇ on the line bundle L, whose curvature is curv(∇) = ev∗ H, (72) S1

where H is the curvature of the gerbe G, and ev : LM × S 1 → M is the evaluation map. The hermitian line bundle L with connection is functorial in the following sense. Proposition 4.3. For a hermitian bundle gerbe G over M with connection and curving, denote the associated hermitian line bundle with connection by LG . i) Any isomorphism of gerbes G → G induces an isomorphism LG → LG of line bundles. ii) For a smooth map f : X → M denote the induced map on loop spaces by Lf : LX → LM . Then, the line bundles (Lf )∗ LG and Lf ∗ G are canonically isomorphic. iii) For the dual gerbe G ∗ we obtain a canonical isomorphism L∗G ∼ = LG ∗ . Let us also have a view on the cohomological counterpart of the construction of a line bundle over LM from a bundle gerbe over M . For this purpose, one has to extend the usual fibre integration of differential forms, : Ω k+1 (X × S 1 ) → Ω k (X), (73) S1

to the Deligne cohomology groups. Here, X can be any smooth and possibly infinite-dimensional manifold. Such extensions have been described in various ways [Gaw88, Bry93, GT01]. Then, for X = LM , the concatenation of this extension with the pull-back along the evaluation map ev : LM × S 1 → M gives a group homomorphism ◦ ev∗ : H3 (M, D(3)) → H2 (LM, D(2)). S1

(74)

(75)


357

One can show that the image of the Deligne class [(g, A, B)] of a hermitian bundle gerbe G with connection and curving under this group homomorphism gives exactly the Deligne class of the line bundle L with the connection ∇ that we have constructed above in a direct geometric way. The construction of the line bundle L can in particular be applied to the bundle gerbes over compact Lie groups from Section 3. In this case, we obtain a sequence / U (1) / L p / LG /1 (76) 1 of smooth maps, where U (1) is mapped to the fibre of L over the loop which is constantly 1 ∈ G. Now a natural question is whether one can equip the total space L with a group structure, such that the sequence (76) is an exact sequence of groups. This would provide a geometric construction of central extensions of loop groups. For simply connected groups there exist definitions of group structures on L [Bry93], while in the general case the geometric construction of loop group extensions from bundle gerbes is still an open problem.

5 Algebraic structures for gerbes There are several additional structures for bundle gerbes, some of which we introduce in this section. We describe the particular case of those additional structures on bundle gerbes over compact connected Lie groups. 5.1 Bundle gerbe modules Bundle gerbe modules, also known as twisted vector bundles, have been introduced in [BCM+ 02] in order to realize twisted K-theory geometrically. They are also the appropriate structure to extend the definition of holonomy to surfaces with boundary [CJM02]. Definition 5.1. Let G be a bundle gerbe over M . A G-module is a 1-morphism E : G → Iω for some 2-form ω ∈ Ω 2 (M ). The 2-form ω is called the curvature of the gerbe module. Let us compare this definition with the original definition of bundle gerbe modules in [BCM+ 02]. A left G-module E : G → Iω consists of a vector bundle E over Y and of an isomorphism : L ⊗ π2∗ E → π1∗ E of vector bundles over Y [2] which satisfies ∗ ∗ ∗ π13 ◦ (µ ⊗ id) = π23 ◦ π12 . (77) The similarity with an action of L on E justifies the notion of gerbe module. The curvature of E is restricted by (34) to

with n the rank of E.

1 tr(curv(E)) = π ∗ ω − C n

(78)

358


In terms of local data, a rank-n bundle gerbe module E: G → Iω is described by a collection (Gij , Πi ) of smooth functions Gij : Ui ∩ Uj → U (n) and u(n)-valued 1-forms Πi ∈ Ω 1 (Ui ) ⊗ u(n) which relate the local data of the bundle gerbes G and Iω in the following way: −1 1 = gijk · Gik G−1 jk Gij

G−1 ij Πi

0 = Aij + Πj − 1 ω = Bi + tr(dΠi ) n

on Ui ∩ Uj ∩ Uk , Gij −

G−1 ij

dGij

on Ui ∩ Uj ,

(79)

on Ui .

Note that the derivative of the last equality gives dω = dBi = curv(G).

(80)

Also note that if the bundle gerbe G is itself trivial, i.e., has local data (1, 0, C|Ui ) for a globally defined 2-form C ∈ Ω 2 (M ), then (Gij , Πi ) are the local data of a rank-n vector bundle over M with curvature of trace n (ω−C). This explains the terminology “twisted” vector bundle in the non-trivial case. According to (80), a necessary condition for the existence of a bundle gerbe module is that the curvature is an exact form. However, this is not the case in many situations, for example, for the bundle gerbes on compact Lie groups we have constructed in Section 3, whose curvature is the canonical 3-form H. For this reason, one often considers a pair (Q, E) of a submanifold Q ⊂ M together with a gerbe module E : G|Q → Iω for the restriction of the gerbe to this submanifold. In conformal field theory, the pair (Q, G) is also called a D-brane. In particular, we can consider this situation for the bundle gerbes over Lie groups G constructed in Section 3. In this case, the important submanifolds are conjugacy classes Q = Cλ , and we already know that the curvature curv(G) = H becomes exact when restricted to a conjugacy class, H|Cλ = dωλ . So the necessary condition (80) is satisfied. One can furthermore show [Gaw05] that precisely for integrable weights λ there exists a G|Cλ -module with curvature ωλ . This is the appropriate description of “flux stabilization of D-branes” in string theory [BDS00]. 5.2 Bundle Gerbe bimodules Bundle gerbe bimodules generalize bundle gerbe modules for one bundle gerbe G to a structure for two bundle gerbes. Definition 5.2 ([FSW]). Let G1 and G2 be bundle gerbes over M . A G1 -G2 bimodule is a morphism D : G1 → G2 ⊗ I (81) for some 2-form ∈ Ω 2 (M ). The 2-form is called the curvature of the bimodule.


359

This definition is related to the one of a gerbe module in the sense that—using the appropriate notion of duality for bundle gerbes [Wal07]—a G1 -G2 -bimodule is the same as a (G1 ⊗G2∗ )-module. We drop the discussion of the local data of a gerbe bimodule. However, it is clear that there is, analogously to equation (80), a necessary condition on the curvature of the two bundle gerbes, H1 = H2 + d.

(82)

It is again useful to consider bimodules for restrictions of the bundle gerbes to a submanifold Q ⊂ M . In particular, if M is the direct product of two manifolds M1 and M2 , each equipped with a bundle gerbe G1 and G2 , respectively, one considers p∗1 G1 |Q -p∗2 G2 |Q -bimodules, where pi : M → Mi is the projection on the ith factor. This is the setup to define holonomies around surfaces with defect lines [FSW]. Examples for such bimodules are again provided by compact Lie groups and the basic bundle gerbes thereon [FSW]. The study of such examples leads to the following relevant submanifolds Q of G × G, called the biconjugacy classes −1 Bh1 ,h2 := (x1 h1 x−1 (83) 2 , x1 h2 x2 ) ∈ G × G | x1 , x2 ∈ G for any pair (h1 , h2 ) ∈ G × G. Biconjugacy classes inherit two commuting actions of G from the diagonal left and diagonal right actions of G on G × G. One observes that the smooth map µ ˜ : G × G → G : (g1 , g2 ) → g1 g2−1 (84) intertwines the diagonal left and diagonal right action of G on G × G and the adjoint and trivial actions of G on itself, respectively. It now follows that a biconjugacy class in G × G is the preimage of a conjugacy class in G under the projection µ ˜: Bh1 ,h2 = µ (85) ˜−1 (Ch h−1 ) = (g1 , g2 ) ∈ G × G | g1 g2−1 ∈ Ch h−1 . 1

2

1

2

We introduce the 2-form ˜ ∗ ωh h1 ,h2 := µ

−1 1 h2

−

k ∗ p θ ∧ p∗2 θ 2 1

(86)

on Bh1 ,h2 , where both summands are restricted to the submanifold Bh1 ,h2 of G × G. From the intertwining properties of µ ˜ and the bi-invariance of ω, it follows that the 2-form is also bi-invariant. One can show that it satisfies p∗1 H = p∗2 H + dh1 ,h2

(87)

on a biconjugacy class Bh1 ,h2 , which is the necessary condition (82) [FSW].

360


5.3 Jandl structures Another structure we want to introduce is a Jandl structure on a bundle gerbe G. Jandl structures extend the definition of holonomy we gave in Section 2 to unoriented and, in particular, unorientable surfaces. Definition 5.3 ([SSW07]). A Jandl structure J on a bundle gerbe G over M is an involution k of M together with an isomorphism

and a 2-morphism

A : k∗ G → G ∗

(88)

ϕ : k ∗ A ⇒ A∗

(89)

which satisfies the equivariance condition k ∗ ϕ = ϕ∗ .

(90)

To give an impression of the details of a Jandl structure, recall that an isomorphism such as A = (A, α) consists of a line bundle A over the space Z which is built up from the two surjective submersions of the bundle gerbes k ∗ G and G ∗ . In this particular situation, there is a canonical lift k˜ of the involution k into the space Z, and it is in fact easy to work out that the ˜ 2-morphism ϕ defines a k-equivariant structure on the line bundle A, which is compatible with the isomorphism α. Summarizing, a Jandl structure J on G is an isomorphism A : k∗ G → G ∗ (91) ˜ whose line bundle A is equivariant with respect to the involution k on Z [SSW07]. Recall that we introduced equivariant structures on bundle gerbes in Section 3 in order to produce bundle gerbes over quotients of a manifold M by a discrete group Z. One can combine equivariant structures and Jandl structures into Z-equivariant Jandl structures, leading to the mathematically appropriate description of what are called orientifolds in string theory [GSW]. The idea behind this combination is that a bundle gerbe G over M with a Γ equivariant Jandl structure defines a bundle gerbe G with Jandl structure over the quotient M/Z. For a cohomological description, one modifies the action of Z on the Deligne cohomology group to an action of the semidirect product Γ := Z2 Z [GSW]. For bundle gerbes over compact Lie groups, where Z is a subgroup of the center of G, the relevant involutions are given by kz : G → G : g → (zg)−1

(92)

for any z in the center. Again, using the basic bundle gerbes constructed in Section 3, one can classify all equivariant Jandl structures over all these bundle gerbes [GSW].


361

6 Applications to conformal field theory Let us explain the relation between conformal field theory and Lie theory, which arises in the study of non-linear sigma models on a Lie group G. Such a model can be defined by amplitudes A(φ) for some path integral, where φ is a map from a closed complex curve Σ—the world sheet—into the target space G of the model. In [Wit84], Witten gives the following definition for G = SU (2). Σ is the boundary of a three-dimensional manifold B, and because the homotopy groups πi (SU (2)) vanish for i = 1, 2, every map φ : Σ → M can be extended into the interior B to a map Φ : B → G. Witten showed that, due to the integrality of the canonical 3-form H, A(φ) := exp Skin (φ) + Φ∗ H (93) B

depends neither on the choice of B nor on the choice of the extension Φ, so that one obtains a well-defined amplitude. Here Skin (φ) is a kinetic term, and, with a certain relative normalization of the two terms in (93), this model is called the Wess–Zumino–Witten model on G at level k. For non simply connected Lie groups, the extension Φ of φ to B does not exist in general. In these cases, the second summand of (93) has to be generalized. Proposition 6.1. Let G be a hermitian bundle gerbe over G with connection and curving of curvature H. For a three-dimensional oriented manifold B with boundary and a map Φ : B → G, we have holG (Φ|∂B ) = exp Φ∗ H . (94) B

Proof. Remember that for any trivialization T : Φ∗ G|∂B → Iρ we have Φ∗ H|∂B = dρ. The rest follows by Stokes’ theorem. This way we reproduce the amplitude of the coupling term of the Wess– Zumino–Witten model by A(φ) = exp (Skin (φ)) · holG (φ).

(95)

Notice that using bundle gerbes we did not impose any condition on the topology of the target space G. For compact, connected and simply connected Lie groups, it reproduces Witten’s original definition. However, for general target spaces there may be bundle gerbes with same curvature, which are not isomorphic. This occurs, for instance, for the Lie group Spin(4n)/(Z2 × Z2 ). Here, the theory of bundle gerbes over Lie groups has brought new insights into the Lagrangian description of Wess–Zumino–Witten models. In conformal field theory, many applications require us to consider surfaces with boundary. For them, we are not able to apply Definition 2.5 of the

362


holonomy of a bundle gerbe G: for a change of the chosen trivialization, a boundary term emerges which has to be compensated to achieve a holonomy independent of the choice of the trivialization. The compensating term is provided by the choice of a symmetric D-brane (Cλ , Eλ ): a conjugacy class Cλ for an integrable weight λ, together with a G|Cλ -module Eλ of curvature ωλ [Gaw05]. Another class of conformal field theories involves unoriented world sheets Σ. Again, Definition 2.5 has to be generalized, since it involves the integral of a differential form over Σ. It has been shown in [SSW07] that the choice of a Jandl structure on the bundle gerbe G again makes the holonomy well defined. The precise classification of Jandl structures leads to a complete classification of unoriented Wess–Zumino–Witten models [SSW07, GSW]. Let us also indicate the relevance of the line bundle L over the loop group LG that we have constructed in Section 4. The Hilbert space of holomorphic sections in L (completed with respect to its hermitian metric) serves as the space of states for the quantized theory [GR02]. The choice of additional structures, like bundle gerbe modules or Jandl structures, has implications on this space. For example, a Jandl structure on the bundle gerbe G implies by Proposition 4.3 an isomorphism ϕ : Lk ∗ L → L∗ which satisfies Lk ∗ ϕ = ϕ.

7 Open questions We conclude this contribution with the discussion of some lines for further research. One obvious direction is to extend the results explained here to gauged Wess–Zumino–Witten models, i.e., the coset theories. Those models having fixed points under the action of the group implementing field identifications for which the untwisted stabilizer is strictly smaller than the stabilizer [FRS04] should be particularly interesting: in this case, simple gerbe modules and bimodules of rank strictly bigger than one appear naturally. A precise understanding of such theories requires the notion of an H-equivariant gerbe (bi-)module on the ambient group G. Another subtle issue is the generalization of our results to non-compact Lie groups; this is partially due to the fact that much less is known about these theories in algebraic approaches. The following two points seem to be conceptually appealing questions: our initial motivating question in this paper was about central extensions of Lie groups. While hermitian bundle gerbes naturally account for a line bundle L on loop space, a more specific structure on the gerbe is needed to obtain a group structure on the line bundle L. Similarly, one wishes to find a structure on a gerbe module E that endows the associated bundles E over loop and interval spaces with a natural L-module structure. Finally, we point out that gerbe bimodules have a natural operation of fusion which is very much in the spirit of a convolution of correspondences.


363

Imposing additional conditions on gerbe bimodules over a compact Lie group (that in physical applications ensure the existence of enough conserved quantities), one should be able to single out interesting subcategories of gerbe bimodules that are semisimple tensor categories. If G is simply connected, the fusion ring of gerbe bimodules can be expected to be the corresponding Verlinde algebra; for non simply connected groups, we expect interesting cousins of the Verlinde algebra.

References [BCM+ 02] P. Bouwknegt, A. L. Carey, V. Mathai, M. K. Murray and D. Stevenson, Twisted K-Theory and K-Theory of Bundle Gerbes, Commun. Math. Phys. 228(1), 17–49 (2002), hep-th/0106194. [BDS00] C. Bachas, M. Douglas and C. Schweigert, Flux Stabilization of D-Branes, JHEP 0005(048) (2000), hep-th/0003037v2. ´ ements de Mathématique. Fasc. XXXIV. Groupes et [Bou68] N. Bourbaki, El´ algèbres de Lie. Chapitre IV–VI, Hermann, Paris, 1968. [BRS01] P. Bordalo, S. Ribault and C. Schweigert, Flux Stabilization in Compact Groups, JHEP 0110(036) (2001). [Bry] J.-L. Brylinski, Gerbes on Complex Reductive Lie Groups, math/0002158. [Bry93] J.-L. Brylinski, Loop Spaces, Characteristic Classes and Geometric Quantization, volume 107 of Progress in Mathematics, Birkh¨ auser, 1993. [Cha98] D. S. Chatterjee, On the Construction of Abelian Gerbes, Ph.D. thesis, Cambridge Univ., Cambridge, UK, 1998. [CJM02] A. L. Carey, S. Johnson and M. K. Murray, Holonomy on D-Branes, J. Geom. Phys. 52(2), 186–216 (2002), hep-th/0204199. [FRS04] J. Fuchs, I. Runkel and C. Schweigert, TFT Construction of RCFT Correlators III: Simple Currents, Nucl. Phys. B 694, 277–353 (2004), hep-th/0403157. [FSW] J. Fuchs, C. Schweigert and K. Waldorf, Bi-Branes: Target Space Geometry for World Sheet Topological Defects, hep-th/0703145, J. Geom. Phys. 58, 576–598 (2008). [Gaw88] K. Gaw¸edzki, Topological Actions in Two-Dimensional Quantum Field Theories, in Non-perturbative Quantum Field Theory, edited by G. Hooft, A. Jaffe, G. Mack, K. Mitter and R. Stora, pages 101–142, Plenum Press, New York, 1988. [Gaw05] K. Gaw¸edzki, Abelian and Non-Abelian Branes in WZW Models and Gerbes, Commun. Math. Phys. 258, 23–73 (2005), hep-th/0406072. [GR02] K. Gaw¸edzki and N. Reis, WZW Branes and Gerbes, Rev. Math. Phys. 14(12), 1281–1334 (2002), hep-th/0205233. [GR03] K. Gaw¸edzki and N. Reis, Basic Gerbe over Non Simply Connected Compact Groups, J. Geom. Phys. 50(1–4), 28–55 (2003), math.dg/0307010. [GSW] K. Gaw¸edzki, R. R. Suszek and K. Waldorf, WZW Orientifolds and Finite Group Cohomology, hep-th/0701071, Commun. Math. Phys. 284 1–49 (2008).

364


[GT01] [Mei02] [MS76] [MS00] [Mur96] [PS86] [SSW07] [Ste00] [Wal07] [Wit84]

K. Gomi and Y. Terashima, Higher-Dimensional Parallel Transports, Math. Research Letters 8, 25–33 (2001). E. Meinrenken, The Basic Gerbe over a Compact Simple Lie Group, Enseign. Math., II. Sr. 49(3–4), 307–333 (2002), math/0209194. J. W. Milnor and J. D. Stasheff, Characteristic Classes, Annals of Mathematical Studies, Princeton University Press, Princeton, NJ, 1976. M. K. Murray and D. Stevenson, Bundle Gerbes: Stable Isomorphism and Local Theory, J. Lond. Math. Soc. 62, 925–937 (2000), math/9908135. M. K. Murray, Bundle Gerbes, J. Lond. Math. Soc. 54, 403–416 (1996), dg-ga/9407015. A. Pressley and G. Segal, Loop Groups, Oxford Univ. Press, Oxford, 1986. U. Schreiber, C. Schweigert and K. Waldorf, Unoriented WZW Models and Holonomy of Bundle Gerbes, Commun. Math. Phys. 274(1), 31–64 (2007), hep-th/0512283. D. Stevenson, The Geometry of Bundle Gerbes, Ph.D. thesis, University of Adelaide, Australia, 2000, math.DG/0004117. K. Waldorf, More Morphisms Between Bundle Gerbes, Theory Appl. Categories 18(9), 240–273 (2007), math.CT/0702652. E. Witten, Nonabelian Bosonization in Two Dimensions, Commun. Math. Phys. 92, 455–472 (1984).

Part C

Representation Theory of Infinite-Dimensional Lie Groups

Functional Analytic Background for a Theory of Infinite-Dimensional Reductive Lie Groups Daniel Beltit¸a˘ Institute of Mathematics “Simion Stoilow” of the Romanian Academy, P.O. Box 1-764, RO-014700 Bucharest, Romania, [email protected] Summary. Motivated by the interesting and yet scattered developments in representation theory of Banach–Lie groups, we discuss several functional analytic issues which should underlie the notion of infinite-dimensional reductive Lie group: norm ideals, triangular integrals, operator factorizations, and amenability. Key words: reductive Lie group, group decomposition, amenable group, operator ideal, triangular integral. 2000 Mathematics Subject Classifications: Primary 22E65. Secondary 22E46, 47B10, 47L20, 58B25.

1 Introduction: What a reductive Lie group is supposed to be We approach the problem of finding an appropriate infinite-dimensional version of a reductive Lie group. The discussion is motivated by the need to have a reasonably general setting where the representation theoretic properties of the classical Lie groups associated with the Schatten ideals of Hilbert space operators—in the sense of [dlH72]—can be investigated in a systematic way. Thus the theory of operator ideals (see, e.g., [GK69], [GK70], [DFWW04], and [KW08]) provides the natural background for the present exposition. The ideas and methods of representation theory of finite-dimensional Lie groups cannot possibly be extended to the setting of Banach–Lie groups in a direct manner. Any attempt to do so fails because of some phenomena specific to the infinite-dimensional Lie groups: there exist Lie algebras that do not arise from Lie groups (see, e.g., Section VI in [Ne06] or Section 3.3 in [Be06]); closed subgroups of Lie groups need not be Lie groups in the relative topology (see Example 3.3(i) in [Ho75] or Example 7.12 in [Up85]); one does not know how to construct smooth structures on homogeneous spaces unless one is able to find a direct complement of the Lie subalgebra in the ambient Lie algebra (see


367

368

Daniel Beltit¸˘ a

Corollary 8.3 in [Up85], the discussions in [BP07] and [Ga06], or Section 4.1 [Be06]); there exists no Haar measure on topological groups which are not locally compact ([We40]); and finally every infinite-dimensional Banach space is the model space of some abelian Lie group without any nontrivial continuous representations (see Theorem 1 in [Ba83]). Nevertheless, the study of representation theoretic properties of some specific Banach–Lie groups has led to a number of interesting results; see for instance the papers [SV75], [Bo80], [Pi90], [Ne98], [NØ98], [BR07], or [BG08]. It seems reasonable to try to find a class of Banach–Lie groups whose representations can be studied in a coherent fashion following the pattern of representation theory of finite-dimensional reductive Lie groups. The aim of the present paper is to survey some of the ideas and notions that might eventually lead to the description of a class of Banach–Lie groups appropriate for the purposes of such a representation theory. The exposition is streamlined by a number of phenomena that play a central role in the classical theory of reductive Lie groups: Cartan decompositions, Iwasawa decompositions, Harish-Chandra decompositions, and existence of invariant measures. By describing appropriate versions of these phenomena in infinite dimensions, the paper provides a self-contained discussion of a number of functional analytic issues which should underlie the notion of an infinite-dimensional reductive Lie group: triangular integrals, operator factorizations, and amenability. Finite-dimensional reductive Lie groups. In order to make clear the structures we shall be looking for in infinite dimensions, we now recall the classical setting. Remarks 1.2 and 1.3 basically concern the matrix Lie algebras/groups, where reductivity means stability under the operation of taking adjoints of matrices. (These remarks will be our main motivation for the discussion of Φ-reductivity in Section 5.) The following general definition is the one used in [Kn96]. Definition 1.1. A finite-dimensional reductive Lie group is actually a 4-tuple (G, K, θ, B), where G is a finite-dimensional Lie group with the Lie algebra g, K is a compact subgroup of G with the Lie algebra k, θ : g → g is an involutive automorphism, and B : g × g → R is a nondegenerate symmetric bilinear form which is Ad(G)-invariant and θ-invariant, such that the following conditions are satisfied: (i) g is a reductive Lie algebra with the complexification denoted gC ; (ii) k = Ker (θ − 1); (iii) if we denote p = Ker (θ + 1), then g = k p and we have B(k, p) = {0}, B is positive definite on p, and negative definite on k; (iv) the mapping K × p → G, (k, X) → k · expG X, is a diffeomorphism; (v) for every g ∈ G the automorphism Ad(g) : gC → gC belongs to the connected component of 1 ∈ Aut (gC ). We say that the Lie group G itself is a reductive Lie group if the other items K, θ, and B are clear from the context.

Infinite-Dimensional Reductive Lie Groups

369

Remark 1.2. Let g be a real finite-dimensional Lie algebra. Then g is a reductive Lie algebra if and only if it is isomorphic to a real Lie subalgebra g1 of the matrix algebra Mn (C) for some integer n ≥ 1 such that for every X ∈ g1 we have X ∗ ∈ g1 . In addition, if g were a complex Lie algebra, then g1 could be chosen to be a complex Lie subalgebra of Mn (C). Remark 1.3. Let G be a connected closed subgroup of GL(n, C) for some n ≥ 1 such that for every g ∈ G we have g ∗ ∈ G. Then G is a reductive Lie group with K = G ∩ U(n), θ(X) = −X ∗ for all X ∈ g = L(G) ⊆ Mn (C), and B(X, Y ) = Re (Tr (XY )) for all X, Y ∈ g (see Example 2 in Section 2 of Chapter VII of [Kn96]). It is not difficult to prove that coverings with finitely many sheets of any matrix group G as above are reductive Lie groups —the covering groups of this type are precisely the connected reductive groups in the sense of Definition 2.5 of [Vo00]. (See Proposition 4.2 below for the way the Cartan decompositions lift to covering groups.) In particular, every connected semisimple Lie group with finite center is a reductive Lie group in the sense of Definition 1.1 above (Example 1 in Section 2 of Chapter VII of [Kn96]). Some conventions and notation. Throughout the paper we denote by H an infinite-dimensional complex separable Hilbert space, by B(H) the set of all bounded linear operators on H, and by F the two-sided ideal of B(H) consisting of all finite-rank operators. Some convenient references for infinitedimensional Lie groups with a functional analytic flavor are [dlH72], [Up85], [Ne04], [Ne06], and [Be06]. As in the latter reference, we shall always denote by L(·) the Lie functor from the category of Banach–Lie groups into the category of Banach–Lie algebras. We also adopt the convention that Lie groups are denoted by Roman capitals and their Lie algebras are denoted by the corresponding Gothic lower case letters.

2 Triangular integrals and factorizations Norm ideals. The norm ideals will play a critical role for the present exposition. In fact, our candidates for infinite-dimensional reductive groups will be Banach–Lie groups whose model spaces are norm ideals. Definition 2.1. By norm ideal we mean a two-sided ideal I of B(H) equipped with a norm · I satisfying T ≤ T I = T ∗ I and AT B I ≤ A T I B whenever A, B ∈ B(H). Definition 2.2. Let c be the vector space of all sequences of real numbers {ξj }j≥1 such that ξj = 0 for all but finitely many indices. A symmetric norming function is a function Φ : c → R satisfying the following conditions: (i) the function Φ(·) is a norm on the linear space c and Φ((1, 0, 0, . . . )) = 1; (ii) Φ({ξj }j≥1 ) = Φ({ξπ(j) }j≥1 ) whenever {ξj }j≥1 ∈ c and π : N\{0} → N\{0} is bijective.

370

Daniel Beltit¸˘ a (0)

Any symmetric norming function Φ gives rise to two norm ideals SΦ and SΦ as follows. For every bounded sequence of real numbers ξ = {ξj }j≥1 define Φ(ξ) := supn≥1 Φ(ξ1 , ξ2 , . . . , ξn , 0, 0, . . . ) ∈ [0, ∞]. For all T ∈ B(H) denote T Φ := Φ({sj (T )}j≥1 ) ∈ [0, ∞], where sj (T ) = inf{ T − F | F ∈ B(H), rank F < j} whenever j ≥ 1. With this notation we can define (0)

SΦ := F

·Φ

⊆ {T ∈ B(H) | T Φ < ∞} =: SΦ ;

(0)

that is, SΦ is the · Φ -closure of the ideal of finite-rank operators F in SΦ . (0) Then · Φ is a norm making SΦ and SΦ into norm ideals (see Section 4 in Chapter III in [GK69]). We say that Φ is a mononormalizing symmetric (0) norming function if SΦ = SΦ . Let 1 ≤ p ≤ ∞ and Φp (ξ) = ξ p whenever ξ∈ c defines a symmetric norming function (mononormalizing if p < ∞) and (0) the corresponding norm ideal Sp := SΦp is the p-th Schatten ideal. In the special case p = 2 we call S2 the Hilbert–Schmidt ideal. (0)

Remark 2.3. Every separable norm ideal is equal to SΦ for some symmetric norming function Φ (see Theorem 6.2 in Chapter III in [GK69]). Remark 2.4. For every symmetric norming function Φ : c → R there exists a unique symmetric norming function Φ∗ : c → R such that ∞ 1 ξj ηj Φ∗ (η) = sup Φ(ξ) j=1

c and ξ1 ≥ ξ2 ≥ · · · ≥ 0 ξ = {ξj }j≥1 ∈

whenever η = {ηj }j≥1 ∈ c and η1 ≥ η2 ≥ · · · ≥ 0. The function Φ∗ is said to be adjoint to Φ and we always have (Φ∗ )∗ = Φ (see Theorem 11.1 in Chapter III in [GK69]). For instance, if 1 ≤ p, q ≤ ∞, 1/p + 1/q = 1, Φp (ξ) = ξ p , and Φq (ξ) = ξ q whenever ξ ∈ c, then (Φp )∗ = Φq . If Φ is any symmetric norming (0) function, then the topological dual of the Banach space SΦ is isometrically (0) isomorphic to SΦ∗ by the duality pairing SΦ∗ × SΦ → C, (T, S) → Tr (T S) (see Theorems 12.2 and 12.4 in Chapter III in [GK69]). The next definition describes the Boyd indices of a symmetric norming function. Some convenient references for this notion are Section 3 in [Ar78] and subsections 2.17–19 in [DFWW04]. Definition 2.5. Let Φ be a symmetric norming function. For each m ≥ 1 define the linear operators Dm : c→ c and D1/m : c→ c by m 2m

1 1 Dm ξ = (ξ1 , . . . , ξ1 , ξ2 , . . . , ξ2 , . . . ) and D1/m ξ = ξi , ξi , . . . m i=1 m i=m+1 m times

m times


371

for arbitrary ξ = (ξ1 , ξ2 , . . . ) ∈ c. We shall think of Dm and D1/m as linear operators on the space c equipped with the norm Φ(·), so that it makes sense to speak about the norms of these operators. The Boyd indices of the symmetric norming function Φ are defined by log m m≥1 log Dm

pΦ = sup

and

log(1/m) m≥1 log D1/m

qΦ = inf

and we have 1 ≤ pΦ ≤ qΦ ≤ ∞. We shall say that these indices are nontrivial if 1 < pΦ ≤ qΦ < ∞. Remark 2.6. The Boyd indices of any symmetric norming function Φ are related to the Boyd indices of its adjoint Φ∗ by the equations 1/pΦ + 1/qΦ∗ = 1/pΦ∗ + 1/qΦ = 1. If 1 < r < ∞ and Φ(·) = · r then for all m ≥ 1 we have Dm = m1/r and D1/m = m−1/r , hence in this case pΦ = qΦ = r. The Boyd indices of symmetric norming functions are important for our present purposes because of the following interpolation theorem. Theorem 2.7. Let Φ be a mononormalizing symmetric norming function and assume that 1 ≤ p < pΦ ≤ qΦ < q ≤ ∞. Then the following assertions hold: (a) We have Sp ⊆ SΦ ⊆ Sq . (b) There exists a constant MΦ > 0 with this property: if T : Sq → Sq is a bounded linear operator such that T (Sp ) ⊆ Sp , then T (SΦ ) ⊆ SΦ and T |SΦ ≤ MΦ max{ T |Sp , T |Sq }. Proof. See Corollary 3.4(i) in [Ar78].

Triangular integrals. The triangular integral is a suitable infinite-dimensional version of the operation of taking the upper triangular part (the lower diagonal part, or the diagonal, respectively) of a square matrix. The classical reference is [GK70] (see also [Er78]). We are going to describe this idea in a setting which is slightly more general than the classical one. Definition 2.8. Let B be a unital associative involutive Banach algebra and I a contractive B-bimodule; that is, I is a Banach space equipped with an involutive isometric antilinear map I → I, X → X ∗ and a trilinear map B × I × B → I,

(b1 , X, b2 ) → b1 Xb2 ,

such that I is an algebraic B-bimodule and in addition (b1 Xb2 )∗ = b∗2 X ∗ b∗1 and b1 Xb2 ≤ b1 X b2 whenever b1 , b2 ∈ B and X ∈ I. Also let E be a totally ordered set of self-adjoint idempotent elements in B such that 0, 1 ∈ E, and denote by Part(E) the set of all partitions of E, that is, the finite families P = {pi }0≤i≤n in E such that 0 = p0 < p1 < · · · < pn = 1. For such a partition P of E we define the diagonal truncation operator DP : I → I,

DP (X) =

n i=1

(pi − pi−1 )X(pi − pi−1 ),

372

Daniel Beltit¸˘ a

the strictly upper triangular truncation operator UP : I → I,

UP (X) =

n

pi−1 X(pi − pi−1 ),

i=1

and the strictly lower triangular truncation operator LP : I → I,

LP (X) =

n

(pi − pi−1 )Xpi−1 .

i=1

Now think of Part(E) as a directed set with respect to the partial order defined by P ≤ Q if and only if P is a subfamily of Q. If X ∈ I and the net {UP (X)}P∈Part(E) is convergent in I, then the corresponding limit is denoted by U(X) and is called the strictly upper triangular integral of X. One similarly defines the strictly lower triangular integral L(X) and the diagonal integral D(X) whenever they exist. Proposition 2.9. Let B be a unital associative involutive Banach algebra, I a contractive B-bimodule, and E a totally ordered set of self-adjoint idempotents in B such that 0, 1 ∈ E. Assume that the following conditions are satisfied. (i) Either the Banach space underlying I is reflexive or the integrals D and U exist on dense subsets of I. (ii) Both families of operators {DP }P∈Part(E) and {UP }P∈Part(E) are uniformly bounded. Then the following assertions hold. (a) The strictly upper triangular integral, the strictly lower triangular integral, and the diagonal integral exist throughout I and the corresponding mappings are bounded linear idempotent operators U, L, D : I → I. (b) There exists the direct sum decomposition I = Ran L Ran D Ran U, and the corresponding decomposition of an arbitrary element X ∈ I is X = L(X) + D(X) + U(X). In addition, L(X ∗ ) = U(X)∗ and D(X ∗ ) = D(X)∗ for all X ∈ I. Proof. If the integrals D and U are convergent on dense subsets of I, then they exist everywhere because of hypothesis (ii). Now let us assume that the Banach space underlying I is reflexive. It is easy to check that for all P ∈ Part(E) all of the operators LP , DP , and UP are idempotent, their mutual products are equal to 0, and their sum is the identity mapping on I. In addition, the nets {UP (X)}P∈Part(E) and {LP (X)}P∈Part(E) are increasing, while {DP (X)}P∈Part(E) is decreasing, in the sense that if P, Q ∈ Part(E) and P ≤ Q, then UP UQ = UQ UP = UP , LP LQ = LQ LP = LP , and also


373

DP DQ = DQ DP = DQ . Now hypotheses (i) and (ii) ensure that Theorem 2.1 of [Er78] applies, whence assertion (a) follows. For assertion (b) note that X = LP (X) + DP (X) + UP (X) and then take the limit with respect to P ∈ Part(E), for each X ∈ I. For all P, Q ∈ Part(E) we have LP DP = 0, whence by taking the limit we get LD = 0. Similarly, all of the mutual products of the operators L, D, and U, are equal to 0, whence the asserted direct sum decomposition follows. To complete the proof note that for all X ∈ I and P ∈ Part(E) we have LP (X ∗ ) = UP (X)∗ and DP (X ∗ ) = DP (X)∗ , and then again take the limit with respect to the partition P of E. Example 2.10. (See Theorem 4.1 in [Ar78].) Let Φ be a mononormalizing symmetric norming function whose Boyd indices are nontrivial and denote I = SΦ ⊆ B(H). Denote by E the set of orthogonal projections associated with a maximal, totally ordered set of closed subspaces of H. Then the corresponding strictly upper triangular, the strictly lower triangular, and the diagonal integral exist throughout I, and they define bounded linear idempotent operators U, L, D : I → I whose mutual products are equal to 0 and U + L + D = 1. Indeed, condition (i) in Proposition 2.9 is satisfied since the integrals of finite-rank operators are clearly convergent. Also, it follows by Theorem 3.2 in [Er78] that the families of operators {DP }P∈Part(E) and {UP }P∈Part(E) are uniformly bounded on each Schatten ideal Sp if 1 < p < ∞. Now Theorem 2.7 shows that both these families are uniformly bounded on I = SΦ as well, since the Boyd indices of Φ are nontrivial. Thus condition (ii) in Proposition 2.9 is satisfied as well, and it then follows that the integrals U, D, and L are convergent throughout I. Triangular factorizations. Our next purpose is to survey some of the methods that allow one to find operator theoretic versions of the well-known LU factorization from linear algebra, that is, the fact that every invertible matrix factorizes as the product of a unitary matrix and a triangular one. Definition 2.11. Let E be the set of orthogonal projections associated with a maximal, totally ordered set of closed linear subspaces of H. Then the nest algebra of E is Alg E = {b ∈ B(H) | (∀e ∈ E) be = ebe}, that is, the set of all operators which leave invariant each subspace in the family {e(H)}e∈E . Definition 2.12. Let SI and SII be norm ideals. We shall say that (SI , SII ) is a pair of associated norm ideals if the following conditions are satisfied: (i) SI is separable; (ii) SI ⊆ SII ; (iii) for every maximal, totally ordered set of closed linear subspaces of H and every X ∈ SI , the corresponding strictly upper triangular integral is convergent in the contractive B(H)-bimodule I = B(H) and U(X) ∈ SII .

374

Daniel Beltit¸˘ a

(See [GK70] and [Er72] for the original version of this definition.)

Example 2.13. Let Φ be a mononormalizing symmetric norming function whose Boyd indices are nontrivial, and denote SI = SII = SΦ . Then Example 2.10 shows that (SI , SII ) is a pair of associated norm ideals. As a special case, it follows by Remark 2.6 that each pair (Sp , Sp ) consisting of the p-th Schatten ideal and itself is a pair of associated norm ideals if 1 < p < ∞. Theorem 2.14. Let (SI , SII ) be a pair of associated norm ideals and consider the triangular and diagonal integrals on the contractive B(H)-bimodule I = B(H) with respect to the set E of orthogonal projections associated with some maximal, totally ordered set of closed linear subspaces in H. Also assume that 0 ≤ a ∈ GL(H). Then the following conditions are equivalent: (i) U(a−1 ) exists and U(a−1 ) ∈ SI . (ii) D(a−1 ) exists and a−1 − D(a−1 ) ∈ SI . If one of these conditions holds, then there exist uniquely determined elements r ∈ SII ∩ Alg E and d = d∗ ∈ Alg E such that a = (1 + r)d(1 + r∗ ) and the spectrum of r is equal to {0}. Proof. See Theorem 4.2 and Lemma 2.5(i) in [Er72].

Here are two corollaries concerning the group GLΦ (H) of Example 5.2 below. In connection with these corollaries as well as with Theorem 2.14 above it is worth noting that for every element r whose spectrum in some unital associative Banach algebra is equal to {0}, the element 1 + r is invertible in that Banach algebra. This remark applies in particular for elements in the nest algebra Alg E. Corollary 2.15. Let Φ be a mononormalizing symmetric norming function whose Boyd indices are nontrivial, and denote by E the set of orthogonal projections associated with a maximal, totally ordered set of closed linear subspaces of H. If 0 ≤ a ∈ GLΦ (H), then there exist uniquely determined operators d, r ∈ B(H) such that 0 ≤ d ∈ GLΦ (H) ∩ Alg E, r ∈ SΦ ∩ Alg E, the spectrum of r is equal to {0}, and a = (1 + r∗ )d(1 + r). Proof. See Theorem 2.14 and Example 2.13.

Corollary 2.16. Assume the setting of Corollary 2.15. Then for every g ∈ GLΦ (H) there exist the operators b, u ∈ GLΦ (H) such that b ∈ Alg E, u∗ u = 1, and g = ub. Proof. Just apply Corollary 2.15 for a = g ∗ g. See for instance the proof of Corollary A.2 in [Be09] for more details.


375

3 Invariant means on groups Amenable groups. We shall briefly discuss the invariant means on topological groups. They can be thought of as weak versions of Haar measures, although they have two main drawbacks: they may not be faithful, in the sense that the mean of a nonzero function with nonnegative values can be equal to zero; and not every Lie group admits an invariant mean. On the other hand, we shall see that many Banach–Lie groups do have invariant means; see for instance Remark 5.6. Classical references for amenability are [Pa88] and [Ey72]. See [Pe06], [Ga06], and [BP07] for some recent developments. Definition 3.1. Let G be a topological group. Consider the commutative unital C ∗ -algebra ∞ (G) = {ψ : G → C | ψ ∞ := supG |ψ(·)| < ∞} and its automorphisms Lx , Rx : ∞ (G) → ∞ (G), (Lx ψ)(y) = ψ(xy), and (Rx ψ)(y) = ψ(yx) whenever y ∈ G and ψ ∈ ∞ (G), defined for arbitrary x ∈ G. The space of right uniformly continuous bounded functions on G is RUCb (G) = {ψ ∈ ∞ (G) | the map G → ∞ (G), x → Lx ψ, is continuous}. Similarly, the space of left uniformly continuous bounded functions on G is the set LUCb (G) of all functions ψ ∈ ∞ (G) such that the mapping G → ∞ (G), x → Rx ψ, is continuous. We say that the topological group G is amenable if there exists a linear functional µ : RUCb (G) → C satisfying the following conditions: (i) µ(1) = 1, (ii) 0 ≤ µ(ψ) if 0 ≤ ψ ∈ RUCb (G), and (iii) µ(Lx ψ) = µ(ψ) for all ψ ∈ RUCb (G) and x ∈ G. In this case the linear functional µ is called a left invariant mean on G.

Remark 3.2. If the topological group G is amenable, then every left invariant mean µ is continuous on RUCb (G) and µ = 1. On the other hand, the space RUCb (G) is a unital C ∗ -subalgebra of ∞ (G) which consists only of continuous functions, and it is invariant under the automorphisms Lx for all x ∈ G. Thus the left invariant means on G are precisely the states of the commutative unital C ∗ -algebra RUCb (G) which are invariant under the automorphism group defined by the mappings Lx for arbitrary x ∈ G. Example 3.3. If the topological group G is either compact or abelian, then it is amenable. In fact, let C(G) denote the space of all continuous functions on G. If G is compact, then the probability Haar measure defines a linear functional µ : C(G) → C. Compactness of G implies that RUCb (G) = C(G), and then the basic properties of the Haar measure show that µ is a left invariant mean on G. On the other hand, if G is abelian, denote by Gd the group G endowed with the discrete topology. Then RUCb (Gd ) = ∞ (G), and it follows by (0.15) in [Pa88] that there exists a left invariant mean µ : ∞ (Gd ) → C on the discrete group Gd . Now the restriction of µ to RUCb (G) defines a left invariant mean on G.

376

Daniel Beltit¸˘ a

Example 3.4. Assume that G is a topological group such that there exists a directed system of amenable topological subgroups {Gα }α∈A whose union is dense in G. Then G is amenable. To see this, we shall say that a linear functional µ : RUCb (G) → C is a mean on G if µ(1) = 1 and 0 ≤ µ(ψ) whenever 0 ≤ ψ ∈ RUCb (G). In this case µ is continuous and µ = 1. Now for every α ∈ A denote Λα = {µ | µ is a mean on G and µ ◦ Lx |RU Cb (G) = µ if x ∈ Gα }, which is a w∗ -compact subset of the unit ball in the topological dual space (RUCb (G))∗ . In addition, Λα = ∅. In fact, any left invariant mean µα on Gα gives rise to an element µ α ∈ Λα defined by µ α (ψ) = µα (ψ|Gα ) for all ψ ∈ RUCb (G). On the other hand, if α, β ∈ A and Gα ⊆ Gβ , then Λα ⊇ Λβ . Thus {Λα }α∈A is a family of w∗ -compact subsets of the unit ball in (RUCb (G))∗ with the property that the intersection of each finite subfamily is nonempty.

Therefore Λα = ∅, and any element µ in this nonempty intersection is a α∈A

left invariant mean on G. In fact, we already know that µ is a mean on G. To check that it is left invariant, let x ∈ G be arbitrary. Since the union of the family {Gα }α∈A is dense in G, there exists a net {xi }i∈I in that union such that lim xi = x. Then for every ψ ∈ RUCb (G) we have lim Lxi ψ = Lx ψ i∈I

i∈I

in ∞ (G); hence µ(Lx ψ) = lim µ(Lxi ψ) = lim µ(ψ) = µ(ψ), where the second i∈I i∈I

equality follows since µ ∈ Λα and xi ∈ Gα for all i ∈ I. α∈A

α∈A

Note that the above reasoning leads to the same conclusion under a more general hypothesis—namely that {Gα }α∈A is a directed system of abstract subgroups of G whose union is dense in G, such that each of these subgroups is endowed with a group topology that makes it an amenable topological group continuously embedded into G such that the inclusion map Gα → Gβ is continuous if α ≤ β in A. Example 3.5. Let G be a finite-dimensional Lie group with finitely many connected components. Denote by R the radical of G (i.e., the connected subgroup of G corresponding to the largest solvable ideal of the Lie algebra of G) and by Gd the group G endowed with the discrete topology. (a) The group G is amenable if and only if it has either of the following properties: (i) the group G/R is compact; (ii) there exists no closed subgroup of G isomorphic to the free group F2 with two generators. (b) The discrete group Gd is amenable if and only if either of the following conditions is satisfied: (j) the group G/R is finite (i.e., the connected component of 1 ∈ G is a solvable Lie group); (jj) there exists no subgroup of Gd isomorphic to F2 . We refer to Theorems (3.8) and (3.9) in [Pa88] for proofs of these facts.


377

→ G is a surjective homomorphism of topologRemark 3.6. Assume that φ : G is amenable ical groups such that Ker φ is an abelian group. Then the group G if and only if G is amenable. In fact, it follows by Example 3.3 that the topo is amenable, and then the assertion follows for logical subgroup Ker φ of G ◦ instance by remark 2 in Section 3 of Exposé no .1 of [Ey72]. and every continuWe note that for every amenable topological group G ous homomorphism φ : G → G with dense range it follows that G is amenable → C is a left without any extra condition on Ker φ. In fact, if µ : RUCb (G) then it is straightforward to check that the corresponinvariant mean on G, dence f → µ (f ◦ φ) defines a left invariant mean on G. Mimicking the group algebras of compact groups Definition 3.7. Let G be a topological group and consider the duality pairing ·, · : (RUCb (G))∗ × RUCb (G) → C. There exists a bounded bilinear map (RUCb (G))∗ × RUCb (G) → RUCb (G),

(µ, ψ) → µ · ψ

defined by (µ · ψ)(x) = µ, Lx ψ for all x ∈ G (by (2.11) in [Pa88]). The Arens-type product on (RUCb (G))∗ is the bounded bilinear mapping (RUCb (G))∗ × (RUCb (G))∗ → (RUCb (G))∗ ,

(µ, ν) → µ · ν

defined by µ · ν, ψ := µ, ν · ψ for all ψ ∈ RUCb (G). Similarly, by using the duality pairing ·, · : (LUCb (G))∗ × LUCb (G) → C, one defines a bounded bilinear map (LUCb (G))∗ × LUCb (G) → LUCb (G),

(µ, ψ) → µ · ψ

by (µ · ψ)(x) = µ, Rx ψ for all x ∈ G (by the version of (2.11) in [Pa88] for left uniformly continuous functions). The Arens-type product on (LUCb (G))∗ is the bounded bilinear mapping (LUCb (G))∗ × (LUCb (G))∗ → (LUCb (G))∗ ,

(µ, ν) → µ · ν

defined by µ · ν, ψ := ν, µ · ψ for all ψ ∈ LUCb (G).

Remark 3.8. Let G be a topological group, consider the space Cb (G) consisting of the continuous elements of ∞ (G), and define σ : Cb (G) → Cb (G), (σ(ψ))(x) = ψ(x−1 ) whenever x ∈ G and ψ ∈ Cb (G). The mapping σ is an antilinear isometric ∗-endomorphism of the unital C ∗ -algebra Cb (G) which satisfies σ 2 = idCb (G) and has the following additional properties: (1) For every x ∈ G we have Lx ◦ σ = σ ◦ Rx−1 . (2) We have σ(RUCb (G)) = LUCb (G) and σ(LUCb (G)) = RUCb (G).

378

Daniel Beltit¸˘ a

(3) If we define Σ : (RUCb (G))∗ → (LUCb (G))∗ as the anti-dual map of σ : LUCb (G) → RUCb (G), that is, Σ(µ), φ = µ, σ(φ) for all µ ∈ (RUCb (G))∗ and φ ∈ LUCb (G), then the diagram (RUCb (G))∗ × RUCb (G) −−−−→ RUCb (G) ⏐ ⏐ ⏐ ⏐σ Σ×σ (LUCb (G))∗ × LUCb (G) −−−−→ LUCb (G) is commutative,where the horizontal arrows are the maps introduced in Definition 3.7. (4) The mapping Σ : (RUCb (G))∗ → (LUCb (G))∗ is antilinear, isometric, bijective, and Σ(µ1 · µ2 ) = Σ(µ2 ) · Σ(µ1 ) for all µ1 , µ2 ∈ (RUCb (G))∗ . In fact, (1) follows by a straightforward computation, and it implies property (2) at once. For property (3) note that for all µ ∈ (RUCb (G))∗ and ψ ∈ RUCb (G) we have (Σ(µ) · σ(ψ))(x) = Σ(µ), Rx (σ(ψ)) = Σ(µ), σ(Lx−1 (ψ)) = µ, σ 2 (Lx−1 (ψ)) = µ, Lx−1 (ψ) = σ(µ · ψ)(x) whenever x ∈ G. To check property (4), compute (3)

Σ(µ2 ) · Σ(µ1 ), ϕ = Σ(µ1 ), Σ(µ2 ) · ϕ = Σ(µ1 ), σ(µ2 · σ(ϕ)) = µ1 , µ2 · σ(ϕ) = µ1 · µ2 , σ(ϕ) = Σ(µ1 · µ2 ), ϕ

for every ϕ ∈ LUCb (G).

The facts described in the following theorem can be found in Theorems 2.5 and 4.1 of [Gr]. See the paper [FN07] and the references therein for another line of research on these Banach algebras. Theorem 3.9. Every topological group G has the following properties. (a) The Arens-type products make (RUCb (G))∗ and (LUCb (G))∗ into associative Banach algebras. (b) Each continuous unitary representation π : G → B(Hπ ) gives rise to two representations of Banach algebras, π R : (RUCb (G))∗ → B(Hπ ) and π L : (LUCb (G))∗ → B(Hπ ), by means of the formulas πL (ν)ξ | η) = ν, (π(·)ξ | η) ( πR (µ)ξ | η) = µ, (π(·)ξ | η) and ( ∗

(1)

∗

for all ξ, η ∈ H, µ ∈ (RUCb (G)) , and ν ∈ (LUCb (G)) . These representations are related by the commutative diagram π

(RUCb (G))∗ −−−R−→ B(Hπ ) ⏐ ⏐ ⏐ ⏐ Σ S π

(LUCb (G))∗ −−−L−→ B(Hπ ) where S : B(Hπ ) → B(Hπ ), b → b∗ .

(2)


379

Proof. Assertion (a) follows at once by (2.8) and (the left-sided version of) (2.11) in [Pa88]. See also Example (19.23)(b) in [HR63] and [Bu50]. For assertion (b), firstly note that the matrix coefficients ψξ,η = (π(·)ξ | η) belong to the function space RUCb (G) ∩ LUCb (G) for arbitrary ξ, η ∈ Hπ . To see this, just note that for all x ∈ G we have Lx (ψξ,η ) = ψξ,π(x)∗ η and Rx (ψξ,η ) = ψπ(x)ξ,η , and then use the continuity of the representation π. Thus the right-hand sides of both equalities in (1) make sense, and then by means of the estimate ψξ,η ∞ ≤ ξ η we get π R (µ), π L (ν) ∈ B(Hπ ) and πR (µ) ≤ µ and πL (ν) ≤ ν . In addition, since σ(ψξ,η ) = ψη,ξ , we get πR (µ)η | ξ) = ( πL (Σ(µ))ξ | η) = Σ(µ), ψξ,η = µ, σ(ψξ,η ) = µ, ψη,ξ = ( ( πR (µ)∗ ξ | η), whence π L (Σ(µ)) = π R (µ)∗ , and thus the diagram (2) is commutative. To conclude the proof of (b), let µ1 , µ2 ∈ (RUCb (G))∗ . Then ( πR (µ1 · µ2 )ξ | η) = µ1 · µ2 , ψξ,η = µ1 , µ2 · ψξ,η = µ1 , ψπR (µ2 )ξ,η = ( πR (µ1 ) πR (µ2 )ξ | η), where the next-to-last equality holds since for every x ∈ G we have (µ2 · ψξ,η )(x) = µ2 , Lx (ψξ,η ) = µ2 , ψξ,π(x)∗ η = ( πR (µ2 )ξ | π(x)∗ η) = (π(x) πR (µ2 )ξ | η) = ψπR (µ2 )ξ,η (x). Thus π R is an algebra representation. Similarly, for ν1 , ν2 ∈ (LUCb (G))∗ we can check that ν1 · ψξ,η = ψξ,πL (ν1 )∗ η , whence ( πL (ν1 · ν2 )ξ | η) = ( πL (ν1 ) πL (ν2 )ξ | η) just as above, and the proof ends. Remark 3.10. In the setting of Theorem 3.9, RUCb (G) and LUCb (G) are commutative unital isomorphic C ∗ -algebras, hence there exist ∗-isomorphisms RUCb (G) C(M0 (G)) LUCb (G), where M0 (G) is a compact topological space. See the main theorem of [Tu95] for a description of this compact space (called the universal ambit of G) when G is the additive group (R, +) with the usual topology. In the special case when the group G is compact, we have M0 (G) = G, C(G) = RUCb (G) = LUCb (G), and the involutive Banach algebra (RUCb (G))∗ is the convolution measure algebra of G. In the general case, the topological duals of RUCb (G) and LUCb (G) still consist of the (not necessarily positive) measures on the compact space M0 (G); however, the Arens-type products may not be defined by convolution formulas by (cf. the comment preceding Proposition (2.25) in [Pa88]) reasonable convolution formulas involving M0 (G). See however the general theory of convolutions of functionals developed in Chapter 5 of [HR63] or the other references mentioned in connection with the proof of Theorem 3.9(a) above. Another problem to be dealt with is to find a method to distinguish, in the set of pairs of representations of the Banach algebras (RUCb (G))∗ and (LUCb (G))∗ that make the diagram (2) commutative, the ones that come from unitary representations of G by means of (1). In this connection, let us recall that there exists a promising approach to an axiomatic theory of group algebras by means of the host algebras, that is, C ∗ -algebras whose representations correspond in a one-to-one fashion with the unitary representations of a given group; see [Gr05], [GN09], and the references therein.

380

Daniel Beltit¸˘ a

Remark 3.11. Assume that G is an amenable topological group and pick a left invariant mean µ : RUCb (G) → C. Then µ is a state of the C ∗ -algebra RUCb (G), and the corresponding Gelfand–Naimark–Segal construction leads to a cyclic ∗-representation ιµ : RUCb (G) → B(H(µ) ). Recall that the Hilbert space H(µ) is obtained out of RUCb (G) as a quotient followed by a completion with respect to the nonnegative definite, sesquilinear form RUCb (G) × RUCb (G) → C, (ψ, χ) → µ(ψχ∗ ). For each x ∈ G we have µ ◦ Lx−1 |RU Cb (G) = µ; hence the mapping RUCb (G) → RUCb (G), ψ → Lx−1 ψ induces a unitary representation λµ : G → B(H(µ) ), which is easily seen to be continuous. In the special case when G is compact and µ is the probability Haar measure on G, we have H(µ) = L2 (G, µ) and λµ is the regular representation of G. For this reason, in the general case of an amenable group G, one can think of λµ as a regular representation associated with the left invariant mean µ. We note that the universal ambit M0 (G) mentioned in Remark 3.10 carries both a continuous action of G (see Sections 2.1 and 2.3 in [Pe06]) and a probability Borel measure µ induced by mean µ; this measure is invariant under the action of G, and the corresponding unitary representation of G on L2 (M0 (G), µ) is unitarily equivalent to the representation λµ . We have to point out that it may happen that dim H(µ) = 1. For instance, this is the case if G is an extremely amenable group and the left invariant mean µ is chosen to be multiplicative. (See [Pe06] for specific examples and for details on the latter notions.) It may also happen that the regular representation is trivial, in the sense that λµ (x) = 1 for all x ∈ G. This is the case for the exotic Banach–Lie groups (see Theorem 1 in [Ba83]), which are abelian topological groups that do not have any non-trivial continuous representation. It is not difficult to verify that the regular representation λµ is trivial if and only if µ(χψ) = µ(χLx ψ) for all χ, ψ ∈ RUCb (G) and x ∈ G.

4 Lifting group decompositions to covering groups This section has a technical character, and its main purpose is to provide tools for enriching the class of reductive Banach–Lie groups to be set forth in the next section. In the proof of the following statements we use some ideas from the proofs of Theorem 6.31 and 6.46 in [Kn96]. are Banach–Lie groups, and e : G → G is Lemma 4.1. Assume that G and G a covering homomorphism. Let K be any Banach–Lie subgroup of G, denote K → G/K, → e( := e−1 (K), and define ψ : G/ gK g )K. Then the following K assertions hold. is a Banach–Lie subgroup of G and the mapping e| : K → K is a (i) K K covering homomorphism. (ii) The mapping ψ is a well-defined diffeomorphism. is connected and the smooth homogeneous space G/K is (iii) If the group G and K are connected. simply connected, then both K


381

→ L(G) is an isomorphism of Banach–Lie alProof. Since L(e) = T1 e : L(G) is a Banach–Lie subgroup gebras, it follows by Proposition 4.8 in [Be06] that K → T1 K. Then Remark C.13(b) in of G and the tangent map L(e)|T1 K : T1 K [Be06] shows that e|K : K → K is a covering map. This completes the proof of assertion (i). = e−1 (K), and ψ is To prove assertion (ii), note that ψ is injective since K surjective because e is surjective. On the other hand, we have a commutative diagram −−−e−→ G G ⏐ ⏐ ⏐ ⏐ ψ K −−− G/ −→ G/K

where the vertical arrows (which are the quotient maps) are submersions. Since the covering map e is a local diffeomorphism, it follows from this commutative diagram that ψ is a local diffeomorphism as well. Then ψ is actually a diffeomorphism since we have already seen that it is bijective. is a covering group of K, so it For assertion (iii), recall from (i) that K will be enough to show that K is connected. And the latter property follows from the long exact sequence of homotopy groups ← π0 (K) ← π1 (G/ K) ← π1 (G) ← π1 (K) ← ··· π0 (G) = {0} by the assumption on G, while π1 (G/ K) = {0} by the since π0 (G) K assumption on G/K along with the fact that G/K is homeomorphic to G/ according to assertion (ii). We now come to a proposition to the effect that the Cartan decompositions lift to the covering groups. be two Banach–Lie groups, and assume that Proposition 4.2. Let G and G → G is a covering homomorphism. Now let K be any Banach–Lie sube: G := e−1 (K). Denote L(G) = g and L(K) = k, and group of G and denote K assume that p is a closed linear subspace of g such that g = k p and the mapping ϕ : K × p → G, (k, X) → k · expG X is a diffeomorphism. Moreover, × ( → denote p := L(e)−1 (p). Then the mapping ϕ : K p → G, k, X) k · expG X is a diffeomorphism as well. is a Banach–Lie subgroup of G by Lemma 4.1, Proof. First note that K → L(G) is an isomorphism of Banach–Lie algebras, so that and L(e) : L(G) L(e) k = k. Now note that there exists a commutative diagram ϕ × K p −−−−→ ⏐ e|K ×L(e)|p ⏐ ϕ

G ⏐ ⏐e

K × p −−−−→ G

382

Daniel Beltit¸˘ a

whose vertical arrows are covering maps (see Lemma 4.1(i)). Since ϕ is a diffeomorphism by assumption, it then follows that ϕ is a local diffeomorphism. To get the wished-for conclusion, we still have to prove that ϕ is a bijective map. j and X j ∈ To check that ϕ is injective, let kj ∈ K pj for j = 1, 2 such that k1 · expG X1 = k2 · expG X2 . By applying the map e to both sides of the latter equality, and using the commutation relation between the exponential maps and group homomorphisms (see, e.g., Remark 2.34 in [Be06]), we get 1 ) = e( 2 ). Since ϕ is injective, it then e( k1 ) · expG (L(e)X k2 ) · expG (L(e)X k2 ). The first of these equalities follows that L(e)X1 = L(e)X2 and e(k1 ) = e( 1 = X 2 , whence j and implies that X k1 = k2 by the assumption on kj ∈ K j ∈ arbitrary. X pj for j = 1, 2. Now, to prove that ϕ is surjective, let g∈G Then e( g ) ∈ G, hence there exist k ∈ K and X ∈ p such that e( g ) = k·expG X := since ϕ is surjective. Further on, pick k0 ∈ e−1 (k) arbitrary and denote X −1 L(e) X. Then e( g) = e(k0 ) · expG (L(e)X) = e(k0 · expG X), so that by Thus · denoting z := k0 ·expG X k := z −1 g −1 we have z ∈ e−1 (1) ⊆ K. k0 ∈ K which concludes the proof. and we have g = k · expG X, The next proposition shows that the familiar integration of local Cartan involutions to global Cartan involutions carries over to the setting of Banach– Lie groups. Proposition 4.3. Let G be a connected Banach–Lie group with the Lie algebra L(G) = g, K a Banach–Lie subgroup of G, and assume that there exists an automorphism θ ∈ Aut (g) such that θ2 = idg , L(K) = Ker (θ − idg ), and the mapping ϕ : K × p → G, (k, X) → k · expG X is a diffeomorphism. Then there exists a unique automorphism Θ ∈ Aut (G) such that L(Θ) = θ and K = {g ∈ G | Θ(g) = g}. → G be the universal covering of G, and denote k = L(K), Proof. Let e : G := e−1 (K), p := Ker (θ + idg ), K g := L(G), p := L(e)−1 (p), and θ := −1 L(e) ◦ θ ◦ L(e) ∈ Aut ( g). Then Proposition 4.2 shows that the mapping is a diffeomorphism. ϕ : K × p → G, (k, X) → k · expG X is connected and simply connected, On the other hand, since the group G : G →G such it follows that there exists a unique smooth homomorphism Θ 2 that L(Θ) = θ. (See for instance Remark 3.13 in [Be06].) Since θ = idg , it 2 = id , and in particular Θ ∈ Aut (G). then follows that Θ G Now we use hypothesis (iii) to see that the mapping τ : p → G/K, X → (expG X)−1 K is a diffeomorphism. In fact, it is clear that this is a smooth map, and its inverse is the smooth well-defined map G/K → p, gK → prp (g −1 ), where prp : G → p is the projection onto p defined by means of the diffeomorphism ϕ−1 : G → K × p. Thus τ : p → G/K is a diffeomorphism, and in particular G/K is simply connected since p is. Then Lemma 4.1 are connected. Since θ| = id and K is connected, it shows that both K and K k k


383

= id . In particular, the subgroup e−1 (1) is invariant under follows that Θ| K K −1 (since e (1) ⊆ K). Then there exists a unique automorphism Θ ∈ Aut (G) Θ it follows that Θ2 = Θ. In addition, since Θ 2 = Θ, such that Θ ◦ e = e ◦ Θ. It remains to prove that K = {g ∈ G | Θ(g) = g}. The inclusion ⊆ is clear = id and K = K/e −1 (1). Conversely, let g ∈ G such that Θ(g) = since Θ| K K g. By hypothesis (iii), there exist k ∈ K and X ∈ p such that g = k · expG X. Then k · expG X = g = Θ(g) = Θ(k) · expG (θ(X)) = k · expG (−X), so that expG (2X) = 1. Thus ϕ(1, 2X) = ϕ(1, 0), and then X = 0 since ϕ : K × p → G is injective. Consequently, g = k ∈ K, and the proof ends. We now come to a proposition that allows us to lift the global Iwasawa decompositions to covering groups. Proposition 4.4. Let G be a connected Banach–Lie group, and K, A, and N be connected Banach–Lie subgroups of G such that the multiplication map m : K × A × N → G is a diffeomorphism. In addition, assume that A and N are simply connected and AN = N A. is a connected Banach–Lie group with a covering hoNow assume that G → G, and define K := e−1 (K) and X the connected commomorphism e : G −1 are connected ponent of 1 ∈ e (X) for X = A or X = N . Then K, A, and N ×N →G : K ×A Banach–Lie subgroups of G, and the multiplication map m is a diffeomorphism. A, and N are Banach–Lie subgroups Proof. It follows by Lemma 4.1(i) that K, → X is a covering map when X is either of the groups K, of G and e|X : X A, N , or G. Since the groups A and N are simply connected, it then follows → X is actually a diffeomorphism if X = A or X = N that the map e|X : X and N (see for instance Theorem C.18 in [Be06]). In particular, the groups A are connected and simply connected. is connected as well, we first show that G/K is homeTo prove that K omorphic to the simply connected group B := AN A × N , where the diffeomorphism A × N → B is defined by the multiplication in G, as an easy consequence of the assumption. Now let prK : G → K and prB : G → B be the smooth projections given by the inverse of the diffeomorphism K × B → G. Then the continuous map τ : B → G/K, b → b−1 K is bijective, and its inverse is also continuous since it is given by τ −1 : G/K → B, gK → prB (g −1 ). Thus τ : B → G/K is a homeomorphism. By taking into account the homeomor K → G/K provided by Lemma 4.1(ii), it then follows that the phism ψ : G/ K is simply connected, since B ( A × N ) is simply connected. quotient G/ is connected. Then Lemma 4.1(iii) shows that K × A× N → G. :K We now prove the assertion on the multiplication map m Using locally defined inverses of the covering map e : G → G, we see that m ×A × N. It remains to prove is a local diffeomorphism at every point of K arbitrary. Since is bijective. To prove that m is surjective, let g ∈ G that m m : K×A×N → G is surjective, there exist k ∈ K, a ∈ A, and n ∈ N such that

384

Daniel Beltit¸˘ a

e( g ) = kan. Denote a := (e|A)−1 (a) ∈ A and n := (e|N )−1 (n) ∈ N , and take −1 k0 ∈ e (k) arbitrary, so that e( k0 an ) = kan = e( g ). Denoting z := k0 an g −1 , Then and k := z −1 k0 ∈ K we have e( z ) = 1, so z ∈ e−1 (1) ⊆ e−1 (K) = K. : K ×A×N → G g = k an ∈ K AN . Since g ∈ G is arbitrary, it follows that m is surjective. we have is injective, first note that for any To prove that m a ∈ A −1 −1 −1 e( aN a ) = e( a)N e( a) ⊆ N , since AN = N A. Thus aN a ⊆ e−1 (N ). −1 The first of these sets is connected and contains 1 ∈ e (N ), hence actu . Consequently, A N =N A =: B is a subgroup of G. Since a−1 ⊆ N ally aN → X is a bijection if X = A or X = N and A ∩ N = {1}, we get e|X : X N = {1} as well. Now assume that and n for j = 1, 2, A∩ kj ∈ K, aj ∈ A, j ∈ N −1 −1 −1 ∩ B, so that it will and k1 k1 = a1 n 1 = k2 a2 n 2 . Then k2 a2 n 2 ( a1 n 1 ) ∈ K ∩B = {1}, that is, K ∩A N = {1}. To prove the be enough to check that K ∩A N arbitrary. Then e( latter equality, let x ∈K x) ∈ K ∩ AN = {1}. On and n , and e( the other hand, we have x = an for some a∈A ∈N a)e( n) = 1. Since A ∩ N = {1}, it follows that e( a) = e( n) = 1. Using the fact that → X is a bijective homomorphism if X = A or X = N , we get e|X : X ∩A N = {1}, and this completes the a=n = 1, whence x = an = 1. Thus K ×A ×N →G is bijective. :K proof of the fact that the map m

5 What a reductive Banach–Lie group could be Reductivity relative to a symmetric norming function. We are going to describe a class of connected Banach–Lie groups that is wide enough to include all the finite-dimensional connected reductive Lie groups of Definiton 1.1 and the connected components of the identity in most of the classical Banach–Lie groups of operators (see [dlH72]). This is an attempt to provide an appropriate framework for the interesting developments from representation theory of these infinite-dimensional Lie groups; see for instance the works [Bo80], [Ne98], or [NØ98]. We chose to single out a class of connected Banach– Lie groups that does not include certain interesting groups, like the restricted groups or unitary groups of operator algebras, because of certain unpleasant properties that they possess; see Section 6 in [BRT07] and Proposition 1.1 in [Be09]. As mentioned above, the following definition is suggested by Remarks 1.2 and 1.3. Definition 5.1. Let Φ be a symmetric norming function. By Φ-reductive Banach–Lie algebra we mean any closed real Lie subalgebra of SΦ satisfying the following conditions. (i) For every X ∈ g we have X ∗ ∈ g. (ii) The set g ∩ F of finite-rank operators in g is dense in g with respect to the (0) norm · Φ . Thus g ⊆ SΦ actually.


385

(0)

In this case the connected Banach–Lie group G (⊆ 1 + SΦ ) corresponding to (0) the Lie subalgebra g of SΦ (H) is said to be a Φ-reductive linear Banach–Lie group. By Φ-reductive Banach–Lie group we shall mean any covering group of a Φ-reductive linear Banach–Lie group. In the above setting, the closure of g ∩ F with respect to the Hilbert– Schmidt norm · 2 will be called the L∗ -algebra associated with g and will be denoted by g2 . Example 5.2. (See [dlH72] and also Definitions 3.2 and 3.2 in [Be09].) Let Φ be any symmetric norming function and denote by GL(H) the group of all invertible bounded linear operators on the complex Hilbert space H. The classical complex Banach–Lie groups and Banach–Lie algebras associated with Φ are defined as follows: (0)

(0)

(A) GLΦ (H) = GL(H)∩(1+SΦ (H)) with the Lie algebra glΦ (H) = SΦ (H); (B) OΦ (H) = {g ∈ GLΦ (H) | g −1 = Jg ∗ J −1 } (where J is a conjugation, that is, an antilinear isometry satisfying J 2 = 1) with the Lie algebra (0) oΦ (H) = {x ∈ SΦ (H) | x = −Jx∗ J −1 }; ∗ J −1 } (where J an anti-conjugation, (C) SpΦ (H) = {g ∈ GLΦ (H) | g −1 = Jg i.e., an antilinear isometry satisfying J 2 = −1), with the Lie algebra (0) ∗ J −1 }. spΦ (H) = {x ∈ SΦ (H) | x = −Jx The classical real Banach–Lie groups and Banach–Lie algebras associated with the symmetric norming function Φ are the following: (AI) GLΦ (H; R) = {g ∈ GLΦ (H) | gJ = Jg} with the Lie algebra (0) glΦ (H; R) = {x ∈ SΦ (H) | xJ = Jx}, where J : H → H is a conjugation; with the Lie algebra (AII) GLΦ (H; H) = {g ∈ GLΦ (H) | g J = Jg} (0) glΦ (H; H) = {x ∈ SΦ (H) | xJ = Jx}, where J : H → H is an anticonjugation; (AIII) UΦ (H+ , H− ) = {g ∈ GLΦ (H) | g ∗ V g = V } with the Lie algebra (0) uΦ (H ) = {x ∈ SΦ (H) | x∗ V = −V x}, where H = H+ ⊕ H− and + , H− 1 0 with respect to this orthogonal decomposition of H; V = 0 −1 (BI) OΦ (H+ , H− ) = {g ∈ GLΦ (H) | g −1 = Jg ∗ J −1 and g ∗ V g = V } with (0) ∗ −1 oΦ (H+ , H− ) = {x ∈ S and x∗ V = −V x}, where | x = −Jx J Φ (H) 1 0 with respect to this decomposition of H, H = H+ ⊕ H− , V = 0 −1 and J : H → H is a conjugation such that J(H± ) ⊆ H± ; with the Lie (BII) O∗Φ (H) = {g ∈ GLΦ (H) | g −1 = Jg ∗ J −1 and g J = Jg} (0) algebra o∗Φ (H) = {x ∈ SΦ (H) | x = −Jx∗ J −1 and xJ = Jx}, where J : H → H is a conjugation and J : H → H is an anti-conjugation such that J J = JJ;

386

Daniel Beltit¸˘ a

∗ J −1 and gJ = Jg} with the (CI) SpΦ (H; R) = {g ∈ GLΦ (H) | g −1 = Jg (0) ∗ J −1 and xJ = Lie algebra spΦ (H; R) = {x ∈ SΦ (H) | −x = Jx Jx}, where J : H → H is any anti-conjugation and J : H → H is any conjugation such that J J = JJ; ∗ J −1 and g ∗ V g = V } with (CII) SpΦ (H+ , H− ) = {g ∈ GLΦ (H) | g −1 = Jg (0) ∗ J −1 and x∗ V = −V x}, spΦ (H+ , H− ) = {x ∈ SΦ (H) |x = −Jx 1 0 with respect to this decomposition where H = H+ ⊕ H− , V = 0 −1 ± ) ⊆ H± . of H, and J : H → H is an anti-conjugation such that J(H As a by-product of the classification of the L∗ -algebras (see for instance Theorems 7.18 and 7.19 in [Be06]), every (real or complex) topologically simple L∗ -algebra is isomorphic to one of the classical Banach–Lie algebras associated with the Hilbert–Schmidt ideal S2 (H). We refer to [dlH72] and [Ne02a] for information on the homotopy groups of the classical Banach–Lie groups associated with the Schatten ideals. The corresponding description of homotopy groups actually holds true for the classical Banach–Lie groups associated with any symmetric norming function Φ. It is is clear that the connected 1-component of any classical Banach–Lie group associated with Φ is a Φ-reductive linear Banach–Lie group. It would be interesting to understand how the classical Lie algebras associated with an operator ideal fit in the framework of Φ-reductive Lie algebras. Here is a proposition in this connection. Recall that a subset A ⊆ B(H) is irreducible if {0} and H are the only closed linear subspaces of H which are invariant under all operators in A. Proposition 5.3. Let Φ be a symmetric norming function, g a Φ-reductive Lie algebra, and g2 the L∗ -algebra associated with g. Then the Lie algebra g is irreducible if and only if g2 is irreducible. In addition, if SΦ = S1 , then g is one of the classical (real or complex) Banach–Lie algebras associated with the norm ideal SΦ if and only if g2 is one of the classical (real or complex) Banach–Lie algebras associated with the Hilbert–Schmidt ideal S2 . If this is the case, then g and g2 are classical Lie algebras of the same type. Proof. Both algebras g and g2 are closed under taking the adjoints, hence any of them is irreducible if and only if every operator that commutes with that algebra is a scalar multiple of the identity operator on H. Therefore it will be enough to prove the following assertion: (∀T ∈ B(H))

[T, g] = {0} ⇐⇒ [T, g2 ] = {0}.

(3)

In fact, if [T, g] = {0}, then in particular [T, g ∩ F] = {0}. Since the set of finite-rank operators g ∩ F is dense in g2 , it follows that [T, g2 ] = {0}. Conversely, if the operator T satisfies the latter condition, then we have in particular [T, g ∩ F] = {0}. By property (ii) of a Φ-reductive Lie algebra, it then follows that [T, g] = {0}.


387

Now assume SΦ = S1 and let g be the closure of g2 ∩ F with respect to the norm · Φ . Since g ∩ F ⊆ g2 ∩ F and g ∩ F is dense in g, it then follows that g ⊆ g. We are going to prove that g is actually a closed ideal of the Lie algebra g. For this purpose it will be enough to show that [g ∩ F, g2 ∩ F] ⊆ g ∩ F. In fact, let X ∈ g ∩ F and Y ∈ g2 ∩ F arbitrary. According to the definition of g2 , there exists a sequence {Xj }j≥1 in g ∩ F such that Xj − Y 2 → 0 as j → ∞. Then [X, Xj ] − [X, Y ] 2 → 0 as j → ∞. On the other hand, the commutators {[X, Xj ]}j≥1 have the ranks uniformly bounded by 2(rank X) since the dimension of linear spaces decreases under linear transformations, hence the rank of a product of operators is less than the rank of any of the factors. Now Lemma 4.3 in [BR05] implies that in the topology of · Φ we also have [X, Xj ] → [X, Y ] as j → ∞, whence [X, Y ] ∈ g ∩ F. Now it is easy to see that the assertion holds. In fact, if g2 is one of the classical (real or complex) Banach–Lie algebras associated with the Hilbert– Schmidt ideal S2 , then g is the similar classical Banach–Lie algebra associated with the norm ideal SΦ . Since SΦ = S1 , it follows as in Proposition 8 on page 92 in [dlH72] that g has no non-trivial closed ideals, whence g = g. Conversely, if g is one of the classical (real or complex) Banach–Lie algebras associated with the norm ideal SΦ , then it is obvious that the L∗ -algebra associated with g is one of the classical L∗ -algebras. Group decompositions. The following theorem supplies a Cartan decomposition for some Φ-reductive Banach–Lie groups. In this connection, we refer to [Ne02b] for a systematic investigation of polar decompositions in infinite dimensions (that is, Cartan decompositions for linear groups). Theorem 5.4. Assume that Φ is a symmetric norming function and G is a Φ-reductive Banach–Lie group whose Lie algebra g ⊆ SΦ (H) is one of the classical Lie algebras associated with Φ. Denote k = {X ∈ g | X ∗ = −X}, p = {X ∈ g | X ∗ = X}, and let K be the connected Banach–Lie group corresponding to the Lie subalgebra k of g. Then K is a Banach–Lie subgroup of G and the mapping ϕ : K × p → G, (k, X) → k expG X, is a diffeomorphism. Moreover, there exists a unique automorphism Θ ∈ Aut (G) such that L(Θ)X = −X ∗ for all X ∈ g and K = {g ∈ G | Θ(g) = g}. Proof. It follows by Proposition 4.2 that we may assume that G is a Φ(0) reductive linear Banach–Lie group. Thus let G ⊆ 1 + SΦ (H) be the connected 1-component of the classical Banach–Lie group associated with the classical Lie algebra g ⊆ SΦ (H). In this case the assertion can be proved by a method similar to the one used in Proposition III.8 of [Ne02a] in the case of the classical groups associated with the Schatten ideals. The existence of the automorphism Θ ∈ Aut (G) as asserted follows by an application of Proposition 4.3. The following theorem states that there exist Iwasawa decompositions for the Φ-reductive Banach–Lie groups corresponding to the classical Lie algebras,

388

Daniel Beltit¸˘ a

provided that the symmetric norming function Φ satisfies some reasonable conditions. Theorem 5.5. Let Φ be a mononormalizing symmetric norming function whose Boyd indices are nontrivial and let G be a Φ-reductive Banach–Lie group whose Lie algebra g ⊆ SΦ (H) is one of the classical Lie algebras associated with Φ. Denote k = {X ∈ g | X ∗ = −X}, p = {X ∈ g | X ∗ = X}, and let K be the connected Banach–Lie group corresponding to the Lie subalgebra k of g. Then there exists X0 ∈ p with the following properties. (a) The set EX0 of spectral projections of the self-adjoint operator X0 corresponding to the intervals of the form (0, t] with t ∈ R determines a direct sum decomposition g = k aX0 nX0 , where aX0 = {X ∈ p | [X, X0 ] = 0} and nX0 = {X ∈ g ∩ Alg (EX0 ) | the spectrum of X is equal to {0}}. (b) If we denote by A and N the connected Banach–Lie groups corresponding to the closed subalgebras aX0 and nX0 of g, then the multiplication map m : K × A × N → G is a diffeomorphism. Moreover, A and N are simply connected Banach–Lie subgroups of G and AN = N A. Proof. Proposition 4.4 allows us to assume that G is actually a Φ-reductive linear Banach–Lie group. The idea of the proof in this case is to start by studying the group G = GLΦ (H), which is the largest classical group. For this group, the construction of a local Iwasawa decomposition of its Lie algebra (assertion (a)) relies on local spectral theory (see [BS01]) along with the properties of triangular integrals established in Example 2.10. As regards the corresponding global Iwasawa decomposition (assertion (b)), one uses Corollary 2.16. See Proposition 4.3 and Theorem 4.5 in [Be09] for details of the proof in the case G = GLΦ (H). The construction of Iwasawa decompositions for all types of Φ-reductive linear Banach–Lie groups is discussed in Sections 4–6 of the paper [Be09]. Remark 5.6. The fundamental groups of topological groups are always abelian; see for instance Corollary 1.7.10 in [Sp81] for the more general statement on homotopy groups of H-spaces. (See also [dlH72] and [Ne02a] for a description of the fundamental groups of the classical Banach–Lie groups associated with Φ.) It then easily follows by Remark 3.6 along with Examples 3.3 through 3.5 that all of the groups K, A, and N that occur in Theorem 5.5 are amenable, although the group G itself may not be amenable; see Example 3.5(a). Harish-Chandra decompositions. We are going to draw a little closer to representation theory, which was the main motivation of the present exposition. For this purpose we borrow the following definition of infinitedimensional Lie groups of Harish-Chandra type from [NØ98]. Some good references for such Harish-Chandra decompositions in the setting of finitedimensional Lie groups are [Sa80], [Kn96], and [Ne00]. Definition 5.7. By Banach–Lie group of Harish-Chandra type we actually mean a 4-tuple (G, GC , K, K C ) consisting of a connected complex Banach–Lie


389

group GC and three connected Banach–Lie subgroups G, K, and K C of GC such that the following conditions are satisfied: (i) the Lie algebra gC of GC is the complexification of the Lie algebra g of G; (ii) K C is a complex Banach–Lie subgroup of GC and the Lie algebra kC of K C is the complexification of the Lie algebra k of K; (iii) there exist connected complex Banach–Lie subgroups P ± of GC whose Lie algebras p± have the properties (ad p± )n gC = {0} for some integer n ≥ 1 and p± ∩ z(g) = {0}, where z(gC ) denotes the center of gC , and moreover: (HC1) we have the direct sum decomposition gC = p+ kC p− , and in addition [kC , p± ] ⊆ p± and p− = p+ , where X → X, gC → gC , is the antilinear involutive map whose fixed-point set is g; (HC2) the multiplication P + × K C × P − → GC , (p+ , k, p− ) → p+ kp− is a biholomorphic diffeomorphism onto its open image; (HC3) we have G ⊆ P + K C P − and G ∩ K C P − = K. If the groups GC , K, and K C are singled out in a certain context, then we may say that the group G itself is a Banach–Lie group of Harish-Chandra type. Example 5.8. Let Φ be a symmetric norming function and H = H+ ⊕ H− . Then the corresponding classical real Banach–Lie group of type (AIII)—in the terminology of Example 5.2—provides an example of a group of HarishChandra type. Specifically, we mean the 4-tuple (G, GC , K, K C ), where G = UΦ (H+ , H− ), C

G = GLΦ (H),

K = {k ∈ UΦ (H+ , H− ) | k(H± ) ⊆ H± } K C = {g ∈ GLΦ (H) | g(H± ) ⊆ H± }.

Then the conditions of Definition 5.7 are satisfied with the connected complex Banach–Lie subgroups P ± = {g ∈ GLΦ (H) | (g − 1)H∓ ⊆ H± }. If we write the operators on H as 2 × 2 block matrices with respect to the orthogonal decomposition H = H+ ⊕ H− , then AB + C − ∈ GLΦ (H) | D ∈ GL(H− ) , G⊆P K P = g= CD and every element in this set can be factorized as 1 BD−1 1 0 AB A − BD−1 C 0 = , D−1 C 1 CD 0 D 0 1 where the factors on the right-hand side belong to P + , K C , and P − , respectively. We refer to [NØ98] for more details as well as for similar examples provided by groups of type (BII) and (CI) (again in the terminology of Example 5.2). Remark 5.9. As noted in [NØ98] (see also [Ne00]), the objects involved in the definition of a group of Harish-Chandra type (Definition 5.7 above) have the following additional properties:

390

Daniel Beltit¸˘ a

(a) If we denote L(P ± ) = p± , then the exponential maps expP ± : p± → P ± are biholomorphic diffeomorphisms and the complex Banach–Lie groups P ± are nilpotent and simply connected. In particular, there exist the logarithm maps logP ± = (expP ± )−1 : P ± → p± . (b) There exists an open connected K-invariant subset Ω ⊆ p+ such that the mapping Ω × K C P − → GK C P − , (Z, p) → (expGC Z)p, is a well-defined biholomorphic diffeomorphism. (c) Denote Ξ = {(g, Z) ∈ GC × p+ | g expGC Z ∈ P + K C P − }, which is an open neighborhood of G × Ω in GC × p+ . Let ζ ± : P + K C P − → P ± and κ : P + K C P − → K C be the natural projections (see condition (HC2) in Definition 5.7), and define Ξ → p+ , (g, Z) → g.Z = logP + (ζ + (g expGC Z)) and J : Ξ → K, (g, Z) → J(g, Z) = κ(g expGC Z). Then the mapping (g, Z) → g.Z defines a transitive action of G upon Ω by biholomorphic diffeomorphisms and J has a cocycle property with respect to this action. (d) We have 0 ∈ Ω and the isotropy group of 0 is equal to K. Thus the aforementioned action leads to a G-equivariant diffeomorphism G/K Ω. These are some of the ideas that allow one to define as in [NØ98] natural reproducing kernels on Ω out of certain representations of the Banach–Lie group K. In this way one ends up with the corresponding reproducing kernel Hilbert spaces of holomorphic functions on Ω which carry representations of the bigger Banach–Lie group G of Harish-Chandra type. Remark 5.10. Applications of a different type of reproducing kernels in the representation theory of Banach–Lie groups can be found in [BR07] and [BG08]. The viewpoint held in the first of these papers is somehow dual to the one of Remark 5.9 (in the sense of duality theory of symmetric spaces); i.e., the bases of the corresponding vector bundles are homogeneous spaces of compact type. The complexification of this picture of compact type is analyzed in [BG08] along with the relationship to Stinespring dilation theory, which eventually leads to geometric models for representations of operator algebras. Acknowledgements: We wish to thank Professor José Galé for drawing our attention to some pertinent references, Professor Hendrik Grundling for kindly sending us his unpublished manuscript, Professor Mihai S ¸ abac and Professor Gary Weiss for several useful comments, and finally Professor Karl-Hermann Neeb for kind encouragement and valuable suggestions and the referee for many helpful remarks. Partial support from the grant GR202/2006 (CNCSIS code 813) is acknowledged.

References [Ar78] J. Arazy, Some remarks on interpolation theorems and the boundness of the triangular projection in unitary matrix spaces, Integral Equations Operator Theory 1 (1978), no. 4, 453–495. [Ba83] W. Banaszczyk, On the existence of exotic Banach–Lie groups, Math. Ann. 264 (1983), no. 4, 485–493.


391

[Be06] D. Beltit¸˘ a, Smooth Homogeneous Structures in Operator Theory, Monogr. and Surveys in Pure and Appl. Math., 137. Chapman & Hall/CRC Press, Boca Raton-London-New York-Singapore, 2006. [Be09] D. Beltit¸˘ a, Iwasawa decompositions of some infinite-dimensional Lie groups. Trans. Amer. Math. Soc. 361 (2009), no. 12, 6613–6644. [BG08] D. Beltit¸˘ a, J.E. Galé, Holomorphic geometric models for representations of C ∗ -algebras. J. Funct. Anal. 255 (2008), no. 10, 2888–2932. [BP07] D. Beltit¸a ˘, B. Prunaru, Amenability, completely bounded projections, dynamical systems and smooth orbits, Integral Equations Operator Theory 57 (2007), no. 1, 1–17. [BR05] D. Beltit¸˘ a, T.S. Ratiu, Symplectic leaves in real Banach Lie-Poisson spaces, Geom. Funct. Anal. 15 (2005), no. 4, 753–779. [BR07] D. Beltit¸˘ a, T.S. Ratiu, Geometric representation theory for unitary groups of operator algebras, Adv. Math. 208 (2007), no. 1, 299–317. [BRT07] D. Beltita, T.S. Ratiu, A.B. Tumpach, The restricted Grassmannian, Banach Lie-Poisson spaces, and coadjoint orbits, J. Funct. Anal. 247 (2007), no. 1, 138–168. [BS01] D. Beltit¸˘ a, M. S ¸ abac, Lie Algebras of Bounded Operators, Operator Theory: Advances and Applications, 120. Birkh¨ auser Verlag, Basel, 2001. [Bo80] R.P. Boyer, Representation theory of the Hilbert-Lie group U(H)2 , Duke Math. J. 47 (1980), no. 2, 325–344. [Bu50] R.C. Buck, Generalized group algebras, Proc. Nat. Acad. Sci. USA 36 (1950), 747–749. [DFWW04] K. Dykema, T. Figiel, G. Weiss, M. Wodzicki, Commutator structure of operator ideals, Adv. Math. 185 (2004), no. 1, 1–79. [Er72] J.A. Erdos, The triangular factorization of operators on Hilbert space, Indiana Univ. Math. J. 22 (1972/73), 939–950. [Er78] J.A. Erdos, Triangular integration on symmetrically normed ideals, Indiana Univ. Math. J. 27 (1978), no. 3, 401–408. [Ey72] P. Eymard, Moyennes Invariantes et Représentations Unitaires, Lecture Notes in Math. 300, Springer-Verlag, Berlin, 1972. [FN07] S. Ferri, M. Neufang, On the topological centre of the algebra LUC(G)∗ for general topological groups, J. Funct. Anal. 244 (2007), no. 1, 154–171. [Ga06] J.E. Galé, Geometr´ıa de ´ orbitas de representaciones de grupos y ´ algebras promediables, Rev. R. Acad. Cienc. Exactas F´ıs. Qu´ım. Nat. Zaragoza (2) 61 (2006), 7–46. [GK69] I.C. Gohberg, M.G. Kre˘ın, Introduction to the Theory of Linear Nonselfadjoint Operators, Transl. Math. Monogr., 18, Amer. Math. Soc., Providence, RI, 1969. [GK70] I.C. Gohberg, M.G. Kre˘ın, Theory and Applications of Volterra Operators in Hilbert Space, Transl. Math. Monogr., 24, Amer. Math. Soc., Providence, RI, 1970. [Gr] H. Grundling, unpublished manuscript. [Gr05] H. Grundling, Generalising group algebras, J. London Math. Soc. (2) 72 (2005), no. 3, 742–762. [GN09] H. Grundling, K.-H. Neeb, Full regularity for a C ∗ -algebra of the canonical commutation relations. Rev. Math. Phys. 21 (2009), no. 5, 587–613. [dlH72] P. de la Harpe, Classical Banach–Lie Algebras and Banach–Lie Groups of Operators in Hilbert Space, Lecture Notes in Math. 285, Springer-Verlag, Berlin, 1972.

392

Daniel Beltit¸˘ a

[HR63] E. Hewitt, K.A. Ross, Abstract Harmonic Analysis, vol. I. Grundl. Math. Wiss., Bd. 115. Springer-Verlag, Berlin, 1963. [Ho75] K.H. Hofmann, Analytic groups without analysis, Inst. Nat. Alta Mat. Symp. Math. 16 (1975), 357–374. [KW08] V. Kaftal, G. Weiss, A survey on the interplay between arithmetic mean ideals, traces, lattices of operator ideals, and an infinite Schur-Horn majorization theorem, in: Hot Topics in Operator Theory, Theta Ser. Adv. Math., 9, Theta, Bucharest, 2008, pp. 101–135. [Kn96] A.W. Knapp, Lie Groups Beyond an Introduction, Progr. Math., 140, Birkh¨ auser-Verlag, Boston-Basel-Berlin, 1996. [Ne98] K.-H. Neeb, Holomorphic highest weight representations of infinite-dimensional complex classical groups, J. Reine Angew. Math. 497 (1998), 171–222. [Ne00] K.-H. Neeb, Holomorphy and Convexity in Lie Theory, Expositions in Mathematics, 28. Walter de Gruyter & Co., Berlin-New York, 2000. [Ne02a] K.-H. Neeb, Classical Hilbert-Lie groups, their extensions and their homotopy groups, in: Geometry and Analysis on Finite and Infinite-Dimensional Lie Groups (B¸edlewo, 2000), Banach Center Publ., 55, Polish Acad. Sci. Warsaw, 2002, pp. 87–151. [Ne02b] K.-H. Neeb, A Cartan-Hadamard theorem for Banach-Finsler manifolds, Geom. Dedicata 95 (2002), 115–156. [Ne04] K.-H. Neeb, Infinite-dimensional groups and their representations, in: Lie Theory, Progr. Math., 228, Birkh¨ auser, Boston, MA, 2004, pp. 213–328. [Ne06] K.-H. Neeb, Towards a Lie theory of locally convex groups, Japan. J. Math. (3rd series) 1 (2006), no. 2, 291–468. [NØ98] K.-H. Neeb, B. Ørsted, Unitary highest weight representations in Hilbert spaces of holomorphic functions on infinite-dimensional domains, J. Funct. Anal. 156 (1998), no. 1, 263–300. [Pa88] A.L.T. Paterson, Amenability, Mathematical Surveys and Monographs, 29. American Mathematical Society, Providence, RI, 1988. [Pe06] V. Pestov, Dynamics of Infinite-Dimensional Groups. The Ramsey-Dvoretzky-Milman Phenomenon, University Lecture Series, 40. Amer. Math. Soc. Providence, RI, 2006. [Pi90] D. Pickrell, Separable representations for automorphism groups of infinite symmetric spaces, J. Funct. Anal. 90 (1990), no. 1, 1–26. [Sa80] I. Satake, Algebraic Structures of Symmetric Domains, Kano Mem. Lectures, 4. Publ. Math. Soc. Japan, 14. Iwanami Shoten & Princeton Univ. Press, Tokyo-Princeton, NJ, 1980. [Sp81] E.H. Spanier, Algebraic Topology. Corrected reprint. Springer-Verlag, New York-Berlin, 1981. [SV75] S ¸ . Str˘ atil˘ a, D. Voiculescu, Representations of AF-algebras and of the Group U(∞), Lecture Notes in Math., 486. Springer-Verlag, Berlin-New York, 1975. [Tu95] S. Turek, Universal minimal dynamical system for reals, Comment. Math. Univ. Carolin. 36 (1995), no. 2, 371–375. [Up85] H. Upmeier, Symmetric Banach Manifolds and Jordan C ∗ -Algebras, Math. Stud., 104, Notas de Mat., 96, North-Holland, Amsterdam, 1985. [Vo00] D.A. Vogan, Jr., The method of coadjoint orbits for real reductive groups, in: Representation Theory of Lie Groups (Park City, 1998), IAS/ Park City Math. Ser. 8, Amer. Math. Soc., Providence, RI, 2000, pp. 177-238. [We40] A. Weil, L’Intègration dans les Groupes Topologiques et ses Applications, Actual. Sci. Ind., 869. Hermann et Cie., Paris, 1940.

Heat Kernel Measures and Critical Limits Doug Pickrell Mathematics Department, University of Arizona, Tucson, AZ, USA 85721, [email protected]

Summary. This article is an exposition of several questions linking heat kernel measures on infinite-dimensional Lie groups, limits associated with critical Sobolev exponents, and Feynman–Kac measures for sigma models. The first part of the article concerns existence and invariance issues for heat kernel measure classes. The main examples are heat kernel measures on groups of the form C 0 (X, F), where X is a Riemannian manifold and F is a finite-dimensional Lie group. These measures depend on a smoothness parameter s > dim(X)/2. The second part of the article concerns the limit s ↓ dim(X)/2, especially dim(X) ≤ 2, and how this limit is related to issues arising in quantum field theory. In the case of X = S 1 , we conjecture that heat kernel measures converge to measures which arise naturally from the Kac–Moody–Segal point of view on loop groups, as s ↓ 1/2. Key words: heat kernel measure, abstract Wiener space, abstract Wiener group, critical Sobolev exponent, loop group, sigma model, Feynman–Kac measure. 2000 Mathematics Subject Classifications: 58D20, 22E65, 22E67.

1 Introduction Given a finite-dimensional real Hilbert space f , there is an associated Lebesgue measure λf , a (positive) Laplace operator ∆f , and a convolution semigroup of heat kernel measures t

1

2

νtf = e− 2 ∆f δ0 = (2πt)− dim(f )/2 e− 2t |x|f dλf (x).

(1)

The map f → {νtf } is functorial, in the sense that if P : f1 → f2 is an orthogonal projection, then P∗ νtf1 = νtf2 . (2) More generally, given a finite-dimensional Lie group F, with a fixed inner product on its Lie algebra f , there is aninduced left invariant Riemannian metric, a Laplace type operator ∆F = − Xi2 , where {Xi } is an orthonormal K.-H. Neeb and A. Pianzola (eds.), Developments and Trends in Infinite-Dimensional Lie Theory, Progress in Mathematics 288, DOI 10.1007/978-0-8176-4741-4_12, © Springer Basel AG 2011

393

394

Doug Pickrell

basis of left invariant vector fields, and a convolution semigroup of heat kernel measures, f νtF = e−t∆F /2 δ1 = lim (exp∗ (νt/N ))N , (3) N →∞

where the power refers to convolution of measures on F (see Sections 4.7–4.8 of [McK]). Unlike the other structures, the heat kernel measures are always inversion invariant. The map F → {νtF } is functorial, in the sense that if P : F1 → F2 is a map of Lie groups, and the induced map of Lie algebras f1 → f2 is an orthogonal projection, then P∗ νtF1 = νtF2 .

(4)

In Section 2 we will recall how this construction can be generalized to infinite dimensions. The form of the linear generalization is well known: given a separable real Hilbert space h, there is an associated convolution semigroup of Gaussian measures {νth }. An essential complication arises when dim(h) = ∞, namely the νth are only finitely additive, viewed as cylinder measures on h (see [St] for orientation on this point). To overcome this, following Gross ([Gr]), we consider an abstract Wiener space h → b; in this framework, the Gaussians are realized as countably additive measures νth⊂b on b, a Banach space completion of h. Geometrically speaking, one imagines that the support of the heat kernel measures diffuses into the enveloping space b. In the larger group category, the objects are abstract Wiener groups. Such an object consists of an inclusion of a separable Hilbert–Lie group into a Banach–Lie group, H → G, (5) with a fixed compatible inner product on the Lie algebra h, such that the induced map of Lie algebras h → g is an abstract Wiener space. A morphism in this category is a map of pairs of Banach–Lie groups P : (H1 , G1 ) → (H2 , G2 ) such that the induced map of Lie algebras P : h1 → h2 is an orthogonal projection. With some possible restrictions, the corresponding heat kernel measures νtH⊂G can be constructed as in the finite-dimensional case, and they form a convolution semigroup of inversion invariant probability measures on G (this use of Ito’s ideas in an infinite-dimensional context apparently originated in [DS], and more concretely in [M]). This construction is also functorial, a point which we will emphasize. In Section 3 we consider invariance properties of heat kernel measure classes. The Cameron–Martin–Segal theorem asserts that for t > 0, the measure class [νth⊂b ] is invariant with respect to translation by h ∈ b if and only if h ∈ h. In the larger group category, we conjecture that [νtH⊂G ] is invariant with respect to translation by h ∈ G if and only if h ∈ H0 and Ad(h)∗ Ad(h) − 1 is a Hilbert–Schmidt operator on h. The evidence in favor of this conjecture is based on deep results of Driver. In Section 4, and in the remainder of the paper, we focus on groups of maps. Suppose that X is a compact manifold and W is a real Hilbert–Sobolev

Heat Kernel Measures

395

space corresponding to degree of smoothness s > dim(X)/2. In this context the Sobolev embedding W → C 0 (X; R) (6) is a canonical example of an abstract Wiener space. If F is a finite-dimensional Lie group with a fixed inner product on its Lie algebra f , then in a natural way H = W (X, F) → G = C 0 (X, F) (7) is an abstract Wiener group. In this example the measure νtH⊂G is determined by its distributions corresponding to evaluation at a finite number of points of X; these distributions involve a nonlocal Green’s function interaction between pairs of points. When X = S 1 , F = K, a simply connected compact Lie group with Ad(K)-invariant inner product, and s = 1, Driver and collaborators have proven that the measure class of νtH⊂G equals the measure class of a Wiener measure. This latter measure has a heuristic expression in terms of a local energy functional 1 exp(− g −1 dg ∧ ∗g −1 dg) dλ(g(θ)) (8) 2t X X

(see [Dr2]). When dim(X) = 2, s = 1 is the critical exponent, the Sobolev embeddings (6)–(7) and associated heat kernel measures are defined only for s > 1, and the heuristic expression (8) is the Feynman–Kac measure for the sigma model with target K. This motivates the following question, considered in Sections 5 and 6: is there a sense in which heat kernel measures have nontrivial limits as s ↓ dim(X)/2, the critical exponent? In the case of X = S 1 , for various reasons (e.g., the Kac–Moody–Segal point of view on loop groups) one is led to consider, in place of (7), the hyperfunction completion W 1/2 (S 1 , K) → Hyp(S 1 , G),

(9)

where G denotes the complexification of K. A generic point in the hyperfunction completion is represented by a formal Riemann–Hilbert factorization g = g− · g0 · g+ ,

(10)

where g− ∈ H 0 (∆∗ , ∞; G, 1) (i.e., a holomorphic function on ∆∗ with values in G sending ∞ to 1 ∈ G), g0 ∈ G, g+ ∈ H 0 (∆, 0; G, 1), and ∆ (∆∗ ) is the open unit disk centered at 0 (∞). There is a family of measures on Hyp(S 1 , G), indexed by a level l, having heuristic expressions involving Toeplitz determinants. The associated measure classes are invariant with respect to left and right translations, by real analytic loops in K, and conjecturally with respect to real analytic reparameterization. Whereas heat kernel measures (and

396

Doug Pickrell

Wiener measures) are determined by distributions corresponding to evaluating continuous K-valued loops at points on S 1 , these latter measures are determined by distributions corresponding to g0 , and to evaluating the holomorphic functions g± at points off S 1 . It seems likely that these distributions are computable, but at present very little is known rigorously. We conjecture that the heat kernel measures parameterized by s and t converge to a measure with level l = 1/2t as s ↓ 1/2. This is nontrivial for non-abelian K, because the change of variables from values of a loop around S 1 to coefficients of the Riemann–Hilbert factorization is complicated. a closed Riemannian surface, because The case of most interest is X = Σ, of (8). In Section 6 I have adopted the point of view of quantum field theory in trying to understand the right questions to ask about the critical limit as s ↓ 1. This point of view stresses that the two-dimensional case is tightly inter → W 1/2 (S), twined with the one-dimensional case, by the trace map W 1 (Σ) when S is a 1-manifold (space) embedded in Σ (a Euclidean space-time). This critical limit is also possibly related to longstanding questions concerning the irreducibility of energy representations in two dimensions (see Section 3.5 of [A]). Surveys of related topics, with more extensive bibliographies, include [Dr2] (for Wiener measure and heat kernel measures on loop groups), [A] (for energy representations), and [Ga2] (for sigma models). I should note that there are important examples of heat-kernel-like measures which do not fit into a Banach space framework; see [AM] and the references in that paper. Notation 1 (1) Throughout this paper, all spaces are assumed to be separable and real, unless explicitly noted otherwise. (2) We will frequently encounter continuous maps of Hilbert spaces h → f such that the quotient Hilbert space equals the Hilbert space structure on the target. If f is a subspace of h itself, then such a map is an orthogonal projection. In this paper we will refer to such a map h → f as a projection. (3) Suppose that h is a Hilbert space. Let L2 denote the symmetrically normed ideal of Hilbert–Schmidt operators on h. For an invertible operator A on h, the two conditions AAt ∈ 1 + L2 ,

and

A − A−t ∈ L2

(11)

are equivalent. We let GL(h)(L2 ) denote the group of invertible operators satisfying these two conditions. In the same way, we can define a restricted general linear group GL(h)(I) , for any symmetrically normed ideal I. (4) Given a measure ν on a space X, and a Borel space Y , a map f : X → Y will be said to be ν-measureable if f −1 (E) is ν-measureable for each Borel set E ⊂ Y . The measure class of ν will be denoted by [ν]. (5) Lebesgue measure will be denoted by dλ, and (left) Haar measure will be denoted by dλG , for a group G.


397

2 General constructions 2.1 Abstract Wiener spaces and Gaussian measures Suppose that h is a Hilbert space. The corresponding convolution semigroup of Gaussian measures, {νth }t>0 , has the heuristic expression 1

2

dνth = (2πt)− dim(h)/2 e− 2t |x|h dλh (x),

(12)

where λh denotes the “Lebesgue measure for the Hilbert space h.” The measure λh and the expression (12) have literal meaning only when h is finite dimensional. To go beyond this, following Gross, suppose that h → b is a continuous dense inclusion of h into a Banach space. We can consistently define a finitely additive probability measure νth⊂b on b in the following way. A cylinder set is a Borel subset of b of the form p−1 (E), where p : b → f , p|h : h → f is a finite rank projection, and E is a Borel subset of f . The set of cylinder sets is an algebra; it is a σ-algebra if and only if h = b is finite dimensional. The functorial property of finite-dimensional Gaussian measures, (2), implies that there exists a well-defined finitely additive measure νth⊂b on the algebra of cylinder sets satisfying p∗ νth⊂b = νtf ,

(13)

for all p as above. Definition 2.1. (a) The inclusion h → b is called an abstract Wiener space if the finitely additive measure νth⊂b has a (necessarily unique) countably additive extension to a Borel probability measure √ on b, for some t > 0 (and hence all t > 0, because νth⊂b (E) = ν1h⊂b (E/ t)). (b) A map of abstract Wiener spaces, (h1 ⊂ b1 ) → (h2 ⊂ b2 )

(14)

is a map b1 → b2 such that the restriction to h1 is a projection h1 → h2 . For a map of abstract Wiener spaces, the corresponding Gaussian semigroups push forward. In many ways, the restriction to Banach space completions of h is artificial. However, the Banach framework seems elegant and natural, in part because of the following characterization and examples. Theorem 2.2. The inclusion h → b is an abstract Wiener space if and only if for each > 0 there exists a finite-dimensional subspace f ⊂ h such that νth⊂b {x ∈ b : |px|f > } <

(15)

for all p : b → f vanishing on f , where p : h → f is a finite rank projection. If b is itself a Hilbert space, this is equivalent to the condition that the inclusion h → b is a Hilbert–Schmidt operator.

398

Doug Pickrell

For a discussion of this theorem, see Section 3.9 of [B]. Example 2.3. (a) If f is a finite-dimensional Hilbert space, then f = h = b is an abstract Wiener space, and νtf is given literally by (12). (b) Suppose that X is a compact manifold, possibly with boundary, and W is a Hilbert Sobolev space of degree of smoothness s > dim(X)/2. Then the Sobolev embedding W → C 0 (X; R) (16) is an abstract Wiener space (see [CL]). This is closely related to the fact that W → W 0 (X) = L2 (X) is a Hilbert–Schmidt operator precisely when s is above the critical exponent. (c) For W as in (b), if h → b is an abstract Wiener space, then W (X, h) = W ⊗ h → C 0 (X, b)

(17)

is also an abstract Wiener space (see 3.11.29 of [B]). A special case of (c) is the Wiener space analogue of the path functor, W ⊗ h = W 1 ([0, T ], 0; h, 0) → C 0 ([0, T ], 0; b, 0)

(18)

(i.e., continuous b-valued paths beginning at 0 ∈ b) with inner product T x, yW 1 = x, ˙ y ˙ L2 = x(τ ˙ ), y(τ ˙ )h dτ. (19) 0

To stay within the Banach category, we have chosen T < ∞. By letting T → ∞, we obtain a Gaussian semigroup on the semi-infinite path space C 0 ([0, ∞), 0; b, 0).

(20)

For notational simplicity, we will write W ⊗h⊂C 0 ([0,∞),0;b,0)

ν h⊂b = ν1

,

(21)

and we will refer to this as the Brownian motion associated to h ⊂ b. Given a partition V : 0 < t1 < · · · < tn , there is an evaluation map EvalV : C 0 ([0, ∞), 0; b, 0) →

n

b : x → (xi ),

xi = x(ti ).

(22)

1

In terms of the coordinates xi , (EvalV )∗ ν h⊂b = dνth⊂b (x1 ) × dνth⊂b (x2 − x1 ) × · · · × dνth⊂b (xn − xn−1 ). 1 1 −t2 n −tn−1 (23) In particular, the distribution of ν h⊂b at time t is νth⊂b . The semigroup property of νth⊂b is equivalent to the consistency of these distributions for ν h⊂b when the partition is refined. The path space construction can be iterated. Thus given h ⊂ b, there is a Brownian motion, a Brownian sheet, etc., associated to h ⊂ b.


399

2.2 Abstract Wiener groups and heat kernel measures Definition 2.4. (a) An abstract Wiener group is an inclusion of a separable Hilbert–Lie group H into a Banach–Lie group G, with a fixed compatible inner product on the Lie algebra h, such that the associated Lie algebra map h → g is an abstract Wiener space. (b) A map of abstract Wiener groups is a map of pairs of Lie groups (H1 , G1 ) → (H2 , G2 ) such that the induced Lie algebra map h1 → h2 is a projection. (c) An abstract Wiener group H ⊂ G with the property that finite rank maps of Wiener groups p : G → F separate the topology of G is called local. Given an abstract Wiener group H ⊂ G, there is an induced left invariant Riemannian metric on H, and a left invariant Finsler metric on G. These induce separable complete metrics generating the topologies of H and G, respectively. Example 2.5. (a) If F is a finite-dimensional Lie group, with an arbitrary inner product on its Lie algebra f , then F = H = G is an abstract Wiener group. (b) If W is as in (b) of Example 2.3, given an abstract Wiener group H ⊂ G, W (X, H) ⊂ C 0 (X, G) (24) is a local abstract Wiener group. (c) The construction in (b) can be adapted to local gauge transformations of a nontrivial principal bundle P → X, provided that the structure group K is a compact Lie group with a fixed Ad(K)-invariant inner product. (d) There are also nonlocal examples involving infinite classical matrix groups; see [Go]. The following would be the ideal existence result. Conjecture 2.6. For fixed t ≥ 0, the sequence of probability measures h⊂g h⊂g exp∗ (νt/N ) ∗ · · · ∗ exp∗ (νt/N )

(25)

(an N -fold convolution of probability measures on G) has a weak limit with respect to BC(G), bounded continuous functions. The limits, denoted νtH⊂G , form a convolution semigroup of inversion invariant probability measures on G. Given a map P : (H1 ⊂ G1 ) → (H2 ⊂ G2 ) (26) of abstract Wiener groups, P∗ νtH1 ⊂G1 = νtH2 ⊂G2 , for each t ≥ 0. This conjecture is known to be true if one additionally assumes that the inclusion h ⊂ g is 2-summable (see [BD]). I have not stated this result in a standard form, and this obscures an important idea. To explain this idea, fix t > 0. Consider the map C 0 ([0, ∞), 0; g, 0) → G : x → gN (t) = ex1 ex2 −x1 · · · exN −xN −1 ,

(27)

400

Doug Pickrell

where xi = x(it/N ). The gN (t)-distribution of the Brownian motion measure ν h⊂g is the N -fold convolution (25), because the xi − xi−1 are independent (g-valued) Gaussian variables with variance t/N . Thus the conjecture is equivalent to the assertion that gN (t) has a limit as N → ∞, say g(t), with ν h⊂g -probability one. The latter statement implies that the heat kernel measure is an image of a Gaussian measure, an insight due to Ito. Example 2.7. (a) If F is a finite-dimensional real Lie group, with an arbitrary inner product on the Lie algebra f , νtF can be expressed as in (3). This measure is absolutely continuous with respect to the Haar measure, and it has a real analytic density. (b) Consider the path construction H = W 1 (I, 0; F, 1) → G = C 0 (I, 0; F, 1),

(28)

where I = [0, T ], and h has the norm (19) (with f in place of h in (18)). In this case νtH⊂G is the Brownian motion for F, with variance t. The path space for G is the double path space C 0 (I × I, ({0} × I) ∪ (I × {0}); F, 1),

(29)

and ν H⊂G is the F-valued Brownian sheet . The Brownian sheet is relevant to physics in the following way. Supppose that F = K, a compact Lie group, and the inner product is Ad(K)-invariant. Suppose that Σ is a closed surface with an area element. The Yang–Mills measure on the space of K-connections has the heuristic expression 1

e− 2

FA ∧∗FA

dλ(A)

(30)

for A ∈ Ω 1 (Σ; k), where FA is the curvature for the connection d + A, and dλ denotes a heuristic Lebesgue measure on the linear space of k-valued one forms. This expression can be interpreted as a finitely additive measure in the following way. Let pt (g)dλ(g) denote the heat kernel measure for K. Given a triangulation of Σ, one considers the projection Ω 1 (Σ, k) → K E : A → (ge ),

(31)

where E denotes the set of edges for the triangulation, and ge ∈ K represents parallel translation for the connection d + A. The image of the Yang–Mills measure is pt Area(f ) (g∂f ) dge , (32) F

E

where g∂f denotes the holonomy around the face f ∈ F . The fact that these measures are coherent with respect to refinement of the triangulation follows directly from the semigroup property of the heat kernel measures. When projected to gauge equivalence classes, this measure can be expressed as a finite-dimensional conditioning of the Brownian sheet (see [Sen]).


401

3 Invariance questions The linear invariance properties of Gaussian measure classes are summarized by the following. Theorem 3.1. Suppose that h → b is an abstract Wiener space. Let ν = νth⊂b for some t > 0. (a) The subset of translations in b which fix [ν] is h. (b) The group of linear transformations which fix [ν] is GL(h)(L2 ) , in the following sense: given a ν-measureable linear map T : b → b which fixes [ν], T = T|h ∈ GL(h)(L2 ) ; conversely, given T ∈ GL(h)(L2 ) , there exists a ν-measureable linear map T : b → b which fixes [ν] and satisfies T|h = T . These results, and their history, are discussed in a very illuminating way in [B] (where one can also find expressions for the Radon–Nikodym derivatives). Now suppose that H ⊂ G is an abstract Wiener group, and g ∈ G. When is [(Lg )∗ νtH⊂G ] = [νtH⊂G ], (33) where Lg is left translation? Necessarily g ∈ G0 , the identity component (because νtH⊂G is supported in the identity component). The abelian case suggests that g ∈ H. Because νtH⊂G is inversion invariant, translation invariance implies that [νtH⊂G ] is g-conjugation invariant. Thus if (33) holds for all t, this suggests that [νth⊂g ] is Ad(g)-invariant, and this implies that, as an operator on h, Ad(g) ∈ GL(h)(L2 ) . These considerations suggest the following. Conjecture 3.2. For T > 0 and h ∈ G, [νTH⊂G ] is left and right h-invariant if and only if h ∈ H0 and Ad(h) ∈ GL(h)(L2 ) . Hilbert–Schmidt criteria are universal in unitary representation-theoretic questions of this type. In the next section we will consider evidence in favor of this conjecture. Before leaving this abstract setting, I will mention two other related questions. When t → ∞, a Gaussian measure is asymptotically invariant in the following sense introduced in [MM]: given νt = νth⊂b and h ∈ h, for each p < ∞, dνt (b + h) p (34) dνt (b) → 0 1 − dνt (b) as t → ∞ (in this linear context, this integral reduces to a one-dimensional integral, which is easily estimated). Now suppose that Conjecture 3.2 is true, and let νt = νtH⊂G and h satisfy the conditions in Conjecture 3.2. Is it true that for each p < ∞, dνt (gh) p (35) 1 − dνt (g) → 0 dνt (g)

402

Doug Pickrell

as t → ∞? In [MM] it is proved that the Wiener measure on C 0 (S 1 , K) is asymptotically invariant in this sense (see Chapter 4, Part III, of [Pi1] for a quantitative version of this result, and an example of how this is used to prove existence of an invariant measure for the loop group). The second question is the following. Suppose that we change the inner product on h in a way which (by (b) of Theorem 3.1) does not change the measure class of νth⊂g : x, y1 = Ax, y,

x, y ∈ h

(36)

where A − 1 ∈ L2 (h). Write h1 for h with this inner product depending on A. Is [νtH1 ⊂G ] = [νtH⊂G ], for each t?

4 The Example of M ap(X, F) (see [M]) Let X denote a compact Riemannian manifold, and fix a Sobolev space (with a fixed inner product) of continuous real-valued functions W → C 0 (X),

(37)

necessarily corresponding to some degree of smoothness s > dim(X)/2. For clarity, we will suppose that ∂X is empty, and the inner product for W is of the form φ1 , φ2 W = (P s/2 (φ1 ))(P s/2 (φ2 ))dV, (38) X

where P denotes a positive elliptic pseudodifferential operator of order 2, e.g., P = (m20 + ∆), where ∆ is a (positive) Laplace-type operator. We will write G(x, y) for the Green’s function of P s ; thus Pys (G(x, y)dV (y)) = δx (y)

(39)

in the sense of distributions. Evaluation at x ∈ X defines a continuous linear functional δx : W → R : φ → φ(x); (40) this functional is represented by G(x, ·) ∈ W : φ(x) = φ, G(x, ·)W .

(41)

Let F denote a finite-dimensional abstract Wiener group. Then H = W (X, F) → G = C 0 (X, F)

(42)

is an abstract Wiener group, where h = W ⊗ f , as a Hilbert space. It is local because we can evaluate at points of X. Given a finite set of points V ⊂ X, let EvalV : C 0 (X, F) → FV : g → (g(v))

(43)


403

denote the evaluation map. This is a map of abstract Wiener groups, where f V has the Hilbert space structure induced by the Lie algebra map EvalV : W (X, f ) → f V .

(44)

It follows from (41) that this inner product on f V is given by Gv,w xv , yw f , (xv ), (yw ) =

(45)

v,w∈V

where (Gv,w ) is the matrix inverse to the covariance matrix (G(v, w))v,w∈V . More generally, given an embedded submanifold S ⊂ X, restriction induces a map of abstract Wiener groups C 0 (X, F) → C 0 (S, F),

(46)

where the inner product on W (S, f ) corresponds to a degree of smoothness s − codim(S)/2. This is very reminiscent of functorial properties which are exploited at a heuristic level for Feynman–Kac measures (see Section 6). (v) If f1 , . . . , fN denotes an orthonormal basis for f , and if fi denotes the V left invariant vector field on F corresponding to the v-coordinate, then the left invariant Laplacian on FV determined by the inner product (45) is given by N (v) (w) ∆V = G(v, w)fj fj , (47) v,w∈V j=1

and

(FV )

(EvalV )∗ νtH⊂G = e−t∆V /2 δ1

.

(48) νtH⊂G .

We will These finite-dimensional distributions determine the measure compare this with the Wiener measure below, when X = S 1 . We now want to understand the content of Conjecture 3.2. Suppose that g ∈ W (X, F). We need to compute the adjoint for Ad(g), denoted Ad(g)∗W , acting on the Hilbert–Lie algebra W (X, f ). Let Ad(g)t denote the adjoint for Ad(g) acting on L2 (X, dV ) ⊗ f . Then ∗W (49) Ad(g) (f1 ), f2 h = P s f1 Ad(g)(f2 )dV = Thus and

P s (P −s Ad(g)t P s )(f1 )f2 dV.

(50)

Ad(g)∗W = P −s ◦ Ad(g)t ◦ P s ,

(51)

Ad(g) Ad(g)∗W = Ad(g)P −s Ad(g)t P s .

(52)

404

Doug Pickrell

In general this is simply a zeroth order operator. However, suppose that Ad(g) has values in O(f ), the orthogonal group. Then Ad(g) Ad(g)∗W = (Ad(g)P −s − P −s Ad(g) + P −s Ad(g)) Ad(g)−1 P s = [Ad(g), P −s ] Ad(g)−1 P s + 1.

(53) (54)

Because of the commutator, this differs from the identity by an operator of order −1 (see [Fr]). To formulate this in operator language, let L+ d denote the ideal of compact operators T on h satisfying N

sup N

1 sn (T )d < ∞ log(N ) 1

(55)

where s1 ≥ s2 ≥ · · · are the eigenvalues of |T |. The fact that (54) has order −1 implies that it belongs to L+ d . This implies the following statement. Theorem 4.1. Let d = dim(X). (a) Suppose that g ∈ W (X, f ) and Ad(g(x)) ∈ O(f ), for each x ∈ X. Then Ad(g) ∈ GL(h)(L+ ) . d

(56)

(b) Suppose that F = K, or F = K C , where K is a compact Lie group, and the inner product on f is Ad(K)-invariant. Assuming the truth of Conjecture 3.2, [νtH⊂G ] is W (X, K)-invariant when X = S 1 , and noninvariant for dim(X) > 1. The important point is that dim(X) = 2 is the critical case, where the criterion marginally fails. We now want to consider evidence in support of Conjecture 3.2. ∂ 2 Suppose that X = S 1 , and P = m20 − ( ∂θ ) . For s > 1/2, G(θ1 , θ2 ) = G(θ1 − θ2 ) =

1 2 (m0 + n2 )−s cos(n(θ1 − θ2 )). π

(57)

n≥0

We will write νts,m0 for the corresponding heat kernel measures. When s = 1, we can compare νts,m0 with Wiener measure ωt . This latter measure has distributions −1 (EvalV )∗ ωt = pt (gv gw ) dλK (gv ), (58) E

V

where E denotes the set of edges (between vertices in V ), and νtK = pt dλK . The important point is that this is a nearest neighbor interaction, which is far more elementary than the interaction involving all pairs of points for heat kernel measures, as in (48). The following summarizes some of the deep results of Driver and collaborators on heat kernel measures for loop groups.


405

Theorem 4.2. Suppose that X = S 1 and F = K, a simply connected compact Lie group with Ad(K)-invariant inner product on f . (a) For s = 1, [νts,m0 ] = [ωt ]. (b) For all s > 1/2, [νts,m0 ] is W (S 1 , K)-biinvariant. A relatively short and illuminating discussion of (a) can be found in [Dr2]. Part (b) is a long story. For s = 1, part (b) follows from (a) and the relatively well-understood fact that [ωt ] is W 1 (S 1 , K)-biinvariant. Driver gave a direct proof of (b), for s = 1 ([Dr1]), and the ideas were shown to extend to all s > 1/2 in [I]. In support of Conjecture 3.2, one can also directly analyze the Brownian motion construction (b) of Example 2.7, using a stochastic Ito map (which essentially linearizes the problem). When X is a closed Riemannian surface, we expect that the heat kernel measure classes for W s,m0 (X, K) → C 0 (X, K) do just miss being translation invariant, for all s > 1. If this is the case, it is interesting to ask whether there is another natural construction which yields an invariant measure class. This should be compared with Shavgulidze’s construction of invariant measure classes for Dif f (X), for X of any dimension (see page 312 of [B]).

5 The critical Sobolev exponent and X = S 1 For s = dim(X)/2, the Sobolev embedding fails, and in general a W dim(X)/2 function on X is not bounded. This leads to a subtle situation. On the one hand, while W dim(X)/2 (X, k) is a Hilbert space, unless k is abelian, it is not a Lie algebra. On the other hand, because K is assumed compact, a map g ∈ W dim(X)/2 (X, K) is automatically bounded. Consequently, W dim(X)/2 (X, K), with the Sobolev topology (we do not impose uniform convergence), is a topological group, but it is not a Lie group (for the algebraic topology of mapping spaces with topologies defined by critical Sobolev norms, see [Br]). For non-abelian k, in general it is simply not clear how to form an analogue of (42), when s = dim(X)/2. In the rest of this section we will focus on the case X = S 1 . We will freely use facts about loop groups, as presented in [PS]. For simplicity of exposition, we will assume that K is simply connected, k is simple, and the inner product (on the dual) satisfies θ, θ = 2, where θ is a long root (the abelian case is essentially trivial, but requires qualifications). The complexification of K will be denoted by G. For each s > 1/2, there is a (Kac–Moody) universal central extension of Lie groups s (S 1 , K) → W s (S 1 , K) → 0. 0→T→W (59) This extension exists for s = 1/2 as well (as an extension of a topological group), and this is the maximal domain for the extension. This extension is

406

Doug Pickrell

realized using operator methods in [PS]. If we fix a faithful representation K ⊂ U (N ), a loop g : S 1 → K can be viewed as a unitary multiplication operator Mg on H = L2 (S 1 , CN ). Relative to the Hardy polarization H = H+ ⊕ H− , where H+ consists of functions with holomorphic extension to the disk, AB Mg = , (60) CD where B (or C) is the Hankel operator, and A (or D) is the Toeplitz operator, associated to the loop g. The condition g ∈ W 1/2 is equivalent to B ∈ L2 . The Toeplitz operator defines a map A : W 1/2 (S 1 , K) → Fred(H+ ).

(61)

The lift of the left action of W 1/2 (S 1 , K) on itself to A∗ Det, the pull-back of the determinant line bundle Det → Fred(H+ ), induces (a power of) the extension (59). The upshot is that, from an analytic point of view, Kac–Moody theory is intimately related to the critical exponent and Toeplitz determinants. The group C ω (S 1 , G) (of real analytic loops) is a complex Lie group. An open, dense neighborhood of the identity consists of those loops which have a unique (triangular or Birkhoff or Riemann–Hilbert) factorization g = g− · g0 · g+ ,

(62)

where g− ∈ H 0 (D∗ , ∞; G, 1), g0 ∈ G, g+ ∈ H 0 (D, 0; G, 1), and D and D∗ denote the closed unit disks centered at 0 and ∞, respectively. A model for this neighborhood is H 1 (D∗ , g) × G × H 1 (D, g),

(63)

−1 (∂g+ ), θ− = where the linear coordinates are determined by θ+ = g+ −1 (∂g− )g− . The (left or right) translates of this neighborhood cover C ω (S 1 , G). The hyperfunction completion, Hyp(S 1 , G), is modelled on the space

H 1 (∆∗ , g) × G × H 1 (∆, g),

(64)

where ∆ and ∆∗ denote the open disks centered at 0 and ∞, respectively, and the transition functions are obtained by continuously extending the transition functions for the analytic loop space. The group C ω (S 1 , G) acts naturally from both the left and right of Hyp(S 1 , G) (see Chapter 2, Part III of [Pi1]). There is a holomorphic line bundle L → Hyp(S 1 , G) with a map 1/2 (S 1 , K) → W L ↓ ↓ W 1/2 (S 1 , K) → Hyp(S 1 , G)

(65)

ω (S 1, K). which is equivariant with respect to natural left and right actions by C ∗ The line bundle L has a distinguished holomorphic section σ0 ; restricted to


407

W 1/2 loops, L∗m = A∗ Det and σ0m = det A, the pull-back of the canonical holomorphic section of Det → Fred(H+ ), where m is the ratio of the CN -trace form and the normalized inner product on k. We will write |L|2 for the line ¯ we can form real powers of this bundle, because the transition bundle L ⊗ L; functions are positive. In the following statement, note that (σ0 ⊗ σ ¯0 )l has ∗ 2l values in the dual bundle |L | . Theorem 5.1. For each l ≥ 0, there exists a C ω (S 1 , K)-biinvariant measure 2l dµ|L| with values in the line bundle |L|2l → Hyp(S 1 , G) such that ¯0 )l dµ|L| dµl = (σ0 ⊗ σ

2l

(66)

is a probability measure. In particular, there exists a C ω (S 1 , K)-biinvariant 0 probability measure dµ = dµ|L| on Hyp(S 1 , G). This is a refinement of the main result in [Pi1] (see the Introduction of [Pi1], and also [Pi2], for further background information and motivation). These bundle-valued measures are conjecturally unique. Uniqueness would imply invariance with respect to the natural action of analytic reparameterizations of S 1 . This independence of scale is the hallmark of the critical exponent. 2l One motivation for constructing the measure µ|L| is to prove a Peter– Weyl theorem of the schematic form H 0 ∩ L2 (L∗⊗l ) = H(Λ) ⊗ H(Λ)∗ , (67) level(Λ)=l

where the H(Λ) are the positive energy representations of level l ∈ Z+ . (This statement requires more explanation. Here we will simply say, Kac and Peterson proved an algebraic generalization of the Peter–Weyl theorem (see Section 1.7, Part I of [Pi1]), and we are seeking an analytic version, suitable for application to sewing rules discussed below.) This remains incomplete. Because |σ0 |2 = det |A|2/m , restricted to W 1/2 (S 1 , K), µl has a heuristic expression 1 dµl = det |A|2(l/m) dµ. (68) Z This begs two questions: (1) why do Toeplitz determinants have anything to do with (limits of) heat kernel measures, and (2) how do we write the background µ in a way which suggests how to think about it analytically? It is instructive to first consider the abelian case. When K = T, the theorem is valid provided l > 0 (the background µ does not exist, reflecting a lack of “compactness” in this flat case), and we consider identity components. An element of Hyp(S 1 , C∗ )0 can be written uniquely as xn z n ) · exp(x0 ) · exp( xn z n ), (69) exp( n0

408

Doug Pickrell

where the sums represent holomorphic functions in the open disks. If (69) represents a loop g ∈ W 1/2 (S 1 , T)0 , then det |A(g)|2l = exp(−l n|xn |2 ) (70) n>0

(the Helton–Howe formula). Consequently, dµl = dλT (exp(x0 ))

1 exp(−lk|xk |2 )dλ(xk ), Zk

(71)

k>0

and there is no diffusion in noncompact directions: x−k = −x∗k a.e. Since T dνts,m0 = dνm −2s (exp(x0 )) t 0

1 1 exp(− (k 2 + m20 )s |xk |2 )dλ(xk ), Zk 2t

(72)

k>0

(71) is the limit of heat kernel measures (with l = 1/2t) as s ↓ 1/2, provided that we also send the mass m0 to zero. Now suppose that K is simply connected. Since K and its loop group fit into the Kac–Moody framework, it is natural to compare µ to Haar measure dλK for K. Fix a triangular decomposition g = n − ⊕ h ⊕ n+ .

(73)

A generic g ∈ K (or G) can be written uniquely as g = lmau,

(74)

where l ∈ N − = exp(n− ), m ∈ T = exp(h ∩ k), a ∈ A = exp(hR ), and u ∈ N + = exp(n+ ). This implies that there is a K-equivariant isomorphism of homogeneous spaces, K/T → G/B + , and l ∈ N − is a coordinate for the top stratum. Harish-Chandra discovered that, in terms of this coordinate, the unique K-invariant probability measure on K/T can be written as a4δ dλ(l) =

|σi (g)|2 dλ(l) =

1 dλ(l), |l · vδ |4

(75)

where (74) implicitly determines a = a(gT ) as a function of l, the σi are the fundamental matrix coefficients, dλ(l) denotes a properly normalized Haar measure for N − , 2δ denotes the sum of the positive complex roots for the triangular factorization (73), and vδ is a highest weight vector in the highest weight representation corresponding to the dominant integral functional δ. The point of the third expression is that it shows that the denominator for the density is a polynomial in l; hence the integrability of (75) is a delicate issue. It also follows from work of Harish-Chandra that for λ ∈ h∗R 2δ, α (76) a−iλ dλK g = 2δ − iλ, α α>0


409

(the right-hand side is Harish-Chandra’s c-function for G/K). In particular, this determines the integrability of powers of a: for > 0, a(1+)2δ dλ(l) = −(d−r)/2 < ∞, (77) d and r denote the dimension and rank of g, respectively. The formula (75) can be derived by calculating a Jacobian in a straightforward way. The formula (76) can be deduced from the Duistermaat–Heckman localization principle (in particular, log(a) is a momentum map), using a (Drinfeld–Evens–Lu) Poisson structure which generalizes to the Kac–Moody framework (see [Pi3]). Now consider the loop situation. Recall that θ− is a coordinate for H 0 (∆∗ , 0; G, 1) (which is similar to N − ). The loop analogue of the formula (75) leads to the following heuristic expression for the θ− distribution of µl , where initially we think of θ− as corresponding to a unitary loop g ∈ W 1/2 (S 1 , K): 1 ˙ |σ0 (g)|2(2g+l) dλ(θ− ) Z 1 ˙ = det |A(g)|2(2g+l)/m dλ(θ− ), Z

(θ− )∗ µl =

(78) (79)

where g˙ is the dual Coxeter number, and dλ(θ− ) is a heuristic Lebesgue background. The important point is the shift by the dual Coxeter number, which reflects a regularization of the “sum over all positive roots”. In a similar manner the Duistermaat–Heckman approach to (76) applies in a heuristic way to compute the g0 distribution of µl :

a(g0 )−iλ dµl =

1 sin( 2(g+l) 2δ, α) ˙

α>0

1 sin( 2(g+l) 2δ − iλ, α) ˙

.

(80)

The function on the right-hand side is an affine analogue of Harish-Chandra’s c-function. The heuristic expression (79) explains why µl is expected to be invariant with respect to P SU (1, 1) (the Toeplitz determinant and the background Lebesgue measure are conformally invariant). It also points to a way of constructing the θ− (or g− ) distribution of µl , by imposing a cutoff and taking a limit: 1 ˙ det |A(g(PN θ− ))|2(2g+l)/m dλ(PN θ− ), N ↑∞ ZPN

d((g− )∗ µl )(θ− ) = lim

(81)

where PN θ− is the projection onto the first N coefficients, g− is related to −1 P θ− by P θ− = (∂g− )g− , and g is a unitary loop having Riemann–Hilbert factorization (62). Many, but not all, of the details of this construction have been worked out. The main idea is that log(det |A(g)|2 ) is part of a momentum map. Consequently, localization can be used to compute integrals involving

410

Doug Pickrell

the determinant against a symplectic volume element, on finite-dimensional approximations. The formula (80) for the g0 distribution remains a conjecture, but it seems to point in a fruitful direction. The zero-mode g0 roughly arises from the projection 0, ∞; G, 1) → Hyp(S 1 , G)generic → G. H 0 (C, (82) More generally, given a closed Riemann surface Σ and a real analytic embed a G-hyperfunction induces a holomorphic G-bundle on Σ, and ding S 1 → Σ, consequently there is a bundle \ S 1 , G) → Hyp(S 1 , G) → H 1 (Σ, OG ), H 0 (Σ

(83)

OG ) denotes the set of isomorphism classes of holomorphic Gwhere H 1 (Σ, OG ) identify with the irreducible points bundles. The stable points of H 1 (Σ, 1 of H (Σ, K), which has a canonical symplectic structure (see [AB]). This suggests the following. Question 5.2. Does dµ project to the (normalized) symplectic volume on the OG )? stable points of H 1 (Σ, Continuous loops have Riemann–Hilbert factorizations, and as a consequence there is a diagram W s (S 1 , K) → C 0 (S 1 , K) ↓ ↓ 1/2 1 W (S , K) → Hyp(S 1 , G) for each s > 1/2. We consider the W s norm (m20 + n2 )s |xn |2 ,

(84)

(85)

n

and νts,m0 , the corresponding heat kernel measures, which we view as measures on Hyp(S 1 , G). Conjecture 5.3. The measures νts,m0 have a limit as s ↓ 1/2, and the measure class of this limit equals the measure class of dµl , where l = 1/2t. When m0 ↓ 0, this limit converges to µl . The intuition is simply that when m0 = 0, the heat kernel measures should relax to a distribution which is conformally invariant, as s ↓ 1/2. As we saw above, the conjecture is true in the abelian case, and in this case the limit is expressed naturally in terms of Toeplitz determinants. In general the heat kernel measure class [νts,m0 ] is biinvariant with respect to W s (S 1 , K), s > 1/2. It is expected that [µl ] will be biinvariant with respect to W 1/2 (S 1 , K), in the sense that the natural induced unitary representation C ω (S 1 , K × K) → U (L2 (dµl ))

(86)

will extend continuously to W 1/2 (S 1 , K × K). These symmetry considerations fit nicely with the conjecture.


411

6 2D quantum field theory The functorial properties of heat kernel measures (see (46)) are reminiscent of properties generally ascribed to path integrals. We will illustrate these properties using what is called a P (φ)2 model. We will then use this as a guide in deciding what is reasonable to ask about heat kernel measures as s ↓ 1. The relevant aspect of quantum field theory (QFT) has been formalized by Segal in a very attractive way. Fix a dimension d, thought of as the dimension of (Euclidean) space-time. As in Section 4 of [Se], let Cmetric denote the category for which the objects are oriented closed Riemannian (d − 1)manifolds, and the morphisms are oriented compact Riemannian d-manifolds with totally geodesic boundaries. Definition 6.1. A primitive d-dimensional unitary QFT is a representation of Cmetric by separable Hilbert spaces and Hilbert–Schmidt operators such that disjoint union corresponds to tensor product, orientation reversal corresponds to adjoint, and Cmetric -isomorphisms correspond to natural Hilbert space isomorphisms. To my knowledge there does not exist an example of a nonfree QFT which has been shown to satisfy Segal’s primitive axioms in dimension d > 2. However, in two dimensions, the space of all theories is definitely large (from one point of view, this space is the configuration space of string theory, and it is expected to have a remarkable geometric structure (see [Ma])). We will begin by recalling how a “generic” theory is constructed. The rationale is that these generic theories are rigorously defined by path integrals (see [Pi4]). Fix (a bare mass) m0 > 0, and a polynomial P : R → R which is bounded the P (φ)2 -action is the local below. Given a closed Riemannian surface Σ, functional 1 →R:φ→ A : F (Σ) ( (|dφ|2 + m20 φ2 ) + P (φ))dA, (87) 2 Σ is the appropriate domain of R-valued fields on Σ for A. where F (Σ) A heuristic expression for the P (φ)2 -Feynman–Kac measure is exp(−A(φ)) dλ(φ(x)). (88) x∈Σ

In this two-dimensional setting, there is a rigorous interpretation of (88) as a finite measure on generalized functions, 1 e− detζ (m20 + ∆)1/2

Σ

:P (φ):C0

dφC ,

(89)

412

Doug Pickrell

where C = (m20 + ∆)−1 , dφC is the Gaussian probability measure with

: covariance C (corresponding to the critical Sobolev space W 1 (Σ)), P (φ) :C0 denotes a regularization of the nonlinear interaction, and detζ denotes the zeta function determinant. Suppose that S is a closed Riemannian 1-manifold. Given an inner product for W 1/2 (S), there is an associated Gaussian measure on generalized functions. Associated to the measure class of this Gaussian measure, say C(S), there is a Hilbert space H(S), the space of half-densities associated to C(S) (see the appendix to [Pi4]). This space is independent of the mass and the length scales of the components of S. Suppose that Σ is a Riemannian surface with totally geodesic boundary = Σ ∗ ◦ Σ obtained by gluing S. We consider the closed Riemannian surface Σ ∗ Σ to its mirror image Σ along S. Because dφC corresponds to the critical exponent s = 1, typical configurations for the Feynman measure (89) are not ordinary functions. However, typical configurations are sufficiently regular so that it makes sense to restrict them to S. Consequently, the Feynman measure can be pushed forward to a finite measure on generalized functions along S. Essentially because the underlying map of Hilbert spaces is the trace map → W 1/2 (S), this finite measure lands in the measure class C(S). By W 1 (Σ) we obtain a taking a square root (which is required because Σ is half of Σ), half-density Z(Σ) ∈ H(S). The main result of [Pi4] is the following. Theorem 6.2. The maps S → H(S) and Σ → Z(Σ) define a representation of Segal’s category. This construction extends to RN -valued fields. But severe problems arise for QFTs with classical fields having values in nonlinear targets. The geometric constraints have to be relaxed to accommodate quantum fluctuations. For example, for the target S n−1 , it is believed that the corresponding quantum sigma model is a limit of theories (in the same universality class) involving potentials of the form P (φ) = λ(|φ|2 − a2 )2 , where φ has values in Rn . These potentials retain the symmetry, but relax the geometry, of the target. This point of view explains some qualitative features of the sigma model (see [K]), but it does not seem to shed light on how to construct the sigma model itself. A basic assumption in what follows is that there should be an intrinsic geometric construction. A naive approach to constructing a projective theory, which might approximate the sigma model with target K, is described by the following rough outline. Fix a degree of smoothness s > 1 and a time t > 0. For a closed Riemannian 1-manifold S, define H(S) to be the space of half-densities corresponding to the measure class of a heat kernel measure (for t) corresponding to W s−1/2 (S, K) (as in the P (φ)2 case, the details of the inner product on the Lie algebra should not affect this measure class and the Hilbert space). the Given a Riemannian surface Σ with geodesic boundary S and double Σ, functorial properties of heat kernel measures imply that the heat kernel mea K) will push forward to a measure in the sure (for t) corresponding to W s (Σ,


413

measure class corresponding to W s−1/2 (S, K). We can thus define Z(Σ) to be the line determined by the square root of this pushforward. While these definitions make sense, one suspects that the homomorphism property fails to hold (projectively), because the heat kernel measures do not have a local character. (It is not clear whether there is some reasonable “why” to quantify this failure.) It is undoubtedly essential to let s ↓ 1 in some sense. The meaning of the Feynman–Kac measure for the sigma model is described in terms of renormalization group analysis (see Lecture 3 of [Ga2]). In this analysis the parameter t in the Feynman–Kac measure, and perhaps other details of the measure, depend dynamically on a momentum scale, which one can think of as s in the present context. This suggests the following question. Suppose that we consider the pushforward of the heat kernel measures cor K) to the hyperfunction completion Hyp(S, G) (or the responding to W s (Σ, larger formal completion considered in [Pi1] and [Pi2]). If we let t depend on s, do these measures have a limit as s ↓ 1? (in the sigma model analysis, t is proportional to ln(s)). It is important to note that the action for the sigma model can be augmented to include a Wess–Zumino–Witten term, so that the Feynman–Kac measure is of the form 1 exp(− g −1 dg ∧ ∗g −1 dg + 2πilW ZW (g)) dλK (g(x)). (90) 2t Σ x When the level l is a positive integer, and t = 1/2l, this (W ZWl ) model is solvable in many senses (see Lecture 4 of [Ga2], and [We]). It is believed that the model satisfies Segal’s axioms (see the Foreword to [Se]). It is also believed that there is a one-parameter family of completely integrable massive theories which interpolates between W ZWl and the original sigma model (see [ORW] and [Sm]). The heuristic expression (90) is not known to have a mathematical interpretation as a measure. However, when t = 1/2l, (90) can be plausibly manipulated, along the same lines as the P (φ)2 path measure. This is explained in [Ga1] and [Ga2]. The upshot is that the Hilbert space of the W ZWl model is given by the right-hand side of (67), which one can interpret as a surrogate for the space of half-densities corresponding to W 1/2 (S, K), taking into account the twisting by the WZW term. The vacuum of the theory, the vector corresponding to a disk, is det(A)l , which one can interpret as a “holomorphic square root” of dµl (the limit of heat kernel measures, according to Conjecture 5.3). For a general compact surface Σ, Z(Σ) can similarly be interpreted as a square root (see (14) and (15) of [Ga1]). The proof of sewing (in the holomorphic sectors, involving the W ZWl modular functor, see page 468 of [Se]), ultimately hinges on a generalized Peter–Weyl theorem; the conjectural analytical version (67) would fit perfectly with Segal’s global approach.

414

Doug Pickrell

References [A] [AB] [AM] [BD] [B] [Br] [DS] [CL] [Dr1] [Dr2] [Fr] [Ga1] [Ga2] [Go] [Gr] [I] [K] [MM] [M]

S. Albeverio, R. Hoegh-Krohn, J. Marion, D. Testard, B. Torresani (1993) Noncommutative Distributions, Pure and Applied Mathematics, Vol. 175, Marcel Dekker, New York. M. Atiyah and R. Bott (1982) The Yang-Mills equations over Riemann surfaces, Phil. Trans. R. Soc. Lond. A 308, 523–615. H. Airault and P. Malliavin (2006) Quasi-invariance of Brownian measures on the group of circle homeomorphisms and infinite dimensional Riemannian geometry, J. Funct. Anal. 241, 99–142. Ya. I. Belopolskaya and Yu. L. Dalecky (1990) Stochastic Equations and Differential Geometry, Kluwer, Dordrecht. V. I. Bogachev (1998) Gaussian Measures, Mathematical Monographs 62, Amer. Math. Soc. H. Brezis (2006) New questions related to topological degree. In: Prog. Math. 244, Birkhauser, 137–154. Yu. L. Daletskii and Ya. I. Snaidermann (1969) Diffusion and quasiinvariant measures on infinite dimensional Lie groups, Funct. Anal. Appl. 3, no. 2, 156–158. P. Colella and O. E. Lanford (1973) Sample field behavior for the free Markov random field, in Constructive Quantum Field Theory, edited by G. Velo and A. Wightman, Springer, New York, 44–70. B. Driver (1997) Integration by parts and quasi-invariance for heat kernel measures on loop groups, J. Funct. Anal. 149, no. 2, 470–547. B. Driver (2003) Analysis of Wiener measure on path and loop groups. In: Finite and Infinite Dimensional Analysis in Honor of Leonard Gross, Contemp. Math. 317, Amer. Math. Soc., 57–85. D. Freed (1987) The geometry of loop groups, J. Diff. Geom. 28, 223–276. K. Gawedzki (1990) Wess–Zumino–Witten conformal field theory. In: Constructive Quantum Field Theory, II (Erice, 1988) NATO Adv. Sci. Inst. Ser. B Phys. 234, Plenum, 89–120. K. Gawedzki (1999) Lectures on conformal field theory. In: Quantum Fields and Strings: A Course for Mathematicians, Vol. 2, Amer. Math. Soc., 727– 805. M. Gordina (2005) Heat kernel analysis on infinite dimensional groups. In: Infinite Dimensional Harmonic Analysis III, World Sci. Publ., Hackensack, NJ, 71–81. L. Gross (1965) Abstract Wiener spaces, Proc. 5th Berkeley Symp. Math. Stat. Prob., Part 1, Univ. CA Press, Berkeley, 31–41. Y. Inahama (2001) Logarithmic Sobolev inequality on free loop groups for heat kernel measures associated with the general Sobolev space, J. Funct. Anal. 179, no. 1, 170–213. C. Kopper (1999) Mass generation in the large N-nonlinear sigma model, Comm. Math. Phys. 202, no. 1, 89–126. M.P. Malliavin and P. Malliavin (1992) Integration on loop groups, III. Asymptotic Peter–Weyl orthogonality, J. Funct. Anal. 108, 13–46. P. Malliavin (1991) Hypoellipticity in infinite dimensions. In: Diffusion Process and Related Problems in Analysis, edited by M.A. Pinsky, Birkhauser, Boston, MA.

Heat Kernel Measures [Ma]

415

E. Martinec (1990) The space of 2d quantum field theories. In: The Interface of Mathematics and Particle Physics, Oxford University Press, Oxford, 107– 115. [McK] H. McKean (1969) Stochastic Integrals, Probability and Mathematical Statistics, No. 5, Academic Press. [ORW] E. Olgievtsky, N. Reshetikhin, P. Weigmann (1987) The principal chiral field in two dimensions on classical Lie algebras. The Bethe-ansatz solution and factorized theory of scattering, Nuclear Phys. B 280, no. 1, 45–96. [Pi1] D. Pickrell (2000) Invariant measures for unitary forms of Kac-Moody groups, Memoirs Amer. Math. Soc., Number 693, 1–144. [Pi2] D. Pickrell (2006) An invariant measure for the loop space of a simply connected compact symmetric space, J. Funct. Anal. 234, 321–363. [Pi3] D. Pickrell (2006) The diagonal distribution for the invariant measure of a unitary type symmetric space, Transf. Groups, vol. 11, no. 4, 705–724. [Pi4] D. Pickrell (2008) P (φ)2 quantum field theories and Segal’s axioms, Comm. Math. Phys. 280, 403–425. [PS] A. Pressley and G. Segal (1986) Loop Groups, Oxford University Press, Oxford. [Se] G. Segal (2003) The definition of conformal field theory. In: Topology, Geometry and Quantum Field Theory, L.M.S. Lect. Note Sers. 308, Cambridge University Press, Cambridge, UK, 423–576. [Sen] A. Sengupta (1997) Gauge theory on compact surfaces, Memoirs Amer. Math. Soc., Number 600, 1–84. [Sm] F. A. Smirnov (1992) Dynamical symmetries of massive integrable models. In: Infinite Analysis, A. Tsuchiya, T. Eguchi and M. Jimbo (eds.), Adv. Ser. Math. Phys. 16, World Scientific, 813–838. [St] D. Stroock (1996) Gaussian measures in traditional and not so traditional contexts, Bull. Amer. Math. Soc. 33, no. 2, 135–155. [We] R. Wendt (2001) A symplectic approach to certain functional integrals and partition functions, J. Geom. Phys. 40, 65–99.

Coadjoint Orbits and the Beginnings of a Geometric Representation Theory Tudor S. Ratiu Section de Mathématiques, Ecole Polytechnique Fédérale de Lausanne, CH-1015 Lausanne, Switzerland, [email protected] Summary. This is a review of Banach Poisson manifolds with special emphasis on Banach–Lie–Poisson spaces. Unlike the finite-dimensional case, the existence of the (weak) symplectic leaves is not guaranteed even for coadjoint orbits in the predual of the associated Lie algebra. A significant part of the paper is hence devoted to examples of such coadjoint orbits. A special class of K¨ ahler orbits is isolated for which the classical Borel–Weil theorem extends to the Banach case by the use of reproducing kernels. Key words: Banach Poisson manifold, Banach–Lie–Poisson space, C ∗ -algebra, W ∗ algebra, predual, coadjoint orbit, operator ideals, restricted unitary algebra, Grassmannian, GNS representation, reproducing kernel, geometric representation theory, Borel–Weil theorem. 2000 Mathematics Subject Classifications: Primary 22E46, 46T05, 53D17, 53D20. Secondary 22E46, 22E65, 22E67, 46E22, 46L10, 46L30, 46L55, 46L65, 47L20, 53D50, 53Z05, 58B12.

1 Introduction Symplectic and Poisson geometry together with Lie group actions on it lie at the foundation of the geometric formulation of physical phenomena. The theory is in very good shape in finite dimensions, and there are numerous books and articles that present this material from various points of view and emphasize different kinds of applications in both pure and applied mathematics, as well as physics, chemistry, biology, robotics, and various engineering sciences (see, for example, [AM78], [Ar89], [AKN06], [Bl03], [BL05], [CJ04], [Co02], [CB97], [CMOI], [Da07], [DCC], [DG05], [Ga83], [GRKSD], [GS90], [HLW], [HZ94], [JSa98], [Ju97], [LM87], [LiR97], [LiM02], [Ma92], [MMOPR], [MMR], [MR94], [McDS], [McIO], [MoR05], [OB98], [ODK], [O02], [Ol86], [OrR04], [PD07], [ST07], [SDM], [SCRa], [SCRb], [Si01], [Si05], [Su03], [T00], [WS88], [YKMK], and references therein). K.-H. Neeb and A. Pianzola (eds.), Developments and Trends in Infinite-Dimensional Lie Theory, Progress in Mathematics 288, DOI 10.1007/978-0-8176-4741-4_13, © Springer Basel AG 2011

417

418

Tudor S. Ratiu

In infinite dimensions, the situation is very unsatisfactory. There are essentially three levels of generalizations of symmetry that have permeated all applications. The simplest ones are Lie algebras and Lie groups whose elements are maps from a manifold to a given finite-dimensional Lie algebra or Lie group. The operations are pointwise. Often one considers certain natural extensions of such objects. The typical, and by far the best developed, such example is the family of Kac–Moody Lie algebras and of the various objects associated to them (underlying Lie groups, representations, theory of coadjoint orbits, etc.) The past decades have seen an explosive growth of work in this area with fundamental achievements that generalize the finite-dimensional theory in a spectacular way. The second level of infinite-dimensional Lie groups and Lie algebras are constructed from operator algebras. At this stage of generalization, all finite-dimensional vestiges have been eliminated, and one has to deal with genuine infinite-dimensional problems using the full power of functional analysis. Nevertheless, these objects still retain one key property: they are genuine Lie algebras and Lie groups and the finite-dimensional theory does serve as a guide to formulate reasonable conjectures regarding their structure and properties of their (linear and non-linear) actions. The third level, and by far the most complicated one, is formed by manifolds of maps, diffeomorphism groups, classes of vector fields, and their actions. Here, one is immediately faced with the necessity to specify topologies and smooth structures. The results heavily depend on these choices. Unfortunately, there is no universal selection of a manifold structure, and the guide is to take the one that is most natural for the examples one wants to consider. Thus, for certain geometric properties, a specific choice may seem very appropriate, but one may find out later, when a partial differential equation of physical significance appears in the study, that a different smooth structure is better adapted to it. The present review is aimed at the second level described above, that is, the study of symplectic and Poisson geometric properties of Lie groups, Lie algebras, and their actions constructed from various algebras of operators. The key applications of such objects lie in quantum mechanics. Due to the heavy functional analytic machinery available, such objects are in the process of being understood. The point of view of this work is geometric, that is, it tries to answer typical symplectic and Poisson geometric questions that one poses to classical functional analytic objects. This review is based on joint work with D. Beltit¸a˘, A. Odzijewicz, and A.B. Tumpach [OR03], [BR05], [BR07], [BRT07] and is written closely following my presentation at the Oberwolfach meeting Infinite Dimensional Lie Theory in December 2006. In §2, the notion of Banach Poisson manifold is introduced and discussed. The special case of Banach–Lie–Poisson spaces is investigated in §3 and particular attention is given to W ∗ -algebras. The symplectic leaves and, in particular, coadjoint orbits of Banach–Lie groups whose Lie algebras admit preduals, are presented in §4. The rest of the paper is

Coadjoint Orbits & the Beginnings of a Geometric Representation Theory

419

entirely devoted to examples. In §5 various coadjoint orbits in certain families of operator algebras are presented. In particular, orbits of the restricted unitary group with special emphasis on Grassmannians and their invariant complex structures are discussed. The beginning of a geometric representation theory modeled on the Borel–Weil theorem is formulated in §6. Acknowledgements: This paper could not have been written without my collaboration with D. Beltit¸˘a, A. Odzijewicz, and A.B. Tumpach. It has been a privilege to work with them on these interesting problems. The partial support of the Swiss National Science Foundation is gratefully acknowledged.

2 Banach Poisson manifolds Sections 2–4 review some of the key results in [OR03]. We begin here with the definition and elementary properties of Banach Poisson manifolds. Recall that in finite dimensions (P, {·, ·}) is said to be a Poisson manifold if the ring of smooth real-valued functions C ∞ (P ) has a Lie algebra bracket {·, ·} satisfying the Leibniz property: {f g, h} = f {g, h} + {f, h}g for all smooth functions on P . For the theory of various aspects of Poisson manifolds see, e.g., [We83], [MR86], [LM87], [MR94], [Va94], [We98], [OrR04], and references therein. 2.1 The definition One could try to adopt the same definition if P is an infinite-dimensional Banach manifold. However, this leads immediately to several problems. Indeed, by the Leibniz property, the value of the Poisson bracket at a given point p ∈ P depends only on the differentials df (p), dg(p) ∈ Tp∗ P , which implies that there is a section of the vector bundle ∧2 T ∗∗ P satisfying {f, g} = (df, dg). We shall assume throughout this paper that is a smooth section, that is, for each p ∈ P the map p : Tp∗ P × Tp∗ P → R is a continuous bilinear antisymmetric map that depends smoothly on the base point p. Let : T ∗ P → T ∗∗ P be the bundle map covering the identity defined by p (dh(p)) := (·, dh)(p), that is, p (dh(p))(dg(p)) = {g, h}(p), for any locally defined functions g and h. The symbol d denotes the exterior derivative on forms and hence on functions. Denote Xf := (·, df ) = (df ), or, as a derivation, Xf = {·, f }. Then Xf is a smooth section of T ∗∗ P and hence is not, in general, a vector field on P . In analogy with the finite-dimensional case, we want Xf to be the Hamiltonian vector field defined by the function f . In order to achieve this, there are several ways to proceed. The simplest one is to assume that the Poisson bracket on P satisfies the condition (T ∗ P ) ⊂ T P ⊂ T ∗∗ P , which leads to the following definition.

420

Tudor S. Ratiu

Definition 2.1. A Banach Poisson manifold is a pair (P, {·, ·}) consisting of a smooth Banach manifold P and a bilinear operation {·, ·} : C ∞ (P ) × C ∞ (P ) → C ∞ (P ) satisfying the following conditions: (i) (C ∞ (P ), {·, ·}) is a Lie algebra; (ii) {·, ·} satisfies the Leibniz identity on each factor; (iii) the vector bundle map : T ∗ P → T ∗∗ P covering the identity satisfies (T ∗ P ) ⊂ T P . By (iii), every h ∈ C ∞ (P ) defines a Hamiltonian vector field Xh ∈ X(P ) by Xh [f ] := df, Xh = {f, h}, where f is an arbitrary smooth locally defined function on P . A smooth map ϕ : (P1 , {·, ·}1 ) → (P2 , {·, ·}2 ) is canonical or Poisson if ϕ∗ {f, g}2 = {ϕ∗ f, ϕ∗ g}1 for any two smooth locally defined functions f and g on P2 . Using (iii), this is equivalent to Xf2 ◦ ϕ = T ϕ ◦ Xf1◦ϕ for any smooth locally defined function f on P2 . Therefore, the flow of a Hamiltonian vector field is a Poisson map and Hamilton’s equations in Poisson bracket formulation f˙ = {f, h} are valid. Classical Poisson reduction in the spirit of [MR86] extends to Banach Poisson manifolds if one takes the closure of the relevant subspaces appearing in the reduction method. For details see [OR03]. We shall not discuss this construction in the present review. A second way to deal with the problem that has values in the bidual tangent bundle instead of the tangent bundle itself is to work with densely defined Hamiltonian vector fields and to deal with harder analytic phenomena that are the non-linear analogues of unbounded operators. The only fundamental attempt along these lines was made in [CM74] and, unfortunately, this approach, eminently suited for the theory of non-linear partial differential equations, has not been further developed in a systematic way, although physically interesting examples abound; see [KP03], [Ku00], and references therein for the rigorous treatment of certain aspects of this program. We shall not comment on this point of view further here because this approach is in need of major development. 2.2 Banach symplectic manifolds Let ω ∈ Ω 2 (P ) be a 2-form on the Banach manifold P . If the continuous linear map vp ∈ Tp P → ω(p)(vp , ·) ∈ Tp∗ P is • injective, ω is said to be weak, • bijective, ω is said to be strong.


421

A strong (weak) symplectic form is a closed strong (weak) form on P . Any strong symplectic Banach manifold (P, ω) is a Banach Poisson manifold: for any f ∈ C ∞ (P ) there exists a unique vector field Xf , the Hamiltonian vector field associated to f , defined by df = ω(Xf , ·). The Poisson bracket is defined by {f, g} = ω(Xf , Xg ) = df, Xg ; thus df = Xf , so (T ∗ P ) ⊂ TP. If (P, ω) is weak, this construction breaks down. This shows that Definition 2.1 generalizes the notion of a strong symplectic manifold. It should be emphasized right away that there is a huge difference between strong and weak symplectic manifolds. For example, the usual Lie transform method (the “Moser trick”) extends without any difficulties from the finite-dimensional case to the strong Banach case to prove the Darboux theorem: any strong symplectic form is locally constant. For weak symplectic manifolds this statement is no longer true; see, for example, Exercise 3.2H in [AM78] for Marsden’s classical counterexample. This and other major differences with the strong symplectic case impose in the weak case, as already mentioned above, a different approach such as restricting the function spaces and working with vector fields that are only densely defined. Thus, it comes as a surprise that the restricted notion of Banach Poisson manifold in Definition 2.1 yields large classes of very interesting examples. One of the goals of this review is to present some of them in the framework of Banach–Lie groups and algebras.

3 Banach–Lie–Poisson spaces This section presents the simplest Banach Poisson manifolds that are not strong symplectic manifolds, namely those whose underlying manifold is a Banach space and whose Poisson bracket is linear. 3.1 Characterization A Banach–Lie algebra (g, [·, ·]) is a Banach space g that is also a Lie algebra such that the Lie bracket [·, ·] : g × g → g is bilinear and continuous. The adjoint and coadjoint maps associated to a given element x ∈ g are denoted, as usual, by adx : y ∈ g → [x, y] ∈ g and ad∗x : g∗ → g∗ , where y, ad∗x b = [x, y], b for all y ∈ g and b ∈ g∗ ; here ·, · : g × g∗ → R denotes the duality pairing between g and its Banach space dual g∗ . The maps adx and ad∗x are linear and continuous. With these preparations we can define Banach–Lie– Poisson spaces. Definition 3.1. A Banach–Lie–Poisson space (b, {·, ·}) is a Banach Poisson manifold with b a Banach space and such that the dual b∗ ⊂ C ∞ (b) is a Banach–Lie algebra under the Poisson bracket operation.

422

Tudor S. Ratiu

For (b, {·, ·}) a Banach–Lie–Poisson space, denote by [·, ·] the restriction of the Poisson bracket {·, ·} from C ∞ (b) to the Lie subalgebra b∗ , that is, [x, y] := {x, y} for all x, y ∈ b∗ ⊂ C ∞ (b). Banach–Lie–Poisson spaces are characterized in the following manner. Theorem 3.2. The Banach space b is a Banach–Lie–Poisson space (b, {·, ·}) if and only if its dual b∗ is a Banach–Lie algebra (b∗ , [·, ·]) satisfying ad∗x b ⊂ b ⊂ b∗∗ for all x ∈ b∗ . Moreover, the Lie–Poisson bracket of f, g ∈ C ∞ (b) is given by {f, g}(b) = [Df (b), Dg(b)], b, (1) where b ∈ b and D denotes the Fréchet derivative. If h is a smooth function on b, the associated Hamiltonian vector field is given by Xh (b) = − ad∗Dh(b) b.

(2)

For example, if b is a reflexive Banach–Lie algebra, then its dual b∗ is a Banach–Lie–Poisson space. In particular, this recovers the classical situation of finite-dimensional Lie–Poisson spaces. A morphism between two Banach–Lie–Poisson spaces b1 and b2 is a continuous linear map φ : b1 → b2 that preserves the Poisson bracket structure, that is, {f ◦ φ, g ◦ φ}1 = {f, g}2 ◦ φ for any f, g ∈ C ∞ (b2 ); φ is also called a linear Poisson map. Note that a continuous linear map φ : b1 → b2 is a morphism of Banach–Lie–Poisson spaces if and only if φ∗ : b∗2 → b∗1 is a Banach–Lie algebra homomorphism. At this point the category of Banach–Lie–Poisson spaces has been constructed, and one can investigate various properties of its objects and morphisms. For details see [OR03]. 3.2 Examples As already mentioned, there are large classes of examples of Banach–Lie– Poisson spaces. We begin by describing the simplest ones, namely various families of operator algebras. One of the earliest references where concrete examples of Banach–Lie groups and Banach–Lie algebras of operators in Hilbert space were studied in a systematic way is [Ha72]. Let H be a complex Hilbert space. Denote by • • • •

S(H) = S(H)1 , the trace class operators on H, HS(H) = S(H)2 , the Hilbert–Schmidt operators on H, K(H), the compact operators on H, B(H) = gl(H), the bounded operators on H.


423

All of them are involutive Banach algebras. In addition, S(H), HS(H), K(H) are self-adjoint ideals in B(H) and we have the continuous inclusions S(H) ⊂ HS(H) ⊂ K(H) ⊂ B(H). On the other hand, the following duality relations exist between these spaces: K(H)∗ ∼ = S(H),

HS(H)∗ ∼ = HS(H),

S(H)∗ ∼ = B(H).

The right-hand sides of these Banach space isomorphisms are all Banach– Lie algebras. These dualities are implemented by the strong non-degenerate pairing x, ρ = Trace (xρ), where x ∈ S(H), ρ ∈ K(H) for the first isomorphism, ρ, x ∈ HS(H) for the second isomorphism, and x ∈ B(H), ρ ∈ S(H) for the third isomorphism. Therefore, the Banach spaces S(H), HS(H), and K(H) are Banach–Lie– Poisson spaces. The Lie–Poisson bracket becomes in this case {F, H}(ρ) = Trace ([DF (ρ), DH(ρ)]ρ), where ρ is an element of S(H), HS(H), or K(H), respectively. The bracket [DF (ρ), DH(ρ)] denotes the commutator bracket of operators. The Hamiltonian vector field associated to H is given by the Lax equation ρ˙ = XH (ρ) = [DH(ρ), ρ]. Some of these examples generalize to a larger family, namely, the preduals of W ∗ -algebras. Recall that a W ∗ -algebra is a C ∗ -algebra m which possesses a predual Banach space m∗ , i.e., m = (m∗ )∗ ; this predual is unique. Since m∗ = (m∗ )∗∗ , the predual Banach space m∗ canonically embeds into the Banach space m∗ dual to m. Thus, we shall always think of m∗ as a Banach subspace of m∗ . Recall that a net {xα }α∈A ⊂ m converges to x ∈ m in the σ-topology if, by definition, limα∈A xα , b = x, b for all b ∈ m∗ . The σ-topology is Hausdorff. The Alaoglu theorem states that the unit ball of m is compact in the σtopology. A fundamental result in the theory of W ∗ -algebras characterizes the predual space in the following way: the predual space m∗ is the subspace of m∗ consisting of all σ-continuous linear functionals (see, for example, [Sa71], [Ta79]). The link between W ∗ -algebras and Banach–Lie–Poisson spaces is given by the following result. Theorem 3.3. The predual m∗ of a W ∗ -algebra is a Banach–Lie–Poisson space.

424

Tudor S. Ratiu

Proof. To show this, we need to check the conditions characterizing a Banach– Lie–Poisson space, namely, that m is a Banach–Lie algebra and ad∗a m∗ ⊂ m∗ for each a ∈ m. Since m is an associative Banach algebra, it is a Banach–Lie algebra relative to the commutator bracket [x, y] = xy − yx of x, y ∈ m. So the first condition is satisfied. To verify the second condition, we proceed in the following way. It is known that left and right multiplication by a ∈ m define uniformly and σ-continuous maps La : m x → ax ∈ m and Ra : m x → xa ∈ m. Let L∗a : m∗ → m∗ and Ra∗ : m∗ → m∗ denote their duals. If v ∈ m∗ , then L∗a (v) and Ra∗ (v) are σ-continuous functionals and therefore, by the characterization of the predual m∗ as the subspace of σ-continuous functionals in m∗ , it follows that L∗a (v), Ra∗ (v) ∈ m∗ . Since ada = [a, ·] = La − Ra we conclude that ad∗a v = L∗a (v) − Ra∗ (v) ∈ m∗ for every v ∈ m∗ . Corollary 3.4. Let a be a C ∗ -algebra. Then its dual a∗ is a Banach–Lie– Poisson space. Proof. The bidual a∗∗ is isomorphic to the universal enveloping W ∗ -algebra of a. The C ∗ -algebra a can be thought of as a C ∗ -subalgebra of a∗∗ by the canonical inclusion a → a∗∗ . Since a∗ is predual to a∗∗ , the previous theorem guarantees that it is a Banach Lie–Poisson space. Let m∗ be the predual of the W ∗ -algebra m and ι : m∗ → m∗ be the canonical inclusion. Then ι is an injective linear Poisson map, and the Poisson structure induced by it from m∗ coincides with the original Lie–Poisson structure on m∗ . The theory of Banach–Lie–Poisson spaces is intimately connected to coherent states quantization. The literature on this subject is vast, and we shall not review it here. The papers that are closest in spirit to this Poisson geometric approach are [Od88], [Od92], [OR03], [OS97]. The references cited in these articles are an excellent guide to the literature.

4 Symplectic leaves The study of the internal structure of Banach–Lie–Poisson spaces is the subject of the rest of this review. For a finite-dimensional Poisson manifold P , the smooth characteristic distribution S whose fiber at a given point p ∈ P is given by the characteristic subspace Sp := {Xf (p) | f ∈ C ∞ (P )} is involutive in the sense of the Stefan–Sussmann theorem (see [St74a, St74b, Su73] or for a discussion of these results [LM87, Va94, OrR04]). Thus, this generalized distribution is integrable and hence P admits a generalized foliation


425

determined by the condition that the tangent space to each leaf is precisely the value at that point of the characteristic distribution. The leaves of this generalized foliation are called symplectic leaves since the Poisson structure on P induces on each one of them a symplectic structure whose Poisson bracket coincides with the given one. The symplectic leaf through a point p ∈ P is the set of all points that can be reached from p by a broken continuous path formed by a finite number of pieces of integral curves of Hamiltonian vector fields, that is, it is the accessible set of p. For Banach Poisson manifolds this approach no longer works, essentially due to the lack of a reasonable infinite-dimensional extension of the Stefan–Sussmann theorem. There have been attempts at generalizing the Stefan–Sussmann theorem to Banach manifolds (see, e.g., [Nu92]), but the few theorems that are available in the literature all have hypotheses that are violated for many examples of Banach Poisson manifolds. There is a generalization of the classical Frobenius theorem to Banach manifolds if the distribution is given by a vector subbundle ([Bo71], [Bo72], [AMR]); in particular, the fiber at each point (that is, the characteristic subspace) is a Banach subspace of the tangent space at that point. As discussed below, this hypothesis is almost never verified for Banach Poisson manifolds, so even Lie’s classical proof giving the decomposition into symplectic leaves of Poisson manifolds whose characteristic distributions are subbundles, does not generalize to the Banach manifold setting. As we shall see below, the situation is quite bad and not much extends from the finite-dimensional situation to the category of Banach Poisson manifolds, in general. Even reasonable examples do not behave well. 4.1 Attempt at constructing symplectic leaves Already the notion of immersion has three possible generalizations to Banach manifolds. A smooth map f : M → N between Banach manifolds is called (i) an immersion if for every m ∈ M the tangent map Tm f : Tm M → Tf (m) N is injective with closed split range; (ii) a quasi immersion if for every m ∈ M the tangent map Tm f : Tm M → Tf (m) N is injective with closed range; (iii) a weak immersion if for every m ∈ M the tangent map Tm f : Tm M → Tf (m) N is injective. Obviously, any immersion is a quasi immersion and any quasi immersion is a weak immersion. The reverse implications clearly do not hold, in general. The notion of immersion is by far the best one, and it is used in all books dealing with Banach manifolds (such as [Bo67], [Bo71], [Bo72], [AMR]). The usual properties of immersions in finite dimensions generalize without any difficulties to Banach manifolds. For example, every immersion is locally an inclusion in one factor of the model of the manifold N which necessarily splits.

426

Tudor S. Ratiu

Quasi immersions retain some of the properties of immersions, but are far less convenient. They appear in modern infinite-dimensional Lie theory when one tries to define the notion of a Lie subgroup of a not necessarily split Lie subalgebra (see §4.2). Weak immersions are rarely used because they do not have good properties. As we shall see, in the category of Banach Poisson manifolds, unfortunately, one needs to deal with this weakest notion. If (P, {·, ·}P ) is a Banach Poisson manifold, we shall adopt the same definitions, conventions, and terminology as in the finite-dimensional case. Thus, the vector subspace Sp := {Xf (p) | f ∈ C ∞ (U ), U open in P, p ∈ U } of Tp P is called the characteristic subspace at p; it is not closed in Tp P , in general. The subset S := ∪p∈P Sp ⊂ T P , is called the characteristic distribution of the Poisson structure on P ; it is not a subbundle of T P , in general. However, S is always a smooth generalized distribution, that is, for every vp ∈ Sp ⊂ Tp P there is a locally defined smooth vector field that is a section of S (namely some Xf ) and whose value at p is vp . Suppose that the characteristic distribution is integrable. Recall that for finite-dimensional manifolds this is always the case. In the Banach Poisson manifold setting this is not automatic, so it has to be assumed. This is the first major difference with the finite-dimensional case. More differences will become apparent below. Let L be a leaf of the characteristic distribution, that is, (i) (ii) (iii) (iv)

L is a connected smooth Banach manifold, the inclusion ι : L → P is a weak injective immersion, Tq ι(Tq L) = Sq for each q ∈ L, if the inclusion ι : L → P is another weak injective immersion satisfying the three conditions above and L ⊂ L , then necessarily L = L, that is, L is maximal. If we assume, in addition, that on the leaf L

(v) there is a weak symplectic form ωL consistent with the Poisson structure on P , then L is called a symplectic leaf. The term “consistent” in point (v) needs an explanation. The linear continuous map p : Tp∗ P → Tp P induces a bijective continuous map [p ] : Tp∗ P/ ker p → Sp . By definition, ωL is consistent with the Poisson structure on P if ωL (q)(uq , vq ) = (ι(q)) ([ι(q) ]−1 ◦ Tq ι)(uq ), ([ι(q) ]−1 ◦ Tq ι)(vq ) , (3) for all q ∈ L, uq , vq ∈ Tq L, where ι : L → P is the inclusion (which is a weak immersion) and is the Poisson structure on P . Recall that is a smooth section of the vector bundle ∧2 T ∗∗ P defined by {f, g} = (df, dg) for any smooth locally defined functions f and g on P . The arguments on


427

the right-hand side of (3) are in Tp∗ P/ ker p , so the expression has to be understood in the following way. Recall that, by definition of a Banach Poisson manifold, p : Tp∗ P → Tp P is defined by αp , p (βp ) := (p) (αp , βp ) for any αp , βp ∈ Tp∗ P . Therefore, (p) induces a continuous bilinear skew-symmetric real form on Tp∗ P/ ker p , which we will denote, by abuse of notation, by the same symbol (p). So, on the right-hand side of (3), (ι(q)) is this bilinear ∗ form on the quotient Tι(q) P/ ker ι(q) . Note that formula (3) shows that the weak symplectic form ωL on L, consistent with the Poisson structure on P , is unique, if it exists. For finite-dimensional Poisson manifolds all leaves are automatically symplectic so assumption (v) above is not necessary. In the infinite-dimensional case it is not known, in general, if the leaves are (weak) symplectic, even for Banach–Lie–Poisson spaces, a situation that will be discussed in great detail in the rest of this review beginning with §4.2. Proposition 4.1. Let ι : (L, ωL ) → P be a symplectic leaf of the Banach Poisson manifold (P, {·, ·}) and U ⊂ P an open subset. Then (i) for any f ∈ C ∞ (U ), q ∈ ι−1 (U ) ∩ L, one has −1 d (f ◦ ι)|ι−1 (U ) (q) = ωL (q) (Tq ι) (Xf (ι(q))), · ; (ii) the subspace ι∗ (C ∞ (P )) of C ∞ (L) consisting of functions that are obtained as restrictions of smooth functions from P is a Poisson algebra relative to the bracket {·, ·}L given by −1 −1 {f ◦ ι, g ◦ ι}L (q) := ωL (q) (Tq ι) (Xf (ι(q))), (Tq ι) (Xg (ι(q))) ; (iii) ι∗ : (C ∞ (P ), {·, ·}P ) → (ι∗ (C ∞ (P )), {·, ·}L ) is a Poisson homomorphism, that is, {f ◦ ι, g ◦ ι}L = {f, g}P ◦ ι, for any f, g ∈ C ∞ (P ). Part (i) of this proposition is remarkable since it guarantees the existence of Hamiltonian vector fields on the weak symplectic manifold L for a large class of functions, namely those that are pull-backs to the symplectic leaf. 4.2 Coadjoint orbits in Banach–Lie–Poisson spaces We begin by recalling a few standard facts from the theory of Banach–Lie groups. A homomorphism f : H → G of Banach–Lie groups is a smooth map that is a group homomorphism. A subset H ⊂ G is a split Lie subgroup of the Banach–Lie group G if H is a submanifold and a subgroup of G. In Chapter 3 of [Bo72] these are called Lie subgroups, but we shall use this term for a weaker notion discussed below. It follows that, relative to the induced group and manifold structures, H is also a Lie group. In addition, it is known ([Bo72], Chapter 3, §1, Proposition 5) that H is necessarily closed and that the

428

Tudor S. Ratiu

inclusion H → G is a Banach–Lie group homomorphism and an embedding (not just an injective immersion or initial map [OrR04]). If H is a subgroup of the Banach–Lie group G, then it is a split Lie subgroup of G if and only if there is an open neighborhood V in G of the identity element such that H ∩ V is a submanifold of G ([Bo72], Chapter 3, §1, Proposition 6). The following criterion for H to be a split Lie subgroup of G is often useful. If K is another Banach–Lie group, a map g : K → H satisfying g(K) ⊂ H is a Lie group homomorphism from K to H if and only if it is a Lie group homomorphism from K to G ([Bo72], Chapter 3, §1, Proposition 5). The quotient G/H of a Banach–Lie group G by a split Banach–Lie subgroup H carries a unique manifold structure making the canonical projection G → G/H a surjective submersion ([Bo72], Chapter 3, §1, Proposition 11). If G is a finite-dimensional Lie group, it is well known that every closed subgroup H is necessarily a Lie subgroup ([Bo72], Chapter 3, §6, Proposition 2). For infinite-dimensional Banach–Lie groups this result is no longer true, in general; counterexamples are given in [Bo72], Chapter 3, §8, Exercise 2, [Ho75], and [Ne06]. Because of this, one introduces the following definition. A closed subgroup H of a Banach–Lie group G which is itself a Lie group relative to the induced topology is called a Lie subgroup. Note that the Lie algebra of a Lie subgroup H is not necessarily split in the Lie algebra of G. The closed subgroups of G which are Banach–Lie groups in their own right are precisely the Lie subgroups of G (this was first observed in [Ho75]; see also Theorem 4.3.3 in [Ne06]). It is not known if quotients by Lie subgroups are smooth manifolds. However, all other major properties for split Lie subgroups also hold for Lie subgroups (see the review [Ne06]). Let us address the question of the existence of symplectic leaves for Banach–Lie–Poisson spaces. To do this, we shall need additional hypotheses that imitate the finite-dimensional situation. Let G be a Banach–Lie group whose Banach–Lie algebra is g. We shall assume that (i) g admits a predual g∗ , that is, (g∗ )∗ = g ; (ii) the coadjoint action of G on the dual g∗ leaves the predual g∗ invariant, that is, Ad∗g (g∗ ) ⊂ g∗ , for any g ∈ G; (iii) if ρ ∈ g∗ , the closed subgroup Gρ := {g ∈ G | Ad∗g ρ = ρ} of G is a split Lie subgroup of G (a submanifold of G and not just injectively immersed). In (iii) one has to assume that Gρ is a split Lie subgroup since, as was discussed previously, closed subgroups of infinite-dimensional-Banach–Lie groups are not necessarily split Lie subgroups. If Gρ is a split Lie subgroup, then one can easily show that the Lie algebra of Gρ equals gρ := {ξ ∈ g | ad∗ξ ρ = 0}. In addition, as mentioned earlier, the quotient topological space G/Gρ := {gGρ | g ∈ G} admits a unique smooth Banach manifold structure, making the canonical projection π : G → G/Gρ a surjective submersion.


429

Theorem 4.2. Under the hypotheses above, the manifold G/Gρ is weak symplectic relative to the 2-form ωρ given by ωρ ([g])(Tg π(Te Lg ξ), Tg π(Te Lg η)) := ρ, [ξ, η] , where ξ, η ∈ g, g ∈ G, [g] := π(g) = gGρ , and ·, · : g∗ × g → R is the canonical duality pairing between g∗ and g. The 2-form ωρ is invariant under the left action of G on G/Gρ given by g · [h] := [gh], for g, h ∈ G. The weak symplectic structure of the coadjoint orbits of G through points of g∗ is given by the following theorem. Theorem 4.3. Let the Banach–Lie group G and the element ρ ∈ g∗ satisfy the hypotheses of the previous theorem. Then the map ι : [g] ∈ G/Gρ → Ad∗g−1 ρ ∈ g∗ is an injective weak immersion of the quotient manifold G/Gρ into the predual space g∗ . Endow the coadjoint orbit O := {Ad∗g ρ | g ∈ G} with the smooth manifold structure making ι into a diffeomorphism. The push forward ωO := ι∗ (ωρ ) of the weak symplectic form ωρ ∈ Ω 2 (G/Gρ ) to O has the expression ωO (Ad∗g−1 ρ) ad∗Adg ξ Ad∗g−1 ρ, ad∗Adg η Ad∗g−1 ρ = ρ, [ξ, η], for g ∈ G, ξ, η ∈ g, and ρ ∈ g∗ . Relative to this orbit symplectic form, the connected components of the coadjoint orbit O are symplectic leaves of the Banach–Lie–Poisson space g∗ . There is a remarkable particular situation that ensures that the symplectic form is strong. Theorem 4.4. Let the Banach–Lie group G and the element ρ ∈ g∗ satisfy the previous hypotheses. The following conditions are equivalent: (i) ι : G/Gρ → g∗ is an injective immersion; (ii) the characteristic subspace Sρ := {ad∗ξ ρ | ξ ∈ g} is closed in g∗ ; (iii) Sρ = g◦ρ , where g◦ρ is the annihilator of gρ in g∗ . Endow the coadjoint orbit O := {Ad∗g ρ | g ∈ G} with the manifold structure making ι a diffeomorphism. Then, under any of the hypotheses (i)–(iii), the weak symplectic form ωO is strong. Next, we turn to the study of concrete examples.

5 Coadjoint orbits in operator spaces In this section we shall give examples of coadjoint orbits of Banach–Lie groups that satisfy the hypotheses of the previous theorems. In addition, we shall inquire when some of these coadjoint orbits are Kähler. Orbits of the restricted unitary group will also be considered.

430

Tudor S. Ratiu

5.1 Symplectic leaves in preduals of W ∗ -algebras The first subsections give concrete examples of coadjoint orbits in various families of operator algebras. They review some of the major results in [BR05]. We begin by recalling standard notation and definitions from the theory of C ∗ -algebras. Let a be a unital C ∗ -algebra. If a ∈ a and ϕ ∈ a∗ denote: {a} = {b ∈ a | ab = ba}, the centralizer of a ∈ a, aϕ = {a ∈ a | ϕ(ab) = ϕ(ba) for all b ∈ a}, the centralizer of ϕ ∈ a∗ , Ga = {g ∈ a | g invertible} ⊂ gl(a), open, so it is a Banach–Lie group, Ua = {u ∈ a | uu∗ = u∗ u = 1} ⊆ Ga is a split Lie subgroup of Ga , ua = {a ∈ a | a = −a∗ } is the Banach–Lie algebra of Ua , asa = {a ∈ a | a = a∗ }, the self-adjoint elements of a, (a∗ )sa = {ϕ ∈ a∗ | ϕ(a∗ ) = ϕ(a)for all a ∈ a}, the self-adjoint functionals. In the special case of a W ∗ -algebra m we will also use the notation: m∗ = {ϕ ∈ m∗ | ϕ is w∗ -continuous}, the (unique) predual of m, ∗ sa msa ∗ = m∗ ∩ (m ) . Recall that an element ϕ ∈ m∗ is called faithful if ϕ(a∗ a) > 0 whenever 0 = a ∈ m and self-adjoint if ϕ = ϕ∗ , in the sense that ϕ(a∗ ) = ϕ(a) for all a ∈ m. As noted in §5.2 in a more general setting, faithful functionals are necessarily self-adjoint. The following statement (Theorem 2.2 in [BR05]) gives information about the unitary orbit of a faithful element in m∗ . It is based on several results in [AV99], [AV02], [Be06], [HK77], [Ta72]. Theorem 5.1. Let m be a W ∗ -algebra and ϕ ∈ m∗ faithful. The unitary orbit Uϕ = {ϕ ◦ Ad(u) | u ∈ Um } Um /Umϕ of ϕ, where Umϕ := {a ∈ Um | ϕ(ab) = ϕ(ba) for all b ∈ m} is the unitary group of the centralizer algebra mϕ , i.e., the unitary elements of mϕ . Then: (i) Uϕ ⊆ m∗ , (ii) The unitary group Umϕ of the centralizer algebra mϕ is a split Lie subgroup of Um , (iii) The unitary orbit Uϕ has a natural structure of weakly immersed submanifold of m∗ and Um acts on it smoothly on the left via (u, ψ) ∈ UM × Uϕ → ψ ◦ Ad(u−1 ) ∈ Uϕ , (iv) The smooth manifold Uϕ is simply connected. If m∗ is separable, it is known (see Remark A.2.2 in [JS97]) that there are faithful elements in m∗ .


431

In particular, if m = B(H) for some complex infinite-dimensional Hilbert space H, then, for any faithful state ϕ ∈ m∗ = S(H) (that is, ϕ = 1) it is known (see [AV99]) that the orbit Uϕ is not locally closed in S(H). Thus, the weakly immersed submanifolds Uϕ are never embedded submanifolds of S(H). Theorem 5.1 concerns the unitary orbits Uϕ of Um passing through a special type of elements ϕ ∈ m∗ , namely the faithful ones. The next theorem extends this result to the unitary orbits through all elements in (um )∗ . It is important to note that the predual (um )∗ consists of all self-adjoint (not skew-adjoint) elements of m∗ . Theorem 5.2. Let m be a W ∗ -algebra and ϕ∗ = ϕ ∈ m∗ . Then: (i) Umϕ is a split Lie subgroup of Um , (ii) The coadjoint orbit of the Lie group Um through ϕ ∈ (um )∗ ⊆ (um )∗ has the structure of a Um -homogeneous weak symplectic manifold which is weakly immersed into (um )∗ . Corollary 5.3. Let m be an arbitrary W ∗ -algebra and consider the corre∗ sponding real Banach–Lie–Poisson space msa ∗ = {ϕ ∈ m∗ | ϕ = ϕ }. Then sa the characteristic distribution of m∗ is integrable, and all its maximal integral manifolds are symplectic leaves and coincide with the Um -coadjoint orbits through all elements of msa ∗ . They are hence weakly immersed weak symplectic submanifolds of msa . ∗ Sometimes these weak symplectic manifolds are strong symplectic. For instance, if m = B(H) for some complex Hilbert space H, the coadjoint orbits of rank-one projections are strong symplectic because one can show that each such orbit is symplectically diffeomorphic to the projectivized Hilbert space P(H), which is known to be strong symplectic. Which coadjoint orbits are embedded? To my knowledge, even in finite dimensions this question is not answered. It is known that for algebraic groups this is the case and there are examples of Lie groups where this is not true (see [MR94], §14.1, Example(f) which presents Kirillov’s classical example). In infinite dimensions we have the following examples of embedded coadjoint orbits due to Bona [Bon00], [Bon04]. Theorem 5.4. The embedded coadjoint orbits of GL(H) through elements of S(H) are precisely those containing finite-rank operators. On unitary orbits of finite-rank operators on Hilbert spaces, the natural quotient topology coincides with the trace-class topology. A similar characterization of the embedded unitary orbits in B(H) and in an arbitrary C ∗ -algebras can be found in [AS89] and [AS91], respectively.

432

Tudor S. Ratiu

5.2 Symplectic leaves in C ∗ -algebras Certain results from the previous subsection can be extended to C ∗ -algebras. We present them below. For the proofs see [BR07] (which uses results from [Be02], [Be05], [Be06], [BR05], [KR97], [OR03]). Let a be a unital C ∗ -algebra. A continuous linear functional ϕ on a is said to be positive (respectively faithful ) if ϕ(a∗ a) ≥ 0 for every a ∈ a (respectively ϕ(a∗ a) > 0 for every 0 = a ∈ a). Continuous positive linear functionals are necessarily self-adjoint. A continuous positive linear functional ϕ ∈ a∗ is tracial if ϕ(a∗ a) = ϕ(aa∗ ) for every a ∈ a. A continuous positive functional ϕ ∈ a∗ with ϕ = 1 is called a state. Let a be a unital C ∗ -algebra and b a unital C ∗ -subalgebra of a. A bounded linear idempotent map E : a → b of norm one is called a conditional expectation. A theorem of Tomiyama (Theorem 1 in [To57] or Theorem 9.1 in [Str81]) states that E has the additional properties: E(a∗ ) = E(a)∗ ,

0 ≤ E(a)∗ E(a) ≤ E(a∗ a),

E(b1 ab2 ) = b1 E(a)b2 ,

for all a ∈ A and b1 , b2 ∈ B. We have the following result. Proposition 5.5. Let a be a unital C ∗ -algebra having a faithful tracial state τ : a → C. Define Θτ : a ∈ a → Θaτ ∈ a∗ , where Θaτ : b ∈ a → τ (ab) ∈ C. Then: τ

(i) We have ker Θτ = {0} and aΘa = {a} for all a ∈ a. (ii) The map Θτ is Ga -equivariant with respect to the adjoint action of Ga on a and the coadjoint action of Ga on a∗ . In particular, the mapping Θτ |asa : asa → (a∗ )sa is Ua -equivariant with respect to the adjoint action of Ua on asa and the coadjoint action of Ua on (a∗ )sa . (iii) For each a ∈ a the mapping Θτ induces a bijection of the adjoint orbit Ga ·a onto the coadjoint orbit Ga ·Θaτ . Thus, if a ∈ asa and there exists a conditional expectation of a onto {a} then UaO ·a L LLL τ LLΘL LLL % / Ua ·Θaτ Ua / U{a} is a commutative diagram of Ua -equivariant diffeomorphisms of Banach manifolds. (iv) If, moreover, a is a W ∗ -algebra and the faithful tracial state τ is normal (that is, τ is continuous on a relative to the weak ∗ topology), then im Θτ ⊆ a∗ and the hypothesis on conditional expectation in (iii) holds for each a ∈ asa . This proposition, together with the results of the previous section, has the following consequences.


433

Corollary 5.6. Let a be a unital C ∗ -algebra having a faithful tracial state τ : a → C and let a = a∗ ∈ a be such that there exists a conditional expectation of a onto {a} . Then the unitary orbit of a is naturally a Ua -homogeneous weak symplectic manifold. Corollary 5.7. In a W ∗ -algebra m that admits a faithful normal tracial state the unitary orbit of each self-adjoint element has a natural structure of a Um -homogeneous weak symplectic manifold. The result in Corollary 5.6 can be strengthened using other hypotheses. Proposition 5.8. Let a be a unital C ∗ -algebra and a = a∗ ∈ a with finite spectrum. Then the unitary orbit of a has a natural structure of a Ua -homogeneous complex Banach manifold. Adding to the hypothesis above the existence of a faithful tracial state yields a much stronger result. Theorem 5.9. If a unital C ∗ -algebra a possesses a faithful tracial state, then the unitary orbit of each self-adjoint element with finite spectrum has a natural structure of a Ua -homogeneous weak K¨ ahler manifold. 5.3 Symplectic leaves in preduals of operator ideals There are results similar to the ones in the previous two subsections for preduals of operator ideals proved in [BR05]. The present subsection reviews some of them. Let H be a separable complex Hilbert space. A Banach ideal is a twosided ideal I of B(H) equipped with a norm · I satisfying T ≤ T I = T ∗ I and AT BI ≤ A T I B whenever A, B ∈ B(H). For every Banach ideal J of B(H) define UJ = U(H) ∩ (1 + J), which is a Banach–Lie group uI = u(H) ∩ J, its Banach–Lie algebra. Proposition 5.10. Let J be a Banach ideal whose underlying Banach space is reflexive and let T ∈ uJ . Then UJ,T := {U ∈ UJ | U T U −1 = T } is a split Lie subgroup of UJ . This proposition has the following important consequence. Corollary 5.11. Let (B, J) be a pair of Banach ideals in B(H) and assume that BJ ⊂ S(H). Suppose that the trace pairing B × J → C,

(T, S) → Trace(T S)

induces a topological isomorphism of J onto the topological dual B∗ .

434

Tudor S. Ratiu

(i) If the underlying Banach spaces of B and J are reflexive, then the characteristic distribution of the real Banach–Lie–Poisson space uB = (uJ )∗ is integrable and all its maximal integral manifolds are symplectic leaves. (ii) Assume that the Banach–Lie group UJ is connected. Let F denote the ideal of all finite-rank operators on H. If T ∈ uB ∩ F, then the homogeneous space UJ /UJ,T has a UJ -invariant weak K¨ ahler structure and this homogeneous space is weakly immersed into uB . Here UJ,T := {U ∈ UJ | U T U −1 = T }. A typical example for which part (i) applies is a pair of Schatten ideals (Sp , Sq ) with p, q ∈ (1, ∞) and 1/p + 1/q = 1. If in part (ii) we take B = J = HS (the Hilbert–Schmidt ideal), then the weak K¨ ahler structure on the homogeneous space UJ /UJ,T guaranteed by the corollary is always strong, as shown in [Ne04]. To understand the next statement we need to introduce the concept of symmetric norming functions. Let c be the vector space of all real sequences {ξj }j≥1 such that ξj = 0 for all but finitely many indices. A symmetric norming function is a function Φ : c → R satisfying the following conditions: (i) (ii) (iii) (iv) (v)

Φ(ξ) > 0 whenever 0 = ξ ∈ c, Φ(αξ) = |α|Φ(ξ) whenever α ∈ R and ξ ∈ c, Φ(ξ + η) ≤ Φ(ξ) + Φ(η) whenever ξ, η ∈ c, Φ((1, 0, 0, . . . )) = 1, Φ({ξj }j≥1 ) = Φ({ξπ(j) }j≥1 ) whenever {ξj }j≥1 ∈ c and π : {1, 2, . . . } → {1, 2, . . . } is bijective.

Any symmetric norming function Φ gives rise to two Banach ideals SΦ (0) and SΦ as follows. For every bounded sequence of real numbers ξ = {ξj }j≥1 define Φ(ξ) := sup Φ(ξ1 , ξ2 , . . . , ξn , 0, 0, . . . ) ∈ [0, ∞]. n≥1

For all T ∈ B(H) denote T Φ := Φ({sj (T )}j≥1 ) ∈ [0, ∞], where sj (T ) = inf{T − F | F ∈ B(H), rank F < j} whenever j ≥ 1. Define SΦ := {T ∈ B(H) | T Φ < ∞}, (0)

SΦ := F (0)

·Φ

⊆ SΦ ,

that is, SΦ is the · Φ -closure of the finite-rank operators F in SΦ . Then (0) · Φ is a norm making SΦ and SΦ Banach ideals. Actually, every separable (0) Banach ideal equals SΦ for some symmetric norming function Φ. Typical examples of such ideals are the Schatten class ideals Sp for 1 ≤ p ≤ ∞. In this case the symmetric norming function is given by Φp (ξ) = ξp for ξ ∈ c.


435

Every symmetric norming function Φ : c → R induces another unique symmetric norming function Φ∗ : c → R, the adjoint of Φ, by ⎧ ⎫ ∞ ⎨ 1 ⎬

ξj ηj ξ = {ξj }j≥1 ∈ c and ξ1 ≥ ξ2 ≥ · · · ≥ 0 Φ∗ (η) := sup ⎩ Φ(ξ) ⎭ j=1

whenever η = {ηj }j≥1 ∈ c and η1 ≥ η2 ≥ · · · ≥ 0. The relation (Φ∗ )∗ = Φ always holds. If Φ is any symmetric norming function, then the topological (0) dual of the Banach space SΦ is isometrically isomorphic to SΦ∗ by means of the duality pairing (0)

SΦ∗ × SΦ → C,

(T, S) → Trace(T S).

For more information on this subject see Sections 11 and 12 of Chapter III in [GK69]. For example, if 1 ≤ p, q ≤ ∞, 1/p + 1/q = 1, Φp (ξ) = ξp , and Φq (ξ) = ξq whenever ξ ∈ c, then (Φp )∗ = Φq . Theorem 5.12. Let Φ and Ψ be symmetric norming functions, J = SΦ , T = T ∗ ∈ F, and UJ (T ) := {V ∗ T V | V ∈ UJ }. Then the following assertions hold: (i) The orbit map π : UJ → F,

V → V ∗ T V,

induces a diffeomorphism of the homogeneous space UJ /UJ,T onto the submanifold UJ (T ) of SΨ . (ii) If, moreover, Ψ ∗ = Φ and the Banach–Lie group UJ is connected, then the orbit UJ (T ) is a UJ -homogeneous weak K¨ ahler manifold. Note that part (i) involves two (completely unrelated to each other) symmetric norming functions. Topologically, this corresponds to the fact that any two symmetric norming functions define the same topology (namely the norm topology) on any unitary orbit of a finite-rank operator. There are many symmetric norming functions that define various types of operator ideals like Schatten, Lorentz, Orlicz and so on. See [DFWW] for a survey of this subject. 5.4 The restricted unitary algebra and its central extension In this subsection we continue to construct examples of weak symplectic leaves in Banach–Lie–Poisson spaces. The Banach–Lie groups and Lie algebras we consider here have a more intricate internal structure. The statements of the next three subsections review the major results in [BRT07]. Let H be a complex Hilbert space and assume that H = H+ ⊕ H− is the orthogonal sum of two closed subspaces H± . This decomposition is fixed throughout the subsequent discussion. Sometimes, additional hypotheses are imposed on H such as separability or the infinite-dimensionality of both direct

436

Tudor S. Ratiu

summands H± . Let p± : H → H± denote the corresponding orthogonal projections. Denote, as usual, the Banach ideal of trace class operators in B(H) by S(H) and the Hilbert ideal of Hilbert–Schmidt operators by HS(H). The norms on these spaces are denoted by · 1 and · 2 , respectively. The constructions that follow involve the Banach–Lie group U(H) = {u ∈ B(H) | u∗ u = uu∗ = I} of unitary operators on H whose Banach– Lie algebra u(H) = {a ∈ B(H) | a∗ = −a} is formed by skew-Hermitian operators. The element d := i(p+ − p− ) ∈ u(H) allows us to introduce the restricted Banach algebra Bres : = {a ∈ B(H) | [d, a] ∈ HS(H)} = {a ∈ B(H) | ares := a + [d, a]2 < ∞}, and the restricted unitary group Ures = {u ∈ U(H) | [d, u] ∈ HS(H)} = U(H) ∩ Bres . The Banach–Lie algebra of Ures is ures = {a ∈ u(H) | [d, a] ∈ HS(H)} = u(H) ∩ Bres . Proposition 5.13. The Banach–Lie algebra (ures )∗ := {ρ ∈ u(H) | [d, ρ] ∈ HS(H), p± ρ|H± ∈ S(H± )} is a predual of the restricted unitary Banach–Lie algebra ures , the duality pairing · , · : (ures )∗ × ures → R being given by (b, c) → Trace(bc). A connected Banach–Lie group with Lie algebra (ures )∗ is U1,2 = {a ∈ U(H) | a − I ∈ HS(H), p± a|H± ∈ I + S(H± )}. Define the Banach–Lie group U1 = {a ∈ U(H) | a − I ∈ S(H)} with Banach–Lie algebra u1 = u(H) ∩ S(H) and the Hilbert–Lie group U2 = {a ∈ U(H) | a − I ∈ HS(H)} with Hilbert–Lie algebra u2 = u(H) ∩ HS(H).


437

˜res = ures ⊕ R of ures relative to Finally, we introduce the central extension u the continuous two-cocycle s(A, B) := Trace(A[d, B]). ˜res is a Banach–Lie algebra endowed with the bracket Thus, u [(A, a), (B, b)]d = ([A, B], −s(A, B)) = ([A, B], − Trace(A[d, B])) . Proposition 5.14. The cohomology class [s] is a generator of the continuous Lie algebra cohomology space H 2 (ures , R). ˜res admits as predual the Banach space (˜ ures )∗ = The Banach–Lie algebra u (ures )∗ ⊕ R. The duality pairing · , ·d between (˜ ures )∗ = (ures )∗ ⊕ R and ˜res = ures ⊕ R is given by u (µ, γ), (A, a)d = µ, A + γa = Trace(µA) + γa for any A ∈ ures , µ ∈ (ures )∗ , and a, γ ∈ R. Theorem 5.15. The Banach space (˜ ures )∗ is a Banach–Lie–Poisson space for the Poisson bracket {f, g}d(µ, γ) := µ, [Dµ f (µ), Dµ g(µ)] − γs(Dµ f, Dµ g), where f, g ∈ C ∞ ((˜ ures )∗ ), (µ, γ) ∈ (˜ ures )∗ , and Dµ denotes the partial Fréchet derivative with respect to µ ∈ (ures )∗ . The Hamiltonian vector field defined by h ∈ C ∞ ((˜ ures )∗ ) is given by Xh (µ, γ) = − ad∗(Dµ h,Dγ h) (µ, γ) = − ad∗Dµ h µ − γ[Dµ h, d], 0 . For each γ ∈ R, (ures )∗ ⊕ {γ} is a Poisson submanifold of (˜ ures )∗ for the Poisson bracket {·, ·}d . More general central extensions of Banach–Lie–Poisson spaces were constructed in [OR04]. The previous central extension (˜ ures )∗ of the Banach–Lie– Poisson space (ures )∗ is a particular case of this more general theory. The considerations above show that (˜ ures )∗ is simultaneously a Banach–Lie algebra and a Banach–Lie–Poisson space. It is therefore reasonable to search res of the Banach–Lie group Ures whose Banach– for a central extension U ˜res and to conjecture, as in the finite-dimensional case, that Lie algebra is u res is a Poisson–Lie group. Part of the problem is to this Banach–Lie group U define in precise functional analytic terms the concept of a Banach Poisson–Lie group, where duals appearing in the finite-dimensional theory of Poisson–Lie groups are replaced by preduals. A second major challenge in this respect is to work with a well-defined Poisson tensor that is a section of the second wedge product of the tangent bundle and not of the second wedge product of the bidual tangent bundle.

438

Tudor S. Ratiu

Theorem 5.16. The restricted unitary group Ures acts on the Banach–Lie– Poisson space (ures )∗ ⊕ {γ} ⊂ (˜ ures )∗ by affine coadjoint action as follows. For g ∈ Ures , g · (µ, γ) := Ad∗ (g −1 )(µ) − γσ(g), γ , where µ ∈ (ures )∗ , γ ∈ R, and where σ : g ∈ Ures → gdg −1 − d ∈ (ures )∗ . The isotropy group of (0, γ) ∈ (ures )∗ ⊕ {γ} for the Ures -affine coadjoint action is a split Lie subgroup of Ures . The smooth affine coadjoint orbits of Ures are tangent to the characteristic distribution of the Banach Poisson manifold (˜ ures )∗ . For every γ = 0, the connected components of the Ures -affine coadjoint orbit O(0,γ) of (0, γ) ∈ (ures )∗ ⊕{γ} are strong symplectic leaves in the Banach– Lie–Poisson space (˜ ures )∗ . Next we turn to the study of the Banach–Lie–Poisson space (ures )∗ . As we have seen above, this is also a Banach–Lie algebra and hence, as before, we conjecture that Ures is a Banach Poisson-Lie group. From the point of view of the theory of Banach–Lie–Poisson spaces, note that (ures )∗ → ures is a continuous inclusion and that [(ures )∗ , ures ] ⊆ (ures )∗ , which implies that the predual (ures )∗ is left invariant by the coadjoint representation of the Banach–Lie algebra ures and hence, by Theorem 3.2, we conclude the following. • •

The predual Banach space (ures )∗ has a natural structure of Banach–Lie– Poisson space. If ρ ∈ (ures )∗ has the property that the isotropy subgroup Ures,ρ := {u ∈ Ures | uρu−1 = ρ} is a split Lie subgroup of Ures , then the connected components of the coadjoint orbit Oρ are integral manifolds of the characteristic distribution of (ures )∗ . Moreover, Oρ is a weak symplectic manifold when equipped with the orbit symplectic structure.

One expects, therefore, that the characteristic distribution of the Banach– Lie–Poisson space (ures )∗ is integrable. Therefore, achieving this reduces to the proof that the isotropy group Ures,ρ of any ρ ∈ (ures )∗ is a split Lie subgroup of Ures . This is in turn equivalent to showing that its Lie algebra is complemented in ures . To prove this equivalence one needs to appeal to the Harris–Kaup theorem [HK77], which is applicable in this case because the isotropy subgroup Ures,ρ is an algebraic subgroup of the invertible elements in the unital Banach algebra glres := Bres ⊂ B(H), whose elements can be alternatively described as operators of the form ρ++ ρ+− ρ= ∈ B(H) ρ−+ ρ−−


439

satisfying ρ±± := p± ρp± ∈ B(H± ), ρ+− := p+ ρp− ∈ HS(H− , H+ ), and ρ−+ := p− ρp+ ∈ HS(H+ , H− ). It turns out that showing that the Lie algebra of Ures,ρ is complemented in ures is not an easy task. Nevertheless, using the averaging method developed in [Ba90] and [BP07] it is possible to construct complements to the Lie algebra of Ures,ρ in ures in many important situations. Proposition 5.17. If ρ ∈ (ures )∗ and [d, ρ] = 0, then the coadjoint isotropy group of ρ is a split Lie subgroup of Ures and the connected components of the corresponding Ures -coadjoint orbit Oρ are smooth leaves of the characteristic distribution of (ures )∗ . So, the Ures -coadjoint orbit of every element ρ ∈ (ures )∗ which commutes with d is a smooth manifold; its connected components are therefore the weak symplectic leaves of the characteristic distribution of (ures )∗ . It follows that the same conclusion holds for every element ρ ∈ (ures )∗ which is Ures -conjugate to an element commuting with d, or equivalently to a diagonal operator with respect to a Hilbert basis compatible with the eigenspaces of d. The set of elements with the latter property is not equal to the whole (ures )∗ ; however, it is dense in (ures )∗ . It turns out that this property is the key modification in the proof of the previous theorem to yield the following result. Proposition 5.18. If ρ ∈ (ures )∗ is a finite-rank operator, then the coadjoint isotropy subgroup of ρ is a split Lie subgroup of Ures , and the connected components of the corresponding Ures -coadjoint orbit Oρ are smooth leaves of the characteristic distribution of (ures )∗ . Recall that every element in the Lie algebra u(n) of skew-Hermitian matrices, the Lie algebra of the unitary group U(n), is U(n)-conjugate to a diagonal matrix with respect to a given basis of Cn . Thus, U(n) acts transitively on the set of Cartan subalgebras of u(n). This is no longer true in the infinitedimensional case. Proposition 5.19. The restricted unitary group Ures does not act transitively on the set of maximal abelian subalgebras of either its Lie algebra ures or its predual (ures )∗ . Since every compact skew-Hermitian operator admits an orthonormal basis of eigenvectors, the set of conjugacy classes of maximal abelian subalgebras in (ures )∗ bijectively corresponds to U(H)/ Ures which is an infinite set. The conjugacy classes of maximal abelian subalgebras are, of course, related to the conjugacy classes of maximal tori. In view of the result above, it is a difficult question to decide whether a given operator ρ in (ures )∗ or ures has the property of being Ures -conjugate to a diagonal operator with respect to a basis adapted to the decomposition H = H+ ⊕ H− . The result below gives a criterion that ensures this property. Proposition 5.20. Assume that ρ ∈ (ures )∗ and that there exist an orthonormal basis {en }n≥1 and real numbers t ∈ (0, 1) and s ∈ (0, 3(1 − t)/100] such that the following conditions are satisfied.

440

Tudor S. Ratiu

(i) We have {en | n ≥ 1} ⊆ H+ ∪ H− . (ii) The matrix (ρmn )m,n≥1 of ρ with respect to the basis {en }n≥1 has the properties |ρm+1,n+1 | ≤ t|ρm,n | whenever m, n ≥ 1, and |ρm,n |2 ≤

s2 |ρmm ρnn | whenever m, n ≥ 1 and m = n. (mn)2

Then the coadjoint isotropy group of ρ is a split Lie subgroup of Ures and the corresponding Ures -coadjoint orbit Oρ is a smooth leaf of the characteristic distribution of (ures )∗ . The following fact is worth noting. For ρ ∈ B(H) each of the following two conditions is equivalent to the existence of a unitary operator u ∈ Ures such that [d, uρu−1 ] = 0. (i) There exists p ∈ B(H) such that p = p∗ = p2 , p − p+ ∈ HS(H), and ρp = pρ. (ii) There exists an element W ∈ Grres such that ρ(W) ⊆ W. In this statement, Grres is the restricted Grassmannian which will be defined and studied in §5.5. Even though it is not known if the characteristic distribution of (ures )∗ is integrable, the situation for the adjoint orbits of Ures in its Lie algebra ures is considerably better, as the following result shows. Proposition 5.21. There exists an open Ures -invariant neighborhood V of d ∈ ures such that V is a union of smooth adjoint orbits of the Banach–Lie group Ures . Now recall that u1,2 , the Lie algebra of the Banach–Lie group U1,2 , is the predual of ures (see Proposition 5.13), so one can pose the same question for the coadjoint orbits of the Banach–Lie group U1,2 . It turns out that a similar result also holds in this case as well as for (˜ ures )∗ . Proposition 5.22. For any γ ∈ R \ {0} there exists an open U1,2 -invariant neighborhood V of γd ∈ ures = u∗1,2 such that V is a union of smooth coadjoint orbits of the Banach–Lie group U1,2 . Proposition 5.23. For any γ ∈ R \ {0} there exists an open Ures -invariant neighborhood W of (0, γ) ∈ (˜ ures )∗ such that W is a union of smooth affine coadjoint orbits of the Banach–Lie group Ures . In view of these properties we can conjecture that Ures is a Poisson–Lie group. As we have seen, all these algebras of operators do not behave quite like their finite-dimensional analogues. The basic reason for this is that the Banach ∗-algebra Bres that contains them is not a C ∗ -algebra. In fact, there are several major differences between them.


441

Proposition 5.24. All of the unitary groups (1 + F ) ∩ U(H), U1,2 , and Ures are unbounded subsets of the unital associative Banach algebra Bres . It is well known that every self-adjoint normal functional in the predual of a W ∗ -algebra can be written as the difference of two positive normal functionals. A similar property holds for the preduals of numerous operator ideals. For example, if J and B are Banach operator ideals in B(H) such that BJ ⊂ S(H) and the trace pairing B × J → C,

(T, S) → Trace(T S)

induces a topological isomorphism of J onto the topological dual B∗ , then for every T = T ∗ ∈ B there exist T1 , T2 ∈ B such that T1 ≥ 0, T2 ≥ 0, and T = T1 − T2 . In fact, we can take T1 = (|T | + T )/2 and T2 = (|T | − T )/2, and we have T1 , T2 ∈ B since |T | ∈ B. (The latter property follows since if T = W |T | is the polar decomposition of T , then |T | = W ∗ T ∈ B.) It turns out that the predual (ures )∗ of the Lie algebra of Ures does not have the similar property of being spanned by its elements ρ with i ρ ≥ 0. In fact, the linear span of these elements is the proper subspace u1 of (ures )∗ . More precisely, we can state the following result. Proposition 5.25. The following assertions hold. (i) If a ∈ (ures )∗ and i a ≥ 0, then a ∈ S(H) and a1 ≤ a(ures )∗ ≤ √ (1 + 2)a1 . (ii) If ρ ∈ (ures )∗ \ u1 , then there exist no ρ1 , ρ2 ∈ (ures )∗ such that i ρ1 ≥ 0, i ρ2 ≥ 0, and ρ = ρ1 − ρ2 . 5.5 The restricted Grassmannian The restricted Grassmannian Grres is defined as the set of subspaces W of the Hilbert space H such that the orthogonal projection from W to H+ (respectively to H− ) is a Fredholm operator (respectively a Hilbert–Schmidt operator). This Hilbert manifold was studied extensively in [PS90] and [Wu01]. Since Grres = Ures / (U(H+ ) × U(H− )), it follows that Grres is a homogeneous space under the natural action of Ures . The connected components of Ures are the sets

U++ U+−

k ∈ Ures index(U++ ) = k for k ∈ Z. Ures = U−+ U−− The pairwise disjoint sets

Grkres = W ∈ Grres index(p+ |W : W → H+ ) = k ,

k∈Z

are the images of Ukres by the projection Ures → Grres , and thus they are the connected components of Grres . So, the connected component of Grres containing H+ is Gr0res .

442

Tudor S. Ratiu

The K¨ ahler form ωGr of Grres is the Ures -invariant 2-form whose value at H+ is given by ωGr (X, Y ) = 2 Im Trace(X ∗ Y ), where X, Y ∈ HS(H+ , H− ) TH+ (Grres ) and Im z denotes the imaginary part of z ∈ C. Equivalently, ωGr is the quotient of the real-valued antisymmetric bilinear form ΩGr on ures which vanishes on u(H+ ) ⊕ u(H− ) and is invariant under the U(H+ ) × U(H− )-action: ΩGr (A, B) = −2s(A, B) = −2 Trace(A[d, B]), where A, B ∈ ures . In this paper, an element A++ −A∗−+ A= ∈ ures A−+ A−− is identified with the vector X = A−+ ∈ HS(H+ , H− ) TH+ (Grres ). The Poisson geometric relationship between the restricted Grassmannian and the characteristic distribution of (˜ ures )∗ is given by the following result. Theorem 5.26. The connected components of the restricted Grassmannian are strong symplectic leaves in the Banach–Lie–Poisson space (˜ ures )∗ . More precisely, for every real γ = 0, the Ures -affine coadjoint orbit O(0,γ) of (0, γ) ∈ (ures )∗ ⊕{γ} is isomorphic to the restricted Grassmannian Grres via the smooth diffeomorphism Φγ : W ∈ Grres → 2 i γ(pW − p+ ) ∈ O(0,γ) , where pW denotes the orthogonal projection on W . The pull-back by Φγ of the orbit symplectic form on O(0,γ) equals (γ/2)-times the symplectic form ωGr on Grres . 5.6 Hilbert–Schmidt skew-Hermitian operators If instead of ures one considers the Hilbert–Lie algebra u2 = u(H) ∩ HS(H), whose underlying Hilbert–Lie group is U2 = {a ∈ U(H) | a − I ∈ HS(H)}, the results of the previous subsections can be significantly improved. Denote by ˜ u2 := u2 ⊕R the central extension of u2 defined by the restriction ˜2 implies of the two-cocycle s to u2 × u2 . The natural isomorphism (˜ u2 )∗ u ˜2 is a Banach–Lie–Poisson space, for the Poisson bracket given by that u {f, g}d(µ, γ) := µ, [Dµ f (µ), Dµ g(µ)] − γs(Dµ f, Dµ g), where f, g ∈ C ∞ (˜ u2 ), (µ, γ) ∈ ˜ u2 , and Dµ denotes the partial Fréchet derivative with respect to µ ∈ u2 . Our first result is the following.


443

Theorem 5.27. The characteristic distribution of the Hilbert–Lie–Poisson space (˜ u2 )∗ ˜ u2 (˜ u2 )∗ is integrable. The (co)adjoint orbits of U2 are the weak symplectic leaves of this distribution. The following has been shown in Theorem 3.5 of [Ca85] and Proposition V.7 in [Ne02a]. Proposition 5.28. The Hilbert–Lie group U2 acts transitively on the connected component Gr0res of the restricted Grassmannian Grres . The action of the split Lie subgroup U1,2 of U2 on Gr0res has been studied in Section 1.3.4 of [Tu05], in which the following result was established; see also [BRT07] for a geometric proof of this statement. Proposition 5.29. The split Lie subgroup U1,2 ⊂ U2 acts transitively on Gr0res . As was the case before, the Grassmannians are important special affine coadjoint orbits. Theorem 5.30. The connected component Gr0res of the restricted Grassman˜2 . More nian is a strong symplectic leaf in the Hilbert–Lie–Poisson space u ˜(0,γ) of (0, γ) ∈ u ˜2 precisely, for every γ = 0, the U2 -affine coadjoint orbit O is diffeomorphic to Gr0res via the map Φγ : W ∈ Gr0res → 2 i γ(pW − p+ ) ∈ O(0,γ) , where pW denotes the orthogonal projection on W . The pull-back by Φγ of the ˜(0,γ) equals (γ/2)-times the symplectic form ωGr orbit symplectic form on O 0 on Grres . The existence of invariant complex structures on the coadjoint orbits is given by the following result. Theorem 5.31. Every symplectic leaf of the Hilbert–Lie–Poisson space ˜u2 is transitively acted upon by U2 by means of the affine coadjoint action and is U2 -equivariantly covered by some complex homogeneous space of U2 . In particular, every simply connected symplectic leaf of the Banach–Lie– ˜2 has a U2 -invariant complex structure. For instance, this Poisson space u is the case for the connected component Gr0res of the restricted Grassmannian ˜2 . viewed as a symplectic leaf of u Remark. Even though not presented here, the theory of coadjoint orbits in L∗ -algebras is well developed and gives rise to a large class of examples with remarkable additional properties. For the development of this subject see [Ne04], [Tu05], [Tu06], and [Tu09] and references therein.

444

Tudor S. Ratiu

6 Geometric representation theory In view of the results of the previous section, one can hope for an infinitedimensional extension of the classical techniques of geometric representation theory. In order to obtain this, one could proceed as in the theory of geometric realizations of Lie group representations (see, e.g., [DF95], [Do97], [N200], [ES02]) by trying to build representation spaces as spaces of sections of certain vector bundles. The basic ingredient in this construction is the reproducing kernel Hilbert space (see, e.g., [Ar50], [Sc64], [Ha82], [Sai88], [Arv98], [BH98], [N200], [CG01], [CG06]). This method can be applied to the case of group representations obtained by restricting Gelfand–Na˘ımark–Segal (GNS) representations to unitary groups of C ∗ -algebras as shown in [BR07]. The goal of this subsection is to review these results. The idea is to build for these representations one-to-one intertwining operators from the representation spaces onto reproducing kernel Hilbert spaces of sections of certain Hermitian vector bundles. The construction of these Hermitian vector bundles is based on a choice of a C ∗ -subalgebra that is related in a very precise manner to the state involved in the GNS construction. In the case of normal states of W ∗ -algebras, this subalgebra is chosen to be the centralizer subalgebra of the state in question, and the base of the corresponding vector bundle is one of the symplectic leaves presented in the previous section. As we have seen, these symplectic leaves are unitary orbits of states. Therefore, the geometric representation theory reviewed here shows that the equivalence class of an irreducible GNS representation only depends on the unitary orbit of the corresponding pure state (see [GK60]). A lot of work has been done on the classification of unitary group representations of operator algebras (see, for example, [Boy80], [Boy88], [Boy93], and references therein). As will be discussed below, some of these representations, namely the ones obtained by restricting GNS representations to unitary groups, can be realized geometrically in the spirit of the classical Borel–Weil theorem for compact groups. 6.1 GNS unital ∗-representation Let a be a unital C ∗ -algebra, and ϕ : a → C a state, that is, a continuous positive linear functional with ϕ = 1. Define the Hilbert space H as the completion of the quotient of a by the null space of the non-negative Hermitian sesquilinear form a × a (a1 , a2 ) → ϕ(a∗2 a1 ) ∈ C. For each a ∈ a, the linear continuous map a ∈ a → aa ∈ a induces a linear continuous map ρ(a) : H → H. This defines the norm-continuous GNS unital ∗-representation ρ : a → B(H).


445

6.2 The fundamental construction Let a be a unital C ∗ -algebra, b a unital C ∗ -subalgebra of a, and ϕ : a → C a state such that there exists a conditional expectation E : a → b with ϕ◦E = ϕ. Let ρ : a → B(H) and ρϕ : b → B(Hϕ ) be the GNS unital ∗representations of a and b corresponding to ϕ and ϕ|b , respectively. Then Hϕ is a closed subspace of H, and for all b ∈ b the diagram ρ(b)

a → H −−−−→ ⏐ ⏐ ⏐PH ⏐ E ϕ

H ⏐ ⏐PH ϕ

ρϕ (b)

b → Hϕ −−−−→ Hϕ commutes, where PHϕ : H → Hϕ denotes the orthogonal projection, while a → H and b → Hϕ are the natural maps. The conditional expectation E : a → b is continuous with respect to the GNS scalar products on a and b: if a ∈ a then E(a)∗ E(a) ≤ E(a∗ a), so ϕ(E(a)∗ E(a)) ≤ ϕ(E(a∗ a)) = ϕ(a∗ a). Its “extension” by continuity is PHϕ . Restricting, we get a norm-continuous unitary representation ρ|Ub : Ub → U(Hϕ ) of the unitary group Ub of b. Since Ub is a Lie group with the topology inherited from the unitary group Ua of a, and the self-adjoint mapping E gives a continuous projection of the Lie algebra of Ua onto the Lie algebra of Ub , it follows that we have a principal Ub -bundle πb : u ∈ Ua → u · Ub ∈ Ua / Ub . Let Πϕ,b : [u, f ] ∈ Dϕ,b := Ua ×Ub Hϕ → u·Ub ∈ Ua / Ub be the associated vector bundle, i.e., (u1 , f1 ) ∼ (u2 , f2 ) if and only if there is some v ∈ Ub such that u1 = u2 v and f1 = ρϕ (v −1 )f2 . This associated bundle Πϕ,b : Dϕ,b = Ua ×Ub Hϕ → Ua / Ub has additional structure. (i) The real analytic Ua -actions on Ua / Ub and on Dϕ,b given by αb : (u , u · Ub ) ∈ Ua × Ua / Ub → u u · Ub ∈ Ua / Ub βϕ,b : (u , [u, f ]) ∈ Ua ×Dϕ,b → [u u, f ] ∈ Dϕ,b make the diagram Ua ×Dϕ,b ⏐ ⏐ idUa ×Πϕ,b

βϕ,b

−−−−→

Dϕ,b ⏐ ⏐Π ϕ,b

α

Ua × Ua / Ub −−−b−→ Ua / Ub commutative. (ii) The vector bundle Πϕ,b : Dϕ,b → Ua / Ub is homogeneous Hermitian. Indeed, there is a bundle Hermitian metric relative to which its fibers are complex Hilbert spaces and for each u ∈ Ua the mapping βϕ,b (u, ·) : Dϕ,b → Dϕ,b is (bounded linear and) fiber-wise unitary.

446

Tudor S. Ratiu

For example, if b = a and E = ida , then the homogeneous vector bundle associated with ϕ and b is just the vector bundle with the base reduced to a single point and with the fiber equal to H. The other extreme case is obtained when b = C1 and E(·) = ϕ(·)1. Let T = {z ∈ C | |z| = 1} be the unit circle in C. In this case, it follows that the homogeneous vector bundle associated with ϕ and b is a line bundle whose base is the Banach–Lie group Ua /T1. If the C ∗ -algebra a is a W ∗ -algebra, then there are natural choices that guarantee the hypotheses of the previous construction. Theorem 6.1. If m is a W ∗ -algebra and 0 ≤ ϕ ∈ m∗ , then there always exists a conditional expectation Eϕ : m → mϕ satisfying ϕ ◦ Eϕ = ϕ. 6.3 Reproducing kernels We need to work with vector bundles in different categories; here we establish terminology and notation. A vector bundle Π : D → T is called (i) continuous (smooth) if D and T are topological spaces (smooth manifolds), Π is continuous (smooth), the fibers Dt := Π −1 (t) for t ∈ T are complex Banach spaces, and Π is locally trivial with fiberwise bounded continuous (smooth) linear trivializations; (ii) Hermitian if the continuous vector bundle D has a continuous Hermitian structure (· | ·) : D ⊕ D → C that makes each Dt into a complex Hilbert space; (iii) holomorphic Hermitian if, in addition, D and T are complex Banach manifolds, Π is holomorphic, and the Hermitian structure is smooth. We denote by C(T, D), C ∞ (T, D), and O(T, D) the spaces of continuous, smooth, and holomorphic sections of the bundle Π : D → T . Let Π : D → T be a continuous Hermitian vector bundle, p1 , p2 : T ×T → T the projections on the first and second factors, and p∗j Π : p∗j D → T × T the pull-back of Π by pj , j = 1, 2, which are in turn Hermitian vector bundles. Also, consider the continuous vector bundle Hom(p∗2 Π, p∗1 Π) → T × T , whose fiber over (s, t) ∈ T ×T is the Banach space of bounded linear maps B(Dt , Ds ) between the corresponding fibers. Definition 6.2. A positive definite reproducing kernel on Π is a continuous section K ∈ C(T × T, Hom(p∗2 Π, p∗1 Π)) having the property that for every integer n ≥ 1 and all choices of t1 , . . . , tn ∈ T and ξ1 ∈ Dt1 , . . . , ξn ∈ Dtn , we have n j,l=1

(K(tl , tj )ξj | ξl )Dtl ≥ 0.


447

If Π is a holomorphic vector bundle, we say that the reproducing kernel K is holomorphic if for each t ∈ T and every ξ ∈ Dt we have K(·, t)ξ ∈ O(T, D). The definition immediately implies that (i) for all t ∈ T we have K(t, t) ≥ 0 in B(Dt ); (ii) for all t, s ∈ T we have K(t, s) ∈ B(Ds , Dt ), K(s, t) ∈ B(Dt , Ds ), and K(t, s)∗ = K(s, t). The following statement is Theorem 4.2 in [BR07]. Theorem 6.3. Let Π : D → T be a continuous Hermitian vector bundle and K ∈ C(T × T, Hom(p∗2 Π, p∗1 Π)). Then K is a reproducing kernel on Π if and only if there exists a linear mapping ι : H → C(T, D), where H is a complex Hilbert space, such that evιt := (ι(·))(t) : H → Dt is a bounded linear operator for all t ∈ T and K(s, t) = evιs (evιt )∗ : Dt → Ds , for all t ∈ T . If this is the case, then the mapping ι can be chosen to be injective. If Π is a holomorphic vector bundle, then the reproducing kernel K is holomorphic if and only if any mapping ι as above has the range contained in O(T, D). 6.4 Reproducing kernels and GNS-representations Let a be a unital C ∗ -algebra, b a unital C ∗ -subalgebra of a, and ϕ : a → C a state such that there exists a conditional expectation E : a → b satisfying ϕ◦E = ϕ. Let ρ : a → B(H) and ρϕ : b → B(Hϕ ) be the GNS representations associated to ϕ and ϕ|b . Let Πϕ,b : Dϕ,b = Ua ×Ub Hϕ → Ua / Ub be the homogeneous Hermitian vector bundle associated to ϕ and b. Define the realization operator ιϕ,b : H → C(Ua / Ub , Dϕ,b ) associated with ϕ and b by ιϕ,b (h)(u Ub ) = [u, PHϕ (ρ(u)−1 h)] for all h ∈ H and u ∈ Ua and the reproducing kernel Kϕ,b ∈ C (Ua / Ub ) × (Ua / Ub ), Hom(p∗2 (Πϕ,b ), p∗1 (Πϕ,b )) associated with ϕ and b by Kϕ,b (u1 Ub , u2 Ub )[u2 , f ] = [u1 , PHϕ (ρ(u−1 1 u2 )f )], where u1 , u2 ∈ Ua , f ∈ Hϕ , and p1 , p2 : (Ua / Ub ) × (Ua / Ub ) → Ua / Ub are the projections on the first and second factors, respectively.

448

Tudor S. Ratiu

Theorem 6.4. The realization operator ιϕ,b : H → C(Ua / Ub , Dϕ,b ) associated with ϕ and b is injective, has the property that (ιϕ,b (·))(s) : H → (Dϕ,b )s is bounded linear for all s ∈ Ua / Ub , and the corresponding reproducing kernel (given by the general theory) is precisely the reproducing kernel Kϕ,b associated with ϕ and b. Moreover, ιϕ,b intertwines the unitary representation ρ|Ua of Ua on H and the natural representation of Ua by linear mappings on C(Ua / Ub , Dϕ,b ). So the representation space has been realized as a space of sections of a homogeneous Hermitian vector bundle. The following result presents the important case when the homogeneous vector bundle occurring in the theorem is holomorphic and the reproducing kernel Hilbert space consists only of holomorphic sections. Corollary 6.5. Let m be a W ∗ -algebra with a faithful normal tracial state τ : m → C. Let a ∈ m have finite spectrum and satisfy 0 ≤ a, τ (a) = 1. Define ϕ : m → C by ϕ(b) = τ (ab). Then the homogeneous vector bundle Πϕ,mϕ : Dϕ,mϕ → Um / Umϕ associated with ϕ and mϕ is holomorphic and the reproducing kernel Kϕ,mϕ is holomorphic as well. Also, if ρ : m → B(H) is the GNS representation corresponding to the normal state ϕ, then the image of the realization operator ιϕ,mϕ : H → C(Um / Umϕ , Dϕ,mϕ ) consists of holomorphic sections. The techniques of the proof of this theorem (see [BR07]) are based on results in [Be02] and [Be05]. For completeness, let us also mention the following related statement, which in some sense provides a complexification of the picture drawn in Theorem 6.4. This is a special case of a theorem proved in [BG07] for Stinespring dilations of completely positive maps. However, in order to avoid introducing extra constructions, we are going to state it in the setting of Theorem 6.4, for the special case of GNS representations (which are just Stinespring dilations of C ∗ -algebra states). Theorem 6.6. In the setting of the beginning of the present subsection, let Ga be the group of invertible elements in the unital C ∗ -algebra a with its split Lie subgroup Gb of invertible elements in the C ∗ -subalgebra b, and define the inclusion mapping θ : Ua /Ub → Ga /Gb by uUb → uGb . If we define p = {X ∈ a | X ∗ = −X}, then the following assertions hold. (i) There exists a real analytic diffeomorphism Ga → Ua × p × G+ b,

a → (u(a), X(a), b(a)),

where G+ b denotes the set of positive elements in Gb , such that a = u(a) exp(iX(a))b(a) for all a ∈ a, which induces a polar decomposition in Ga /Gb given by aGb = u(a) exp(iX(a))Gb for every a ∈ Ga .


449

(ii) The mapping −∗ : Ga /Gb → Ga /Gb , u exp(iX)Gb → u exp(−iX)Gb , is an anti-holomorphic involutive diffeomorphism of Ga /Gb whose fixedpoint set is precisely the image of Ua /Ub under the aforementioned embedding θ. (iii) The projection Ga /Gb → Ua /Ub , u exp(iX)Gb → uUb , has the structure of a vector bundle isomorphic to the tangent bundle of the homogeneous space Ua /Ub , which is in turn isomorphic to the homogeneous vector bundle Ua ×Ub p → Ua /Ub . The corresponding isomorphism is given by u exp(iX)Gb → [(u, X)] for all u ∈ Ua , X ∈ p. (iv) Let HU (E, ϕ) ⊆ C(Ua /Ub , D) be the range of the realization operator ιϕ,b and H(E, ϕ) ⊆ O(Ga /Gb , Ga ×Gb Hϕ ) the range of the holomorphic realization operator γϕ,b : H → O(Ga /Gb , Ga ×Gb Hϕ ),

γϕ,b (h)(aGb ) = [(a, PHϕ (ρ(a)−1 h))].

Denote by µ(a) both operators on the section spaces C(Ua /Ub , D) and O(Ga /Gb , Ga ×Gb Hϕ ) defined by natural multiplication by a ∈ Ua and a ∈ Ga , respectively. Then HU (E, ϕ) and H(E, ϕ) are Hilbert spaces isometric to H by means of the realization operators ιϕ,b and γϕ,b , and the diagram ιϕ,b

HU (E, ϕ) ←−−−− ⏐ ⏐ µ(a) ιϕ,b

γϕ,b

H −−−−→ H(E, ϕ) ⏐ ⏐ ⏐ρ(a) ⏐µ(a) γϕ,b

HU (E, ϕ) ←−−−− H −−−−→ H(E, ϕ) is commutative for every a ∈ Ua . (v) If we denote by T1 µ the tangent map of µ(·)|H(E,ϕ) at the identity 1 ∈ Ga , then T1 µ is a ∗-representation of the C ∗ -algebra a which extends the group representation µ : Ga → B(H(E, ϕ)). Moreover, there exists a bundle section h0 ∈ H(E, ϕ) such that ϕ(a) = (T1 µ)(a)h0 , h0 for every a ∈ a, where ·, · stands for the scalar product in the Hilbert space H(E, ϕ) consisting of holomorphic sections in the vector bundle Ga ×Gb Hϕ → Ga /Gb . 6.5 Example Let us analyze Corollary 6.5 in a very simple finite-dimensional situation in order to get a feeling of what it really says. Let n ≥ 1, m = gl(n, C) (which is a W ∗ -algebra), and let Cn be the complex n-dimensional Hilbert space with the usual scalar product (· | ·). Next, denote ⎛ ⎞ ⎛ ⎞ 1 0 ... 0 1 ⎜0 0 . . . 0⎟ ⎜0⎟ ⎜ ⎟ ⎜ ⎟ and p = ⎜ . . . . ⎟ ∈ m. h = ⎜.⎟ . . . . . ⎝. . . .⎠ ⎝.⎠ 0

0 0 ... 0

450

Tudor S. Ratiu

Now define ϕ : m → C by ϕ(a) = (ah|h) = Trace(ap) = a11 for a = (aij )1≤i,j≤n ∈ m. It is clear that ϕ is a pure normal state of m. We begin by constructing the GNS representation of m with respect to ϕ. The kernel of the sesquilinear form m × m (a1 , a2 ) → ϕ(a∗2 a1 ) ∈ C is ⎧⎛ ⎞⎫ 0 ∗ ··· ∗ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬ ⎨⎜0 ∗ · · · ∗⎟⎪ ⎜ ⎟ ∗ m0 := {a ∈ m | ϕ(a a) = 0} = {a ∈ m | ah = 0} = ⎜ . . . . ⎟ ⎪ ⎝ . . . . ⎠⎪ ⎪ ⎪ ⎪ . . . . ⎪ ⎭ ⎩ 0 ∗ ··· ∗ and ⎛

ϕ(a∗ a) = |a11 |2 + |a21 |2 + · · · + |an1 |2

a11 ⎜ a21 ⎜ for a = ⎜ . ⎝ ..

0 ··· 0 ··· .. . . . .

⎞ 0 0⎟ ⎟ .. ⎟ . .⎠

an1 0 · · · 0 Hence, the completion H of m/m0 with respect to the scalar product induced by (a, b) → ϕ(b∗ a) is just the Hilbert space Cn with the usual scalar product, viewed as the set of column vectors, and the natural mapping m → H is ⎛ ⎞ ⎛ ⎞ a11 a11 a12 · · · a1n ⎜ a21 ⎟ ⎜ a21 a22 · · · a2n ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ .. .. . . .. ⎟ → ⎜ .. ⎟ . ⎝ . ⎠ ⎝ . . ⎠ . . an1 an1 an2 · · · ann Moreover, according to the way matrices multiply, it follows that for each a ∈ m the operator ρ(a) : H → H that makes the diagram b →ab

m −−−−→ ⏐ ⏐

m ⏐ ⏐

ρ(a)

H −−−−→ H commutative is just the natural action of a ∈ m = gl(n, C) on Cn . Thus, the GNS representation of m associated to ϕ is just the natural representation of m = gl(n, C) on Cn . We have ⎧⎛ ⎞⎫ ⎪ ⎪ ⎪ ∗ 0 ··· 0 ⎪ ⎪

⎬ ⎨⎜0 ∗ · · · ∗⎟⎪ z 0

⎟ ⎜ ϕ m = {p} = ⎜ . . . . ⎟ =

z ∈ C, W ∈ gl(n − 1, C) 0W ⎪ ⎝ .. .. . . .. ⎠⎪ ⎪ ⎪ ⎪ ⎪ ⎭ ⎩ 0 ∗ ··· ∗ C × gl(n − 1, C).


451

Hence, mϕ ∩ m0 = (1 − p)m(1 − p) gl(n − 1, C), and thus the space of the GNS representation ρϕ : mϕ → B(Hϕ ) of mϕ corresponding state to the z 0 ϕ is the ϕ|mϕ : m → C is one dimensional, i.e., Hϕ = C. Moreover, ρϕ 0W multiplication-by-z operator on Hϕ for all z ∈ C and W ∈ gl(n − 1, C). Now Um / Umϕ = U(n)/(U(1) × U(n − 1)) = Pn−1 (C), and it follows at once that the homogeneous vector bundle Πϕ,mϕ : Dϕ,mϕ → Um / Umϕ associated with ϕ is dual to the tautological line bundle over the complex projective space Pn−1 (C). Thus, in this special case, the theorem says that the natural representation of U(n) on Cn can be geometrically realized as a representation in the finitedimensional vector space of global holomorphic sections of the dual to the tautological line bundle over the U(n)-homogeneous compact K¨ ahler manifold Pn−1 (C), which is a special case of the Borel–Weil theorem.

References [AM78]

Abraham, R, Marsden, JE (1978) Foundations of Mechanics. Second edition, Addison-Wesley, Reading, MA [AMR] Abraham, R, Marsden, JE, Ratiu, TS (1978) Manifolds, Tensor Analysis, and Applications. Second edition. Volume 75 of Applied Mathematical Sciences. Springer-Verlag, New York [AS89] Andruchow, E, Stojanoff, D (1989) Differentiable structure of similarity orbits. J. Operator Theory 21(2): 349–366 [AS91] Andruchow, E, Stojanoff, D (1991) Geometry of unitary orbits. J. Operator Theory 26(1): 25–41 [AV99] Andruchow, E, Varela, A (1999) Weight centralizer expectations with finite index. Math. Scand. 84(2): 243–260 [AV02] Andruchow, E, Varela, A (2002) Homotopy of state orbits. J. Operator Theory, 48(2): 419–430 [Ar89] Arnold, VI (1978), Mathematical Methods of Classical Mechanics. volume 60 of Graduate Texts in Math. Springer-Verlag, First edition 1978, second edition, 1989 [AKN06] Arnold, VI, Kozlov, VV, Neishtadt AI (2006) Mathematical Aspects of Classical and Celestial Mechanics. Third edition. Volume 3 of Encyclopaedia of Mathematical Sciences. Springer-Verlag, Berlin [Ar50] Aronszajn, N (1950) Theory of reproducing kernels. Trans. Amer. Math. Soc. 68: 337–404 [Arv98] Arveson, W (1998) Subalgebras of C ∗ -algebras. III. Multivariable operator theory. Acta Math. 181(2): 159–228 [Ba90] Baskakov, AG (1990) Diagonalization of operators and complementability of subspaces of Banach spaces. Ukrain. Mat. Zh. 42(7): 867–873; translation in Ukrainian Math. J. 42(7), 763–768 (1991)

452 [Be02]

Tudor S. Ratiu

Beltit¸˘ a, D (2002) Spectra for solvable Lie algebras of bundle endomorphisms. Math. Ann. 324(2): 405–429 [Be05] Beltit¸˘ a, D (2005) Integrability of analytic almost complex structures on Banach manifolds. Ann. Global Anal. Geom. 28(1): 59–73 [Be06] Beltit¸˘ a, D (2006) Smooth Homogeneous Structures in Operator Theory. Monographs and Surveys in Pure and Applied Mathematics, vol 137. Chapman & Hall/CRC Press, Boca Raton-London-New York-Singapore [BG07] Beltit¸˘ a, D, Galé, JE (2007) Holomorphic geometric models for representations of C ∗ -algebras. Preprint arXiv:0707.0806v1 [math.OA]. [BP07] Beltit¸˘ a, D, Prunaru, B (2007) Amenability, completely bounded projections, dynamical systems and smooth orbits. Integral Equations Operator Theory 57(1): 1–17 [BR05] Beltit¸˘ a, D, Ratiu, TS (2005) Symplectic leaves in real Banach–Lie–Poisson spaces. Geom. Funct. Anal. 15(4): 753–779 [BR07] Beltit¸˘ a, D, Ratiu, TS (2007) Geometric representation theory for unitary groups of operator algebras. Advances in Mathematics 208: 299–317 [BRT07] Beltit¸˘ a, D, Ratiu, TS, Tumpach, AB (2007) The restricted Grassmannian, Banach–Lie–Poisson spaces, and coadjoint orbits. J. Funct. Anal. 247: 138–168 [BH98] Bertram, W, Hilgert, J (1998) Reproducing kernels on vector bundles. In: Lie Theory and Its Applications in Physics III : 43–58. World Scientific, Singapore [Bl03] Bloch, AM (2003) Nonholonomic Mechanics and Control. Volume 24 of Interdisciplinary Applied Mathematics. Springer-Verlag, New York. [Bon00] B´ ona, P (2000) Extended quantum mechanics. Acta Physica Slovaca 50 (1): 1–198 [Bon04] B´ ona, P (2004) Some considerations on topologies of infinite dimensional unitary coadjoint orbits. J. Geom. Phys. 51(2): 256–268 [Bo67] Bourbaki, N (1967) Variétés différentielles et analytiques. Fascicule de résultats. Paragraphes 1 a ` 7. Hermann, Paris [Bo71] Bourbaki, N (1971) Variétés différentielles et analytiques. Fascicule de résultats. Paragraphes 8 a ` 15. Hermann, Paris [Bo72] Bourbaki, N (1972) Groupes et Algèbres de Lie, Chapitres 2 et 3. Hermann, Paris [Boy80] Boyer, RP (1980) Representation theory of the Hilbert–Lie group U(H)2 . Duke Math. J. 47(2): 325–344 [Boy88] Boyer, RP (1988) Representation theory of U1 (H). Proc. Amer. Math. Soc. 103(1): 97–104 [Boy93] Boyer, RP (1993) Representation theory of infinite-dimensional unitary groups. In: Representation Theory of Groups and Algebras, Contemp. Math., vol 145: 381–391. Amer. Math. Soc., Providence, RI [BL05] Bullo, F, Lewis AD (2005) Geometric Control of Mechanical Systems. Volume 49 of Texts in Applied Mathematics. Springer-Verlag, New York [Ca85] Carey, AL (1985) Some homogeneous spaces and representations of the Hilbert Lie group U(H)2 . Rev. Roumaine Math. Pures Appl. 30(7): 505–520 [CM74] Chernoff, PR, Marsden, JE (1974) Properties of Infinite Dimensional Hamiltonian Systems. Volume 425 of Lecture Notes in Mathematics. Springer Verlag, New York

Coadjoint Orbits & the Beginnings of a Geometric Representation Theory [CJ04]

453

Chru´sci´ nski, D, Jamiolkowski, A (2004) Geometric Phases in Classical and Quantum Mechanics. Volume 36 of Progress in Mathematical Physics. Birkh¨ auser, Boston, MA [CG01] Constantinescu, T, Gheondea, A (2001) Representations of Hermitian kernels by means of Krein spaces. II. Invariant kernels. Comm. Math. Phys. 216(2): 409–430 [CG06] Constantinescu, T, Gheondea, A (2006) On L. Schwartz’s boundedness condition for kernels. Positivity, 10(1): 65–86 [Co02] Cortés Monforte, J (2002) Geometric Control and Numerical Aspects of Nonholonomic Systems. Volume 1793 of Lecture Notes in Mathematics. Springer-Verlag, New York [CMOI] Cortés Monforte, J, Mart´ınez, S, Ostrowski, JP, McIsaac, KA (2001) Optimal gaits for dynamic robotic locomotion. The International Journal of Robotics Research 20(9): 707–728 [CB97] Cushman, RH, Bates, LM (1997) Global Aspects of Classical Integrable Systems. Birkhh¨ auser, Boston, MA [Da07] Dabiri, JO (2007) Renewable fluid dynamic energy derived from aquatic animal locomotion. Bioinspiration and Biomimetics 2: L1–L3 [DCC] Dabiri, JO, Colin, SP, Costello, JH (2006) Fast-swimming hydromedusae exploit velar kinematics to form an optimal vortex wake. Journal of Experimental Biology 209(11): 2025–2033 [DG05] Dabiri, JO, Gharib, M (2005) The role of optimal vortex formation in biological fluid transport. Proc. Royal Soc. B: Biological Sciences 272: 1557–1560 [DF95] Davidson, MG, Fabec, RC (1995) Geometric realizations for highest weight representations. In: Representation Theory and Harmonic Analysis (Cincinnati, OH, 1994), Contemp. Math., vol 191: 13–31. Amer. Math. Soc., Providence, RI [DFWW] Dykema, K, Figiel, T, Weiss, G, Wodzicki, M (2004) Commutator structure of operator ideals. Adv. in Math. 185(1): 1–79. [Do97] Donley, RW (1997) The Borel–Weil Theorem for U(n). In: Representation Theory and Automorphic Forms (Edinburgh, 1996). Proc. Sympos. Pure Math. vol 61: 115–121. Amer. Math. Soc., Providence, RI [ES02] Eastwood, M, Sawon, J (2002) The Borel–Weil Theorem for complex projective space. In: Invitations to Geometry and Topology, Oxf. Grad. Texts Math., vol 7:126–145. Oxford Univ. Press, Oxford [Ga83] Gallavotti, G (1983) The Elements of Mechanics. Texts and Monographs in Physics. Springer-Verlag, New York [GRKSD] Gharib, M, Rambod, E, Kheradvar, A, Sahn, DJ, Dabiri, JO (2006) Optimal vortex formation as an index of cardiac health. Proceedings of the National Academy of Sciences of the USA 103(16): 6305–6308 [GK60] Glimm, JG, Kadison, RV (1960) Unitary operators in C ∗ -algebras. Pacific J. Math. 10: 547–556 [GK69] Gohberg, IC, Kre˘ın, MG (1969) Introduction to the Theory of Linear Nonselfadjoint Operators. Translations of Mathematical Monographs, vol. 18, American Mathematical Society, Providence, RI [GS90] Guillemin, V, Sternberg. S (1990) Symplectic Techniques in Physics. Second edition. Cambridge University Press, Cambridge, UK

454 [HLW]

Tudor S. Ratiu

Hairer, E, Lubich, C, Wanner, G (2006) Geometric Numerical Integration. Structure-preserving Algorithms for Ordinary Differential Equations. Second edition. Volume 31 of Springer Series in Computational Mathematics. Springer-Verlag, Berlin [Ha82] Halmos, PR (1982) A Hilbert Space Problem Book. Second edition. Volume 19 of Graduate Texts in Mathematics. Encyclopedia of Mathematics and its Applications 17, Springer-Verlag, New York, Berlin [Ha72] de la Harpe, P (1972) Classical Banach–Lie Algebras and Banach–Lie Groups of Operators in Hilbert Space. Volume 285 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, New York [HK77] Harris, LA, Kaup, W (1977) Linear algebraic groups in infinite dimensions. Illinois J. Math. 21(3): 666–674. [HZ94] Hofer, H, Zehnder, E (1994) Symplectic Invariants and Hamiltonian Dynamics. Birkh¨ auser Advanced Texts: Basler Lehrb¨ ucher. Birkh¨ auser Verlag, Basel [Ho75] Hoffmann, KH (1975) Analytic groups without analysis. In Symposia Mathematica 16 (Convegno sui Gruppi Topologici i Lie, INDAM, Toma, 1974):357–374. Academic Press, London [JS97] Jones, V, Sunder, VS (1997) Introduction to Subfactors. Volume 234 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, UK [JSa98] José, JV, Saletan, EJ (1998) Classical Dynamics. A Contemporary Approach. Cambridge University Press, Cambridge, UK [Ju97] Jurdjevic, V (1997) Geometric Control Theory. Volume 51 in Cambridge Studies in Advanced Mathematics, Cambridge University Press, Cambridge, UK [KR97] Kadison, RV, Ringrose, JR (1997) Fundamentals of the Theory of Operator Algebras. Vol. II. Advanced Theory. Corrected reprint of the 1986 original. Volume 16 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI [KP03] Kappeler, T, P¨ oschel, J (2003) KdV & KAM. Volume 45 of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics. Springer-Verlag, Berlin [Ku00] Kuksin, SB (2000) Analysis of Hamiltonian PDEs. Volume 19 of Oxford Lecture Series in Mathematics and Its Applications. Oxford University Press, Oxford [LM87] Libermann, P, Marle, CM (1987) Symplectic Geometry and Analytical Mechanics. Kluwer Academic Publishers, Dordrecbt [LiR97] Littlejohn, RG, Reinsch, M (1997) Gauge fields in the separation of rotations and internal motions in the n-body problem. Rev. Mod. Phys. 69(1): 213–275 [LiM02] Littlejohn, RG, Mitchell, KA (2002) Gauge theory of small vibrations in polyatomic molecules. In Geometry, Mechanics, and Dynamics: 407–428. Springer, New York, 2002 [Ma92] Marsden, JE (1992) Lectures on Mechanics. Volume 174 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, UK [MMOPR] Marsden, JE, Misiolek, G, Ortega, JP, Perlmutter, M, Ratiu, TS (2007) Hamiltonian Reduction by Stages. Volume 1913 of Springer Lecture Notes in Mathematics. Springer-Verlag, New York

Coadjoint Orbits & the Beginnings of a Geometric Representation Theory [MMR]

455

Marsden, JE, Montgomery, R, Ratiu, TS (1990) Reduction, Symmetry and Phases in Mechanics. Volume 436 of Memoirs of the AMS, vol. 436. Amer. Math. Soc., Providence, RI [MR86] Marsden, JE, Ratiu, TS (1986) Reduction of Poisson manifolds. Lett. in Math. Phys. 11: 161–170. [MR94] Marsden, JE, Ratiu, TS (1994) Introduction to Mechanics and Symmetry. Second edition, second printing 2003. Volume 17 of Texts in Applied Mathematics. Springer-Verlag, New York [McDS] McDuff, D, Salamon, D (1998) Introduction to Symplectic Topology. Second edition. Oxford Mathematical Monographs. The Clarendon Press, Oxford University Press, New York [McIO] McIsaac, KA, Ostrowski, JP (2003) A framework for steering dynamic robotic locomotion systems. The International Journal of Robotics Research 22(2): 83–97 [MoR05] Montaldi, J, Ratiu, TS, editors (2005) Geometric Mechanics and Symmetry. The Peyresq Lectures. Volume 306 of London Math. Soc. Lecture Note Series. Cambridge University Press, Cambridge, UK [N200] Neeb, KH (2000) Holomorphy and Convexity in Lie Theory. Volume 28 in de Gruyter Expositions in Mathematics. Walter de Gruyter & Co., Berlin [Ne02] Neeb, KH (2002) Highest weight representations and infinite-dimensional K¨ ahler manifolds. In: Recent Advances in Lie Theory (Vigo, 2000): 367–392. Volume 25 of Res. Exp. Math. Heldermann, Lemgo, 2002 [Ne02a] Neeb, KH (2002a) Classical Hilbert–Lie groups, their extensions and their homotopy groups. In: Geometry and Analysis on Finite- and Infinitedimensional Lie Groups (B¸edlewo, 2000): 87–151. Volume 55 of Banach Center Publications. Polish Acad. Sci., Warsaw [Ne02b] Neeb, KH (2002b) Universal central extensions of Lie groups. Acta Appl. Math. 73(1-2): 175–219 [Ne02c] Neeb, KH (2002c) A Cartan-Hadamard theorem for Banach-Finsler manifolds. Geom. Dedicata 95: 115–156 [Ne04] Neeb, KH (2004) Infinite-dimensional groups and their representations. In: Lie Theory: 213–328. Volume 228 of Progress in Mathematics. Birkh¨ auser, Boston, MA [Ne06] Neeb, KH (2006) Towards a Lie theory of infinite-dimensional groups. Jap. J. Math. 3rd Series 1:2: 291–468. [Nu92] N¨ ubel, F (1992) On integral manifolds for vector space distributions. Math. Ann. 294(1): 1–17. [Od88] Odzijewicz, A (1988) On reproducing kernels and quantization of states. Comm. Math. Phys. 114: 577–597 [Od92] Odzijewicz, A (1992) Coherent states and geometric quantization. Comm. Math. Phys. 150: 385–413 [OR03] Odzijewicz, A, Ratiu, TS (2003) Banach Lie–Poisson spaces and reduction. Comm. Math. Phys. 243: 1–54 [OR04] Odzijewicz, A, Ratiu, TS (2004) Extensions of Banach–Lie–Poisson spaces. J. Funct. Anal. 217(1): 103–125 [OS97] Odzijewicz, A, Swietochowski, M (1997) Coherent states map for MICKepler system. J. Math. Phys. 38: 5010–5030 [O02] Oliva, WM (2002) Geometric Mechanics. Volume 1798 of Lecture Notes in Mathematics. Springer-Verlag, Berlin

456 [Ol86] [OrR04] [OB98] [ODK] [PD07] [PS90] [Sai88] [Sa71] [Sc64] [ST07] [SDM] [SCRa] [SCRb]

[Si01] [Si05] [St74a] [St74b] [Str81] [Su03]

Tudor S. Ratiu Olver, PJ (1986) Applications of Lie Groups to Differential Equations. Volume 107 of Graduate Texts in Mathematics. Springer-Verlag, Berlin Ortega, JP, Ratiu, TS (2004) Momentum Maps and Hamiltonian Reduction. Volume 222 of Progress in Mathematics. Birkh¨ auser, Boston, MA Ostrowski, J, Burdick, J (1998) The geometric mechanics of undulatory robotic locomotion. The International Journal of Robotics Research 17(7): 683–701 Ostrowski, JP, Desai, JP, Kumar, V (2000) Optimal gait selection for nonholonomic locomotion systems. The International Journal of Robotics Research 19(3): 225–237 Peng J, Dabiri, JO (2007) A potential-flow, deformable-body model for fluid-structure interactions with compact vorticity: application to animal swimming measurements. Experiments in Fluids 43(5): 655–664 Pressley, A, Segal, G (1990) Loop Groups. Oxford Mathematical Monographs. Oxford Science Publications. The Clarendon Press, Oxford University Press, Oxford Saitoh, S (1988) Theory of Reproducing Kernels and Its Applications. Volume 189 of Pitman Research Notes in Mathematics Series. Longman Scientific & Technical, Harlow Sakai, S (1971) C ∗ -Algebras and W ∗ -Algebras. 1998 reprint of the 1971 edition. Volume 60 of Ergebnisse der Mathematik und ihrer Grenzgebiete. Springer-Verlag, New York Schwartz, L (1964) Sous-espaces hilbertiens d’espaces vectoriels topologiques et noyaux associés (noyaux reproduisants). J. Analyse Math. 13: 115–256 Sfakiotakis, M, Tsakiris, DP (2007) Biomimetic centering for undulatory robots. The International Journal of Robotics Research 26(11-12): 1267–1282 Shadden, SC, Dabiri, JO, Marsden, JE (2006) Lagrangian analysis of entrained and detrained fluid in vortex rings. Physics of Fluids 18, 1–11 Shammas, EA, Choset, H, Rizzi, AA (2007) Geometric motion planning analysis for two classes of underactuated mechanical systems. The International Journal of Robotics Research 16(10): 1043–1073 Shammas, EA, Choset, H, Rizzi, AA (2007) Towards a unified approach to motion planning for dynamic underactuated mechanical systems with non-holonomic constraints. The International Journal of Robotics Research 26(10): 1075–1124 Singer, SF (2001) Symmetry in Mechanics. A Gentle, Modern Introduction. Birkh¨ auser, Boston, MA Singer, SF (2005) Linearity, Symmetry, and Prediction in the Hydrogen Atom. Undergraduate Texts in Mathematics. Springer-Verlag, New York Stefan, P (1974a) Accessible sets, orbits and foliations with singularities. Proc. London Math. Soc. 29: 699–713 Stefan, P (1974b) Accessibility and foliations with singularities. Bull. Amer. Math. Soc. 80: 1142–1145 Str˘ atil˘ a, S ¸ (1981) Modular Theory in Operator Algebras. Editura Acad., Bucure¸sti; Abacus Press, Tunbridge Wells, 1981 Suris, Y (2003) The Problem of Integrable Discretization: Hamiltonian Approach. Volume 219 in Progress in Mathematics. Birkh¨ auser-Verlag, Basel

Coadjoint Orbits & the Beginnings of a Geometric Representation Theory [Su73]

457

Sussmann, HJ (1973) Orbits of families of vector fields and integrability of distributions. Trans. Amer. Math. Soc. 180: 171–188 [T00] Talman, R (2000) Geometric Mechanics. Wiley, New York [Ta72] Takesaki, M (1972) Conditional expectations in von Neumann algebras. J. Functional Analysis 9: 306–321. [Ta79] Takesaki, M (1979) Theory of Operator Algebras I. Reprint of the first (1979) edition. Volume 124 in Encyclopaedia of Mathematical Sciences. Volume 5 in Operator Algebras and Non-commutative Geometry. SpringerVerlag, Berlin, 2002 [To57] Tomiyama, J (1957) On the projection of norm one in W ∗ -algebras. Proc. Japan Acad. 33: 608–612 [Tu05] Tumpach, AB (2005) Variétés K¨ ahlériennes et Hyperk¨ ahlériennes de ´ Dimension Infinie. Ph.D. Thesis, Ecole Polytechnique, Paris [Tu06] Tumpach, AB (2006) Mostow Decomposition Theorem for a L∗ -group and applications to affine coadjoint orbits and stable manifolds. Preprint math-ph/0605039 (May 2006) [Tu09] Tumpach, AB (2009) On the classification of infinite-dimensional irreducible Hermitian-symmetric affine coadjoint orbits. Forum Mathematicum, 21(3), 375–393. [Va94] Vaisman, I (1994) Lectures on the Geometry of Poisson Manifolds. Volume 118 in Progress in Mathematics. Birkh¨ auser, Boston, MA [We83] Weinstein, A (1983) The local structure of Poisson manifolds. Journ. Diff. Geom 18: 523–557 [We98] Weinstein, A (1998) Poisson geometry. Differential Geom. Appl. 9: 213–238 [WS88] Wilczek, F, Shapere, A (1988) Geometric Phases in Physics. World Scientific Publishing Co. Inc., Singapore [Wu01] Wurzbacher, T (2001) Fermionic second quantization and the geometry of the restricted Grassmannian. In: Huckleberry A, Wurzbacher, T (eds) Infinite Dimensional K¨ ahler Manifolds. Volume 31 in DMV Seminar. Birkh¨ auser, Basel [YKMK] Yanao, T, Koon, WS, Marsden, JE, Kevrekidis, IG (2007) Gyrationradius dynamics in structural transitions of atomic clusters. J. Chem. Physics 126: 1–17

Infinite-Dimensional Multiplicity-Free Spaces I: Limits of Compact Commutative Spaces Joseph A. Wolf Department of Mathematics, University of California, Berkeley, CA 94720–3840, USA, [email protected] Summary. We study direct limits (G, K) = lim (Gn , Kn ) of compact Gelfand pairs. −→ First, we develop a criterion for a direct limit representation to be a multiplicityfree discrete direct sum of irreducible representations. Then we look at direct limits G/K = lim Gn /Kn of compact riemannian symmetric spaces, where we combine −→ our criterion with the Cartan–Helgason theorem to show in general that the regular representation of G = lim Gn on a certain function space lim L2 (Gn /Kn ) is −→ −→ multiplicity-free. That method is not applicable for direct limits of nonsymmetric Gelfand pairs, so we introduce two other methods. The first, based on “parabolic direct limits” and “defining representations”, extends the method used in the symmetric space case. The second uses some (new) branching rules from finite-dimensional representation theory. In both cases we define function spaces A(G/K), C(G/K), and L2 (G/K) to which our multiplicity-free criterion applies. Key words: Lie group, Gelfand pair, commutative space, direct limit representation, multiplicity-free representation. 2000 Mathematics Subject Classifications: 20G05, 22E45, 22E65, 43A85, 43A90.

1 Introduction Gelfand pairs (G, K), and the corresponding “commutative” homogeneous spaces G/K, form a natural extension of the class of riemannian symmetric spaces. We recall some of their basic properties. Let G be a locally compact topological group, K a compact subgroup, and M = G/K. Then the following conditions are equivalent; see [W2007, Theorem 9.8.1]. 1. (G, K) is a Gelfand pair, i.e., L1 (K\G/K) is commutative under convolution. 2. If g, g ∈ G then µKgK ∗ µKg K = µKg K ∗ µKgK (convolution of Dirac measures on K\G/K). 3. Cc (K\G/K) is commutative under convolution.


459

460

Joseph A. Wolf

4. The measure algebra M(K\G/K) is commutative. 5. The representation of G on L2 (M ) is multiplicity-free. If G is a connected Lie group one can also add 6. The algebra of G-invariant differential operators on M is commutative. When we drop the requirement that K be compact, conditions 1, 2, 3, and 4 lose their meaning because integration on M or K\G/K no longer corresponds to integration on G. Condition 5 still makes sense as long as K is unimodular in G. Condition 6 remains meaningful (and useful) whenever G is a connected Lie group; there one speaks of “generalized Gelfand pairs”. In this paper we look at some cases where G and K are not locally compact, in fact are infinite dimensional, and show in those cases that the multiplicityfree condition 5 is satisfied. We first discuss a multiplicity-free criterion that can be viewed as a variation on some of the combinatoric considerations of [DPW2002]; it emerged from some discussions with Ivan Penkov in another context. We then apply the criterion in the setting of symmetric spaces, proving that direct limits of compact symmetric spaces are multiplicity-free. This applies in particular to infinite-dimensional real, complex, and quaternionic Grassmann manifolds, and it uses some basic symmetric space structure theory. In particular, our argument for direct limits of compact riemannian symmetric spaces makes essential use of the Cartan–Helgason theorem, and thus does not extend to direct limits of nonsymmetric Gelfand pairs. In order to extend the multiplicity-free result to at least some direct limits of nonsymmetric Gelfand pairs, we define the notion of “defining representation” for a direct system {(Gn , Kn )}, where the Gn are compact Lie groups and the Kn are closed subgroups. We show how a defining representation for {(Gn , Kn )} leads to a direct system {A(Gn /Kn )} of C-valued polynomial function algebras, a continuous function completion {C(Gn /Kn )}, and a Lebesgue space completion {L2 (Gn /Kn )}. The direct limit spaces A(G/K), C(G/K), and L2 (G/K) are the function spaces on G/K = lim Gn /Kn which −→ we study as G-modules. Next, we prove the multiplicity-free property, for the action of G on A(G/K), C(G/K), and L2 (G/K), when {(Gn , Kn )} is one of several families of Gelfand pairs related to spheres and Grassmann manifolds. We prove the multiplicity-free property for three other types of direct limits of Gelfand pairs. Finally we summarize the results, extending them slightly by including the possibility of enlarging the Kn within their Gn -normalizers without losing the property that {Kn } is a direct system. Our proofs of the multiplicity-free condition, for some direct limits of nonsymmetric Gelfand pairs, use a number of branching rules, new and old, for finite-dimensional representations. This lends a certain ad hoc flavor which I hope can be avoided in the future. Direct limits (G, K) = lim(Gn , Kn ) of riemannian symmetric spaces were −→ studied by Ol’shanskii from a very different viewpoint [Ol1990]. He viewed the

Infinite-Dimensional Multiplicity-Free Spaces, I

461

Gn inside dual reductive pairs and examined their action on Hilbert spaces of Hermite polynomials. Ol’shanskii made extensive use of factor representation theory and Gaussian measure, obtaining analytic results on limit-spherical functions. See Faraut [Fa2006] for a discussion of spherical functions in the setting of direct limit pairs. In contrast to the work of Ol’shanskii and Faraut, we use the rather simple algebraic method of renormalizing formal degrees of representations to obtain isometric embeddings L2 (Gn /Kn ) → L2 (Gn+1 /Kn+1 ). That leads directly to our multiplicity-free results. Acknowledgements: I am indebted to Ivan Penkov for discussions of multiplicities in direct limit representations, which are formalized in Theorem 2 below. I also wish to acknowledge hospitality from the Mathematisches Forschungsinstitut Oberwolfach and support from NSF Grant DMS 04 00420.

2 Direct limit groups and representations We consider direct limit groups G = lim Gn and, their direct limit represen−→ tations π = lim πn . This means that πn is a representation of Gn on a vector −→ space Vn , that the Vn form a direct system, and that π is the representation of G on V = lim Vn given by π(g)v = πn (gn )vn whenever n is sufficiently large −→ that Vn → V and Gn → G send vn to v and gn to g. The formal definition amounts to saying that π is well defined. It is clear that a direct limit of irreducible representations is irreducible, but there are irreducible representations of direct limit groups that cannot be formulated as direct limits of irreducible finite-dimensional representations. This is a combinatoric matter and is discussed extensively in [DPW2002]. The following definition is closely related to those combinatorics but applies to a somewhat simpler situation. Definition 1 We say that a representation π of G is limit aligned if it has form lim πn in such a way that (i) each πn is a direct sum of primary representations, −→ and (ii) the corresponding representation spaces V = lim Vn have the property −→ that every primary subspace of Vn is contained in a primary subspace of Vn+1 . Theorem 2 A limit-aligned representation π = lim πn of G = lim Gn is a −→ −→ direct sum of primary representations. If the πn are multiplicity free, then π is a multiplicity-free direct sum of irreducible representations. Proof. Let V = lim Vn be the representation spaces. Decompose Vn = −→ α∈In Vn,α , where the Vn,α are the subspaces for the primary summands of πn . Write πn,α for the representation of Gn on Vn,α , so πn = α∈In πn,α . Since π is limit aligned, i.e., since each Vn,α ⊂ Vn+1,β for some β ∈ In+1 , we may assume In ⊂ In+1 in such a way that each V n,α ⊂ Vn+1,α for every α ∈ In . Now V = α∈I Vα , discrete sum, where I = In and Vα = Vn,α . The sum is direct, for if u1 + u2 + · · · + ur = 0 where ui ∈ Vαi for distinct

462

Joseph A. Wolf

indices α1 , . . . , αr , then we take n sufficiently large so that each ui ∈ Vn,αi and conclude that u1 = u2 = · · · = ur = 0. Thus π is the discrete direct sum of the representations πα = lim πn,α of G on Vα . −→ Let Cα = {X : Vα → Vα linear | Xπα (g) = πα (g)X for all g ∈ G}, the commuting algebra of πα . If πα fails to be primary, then Cα contains nontrivial commuting ideals Cα and Cα . Then for n large, the stabilizer NCα (Vn,α ) of Vn,α in Cα contains nontrivial commuting ideals NCα (Vn,α ) and NCα (Vn,α ). That is impossible because πn,α is primary. We have proved that π is the discrete direct sum of primary representations πα . If the πn are multiplicity free, then the πn,α are irreducible and it is immediate that the πα = lim πn,α are irreducible. This completes the proof of −→ Theorem 2. A direct limit of irreducible representations is irreducible, but it is not immediate that every irreducible direct limit representation can be rewritten as a direct limit of irreducible representations. With this and Theorem 2 in mind, we extend Definition 1 as follows. Definition 3 A representation π of G = lim Gn is lim irreducible if it has form −→ π = lim πn where each πn is an irreducible representation of Gn . Similarly, π is −→ lim primary if it has form π = lim πn where each πn is a primary representation −→ of Gn . Theorem 4 Consider a representation π = lim πn of G = lim Gn with −→ −→ representation space V = lim Vn . Suppose that each πn is a multiplicity-free −→ direct sum of irreducible highest weight representations. Suppose for n > > 0 that the direct system map Vn−1 → Vn sends Gn−1 -highest weight vectors to Gn -highest weight vectors. Then π is a multiplicity-free direct sum of lim-irreducible representations of G. Proof. By hypothesis each πn is a direct sum of primary representations which, in fact, are irreducible highest weight representations. We recursively choose highest weight vectors so that πn−1 = πλ,n−1 , where πλ,n−1 has highest weight vector vλ,n−1 ∈ Vn−1 that maps to a highest weight vector vλ,n ∈ Vn of an irreducible constituent πλ,n of πn . This exhibits π as a limit-aligned direct sum because it embeds the summand Vλ,n−1 of Vn−1 into the irreducible summand of Vn that contains vλ,n . Now Theorem 2 shows that π is a multiplicity-free direct sum of lim-irreducible representations of G.

3 Limit theorem for symmetric spaces We now apply Theorems 2 and 4 to direct limits of compact riemannian symmetric spaces. Fix a direct system of compact connected Lie groups Gn and subgroups Kn such that each (Gn , Kn ) is an irreducible riemannian symmetric pair. Suppose that the corresponding compact symmetric spaces


463

Mn = Gn /Kn are connected and simply connected. Up to renumbering and passage to a common cofinal subsequence, the only possibilities are as given in the following table.

1 2 3 4 5 6 7 8 9 10 11

compact irreducible riemannian symmetric Mn = Gn /Kn Gn Kn RankMn DimMn SU (n) × SU (n) diagonal SU (n) n−1 n2 − 1 Spin(2n + 1) × Spin(2n + 1) diagonal Spin(2n + 1) n 2n2 + n Spin(2n) × Spin(2n) diagonal Spin(2n) n 2n2 − n Sp(n) × Sp(n) diagonal Sp(n) n 2n2 + n (5) SU (p + q), p = pn , q = qn S(U (p) × U (q)) min(p, q) 2pq (n−1)(n+2) SU (n) SO(n) n−1 2 SU (2n) Sp(n) n − 1 2n2 − n − 1 SO(p + q), p = pn , q = qn SO(p) × SO(q) min(p, q) pq SO(2n) U (n) [ n2 ] n(n − 1) Sp(p + q), p = pn , q = qn Sp(p) × Sp(q) min(p, q) 4pq Sp(n) U (n) n n(n + 1)

Fix one of the direct systems {(Gn , Kn )} of Table 5. Then we have involutive automorphisms θn of Gn such that the Lie algebras decompose into ±1 eigenspaces of the θn , gn = kn + sn in such a way that kn = gn ∩ kn+1 and sn = gn ∩ sn+1 . Then we recursively construct a system of maximal abelian subspaces an : maximal abelian subspace of sn such that an = gn ∩ an+1 . The restricted root systems Σn = Σn (gn , an ) : the system of an -roots on gn form an inverse system of linear functionals: Σ = Σ(g, a) is the system lim Σn ←− of linear functionals on a = lim an . In this inverse system, the multiplicities of −→ the restricted roots will increase without bound, but we can make consistent choices of positive subsystems Σn+ = Σn+ (gn , an ) : system of positive an -roots on gn + so that Σn+ ⊂ Σm |an for m n n0 . Consider the reduced root system

Σ0,n = {α ∈ Σn | 2α ∈ Σn } + and its positive subsystem Σ0,n := Σ0,n ∩ Σn+ . Examining the tables of Araki ([Ar1962], or referring to [He1978, pp. 532–534] or [W1980, pp. 90–93]), we see the following.

464

Joseph A. Wolf

Lemma 6 Suppose that Gn is simple. Then there are only two possibilities. (a) Σ0,n = Σn ; in other words, if α ∈ Σn then 2α ∈ / Σn . (b) Σ0,n = Σn ; there is exactly one simple root ψ1,n for Σn+ such that 2ψ1,n ∈ Σn , and ψ1,n is at the end of the Dynkin diagram of Σn+ opposite to + the end where roots are added to obtain the diagram of Σn+1 . + , which we denote Then the corresponding simple root systems for Σ0,n

Ψn = Ψn (gn , an ) = {ψ1,n , . . . , ψrn ,n } : simple reduced an -roots on gn satisfy Ψn ⊂ Ψm |an for m n n0 as well. Here rn = dim an , rank of Mn . In case (a) of Lemma 6, Ψn is a simple root system for Σn+ , but in case (b) the corresponding simple root system for Σn+ is { 12 ψ1,n , ψ2,n , . . . , ψrn ,n }. In both cases Ψn ⊂ Ψm |an for m n n0 . More precisely, if ψj,n ∈ Ψn and m n, then there is just one element ψ ∈ Ψm with ψ|an = ψj,n . In other words, we may (and do) recursively enumerate the simple root systems Ψn so that if m n and ψj,n ∈ Ψn , then ψj,m ∈ Ψm satisfies ψj,m |an = ψj,n , retaining the convention that in case (b) of Lemma 6 the Later we will use the fact that

1 2 ψ1,n

are roots.

+ in case (b) of Lemma 6, if m n and 12 ψ1,n ∈ Σn+ , then 12 ψ1,m ∈ Σm . (7)

Recursively define θn -stable Cartan subalgebras of hn = tn + an of gn with hn = gn ∩ hn+1 . Here tn is a Cartan subalgebra of the centralizer mn of an in kn . Now recursively construct positive root systems Σ + (mn , tn ) such that if α ∈ Σ + (mn+1 , tn+1 ), then either α|tn = 0 or α|tn ∈ Σ + (mn , tn ). Then we have positive root systems Σ + (gn , hn ) = {α ∈ ih∗n | α|an = 0 or α|an ∈ Σn+ (gn , an )}, the corresponding simple root systems, and the resulting systems of fundamental highest weights. The Cartan–Helgason theorem says that the irreducible representation πλ of gn of highest weight λ gives a summand of the representation of Gn on L2 (Mn ) if and only if (i) λ|tn = 0, so we may view λ as an element of ia∗n , and λ,α (ii) if α ∈ Σn+ (gn , an ) then α,α is an integer 0. Condition (i) persists under restriction λ → λ|hn−1 because tn−1 ⊂ tn . Given (i), condition (ii) says that 1 2 λ belongs to the weight lattice of gn , so its restriction to hn−1 exponentiates to a well-defined function on the corresponding maximal torus of Gn−1 and thus belongs to the weight lattice of gn−1 . Given condition (i) now (7) says that condition (ii) persists under restriction λ → λ|hn−1 . With this in mind, we define linear functionals ξn,j ∈ ia∗n by ξn,i ,ψn,j ψn,j ,ψn,j

= δi,j for 1 j rn , except that

ξn,1 ,ψn,1 ψn,1 ,ψn,1

= 2 if 2ψn,1 ∈ Σn .


465

The weights ξn,j are the class 1 fundamental highest weights for (gn , kn ). We denote Ξn = Ξn (gn , tn , an ) = {ξn,1 , . . . , ξn,rn } . Define Λn = Λ(gn , kn , an ) =

nk ξk | ξk ∈ Ξn and nk ∈ Z, nk 0 .

This is the set of highest weights for representations of Gn on L2 (Mn ), and we have just verified that Λn |an−1 ⊂ Λn−1 . Lemma 8 For n sufficiently large, and passing to a cofinal subsequence, if ξ ∈ Ξn−1 there is a unique ξ ∈ Ξn such that ξ |an−1 = ξ. Proof. In the group manifold cases, lines 1, 2, 3, and 4 of Table 5, express Gn = Ln × Ln , and note that the complexification (Ln−1 )C is the semisimple component of a parabolic subgroup of (Ln )C . The restricted root and weight systems of (Gn , Kn ) are the same as the unrestricted root and weight systems of Ln , and the assertion follows. In the Grassmann manifold cases, lines 5, 8, and 10 of Table 5, we first consider the case where {pn } is bounded. Then we may assume pn = p constant and qn increasing for n > > 0. Thus an−1 = an , Ψn−1 = Ψn (though the multiplicities of the restricted roots will increase), and Ξn−1 = Ξn . The assertion now is immediate. In the Grassmann manifold cases we may now assume that both pn and qn are unbounded. If pn = qn on a cofinal sequence of indices n we may assume pn = qn for all n, so Ψn is always of type Crn . Then we interpolate pairs and renumber so that pn = qn = pn−1 + 1 = qn−1 + 1 for all n and notice that the Dynkin diagram inclusions Cr−1 ⊂ Cr are uniquely determined by the integer r. If pn = qn for only finitely many n and pn < qn on a cofinal sequence of indices n we may assume that rn = pn < qn for all n, so Ψn is always of type Brn . Then we interpolate (pn−1 , qn − 1), (pn−1 , qn ), (pn−1 +1, qn ), . . . , (pn , qn ) and renumber so that we always have rn = rn−1 or rn = rn−1 + 1 and notice that the Dynkin diagram inclusions Br−1 ⊂ Br are uniquely determined by the integer r. If pn = qn for only finitely many n and also pn = qn for only finitely many n, then pn > qn on a cofinal sequence of indices n, and we may assume pn > qn = rn for all n. We interpolate as before, exchanging the rˆ oles of p and q , and we note again that the Dynkin diagram inclusions Br−1 ⊂ Br are uniquely determined by the integer r. Thus in all cases the fundamental highest weights restrict as asserted. In the lower rank cases, lines 6 and 7 of Table 5, Ψn is of type An−1 , so again restriction to an−1 has the required property. In the hermitian symmetric case, line 11 of Table 5, an is a Cartan subalgebra of gn and gn−1 complexifies to the semisimple part of a parabolic subalgebra of (gn )C , so the assertion follows as in the group manifold cases. In the remaining case, line 9 of Table 5, Ψn is of type Cn/2 for n even, type B(n−1)/2 for n odd. Passing to a cofinal subsequence

466

Joseph A. Wolf

we may assume n always even or always odd, and we may interpolate as necessary by pairs so that n increases in steps of 2. Then, again, there is no choice about the restriction, and the assertion follows. In view of Lemma 8, after passage to a cofinal subsequence and renumbering, we may assume the sets Ξn ordered so that Ξn = Ξ(gn , kn , an ) ={ξ1,n , . . . , ξrn ,n } with ξ,n−1 = ξ,n |an−1 for 1 rn−1 .

(9)

Now define In : all rn -tuples I = (i1 , . . . , irn ) of non-negative integers, I = lim In where In → Im by (i1 , . . . , irn ) → (i1 , . . . , irn , 0, . . . , 0), −→ πI,n : rep of Gn with highest weight ξI = i1 ξ1 + · · · + ip ξrn ,

(10)

πI = lim πI,n for I ∈ I. −→ According to the Cartan–Helgason theorem, the πI,n exhaust the representations of Gn on L2 (Mn ). Denote VI,n : representation space for the abstract representation πI,n .

(11)

Then VI,n occurs with multiplicity 1 in the representation of Gn on L2 (Mn ). In effect, the representation of Gn on L2 (Mn ) is multiplicity-free, and L2 (Mn ) ∼ = V as a G -module. However, in the following we must distinguish n I∈I I,n between I∈I VI,n as a Gn -module and L2 (Mn ) as a space of functions. Let U(gn ) denote the (complex) universal enveloping algebra of gn . Let vn+1 be a highest weight unit vector in VI,n+1 for the action of Gn+1 . Then we have the Gn -submodule U(gn )(vn+1 ) ⊂ VI,n+1 ⊂ L2 (Mn+1 ). If u, v ∈ VI,n we write fu,v;I,n for g → u, πI,n (g)v , the matrix coefficient function on Gn . These matrix coefficient functions span a space EI,n that is invariant under left and right translations by elements of Gn . As a (Gn × Gn )∗ module EI,n ∼ . If u∗n is the (unique up to scalar multiplication) = VI,n VI,n ∗ Kn -fixed unit vector in VI,n , then the right Kn -fixed functions in EI,n form Kn ∼ the left Gn -module EI,n = VI,n . = VI,n ⊗ u∗n C ∼ In the following, it is crucial to distinguish between the abstract represenKn tation space VI,n and the space EI,n of functions on Gn /Kn . We normalize the Haar measure on Gn (and the resulting measure in Mn ) to total mass 1. If u, v, u , v ∈ VI,n , then we have the Schur orthogonality relation fu,v;I,n , fu ,v ;I,n |L2 (Gn ) = (deg πI,n )−1 u, u v, v . Kn Theorem 12 The space EI,n of functions on Gn /Kn is Gn -module equivaK

K

Kn n+1 n+1 . We map EI,n into EI,n+1 as follows. lent to U(gn )(vn+1 ⊗ u∗n+1 ) ⊂ EI,n+1 Let {wj } be a basis of VI,n and define

ψn+1,n


cj fwj ,u∗n ;I,n

= (deg πI,n+1 / deg πI,n )1/2

K

n+1 . cj fwj ,u∗n+1 ;I,n+1 ∈ EI,n+1

467

(13)

K

Kn n+1 : EI,n → EI,n+1 is Gn -equivariant and is isometric for L2 Then ψn+1,n norms on Gn /Kn and Gn+1 /Kn+1 . In particular, as I varies with n fixed, ψn+1,n : L2 (Gn /Kn ) → L2 (Gn+1 /Kn+1 ) is a Gn -equivariant isometry.

Proof. We have a(vn+1 ) = ξI (a)vn+1 for all a ∈ a. The inclusion Gn → Gn+1 is Gn -equivariant, so restriction of functions is Gn -equivariant and thus is Aequivariant, and (vn+1 ⊗ u∗n+1 )|Mn is a ξI -weight vector in L2 (Mn ). If α is a positive restricted root for Gn+1 and eα ∈ gn+1 is an α root vector, then eα (vn+1 ) = 0. If α is already a root for Gn and if eα ∈ gn , then we have eα ((vn+1 ⊗ u∗n+1 )|Mn ) = 0. Thus either the restriction (vn+1 ⊗ u∗n+1 )|Mn = 0 Kn or (vn+1 ⊗ u∗n+1 )|Mn is a highest weight vector in EI,n . Suppose that (vn+1 ⊗ u∗n+1 )|Mn = 0 as a function on Mn = Gn /Kn . Denote Vn = U(gn )(vn+1 ). It is a cyclic highest weight module for Gn with highest weight ξI , and (Vn ⊗ u∗n+1 C)|Mn = 0, and it contains a unique (up to scalar multiple) Kn -invariant unit vector un . The coefficient function ϕ(g) := un , πI,n+1 (g)un Vn = Gn (un ⊗ u∗n )(x)(un ⊗ u∗n )(x−1 g)dx is identically zero because the un (x) factor in the integrand vanishes for x ∈ Gn . But ϕ|Gn is the positive definite (Gn , Kn )-spherical function on Gn for the representation πI,n , and in particular ϕ(1) = 1. That is a contradiction. We conclude that Kn . (vn+1 ⊗ u∗n+1 )|Mn = 0, so (vn+1 ⊗ u∗n+1 )|Mn is a highest weight vector in EI,n K Kn ∼ n+1 ∗ In particular, EI,n = (Vn ⊗ un+1 C)|Mn ⊂ EI,n+1 |Mn . That is the equivariant map assertion. The unitary map assertion follows by Schur orthogonality. Theorem 12 gives isometric embeddings ψm,n : L2 (Mn ) → L2 (Mm ) for n m. By construction, ψm,n is Gn -equivariant. Define } : direct limit in the L2 (G/K) = lim{L2 (Gn /Kn ), ψm,n −→ category of Hilbert spaces and unitary injections.

(14)

We emphasize the renormalizations of Theorem 12. Without those renormalizations we lose the Hilbert space structure of L2 (G/K). Theorem 15 The left regular representation of G on L2 (G/K) is a multiplicity-free discrete direct sum of lim-irreducible representations. Specif ically, that left regular representation is π , where πI = lim πI,n is I∈I I − → the irreducible representation of G with highest weight ξI := ir ξr . This applies to all the direct systems of Table 5. In particular, we have the thirteen infinite-dimensional multiplicity-free spaces

468

Joseph A. Wolf

1. SU (∞) × SU (∞)/diag SU (∞), group manifold SU (∞), 2. Spin(∞) × Spin(∞)/diag Spin(∞), group manifold Spin(∞), 3. Sp(∞) × Sp(∞)/diag Sp(∞), group manifold Sp(∞), 4. SU (p + ∞)/S(U (p) × U (∞)), Cp subspaces of C∞ , 5. SU (2∞)/[S(U (∞) × U (∞))], C∞ subspaces of infinite codim in C∞ , 6. SU (∞)/SO(∞), 7. SU (2∞)/Sp(∞), 8. SO(p + ∞)/[SO(p) × SO(∞)], oriented Rp subspaces of R∞ , 9. SO(2∞)/[SO(∞) × SO(∞)], R∞ subspaces of infinite codim in R∞ , 10. SO(2∞)/U (∞), 11. Sp(p + ∞)/[Sp(p) × Sp(∞)], Hp subspaces of H∞ , 12. Sp(2∞)/[Sp(∞) × Sp(∞)], H∞ subspaces of infinite codim in H∞ , 13. Sp(∞)/U (∞). Proof. λ is limit aligned by Theorem 12. Denote VI = VI,n = lim VI,n . −→ Then G acts irreducibly on it by πI = lim πI,n , and the various πI are mutu−→ ally inequivalent because they have different highest weightsξI := ir ξr , and are lim irreducible by construction. Now let V = V . Then I∈I I V = lim L2 (Gn /Kn ) = L2 (G/K). −→

4 Gelfand pairs and defining representations In this section we set the stage for the extension of Theorem 15 to a number of direct systems {(Gn , Kn )} of compact nonsymmetric Gelfand pairs. A glance at [Ya2004] or [W2007] reveals many such pairs, but here we will only consider those for which the compact groups Gn are simple. The following table shows the Kr¨ amer classification of Gelfand pairs corresponding to compact simple Lie groups (see [Kr1979] or [Ya2004] or [W2007, Table 12.7.1]).

1 2 3 4 5 6 7 8 9 10 11 12

Mn = Gn /Hn weakly symmetric Gn Hn Conditions SU (m + n) SU (m) × SU (n)] n > m 1 SO(2n) SU (n) n odd, n 3 E6 Spin(10) SU (2n + 1) Sp(n) n1 SU (2n + 1) Sp(n) × U (1) n1 Spin(7) G2 G2 SU (3) SO(10) Spin(7) × SO(2) SO(9) Spin(7) Spin(8) G2 SO(2n + 1) U (n) n2 Sp(n) Sp(n − 1) × U (1) n 1

Gn /Kn symmetric Kn with Hn ⊂ Kn ⊂ Gn S[U (m) × U (n)] U (n) Spin(10) · Spin(2) U (2n) = S[U (2n) × U (1)] U (2n) = S[U (2n) × U (1)] (there is none) (there is none) SO(8) × SO(2) SO(8) Spin(7) SO(2n) Sp(n − 1) × Sp(1)

(16)


469

This gives us the nonsymmetric direct systems {(Gn , Kn )}, where (a) Gn = SU (pn + qn ) and Kn = SU (pn ) × SU (qn ), pn < qn (b) Gn = SO(2n) and Kn = SU (n), n odd, n 3 (c) Gn = SU (2n + 1) and Kn = Sp(n), n 1 (d) Gn = SU (2n + 1) and Kn = U (1) × Sp(n), n 1 (e) Gn = SO(2n + 1) and Kn = U (n), n 2

(17)

(f) Gn = Sp(n) and Kn = U (1) × Sp(n − 1), n 2. Definition 18 Let {(Gn , Kn )} be a direct system of Lie groups and closed subgroups. Suppose that π = lim πn is a lim-irreducible representation of −→ G = lim Gn , with representation space V = lim Vn , such that (i) πn (Kn ) is −→ −→ the πn (Gn )-stabilizer of a vector vn ∈ Vn and (ii) each vn+1 = vn +wn+1 where πn (Gn ) leaves wn+1 fixed. (Thus the vn give a coherent system of embeddings of the Gn /Kn .) Suppose further that for n > > 0 the πn have the same highest weight vector. Then we say that π = lim πn is a defining representation for −→ {(Gn , Kn )}. Now let’s consider some important examples of defining representations. We will use these examples later. Example 19 Gn = SU (pn + qn ) and Kn = SU (pn ) × SU (qn ), pn < qn , in (17). Let {e1 , . . . , epn +qn } denote the standard orthonormal basis of Cpn +qn . Then Kn is the Gn -stabilizer of e1 ∧· · ·∧epn in the representation πn = Λpn (τ ), where τ is the standard (vector) representation of SU (pn + qn ) on Cpn +qn . In the usual notation, e1 ∧ · · · ∧ epn also is the highest weight vector, and the highest weight is ε1 + · · · + εpn . If the pn are bounded, so that we may assume each pn = p < ∞, then π = lim πn is well defined and is a defining −→ representation for {(Gn , Kn )}. ♦ Example 20 Gn = SU (2n + 1) and Kn = U (1) × Sp(n), n 1, in (17). Again, {e1 , . . . , e2n+1 } is standard orthonormal basis of C2n+1 . Now Kn the n 2 is the Gn -stabilizer of =1 e2 ∧ e2+1 in the representation πn = Λ (τ ), 2n+1 where τ is the standard (vector) representation of SU (2n + 1) on C . Here e1 ∧ e2 is the highest weight vector and the highest weight is ε1 + ε2 . Thus π = lim πn is well defined and is a defining representation for {(Gn , Kn )}. ♦ −→ Example 21 Gn = SO(2n + 1) and Kn = U (n), n 2, in (17). Let 2n+1 {e . Let J =

01 , .1.. , e2n+1 } denote the standard orthonormal basis of R . Then K is the G -stabilizer of diag{0, J, . . . , J} ∈ g in the adjoint n n n −1 0 representation of G ; in other words (in this case), it is the G -stabilizer of n n n 2 e ∧ e in the representation π = Λ (τ ), where τ is the standard 2 2+1 n =1 (vector) representation of SO(2n + 1) on R2n+1 . As in the previous example, e1 ∧ e2 is the highest weight vector and the highest weight is ε1 + ε2 . Thus π = lim πn is well defined and is a defining representation for {(Gn , Kn )}. ♦ −→

470

Joseph A. Wolf

Example 22 Gn = Sp(n) and Kn = U (1) × Sp(n − 1), n 2, in (17). In quaternion matrices, Kn is the Gn -commutator of diag{i, 0, 0, . . . , 0}. In 2n × 2n complex

0 1 matrices, it is the Gn -commutator of diag{J, 0, 0, . . . , 0}, where J = −1 0 . There, Gn consists of all elements g ∈ U (2n) such that t = J, where J = diag{J, J, . . . , J}. Thus gn is given by xJ + J xt = 0, g Jg and in particular diag{J, 0, 0, . . . , 0} ∈ gn . Now Kn is the Gn -stabilizer of diag{J, 0, 0, . . . , 0} in the adjoint representation πn of Gn . That adjoint representation is the symmetric square of the standard (vector) representation of Gn on C2n , so it has highest weight 2ε1 and highest weight vector e21 . Thus π = lim πn is well defined and is a defining representation for {(Gn , Kn )}. ♦ −→

5 Function algebras Fix a defining representation π = lim πn for {(Gn , Kn )}. We are going to −→ define algebras

A(Gn ) and A(G) = A(Gn );

A(Gn /Kn ) A(Gn /Kn ) and A(G/K) = of complex-valued polynomial functions and look at their relations to square integrable functions. Let dn = dimR Vn . Then we can consider Gn to be a group of real dn × dn matrices. Since the Gn are reductive linear algebraic groups, this lets us define A(Gn ) : the algebra of all C-valued functions f |Gn where f : Rdn ×qn → C is a polynomial, rn : A(Gn ) → A(Gn−1 ) : restriction of functions, Sn : kernel of the algebra homomorphism rn ,

(23)

Tn : Gn−1 -invariant complement to Sn in A(Gn ). The following is immediate. Lemma 24 The restriction rn |Tn : Tn → A(Gn−1 ) is a Gn−1 -equivariant vector space isomorphism. In other words we have a Gn−1 -equivariant injection (rn |Tn )−1 : A(Gn−1 ) → A(Gn ) of vector spaces with image complementary to the kernel of the restriction rn : A(Gn ) → A(Gn−1 ) of functions. Lemma 24 gives us A(G) = lim A(Gn ) = −→

A(Gn ).

Taking the right-invariant functions we arrive at


A(Gn /Kn ) := {h ∈ A(Gn ) | h(xk) = h(x) for x ∈ Gn , k ∈ Kn },

A(G/K) = A(Gn /Kn )

471

(25)

= {h ∈ A(G) | h(xk) = h(x) for x ∈ G, k ∈ K}. These are our basic function algebras. The algebra A(Gn ) contains the constants, separates points on Gn , and is stable under complex conjugation. The Stone–Weierstrass theorem is the main component of the following lemma. Lemma 26 The algebra A(Gn ) is dense in C(Gn ), the algebra of continuous functions Gn → C with the topology of uniform convergence. Let Sn and Tn denote the uniform closures of Sn and Tn in C(Gn ). Then rn extends by continuity to the restriction map rn : C(Gn ) → C(Gn−1 ), that extension rn restricts to a Gn−1 -equivalence Tn ∼ = C(Gn−1 ), C(Gn ) is the vector space direct sum of closed Gn−1 -invariant subspaces Sn and Tn , and this identifies C(Gn−1 ) as a Gn−1 -submodule of C(Gn ). Proof. The density is exactly the Stone–Weierstrass theorem in this setting. Since Sn and Tn involve different sets of variables, so do Sn and Tn . Now rn extends to rn as asserted and Sn is the kernel of rn . Similarly, Sn ∩ Tn = 0, and the induced algebra homomorphism rn : C(Gn ) → C(Gn−1 ) restricts to a Gn−1 -equivariant map rn : Tn ∼ = C(Gn−1 ). Finally, Sn + Tn is closed in C(Gn ) and contains A(Gn ). Thus C(Gn ) = Sn ⊕ Tn and we can identify C(Gn−1 ) with the closed Gn−1 -invariant subspace Tn of C(Gn ). Weuse the identifications C(Gn−1 ) ⊂ C(Gn ) of Lemma 26 to form the union C(Gn ). Note that C(Gn ) is the algebra of continuous functions on G that depend on only finitely many variables. Now use the sup norm, and thus the topology of uniform convergence, and define a Banach algebra

C(G) : functions f : G → C in the uniform limit closure of C(Gn ) with sup norm and topology of uniform convergence. Passing to the right Kn -invariant functions we have Banach function algebras C(Gn /Kn ) := {h ∈ C(Gn ) | h(xk) = h(x) for x ∈ Gn , k ∈ Kn and

C(G/K) = C(Gn /Kn )

(27)

= {h ∈ C(G) | h(xk) = h(x) for x ∈ G, k ∈ K}. Here A(Gn /Kn ) is the subalgebra consisting of all Gn -finite functions in C(Gn /Kn ), and consequently A(G/K) is the subalgebra consisting of all Gfinite functions in C(G/K). We pass to L2 limits more or less in the same way as in (25) and (27), except that we must rescale to preserve L2 norms as in Theorem 12. For this we need some machinery from [W2009]. Let {Gn } be a strict direct system

472

Joseph A. Wolf

of compact connected Lie groups, and {(Gn )C } the direct system of their complexifications. Suppose that, for each n, the semisimple part [(gn )C , (gn )C ] of the reductive algebra (gn )C is the semisimple component of a parabolic subalgebra of (gn+1 )C .

(28)

Then we say that the direct systems {Gn } and {(Gn )C } are parabolic and that lim Gn and lim(Gn )C are parabolic direct limits. This is a special case of the −→ −→ definition of parabolic direct limit in [W2005]. Now let {Gn } be a strict direct system of compact connected Lie groups that is parabolic. We recursively construct Cartan subalgebras tn ⊂ gn with t1 ⊂ t2 ⊂ · · · ⊂ tn ⊂ tn+1 ⊂ . . . and simple root systems Ψn = Ψ ((gn )C , (tn )C ) such that each simple root for (gn )C is the restriction of exactly one simple root for (gn+1 )C . Then we may assume that Ψn = {ψn,1 , . . . , ψn,p(n) } in such a way that each ψn,j is the (tn )C -restriction of ψn+1,j and of no other element of Ψn+1 . The corresponding sets Ξn = {ξn,1 , . . . , ξn,p(n) } of fundamental highest weights can be ordered so that they satisfy: ξn+1,j is the unique element of Ξn+1 whose (tn )C -restriction is ξn,j , for 1 j p(n). Exactly as in Theorem 12 this gives us isometric Gn -equivariant injections ψm,n : L2 (Gn ) → L2 (Gm ) for n m. The associated direct limit maps ψn : L2 (Gn ) → lim{L2 (Gn ), ψm,n } define the direct limit in the category of −→ Hilbert spaces and unitary maps as the Hilbert space completion completion L2 (G) = lim {L2 (Gn ), ψm,n } = ψn (L2 (Gn )) . −→ unitary

Lemma 29 Let {(Gn , Kn )} be one of the systems of Examples 19, 20, 21, or 22. Then {Gn } is parabolic and the Gn -equivariant maps ψm,n : L2 (Gn ) → L2 (Gm ) send right-Kn -invariants to right-Kn+1 -invariants, resulting in Gn -equiva riant unitary injections ψm,n : L2 (Gn /Kn ) → L2 (Gm /Km ). Proof. We use the defining relations given in Examples 19, 20, 21, and 22. In each case we look at the subspaces of L2 given by polynomials of degree d; those are finite-dimensional invariant subspaces of the A(Gn /Kn ). We observed above that A(Gn ) → A(Gn+1 ) maps right-Kn -invariants to rightKn+1 -invariants. On each irreducible summand, the L2 (Gn ) → L2 (Gn+1 ) differ only by scale from the corresponding summands of A(Gn ) and A(Gn+1 ), so they also map right-Kn-invariants to right-Kn+1-invariants. Now we have some L2 analogues of (25) and (27). L2 (Gn /Kn ) := {h ∈ L2 (Gn ) | h(xk) = h(x) for x ∈ Gn , k ∈ Kn }, completion L2 (G/K) = ψn (L2 (Gn /Kn )) = {h ∈ L2 (G) | h(xk) = h(x) for x ∈ G, k ∈ K}.

(30)


473

We have A(G/K) ⊂ C(G/K) ⊂ L2 (G/K) for these spaces, and A(G/K) is the set of polynomial elements in L2 (G/K). Theorem 31 Let {(Gn , Kn )} be one of the direct systems of nonsymmetric Gelfand pairs given by Examples 19, 20, 21, and 22. Then the left regular representations of G on A(G/K), C(G/K), and L2 (G/K) are multiplicityfree discrete direct sums of lim-irreducible representations. In the notation of (9), (10), and (11), those left regular representations are I∈I πI , where πI = lim πI,n is the irreducible representation of G with highest weight ξI := ir ξr . −→ Thus we have the infinite-dimensional multiplicity-free spaces (1) SU (p + ∞)/(SU (p) × SU (∞)) for 1 p ∞, (2) SU (1 + 2∞)/(U (1) × Sp(∞)), (3) SO(1 + 2∞)/U (∞), and (4) Sp(1 + ∞)/(U (1) × Sp(∞)) Proof. Examples 19, 20, 21, and 22 have defining representations and welldefined function spaces A(G/K) and C(G/K). The same holds for L2 (G/K) by Lemma 29. In these examples {Gn } is parabolic, so the left regular representations are limit aligned by Theorem 12. Now the proof of Theorem 15 holds for these four examples, resulting in the multiplicity-free property for their left regular representations.

6 Pairs related to spheres and Grassmann manifolds In dealing with nonsymmetric Gelfand pairs we have to be very specific about the embeddings Gn−1 → Gn , so we review a few of those embeddings. Orthogonal groups. Let Gn = SO(n0 + 2n), the special orthogonal group n +2n ui vi . The embeddings are given by for the bilinear form h(u, v) = 1 0 x 00 0 10 0 01

. Then G = lim Gn is the classical direct −→ limit group SO(∞). It doesn’t matter what n0 is here, but sometimes we have to distinguish between the cases of even or odd n0 , and in any case we want {Gn } to be parabolic, so we jump by two 1’s instead of just one. Specifically, this direct system consists either of groups of type B (when the n0 + 2n are odd) or of type D (when the n0 + 2n are even). In this section Kn = { ( 10 x0 )| x ∈ SO(n0 + 2n − 1)} ⊂ Gn . Then Gn /Kn is the sphere S n0 +2n−1 , G = lim Gn = SO(∞), and we express K = lim Kn as SO(1) × SO(∞ − 1) to −→ −→ indicate the embedding K → G. A defining representation for {(Gn , Kn )} is given by the family of standard (vector) representations πn of SO(n0 + 2n) on Rn0 +2n . Here {SO(n0 + 2n)} is a parabolic direct system. In the standard orthonormal basis the πn all have the same highest weight vector e1 and highest weight ε1 . Following the considerations of Section 5, this defining representation π = lim πn −→ Gn → Gn+1 given by x →

474

Joseph A. Wolf

defines the function spaces A(Gn /Kn ), C(Gn /Kn ), and L2 (Gn /Kn ). The πn share a highest weight vector so we have natural equivariant inclusions A(Gn−1 /Kn−1 ) → A(Gn /Kn ), C(Gn−1 /Kn−1 ) → C(Gn /Kn ), and L2 (Gn−1 /Kn−1 ) → L2 (Gn /Kn ), and thus the limits A(G/K), C(G/K), and L2 (G/K). Thus we have the regular representation of G = SO(∞) on those limit spaces. Unitary groups. Fix p > 0 and define Gn = SU (p + n), the special p+n unitary group for the complex hermitian form h(u, v) = ui v¯i . The 1 embedding Gn → Gn+1 is given by x → ( x0 10 ). Then G = lim Gn is −→ the classical parabolic direct limit group SU (∞). In this section Kn = { ( 10 x0 )| x ∈ SU (p), y ∈ SU (n)}. Then Gn /Kn is a circle bundle over the Grassmann manifold of p-dimensional linear subspaces of Cp+n , G = lim Gn = −→ SU (∞), and we sometimes express K = lim Kn as SU (p) × SU (∞ − p) to −→ indicate the embedding K → G. If p = 1 then Gn /Kn is the sphere S 2n+1 , the complex Grassmann manifold is a complex projective space, and the circle bundle projection is the Hopf fibration. Here the defining representation is essentially that of Example 19. Let πξ1 denote the usual vector representation of Gn on Cp+n . Write πξp for its pth alternating power, the representation of Gn on Λp (Cp+n ); it is the first representation of Gn with a vector fixed under Kn . That vector is e1 ∧ · · · ∧ ep relative to the standard basis {e1 , . . . , en } of Cn , and Kn is its Gn -stabilizer. Thus the πn = πξp give a defining representation for {(Gn , Kn )}. Note that the πn all have the same highest weight vector e1 ∧ · · · ∧ ep and highest weight ε1 + · · · + εp . Following the considerations of Section 5, this defining representation π = lim πn defines the function spaces A(Gn /Kn ), C(Gn /Kn ), and −→ L2 (Gn /Kn ). The πn share a highest weight vector, so we have natural equivariant inclusions A(Gn−1 /Kn−1 ) → A(Gn /Kn ), C(Gn−1 /Kn−1 ) → C(Gn /Kn ), and L2 (Gn−1 /Kn−1 ) → L2 (Gn /Kn ), and thus the limits A(G/K), C(G/K), and L2 (G/K). That gives us the regular representation of G = SU (∞) on those limit spaces. Symplectic groups. Here nSp(n) is the unitary group of the quaternionhermitian form h(u, v) = 1 ui v¯i on the quaternionic vector space Hn . We then have Gn = Sp(n) × Sp(1), where the Sp(1) acts by quaternion scalars on Hn . We will also look at its subgroup Sp(n) × U (1), where U (1) is any (they are all conjugate) circle subgroup of Sp(1), say {eiθ | θ ∈ R}. In both cases the embeddings Gn → Gn+1 are specified by the maps Sp(n) → Sp(n + 1) given by x → ( x0 10 ). (We are using quaternionic matrices.) Then G = lim Gn is the −→ classical direct limit group Sp(∞) × Sp(1) and G = lim Gn is Sp(∞) × U (1). −→ (We need the Sp(1) or the U (1) factor because otherwise, as we will see below, the multiplicity-free property will fail.) Symplectic 1. First consider the parabolic direct system given by Gn = Sp(n) × Sp(1). Given n we have two Sp(1) groups to deal with at the same time, so we avoid confusion by denoting the Sp(1) factor of Gn as Sp(1)ext,n (ext for external) and the identity component of the centralizer of Sp(n − 1)


475

in Sp(n) by Sp(1)int,n (int for internal). In our matrix descriptions of Gn , the group Sp(1)diag,n is the diagonal subgroup in Sp(1)int,n × Sp(1)ext,n . Then Gn = Sp(n) × Sp(1)ext,n and G = lim Gn = Sp(∞) × Sp(1). Now let −→ Kn = { ( 10 x0 )| x ∈ Sp(n − 1)} × Sp(1)diag,n and K = lim Kn . Then Gn /Kn −→ is the sphere S 4n−1 , in other words, the Hopf fibration 3-sphere bundle over quaternion projective space P n−1 (H). In order to indicate the embedding K → G we express K as {1} × Sp(∞ − 1) × Sp(1). A defining representation for {(Gn , Kn )} is given by the family of standard (vector) representations πn of Sp(n) on C2n tensored with the standard 2-dimensional representation of Sp(1) on C2 . That representation has an invariant real form R4n . Consider the standard orthonormal basis {ei ⊗ fj } of C2n ⊗ C2 . The representations πn of Gn there have the same highest weight vector e1 ⊗ f1 and highest weight (ε1 )Sp(n) + (ε1 )Sp(1) . They give a defining representation for {(Gn , Kn )}. Following the considerations of Section 5, this defining representation π = lim πn defines the function spaces A(Gn /Kn ), −→ C(Gn /Kn ), and L2 (Gn /Kn ). The πn have the same highest weight vector so we have natural equivariant inclusions A(Gn−1 /Kn−1 ) → A(Gn /Kn ), C(Gn−1 /Kn−1 ) → C(Gn /Kn ), and L2 (Gn−1 /Kn−1 ) → L2 (Gn /Kn ), and thus the limits A(G/K), C(G/K), and L2 (G/K). So we have the regular representation of G = Sp(∞) × Sp(1) on those limit spaces. Symplectic 2. Next consider the parabolic direct system given by Gn = Sp(n) × U (1), where the Sp(1) factor of Sp(n) × Sp(1) is replaced by the circle subgroup {eiθ | θ ∈ R}. Given n we have two U (1) groups, the U (1)ext,n that is the U (1) factor of Gn and the corresponding circle subgroup U (1)int,n of Sp(1)int,n . Then of course we have the diagonal U (1)diag,n . As above we define Kn to be the product group { ( 10 x0 )| x ∈ Sp(n − 1)} × U (1)diag,n and we set K = lim Kn . Then Gn /Kn again is the sphere S 4n−1 . We express K −→ as {1} × Sp(∞ − 1) × U (1). A defining representation for {(Gn , Kn )} is given by the family of standard (vector) representations πn of Sp(n) on C2n tensored with the standard 1-dimensional representation of U (1) on C. The representations πn of Gn there have the same highest weight vector e1 ⊗f1 . The corresponding highest weight is (ε1 )Sp(n) +(ε1 )U (1) , and the πn give a defining representation for {(Gn , Kn )}. Following the considerations of Section 5, this defining representation π = lim πn defines the function spaces A(Gn /Kn ), C(Gn /Kn ), and L2 (Gn /Kn ). −→ The πn have the same highest weight vector, so we have natural equivariant inclusions A(Gn−1 /Kn−1 ) → A(Gn /Kn ), C(Gn−1 /Kn−1 ) → C(Gn /Kn ), and L2 (Gn−1 /Kn−1 ) → L2 (Gn /Kn ), and thus the limits A(G/K), C(G/K), and L2 (G/K). So we have the regular representation of G = Sp(∞) × U (1) on those limit spaces. Symplectic 3. A variation on the case just considered is where Kn = { ( z0 x0 )| z ∈ U (1), x ∈ Sp(n − 1)} × U (1), and K = lim Kn . Then the U (1) −→ factor of Gn is contained in Kn so it acts trivially on Gn /Kn . Thus Gn /Kn is a 2-sphere bundle over P n−1 (H) exactly as in the “Symplectic 2” case.

476

Joseph A. Wolf

We express K as U (1) × Sp(∞ − 1) × U (1). The groups Kn are larger than the case “Symplectic 2” just considered, so the present function spaces A(Gn /Kn ), C(Gn /Kn ), and L2 (Gn /Kn ) are subspaces of those of “Symplectic 2”, and the same holds for their limits A(G/K), C(G/K), and L2 (G/K). Now we have the regular representation of G = Sp(∞) × Sp(1) on those limit spaces. Symplectic 4. A variation on the “Symplectic 1” case is where Gn = Sp(n) × Sp(1) and Kn = { ( z0 x0 )| z ∈ U (1), x ∈ Sp(n − 1)} × Sp(1) and K = lim Kn . Then the Sp(1) factor of Gn is contained in Kn , so it acts trivially −→ on Gn /Kn . Thus Gn /Kn = Sp(n)/[U (1) × Sp(n − 1)] is a 2-sphere bundle over P n−1 (H), exactly as in the “Symplectic 3” case above. We express K as U (1) × Sp(∞ − 1) × Sp(1) and we note that the function spaces A(Gn /Kn ), C(Gn /Kn ), and L2 (Gn /Kn ) are exactly the same as those of “Symplectic 3”, so the same holds for their limits A(G/K), C(G/K), and L2 (G/K). Thus we have the regular representation of G = Sp(∞) × Sp(1) on those limit spaces. The classifications of Krämer [Kr1979] and Yakimova [Ya2004] (see [W2007]) show that the six direct systems just described, one orthogonal, one unitary, and four symplectic, all consist of Gelfand pairs.

7 Limits related to spheres and Grassmann manifolds In this section we prove the multiplicity-free property for the direct limits of Gelfand pairs described in Section 6. Theorem 32 Let (G, K) = lim{(Gn , Kn )}, where {(Gn , Kn )} is one of the −→ six systems described in Section 6. Let A(G/K), C(G/K), and L2 (G/K) be as described there. Then the regular representations of G on A(G/K), C(G/K), and L2 (G/K) are multiplicity-free discrete direct sums of lim-irreducible representations. Proof. We run through the proof of Theorem 32 for the three types of limit groups G. In each case we do this by examining the representation of Gn on A(Gn /Kn ), verifying the limit-aligned condition, and applying Theorem 4 to the regular representation of G on A(G/K). We already know the result for the orthogonal group case, where the (Gn , Kn ) are symmetric pairs, but we need the representation-theoretic information from that case in order to deal with the other cases. Orthogonal group case. Here we shift the index so that Gn = SO(n) and Kn = SO(n − 1). Then Gn /Kn is the unit sphere in Rn . The Gn -finite functions on Gn /Kn are just the restrictions of polynomial functions on Rn . Let ψ1;n denote the usual representation of Gn on Rn and let ξ denote its highest weight. Choose orthonormal linear coordinates {x1 , . . . , xn } of that Rn such that the monomial x1 is a highest weight vector. Then the representation of Gn on the space of polynomials of pure degree is of the form ψ;n ⊕ γ;n ,


477

where ψ;n is the irreducible representation of highest weight ξ and highest weight vector x1 . Then γ;n is the sum of the ψ−2j;n for 1 j [/2], and the representation space of that ψ−2j;n consists of the polynomial functions on Rn divisible by ||x||2j but not by ||x||2j+2 . Write E;n for the space of functions on Gn /Kn obtained by restricting those polynomials of degree contained in the representation space for ψ;n . Then A(Gn /Kn ) = 0 E;n . We now verify that the inclusions A(Gn /Kn ) → A(Gn+1 /Kn+1 ) send E;n into E;n+1 , so that the representation of G on A(G/K) is limit aligned and Theorem 4 shows that lim A(Gn /Kn ) is the multiplicity-free direct sum of −→ lim-irreducible G-modules E = lim E;n . For that, note that the restriction −→ A(Gn+1 /Kn+1 ) → A(Gn /Kn ) is obtained by setting xn+1 equal to zero. Thus the inclusions E,n → A(Gn+1 /Kn+1 ) are given by identifying the function x1 : Rn → R with the function x1 : Rn+1 → R and applying Gn -equivariance. Now A(G/K) = lim A(Gn /Kn ) is the direct sum of the E = lim E;n , and the −→ −→ representations of G on the E are the mutually inequivalent lim-irreducible lim ψ;n . That gives an elementary proof for the case G = SO(∞) and K = −→ SO(∞ − 1). Unitary group cases. Here we shift the index so that Gn = SU (n) and Kn = SU (p)×SU (n−p), n > p. So G = SU (∞) and K = SU (p)×SU (∞−p). Without loss of generality assume n > 2p so that the (Gn , Kn ) are Gelfand pairs. Recall the defining representation π = lim πn where πn = πξp , the pth −→ exterior power of the vector representation of Gn on Cn . So Kn is the Gn stabilizer of eI0 := e1 ∧ · · · ∧ ep , resulting in the map Gn /Kn → Λp (Cn ) by gKn → g(eI0 ). We have C-linear functions zI on Λp (Cn ) dual to the basis of Λp (Cn ) consisting of the eI with I = (i1 , . . . , ip ), where 1 i1 < · · · < ip n. (Here I0 = (1, 2, . . . , p).) Their real and imaginary parts generate the algebra A(Gn /Kn ). Relative to the diagonal Cartan subalgebra of gn the eI are weight vectors, and eI0 is the highest weight vector, for πξp . Now the action of Gn on the polynomials of degree in the zI and the zI is r+s= πrξp +sξn−p , where πrξp +sξn−p has highest weight rξp + sξn−p and highest weight vector zIr0 zIs0 . Those representations are mutually inequivalent, using n > 2p, and A(Gn /Kn ) = E 0 r+s= r,s;n , where Gn acts on Er,s;n by πrξp +sξn−p . The A(Gn /Kn ) → A(Gn+1 /Kn+1 ) are given on the level of Er,s;n → Er,s;n+1 by identifying zIr0 zIs0 : Λp (Cn ) → C with zIr0 zIs0 : Λp (Cn+1 ) → C. In view of Theorem 4, it follows that the representation of G on A(G/K) is a limitaligned discrete direct sum of mutually inequivalent lim-irreducible representations. We will need the case p = 1 when we look at the symplectic group cases. There Gn = SU (n) and Kn = {1} × SU (n − 1), and the Gn -finite functions on Gn /Kn are just the restrictions of finite linear combinations of the functions z r z¯s . We saw how to decompose A(S 2n−1 ) into irreducible modules for SO(2n): it is the sum of the spaces E;2n described above with highest weight

478

Joseph A. Wolf

ξ and highest weight vector x1 , where, of course, xj = 12 (zj + z¯j ). In terms of the Dynkin diagram that representation is c ψ : c c q q q c HH Hc and ψ;2n |U (n) = r+s= ψr,s;n , where ψr,s;n has diagram r c

s−r × . r s ¯1 . Let Both ψr,s;n and ψr,s;n |SU(n) have highest weight vector z 1z Er,s;n denote the representation space for ψr,s;n . Now A(Gn /Kn ) = 0 r+s= Er,s;n . Symplectic group cases. First suppose Gn = Sp(n) × U (1). There are two cases: (i) Kn = {1} × Sp(n − 1) × U (1)diag,n and (ii) Kn = U (1) × Sp(n − 1) × U (1). The assertions for case (i) will imply them for case (ii), so we may assume that Kn = {1} × Sp(n − 1) × U (1)diag,n . Then (G, K) = lim{(Gn , Kn )} has defining representation π = lim πn where πn −→ −→ 1 1 b qqq b < b × of Gn on C2n . is the representation b Note that πn factors through the vector representation of U (2n) on C2n . We saw how U (2n) acts irreducibly on the space Er,s;2n by the representation ψr,s;2n , which has diagram r s s−r c c q q q c × . sends the U (1) factor of Gn We now need two facts. First, Gn → U (2n) to the center of U (n). Second, ψr,s;2n |Sp(n) = 0mmin(r,s) ϕr,s,m;n , where c

r + s − 2m m

qqq

s c

c c q q q c < c . That gives us ϕr,s,m;n has diagram ψr,s;2n |Sp(n)U (1) = 0mmin(r,s) ϕr,s,m;n , where ϕr,s,m;n is the representation r + s − 2m m s−r c c q q q c < c of Sp(n)U (1) with diagram × , and ϕr,s,m;n has the same representation space (call it Er,s,m;n ) as ϕr,s,m;n . The Er,s,m;n are irreducible and inequivalent under Sp(n)U (1); in other words, the irreducible representations ϕr,s,m;n all are mutually inequivalent. Note, however, that ϕr,s,m;n ϕr+t,s−t,m;n for all t such that r + t, s − t 0; this reflects the fact that (Sp(n), Sp(n − 1)) is not a Gelfand pair. To trace the inclusions let {z1 , . . . , z2n } be the coordinates of C2n , all weight vectors, where z1 is the highest weight vector, z2 = e−α1 z1 is the next highest, and so and the antisymmetric bilinear invariant of Sp(n) on C2n on, n is vn (z, w) = 1 (z2i−1 w2i − z2i w2i−1 ). Then z1 is the highest weight vector b b qqq b < b and Λ2 C2n is the sum Λ20 C2n ⊕ vn C of its of irreducible component and trivial under the action of Sp(n). component its 0 1 0 1 Here vn has matrix diag −1 . . . and we work with the maximal 0 −1 0 toral subalgebra that consists of all matrices diag {a1 , −a1 ; . . . ; an , −an }; thus the highest weight vector on Λ20 C2n is sn (z, w) = z1 w3 − z3 w1 . Now sm n is the m b b qqq b < b . The corresponding highest weight vector of


479

highest weight vector of ϕr,s,m;n is z1r−m z¯1s−m sm n . Now the restriction A(Gn+1 /Kn+1 ) → A(Gn /Kn ) maps the highest weight vector z1r−m z¯1s−m sm n+1 of ϕr,s,m;n+1 to the highest weight vector z1r−m z¯1s−m sm n . This proves that the representation of G on A(G/K) is limit aligned. Theorem 4 shows that lim A(Gn /Kn ) is the multiplicity-free direct sum of lim-irreducible G-modules −→ Er,s,m := lim Er,s,m;n . −→ Finally, we suppose Gn = Sp(n) × Sp(1). Again there are two cases: (i) that Kn = {1}×Sp(n−1)×Sp(1)diag,n and (ii) Kn = U (1)×Sp(n−1)×Sp(1). The function algebras and group actions in case (ii) are exactly the same as those of the setting (Gn , Kn ) = (Sp(n) × U (1), U (1) × Sp(n − 1) × U (1)) above, where the assertions are proved. Thus we need only consider case (i), Kn = {1} × Sp(n − 1) × Sp(1)diag,n . Then (G, K) = lim{(Gn , Kn )} −→ has defining representation π = lim πn described in “Symplectic 1” above. −→ Those πn satisfy the condition of Theorem 4 because Sp(n) × Sp(1) simply puts together representation spaces Er−m,s−m,m;n of Sp(n) × U (1) on A(Sp(n)U (1)/Sp(n − 1)U (1)). This assembly maintains total degree = (r − m) + (s − m) + 2m, views the U (1) factor of Sp(n) × U (1) as a maximal s−r torus of the Sp(1) factor of Sp(n)× Sp(1), and sums the spaces for the × to form the space for the irreducible representation (call it β ) of Sp(1) of degree b . Now the irreducible spaces for Sp(n) × Sp(1) are + 1. It has diagram the F,m;n := r+s= E r−m,s−m,m;n and the corresponding representations are the ϕ,m,n := r+s= ϕr−m,s−m,m;n . This proves that the representation of G on A(G/K) is limit aligned. Theorem 4 shows that lim A(Gn /Kn ) is the −→ multiplicity-free direct sum of lim-irreducible G-modules F,m := lim F,m;n . −→ We have proved Theorem 32. Remark 33 Alternatively, the systems (d), (e), and (f) from the list (17), and also (a) when the {pn } are bounded, can be treated by the method of Sections 6 and 7. That gives an alternative proof of the multiplicity-free property for the pairs (1) SU (p + ∞)/(SU (p) × SU (∞)) for 1 p ∞, (2) SU (1 + 2∞)/(U (1) × Sp(∞)), (3) SO(1 + 2∞)/U (∞), and (4) Sp(1 + ∞)/(U (1) × Sp(∞)) of Theorem 31.

♦

8 Conclusions We have proved that the regular representations of G on A(G/K), C(G/K), and L2 (G/K), are multiplicity-free discrete direct sums of lim-irreducible representations in the following cases. In addition, in these cases it is always permissible to enlarge the groups Kn , say to F · Kn where F is a closed subgroup

480

Joseph A. Wolf

of the normalizer NGn (Kn ), because A(Gn /[F · Kn ]) is a Gn -submodule of A(Gn /Kn ). Limits of riemannian symmetric spaces. We have the multiplicity-free property for the thirteen cases described in Theorem 15, as well as some obvious variations. The latter include SO(∞) × SO(∞)/diag SO(∞) = lim SO(n) × SO(n)/diag SO(n) and −→ SO(p + ∞)/[S(O(p) × O(∞))] = lim SO(n)/[S(O(p) × O(n − p))] . −→ Limits of a few systems of Gelfand pairs. We have the multiplicity-free property for the four cases described in Theorem 31, (1) SU (p + ∞)/(SU (p) × SU (∞)) for 1 p ∞, (2) SU (1 + 2∞)/(U (1) × Sp(∞)), (3) SO(1 + 2∞)/U (∞), and (4) Sp(1 + ∞)/(U (1) × Sp(∞)). We also have the multiplicity-free property for spaces that interpolate between (SU (p + ∞), SU (p) × SU (∞)) and the limit Grassmannian (U (p + ∞), lim U (p) × U (n). −→ Fix a closed subgroup F of U (1). Then we have the multiplicity-free property for the pairs (G, K) = lim{(Gn , Kn )}, where Gn = SU (p + n) and −→ kn 0 k ∈ U (p), k ∈ SU (n), det k ∈ F . Kn = n n n 0 k n

Limits of Gelfand pairs related to spheres and Grassmann manifolds. We have the multiplicity-free property for the six cases described in Theorem 32, four of which are nonsymmetric, as well as some obvious variations. Fix a closed subgroup F of U (1); it can be any finite cyclic group or the entire circle group U (1). As a result we have the multiplicity-free property for the nonsymmetric pairs SU (∞)/[SU (p) × SU (∞ − p)] = lim SU (n)/[SU (p) × SU (n − p)] , −→ [Sp(∞) × U (1)]/[F × Sp(∞ − 1) × U (1)diag ] = lim[Sp(n) × U (1)]/[F × Sp(n − 1) × U (1)diag ] , −→ [Sp(∞) × Sp(1)]/[{1} × Sp(∞ − 1) × Sp(1)diag ] = lim[Sp(n) × Sp(1)]/[{1} × Sp(n − 1) × Sp(1)diag ] , and −→ [Sp(∞) × Sp(1)]/[{±1} × Sp(∞ − 1) × Sp(1)diag ] = lim[Sp(n) × Sp(1)]/[{±1} × Sp(n − 1) × Sp(1)diag ] . −→ What we don’t have. There is a huge number of direct systems {(Gn , Kn )} of Gelfand pairs where the Gn are compact connected Lie groups. We have


481

only verified the multiplicity-free condition for a few of them. We have not, for example, checked it for the interesting cases Gn = SU (2n + 1) and Kn = F × Sp(n), F ⊂ U (1) finite cyclic, and Gn = SO(2n) and Kn = F × SU (n), n odd, n 3. Also, we have not checked it for the very interesting case Gn = Sp(an ) × Sp(bn ) and Kn = Sp(an − 1) × Sp(1) × Sp(bn − 1), which is a prototype for nonsymmetric irreducible direct systems {(Gn , Kn )} with the Gn semisimple → Gn is given by not simple. In that case Kan −1

k 0 a but 0 1 n (k1 , a, k2 ) → , , so G /K fibers over P (H) × P bn −1 (H) n n 0 k 0 a 2 3 with fiber (Sp(1) × Sp(1))/(diagonal) = S .

References [Ar1962]

S. Araki, On root systems and an infinitesimal classification of irreducible symmetric spaces, J. Math. Osaka City Univ. 13 (1962), 1–34. [DPW2002] I. Dimitrov, I. Penkov & J. A. Wolf, A Bott–Borel–Weil theory for direct limits of algebraic groups, Amer. J. of Math. 124 (2002), 955–998. [Fa2006] J. Faraut, Infinite dimensional harmonic analysis and probability, in “Probability Measures on Groups: Recent Directions and Trends,” ed. S. G. Dani & P. Graczyk, Narosa, New Delhi, 2006. [He1978] S. Helgason, “Differential Geometry, Lie Groups, and Symmetric Spaces,” Academic Press, New York, 1978 [Kr1979] M. Kr¨ amer, Sph¨ arische Untergruppen in kompakten zusammenh¨ angenden Liegruppen, Compositio Math. 38 (1979), 129–153. [Ol1990] G. I. Ol’shanskii, Unitary representations of infinite dimensional pairs (G, K) and the formalism of R. Howe, in “Representations of Lie Groups and Related Topics,” ed. A. M. Vershik & D. P. Zhelobenko, Advanced Studies Contemp. Math. 7, Gordon and Breach, New York, 1990. [W1980] J. A. Wolf, Foundations of representation theory for semisimple Lie groups, in “Harmonic Analysis and Representations of Semisimple Lie groups,” ed. Wolf, Cahen & DeWilde, Mathematical Physics and Applied Mathematics, vol. 5 (1980), Reidel, Dordrecht, pp. 69–130. [W2005] J. A. Wolf, Direct limits of principal series representations, Compositio Mathematica, 141 (2005), 1504–1530. [W2007] J. A. Wolf, Harmonic Analysis on Commutative Spaces, Math. Surveys & Monographs vol. 142, Amer. Math. Soc., 2007. [W2009] J. A. Wolf, Infinite dimensional multiplicity-free spaces II: Limits of commutative nilmanifolds, Contemporary Mathematics 491 (2009), 179–208. [Ya2004] O. S. Yakimova, Weakly symmetric Riemannian manifolds with reductive isometry group, Math. USSR Sbornik 195 (2004), 599–614.

Index

(C • (g, V ), dg ); Chevalley–Eilenberg complex of g, 286 C k -regular Lie group, 269 Fω ; flux homomorphism, 329 M (k, σ1 , . . . , σr ); multi loop algebra, 324 Pρ ; flat bundle, 287 V (k), 283 W ∗ -algebra, 423 [A, A], 78 [a b], 76 Aut(P ), 282, 286 Bun(M, K), 328 Γ E; sections of the bundle E, 286 Gau(P ), 282, 286 aut(P ) = V(P )K , 282 V, 315 z ⊕ω g; central extension defined by ω, 294 gau(P ), 287, 314 Ad(P ), 316 λg , 285 1 Ω (M, V), 315 p Ω (M, V ), 294 [p] ω ; cocycles defined by a differential form, 302 ωη , 290 ωκ , 290, 315 ωκ,η , 290 1 Ω (M, V ), 283 perκ , 297 perκ ; κ invariant bilinear form, 297 perω ; period homomorphism, 329

ρg , 285 σ-topology, 423 Diff(M ), 287 ∧κ , 298 kω -space, 256 xl , left invariant vector field, 285 xr , right invariant vector field, 285 A(P ); affine space of connections on P , 321 V(M ), 282 (LB)-space, 260 (LF)-space, 260 2-morphism betw. morph. of bun. ger., 342

˙ I (root system of type A), 62 A An (finite root system of type A), 63 abstract Wiener group, 399 abstract Wiener space, 397 accessible set, 425 adjoint action, 432 adjoint map, 421 adjoint of a symmetric norming function, 435 adjoint orbit, 432 admissible Cartan matrix, 173 affine coadjoint action, 438, 443 affine coadjoint orbit, 442 affine form, 74 affine Kac–Moody Lie groups, 282 affine Kac–Moody superalgebras, 183 affine Lie algebra, 65 extended, 104

484

Index

locally extended, 108 toral type extended, 110 affine reflection Lie algebra, 96 isomorphic, 96 nullity, 97 affine reflection system, 72 affine form, 74 discrete, 75 morphism, 72 nullity, 72 affine root system, 65 extended, 75 locally extended, 75 Saito’s extended, 75 twisted, 67 untwisted, 65 algebra, 76 k-algebra, 76 base field extension, 79 central, 77 central-simple, 77 centre (associative), 77 centre (Lie), 78 centroid, 76 derivation, 76 graded, see graded algebra, 77 invariant bilinear form, 76 perfect, 76 simple, 76 almost local map, 262 alternative torus, 23, 29 isotope, 29 A(u1 ,u2 ) , 29 isotopic, 30 opposite, 31 amenable group, 375 anisotropic roots, 100 Arens-type product, 377 ARLA, see affine reflection Lie algebra, 96 associated norm ideals, 373 associated vector bundle, 445 associative torus, 23, 24 associative torus with involution, 23, 35 isotope, 37 (A, ι(h) ), 37 isotopic, 39 mod-2 quadratic form, 36 automatic smoothness, 275

automorphic forms on orthogonal groups, 151, 157–160, 162–165, 167 automorphic product, 157–160, 162–165, 167 classification, 164–165, 167 reflective, 160, 163–165, 167 BI (root system of type B), 62 Bn (finite root system of type B), 63 Banach ideal, 433, 434 Banach Poisson manifold, 420 Banach symplectic manifold, 420 Banach–Lie algebra, 421 Φ-reductive, 384 classical complex, 385 classical real, 385 Banach–Lie group Φ-reductive, 385 Φ-reductive linear, 385 classical complex, 385 classical real, 385 of Harish-Chandra type, 388 Banach–Lie–Poisson space, 421 base, 174 BCI (root system of type BC), 62 BCn (finite root system of type BC), 63 biconjugacy class, 359 bilinear form graded, 78 invariant, 61, 76 radical of, 61 Borel–Weil Theorem, 419, 444, 451 Bott–Thurston cocycle, 309 bounded operator, 422 boundedly regular, 249 box, 257 box topology, 257 Boyd indices, 371 Brownian motion, 398 Brownian sheet, 400 bump function, 268 bundle automorphism group; Aut(P ), 282 bundle gerbe, 323, 341, 349, 354 basic, 347 bundle gerbe bimodule, 358 bundle gerbe module, 357

Index CI (root system of type C), 62 Cn (finite root system of type C), 63 canonical map, 420 canonical projection, 72 Cartan map, 290 Cartan matrix, 174 Cartan subalgebra splitting, 64 Cartan–Helgason Theorem, 459 Cartan–Helgason theorem, 464 CDer (centroidal derivations), 82 Cent(A) (centroid of A), 76 (Cent A) ( centroidal transformations of degree ), 79 central, 77 central extension, 90, 339, 357, 435, 437, 442 covering, 90 graded, 90 graded covering, 90 central-simple, 77 centralizer, 432 centralizer of a functional, 430 centralizer of an element, 430 centreless core, 96 Lie algebra, 77 centroid, 9, 76 centroidal grading group, 81 centroidal derivation, 82 centroidal grading group, 9, 81 Γ (L), 9 character, 194 character formula, 151, 153, 154 characteristic distribution, 424, 426, 431, 434, 438–440, 442, 443 characteristic subspace, 424, 426, 429 class 1 fundamental highest weight, 465 classical root systems, 62 classification of locally finite root systems, 62 2-cocyle group, 80 invariant toral, 102 coadjoint action, 428, 432, 438 coadjoint map, 421 coadjoint orbit, 350, 427, 431, 432, 438, 440, 443 coefficient function, 467

485

coherent pre-reflection system, 60 commutative space, 459 compact operator, 422 compact regularity, 246 compactly regular, 246 compatible grading, 77 complex analytic map, 252 complex projective space, 451 conditional expectation, 432, 445–447 cone, 252 conjugacy class, 352, 358 connected component of Re(R), 60 connected real part, 60 connected roots, 60 connection on a bundle gerbe, 345 on a line bundle, 344 consistent, 426 continuous vector bundle, 446 contractive bimodule, 371 contragredient Lie algebra, 169 contragredient Lie superalgebra, 170 core, 96 coroot system, 62 covering, 90 crossed homomorphism of Lie algebras, 304 crossed homomorphism of Lie groups, 306 crossed product algebra, 80 curving, 345 DI (root system of type D), 62 Dn (finite root system of type D), 63 D (degree derivations), 83 D-brane, 362 Dahmen’s Theorem, 261 Darboux Theorem, 421 ∆(E, H), 5 defect, 211 defining representation, 469 degree derivation, 83 Deligne cohomology, 344, 346, 347, 355 denominator identity, 151, 153, 154, 160–163, 165, 166 Der derivation algebra, 76 (Der A) (derivations of degree ), 79 derivation centroidal derivation, 82

486

Index

degree derivation, 83 diffeomorphism group, 243 direct limit, 252, 459 direct limit chart, 246 direct limit properties, 245 direct limit property, 268 direct limit representation, 461 direct limit topology, 254 direct sum of pre-reflection systems, 60 direct system, 252 discrete extended affine Lie algebra, 107 reflection system, 75 divisible root, 62 division-(R, Λ)-graded Lie algebra, 88 division-R-graded Lie algebra, 89 division-graded associative algebra, 79 division-root-graded Lie algebra, 88 Dixmier–Douady class, 340 dual reductive pair, 461 duality pairing, 423, 433, 435–437, 441

isotopy, 14–22 isotropic root, 6 locally, 108 nullity, 6 toral type, 110 type, 6 extended affine root system, 75 locally, 75 Saito’s, 75 extension central extension, 90 graded, 90 datum, 69 of locally finite root system, 73 locally finite root system, 72 canonical projection, 72 of pre-reflection systems, 69 universal central extension, 90 extension datum, 69 of a locally finite root system, 73 extremely amenable group, 380

Eij (matrix unit), 112 EALA, see extended affine Lie algebra, see extended affine Lie algebra EARS, see extended affine root system endomorphism of degree , 78 equivalent crossed homomorphisms, 306 equivariant structure on a bundle gerbe, 353, 360 Euler element, 225 ev (evaluation map), 83 even reflection, 179 evolution map, 247 exact invariant bilinear map, 290 exponential function, 247 exponential function (of a Lie group), 286 exponential map, 247 extended affine Lie algebra, 5, 104 centreless core, 6 construction, 10–13 core, 6 discrete, 107 E(L, D, τ ), 12 family, 13 bijectively isomorphic, 22 isomorphic, 6, 105

faithful, 448 faithful functional, 430, 432 fake monster algebra, 151, 160–163, 166 fgc, 77 fiber integral, 310 finite growth, 173 finite root system, 62 finite-dimensional superalgebras, 182 finite-rank operator, 434 flag geometries, 235 flux homomorphism, 329 formal degree, 461 formal degree renormalization, 466 Fq (the quantum torus associated to q), 85 fragmentation map, 258 function algebra, 470 Fundamental Formula, 227 fundamental highest weight, 464 Γ (centroidal grading group), 81 gl (general linear Lie algebra), 112 gl (general Lie algebra of row- and column-finite matrices), 119 gauge group Gau(P ), 282 Gelfand pair, 459 generalized foliation, 424

Index generalized Kac–Moody algebra, 151–156, 160–166 classification, 164–165, 167 construction, 160–161, 165–167 generalized reductive Lie algebra, 109 root system, 75 geometric representation theory, 444 GIF (graded invariant bilinear forms), 78 GNS representation, 444, 445, 447, 450 grCDer (graded centroidal derivations), 82 grCent (graded centroid), 79 grDer (graded derivation algebra), 79 grEnd (graded endomorphism algebra), 79 grSCDer (graded skew centroidal derivations), 82 graded algebra, 77 compatible grading, 77 crossed product algebra, 80 division-graded (associative), 79 division-graded (Lie), 89 full support, 78 graded-central, 79 graded-central-simple, 79 graded-isomorphic, 78, 89 graded-simple, 78 twisted, 80 homogeneous space, 77 isograded-isomorphic, 78, 89 Lie torus, 89 multiloop, 81 predivision-graded (associative), 79 predivision-graded (Lie), 89 quantum torus, 85 root-graded, 88, 89 division, 88 invariant, 89 predivision, 88, 89 support, 77 torus, 79 Lie, 89 graded-central algebra, 79 graded-central-simple algebra, 79 graded-isomorphic, 5, 78, 89 graded-simple, 78 grading subalgebra, 94

487

Grassmann manifold, 476 Grassmannian, 233 group algebra, 77 twisted, 80 group cohomology, 353 group manifold, 465 group of germs of diffeomorphisms, 244 GRRS, see generalized reductive root system, 75 HC1 (C) (first cyclic homology group of C), 92 Hamilton equations, 420 Hamiltonian vector field, 420–423, 427, 437 Harish-Chandra decompositions, 388 Heisenberg group, 320 Hermitian vector bundle, 446 highest vector, 195 highest weight, 195 highest weight module, 131, 132, 138, 141, 142 highest weight representation, 462 highest weight vector, 462 Hilbert–Schmidt ideal, 370, 434 Hilbert–Schmidt operator, 422, 436 holomorphic Hermitian vector bundle, 442 holomorphic reproducing kernel, 447 holonomy of a bundle gerbe, 348 homogenous space, 77 homomorphism, 427 homotopy group, 246, 273 IDer (inner derivations), 84 IDer (completed inner derivations), 119 IARA, see invariant affine reflection algebra, 100 imaginary root, 59, 181 immersion, 425, 429 indecomposable module, 128, 135, 139, 143 indecomposable pre-reflection system, 60 indivisible root, 62 infinite dimensional Lie group, 285 inflation map, 331 initial Lie subgroup, 274

488

Index

inner ideal, 238 integrable module, 194 integrable root, 64 integrable weight, 196, 209 integral diagonal, 372 strictly lower triangular, 372 strictly upper triangular, 372 integral pre-reflection system, 60 intrinsic subspace, 237 invariant affine reflection algebra, 100 isomorphic, 101 bilinear form, 61, 76 root-graded Lie algebra, 89 toral 2-cocycle, 102 invariant 2-cocycle, 11 invariant differential operator, 460 inverse of an invertible element, 88 invertible, 227 invertible element, 88 irreducible, 62 isograded-isomorphic, 5, 78, 89 ηgr , 5 isomorphic affine reflection Lie algebras, 96 extended affine Lie algebras, 105 invariant reflection algebras, 101 isotope, 90 isotopic, 90 Isotrivilaity, 46 isotropic root, 173 iterated loop algebra, 324 Jandl structure, 360 Jordan algebra, 227 Jordan pair, 225 Jordan torus, 23, 25 isotope, 25 A(u) , 25 isotopic, 25 Jordan triple system, 226 Jordan-H¨ older series, 134 Jordan-Lie functor, 226, 233 k-algebra, 76 K¨ ahler form, 442 K¨ ahler manifold, 451 Kac–Moody Lie superalgebra, 170, 179

M algebras, 169 Kantor pair, 226 k[Λ] (group algebra of Λ), 77 kt [Λ] (twisted group algebra of Λ), 80 Kr¨ amer classification, 468 k(S), 65 Lα (α-root space of L), 64 Lagrangian geometry, 235 LB-space, 258 leaf, 426 leaf of the characteristic distribution, 426 LEALA, see locally extended affine Lie algebra, 108 LEARS, see locally extended affine root system, 75 Leibniz property, 419 LF-space, 258 Lie algebra affine, 65 affine reflection, 96 discrete extended affine, 107 extended affine, 104 generalized reductive, 109 locally extended affine, 108 root-graded, 89 simply connected, 90 Steinberg, 116 toral type extended, 110 Lie algebra (filtered), 236 Lie algebra (graded), 224 Lie algebra (symmetric), 224 Lie group compact, 357 simply-connected, 349, 357 Lie subgroup, 428 Lie torus, 7 bi-isograded-isomorphic, 8 ϕr and ϕe , 8 bi-isomorphic, 8 centreless, 8 coordinatization theorems, 22–24 grading pair, 8 isotope, 13 L(s) , 13 isotopic, 14 multiloop, 41 root grading and external grading, 7

Index type, 7 untwisted, 9 Lie triple system, 224 Lie-(R, Λ)-torus, 89 Lie–Poisson bracket, 422 limit aligned representation, 461 limit irreducible representation, 458 line bundle, 340, 345 over a coadjoint orbit, 350 over the loop space, 355 linear Poisson map, 422 Littlewood-Richardson coefficients, 129 local map, 262 locally kω space, 256 locally convex direct limit, 248 locally convex direct limit topology, 258 locally exponential Lie group, 275, 286 locally extended affine Lie algebra, 108 affine root system, 75 locally finite Lie algebra, 127, 144, 264 locally finite root system, 62 classical, 62 classification, 62 coroot system, 62 divisible root, 62 extension datum, 73 indivisible root, 62 irreducible, 62 normalized form, 63 root basis, 62 Loewy length, 129, 133, 139, 142 logarithmic derivative, (left/right) δ l/r (f ), 286 long root, 65 loop algebra, 324 multiloop 81 twisted, 67 untwisted, 66 loop group, 281 loop space, 355 Mat (finitary matrices), 112 Mat (row- and column-finite matrices), 119 Mackey complete, 247 mean, 376 left invariant, 375 Meyberg’s Theorem, 227

489

mononormalizing, 370 moonshine for Conway’s group, 160–163 morphism, 422 between bundle gerbes, 342, 347 of affine reflection systems, 72 multiloop algebra, 49, 81, 324 multiplicity, 201 multiplicity-free, 459 natural representation, 127, 128, 130, 131, 137, 138, 141, 144, 145 negative root, 172 nest algebra, 373 non-isotropic root, 173 non-regular Lie group, 272 non-symmetrizable superalgebra, 186, 193 nondegenerate pre-reflection system, 60 nontrivial Boyd indices, 371 norm ideal, 369 normal, 432, 441, 448 normal functional, 432 normalized form, 63 normalized matrices, 172 null roots, 100 null-system, 234 nullity of an affine reflection Lie algebra, 97 of an affine reflection system, 72 octonion torus, 29 odd reflection, 176 orbit symplectic form, 429 psl (projective special linear Lie algebra), 114 pair geometry (linear, affine), 231 parabolic direct limit, 472 parabolic direct system, 472 partial section, 68 partition, 129 perfect algebra, 76 period homomorphism, 329 pointed reflection subspace, 70 Poisson bracket, 421 Poisson Lie group, 440 Poisson manifold, 419 Poisson map, 420 Poisson-Lie group, 437, 438

490

Index

polar decomposition, 448 positive definite reproducing kernel, 446 positive functional, 432 positive root, 172 pre-reflection system, 59 affine form, 74 coherent, 60 direct sum, 60 extension, 69 indecomposable, 60 integral, 60 invariant bilinear form, 61 morphism, 60 partial section, 68 nondegenerate, 60 quotient, 69 real part of, 60 connected, 60 connected component, 60 reduced, 60 root string, 61 unbroken, 61 strictly invariant bilinear form, 61 symmetric, 60 tame, 60 predivision-(R, Λ)-graded Lie algebra, 88 predivision-R-graded Lie algebra, 89 predivision-graded associative algebra, 79 predivision-root-graded Lie algebra, 88, 89 predual, 423, 428, 430, 433, 436–438 principal root, 176 product integral, 247 product map, 258 projective elementary Lie algebra, 28 pe3 (A), 28 projective geometry (generalized), 230 projective group (elementary), 230 quantum matrix, 85 quantum torus, 24, 85 quasi immersion, 425 quasisimple, 173 quotient pre-reflection system, 69 quotient root system, 72 (R, Λ)-graded Lie algebra, 88

graded-isomorphic, 89 isograded-isomorphic, 89 R-graded Lie algebra, 89 (pre)division, 89 Ran (anisotropic roots), 100 R char (coroot system of R), 62 R div (divisible roots of R), 62 R im (imaginary roots of R), 59 Rind (indivisible roots of R), 62 R (integrable roots of R), 64 Rre (real roots of R), 59 R× (the non-zero roots of R), 88 × (the non-zero indivisible roots of Rind R), 88 R0 (null roots), 100 Rad (radical of a bilinear form), 61 Re(R) (the real part of R), 60 real analytic map, 252 real part of a pre-reflection system, 60 real root, 59, 181 realization operator, 447–449, 451 reduced pre-reflection system, 60 reduced root system, 7 reductive Lie algebra, 369 reductive Lie group, 368 reflection space, 70 reflection subspace, 70 pointed, 70 symmetric, 70 reflection system, 59 affine, 72 associated to bilinear forms, 63 reflective root, 59 regular Cartan matrix, 173 regular contragredient superalgebra, 176 regular Lie group, 247, 285 regular root, 173 regular weight, 205, 209 reproducing kernel, 390, 446, 447 restricted Banach algebra, 436 restricted general linear group, 396 restricted Grassmannian, 441–443 restricted unitary algebra, 435, 436 restricted unitary group, 436, 438, 439 restriction map, 331 riemannian symmetric space, 462 (R, Λ)-graded Lie algebra isotope, 90

Index isotopic, 90 root, 172 anisotropic, 100 divisible, 62 divisible, 62 imaginary, 59 integrable, 64 long, 65 null, 100 real, 59 reflective, 59 short, 65 root basis, 62 root space decomposition, 64 root string, 61 unbroken, 61 root system, affine, 65 twisted, 67 untwisted, 65 extended affine, 75 finite, 62 generalized reductive, 75 locally finite, 62 quotient, 72 root-graded Lie algebra, 88, 89 grading subalgebra, 94 invariant, 89 root-reductive Lie algebra, 144 S lg (long roots), 65 Ssh (short roots), 65 SS(β, α) (α-root sting through β), 61 slI (A) (special linear Lie Lie algebra), 112 st (Steinberg Lie algebra), 116 sl2 -triple, 64 sα (reflection in α), 59 Saito’s extended affine root system, 75 Schatten ideal, 370, 434 Schur functor, 129 Schur Orthogonality, 466 Schur-Weyl Duality, 133, 135, 136, 146 SEARS, see Saito’s extended affine root system, 75 section, 446 self-adjoint element, 430 self-adjoint functional, 430 Serre’s relations, 180

491

set of weights, 194 short root, 65 Silva space, 260 simple algebra, 76 simple roots, 172 simply connected, 90 singular root, 173 small subgroup, 274 small subgroups, 248 smooth bump function, 268 smooth generalized distribution, 426 smooth vector bundle, 446 smoothly regular space, 268 Sobolev–Lie group, 244 socle filtration, 128, 129 special linear Lie algebra, 32 slr+1 (A), 32 special symplectic Lie algebra, 34 ssp2r (A, ι), 34 spherical function, 467 spin factor, 229 split Lie subgroup, 427 splitting Cartan subalgebra, 64 squeezed bundle, 314 standard base, 175 standard imbedding, 224 state, 237, 431, 432, 444, 445, 447, 448 Steinberg Lie algebra, 116 Stinespring dilation, 390 Stone–Weierstrass theorem, 471 strange twisted affine superalgebra, 185, 193 strict direct system, 254 strictly invariant bilinear form, 61 string, 178 strong symplectic form, 420, 429, 442, 443 strong symplectic leaf, 434, 438, 439 strong symplectic manifold, 421, 431 strongly C k -regular Lie group, 269 subalgebra ad-diagonalizable, 64 toral, 64 subordinated cohomology class, 302 subsystem, 60 support, 77 full, 78 symmetric norming function, 369, 434, 435

492

Index

symmetric pre-reflection system, 60 symmetric reflection subspace, 70 symmetric space, 232 symplectic leaf, 425, 426, 429, 434 tr (trace), 112 tame, 217 tame pre-reflection system, 60 tame toral pair, 96 tensor algebra, 128 tensor representation, 128, 130–132, 135–141, 144 test function group, 244 tier number, 65 Tits-Kantor-Koecher Lie algebra, 24 TKK(A), 24 topological group with Lie algebra, 275 toral pair, 96 centreless core of, 96 core of, 96 tame, 96 toral subalgebra, 64 toral type extended affine Lie algebra, 110 toroidal Lie algebra, 9, 314 Torsor, 46 torus associative, 79 Lie, 89 quantum, 85 trace class operator, 422, 436 tracial, 432, 448 transversal, 230, 235, 238 triangular decomposition, 172 trivialization of a bundle gerbe, 347 truncation diagonal, 371 strictly lower triangular, 372 strictly upper triangular, 372 twisted affine Kac–Moody Lie groups, 282 twisted affine root system, 67 twisted affine superalgebras, 184 twisted loop algebra, 67

two-cocycle, 437 types I and II, 174 typical weight, 200, 209 uce (universal central extension), 90 unbroken root string, 61 uniformly continuous, left, 375 right, 375 unitary group, 430, 436 unitary orbit, 430, 431, 433 universal ambit, 379 universal central extension, 90 untwisted affine root system, 65 untwisted loop algebra, 66 Verma module, 195 vertex algebra, 151, 153, 155, 157, 160, 162, 165, 166 lattice, 156, 160, 162 vertex operator algebra, 155–157, 165, 166 WZW model, 156 Virasoro group, 282, 308 W (R), 60 weak direct limit chart, 246 weak direct product, 244, 257 weak immersion, 425, 429–431, 434 weak K¨ ahler, 433–435 weak symplectic form, 420, 429, 438 weak symplectic leaf, 435, 439, 443 weak symplectic manifold, 421, 431, 433 weight module, 194 Weil representation, 158, 159, 164 well-filled chart, 272 Wengenroth’s Theorem, 265 Wess–Zumino–Witten model, 361 Weyl group, 60, 179 Weyl groupoid, 176 Yamasaki’s Theorem, 257 Young projector, 129 Z(β, α), 61 Z(.) (the centre of an algebra), 77, 78

Developments and Trends in Infinite-Dimensional Lie Theory (Progress in Mathematics, Volume 288)

Surveys in Number Theory (Developments in Mathematics)

Surveys in Number Theory (Developments in Mathematics)

Graph Theory in Paris (Trends in Mathematics)

Advances in Ring Theory (Trends in Mathematics)

Modules and Comodules (Trends in Mathematics) (Trends in Mathematics)

Naive Lie Theory: Undergraduate Texts in Mathematics

Naive Lie Theory (Undergraduate Texts in Mathematics)

Naive Lie Theory (Undergraduate Texts in Mathematics)

Positivity (Trends in Mathematics)

Developments and Retrospectives in Lie Theory: Algebraic Methods

Recent Developments in Lie Algebras, Groups and Representation Theory

Floer Memorial Volume (Progress in Mathematics)

New Trends and Developments in Automotive Industry

Dendroclimatology: Progress and Prospects (Developments in Paleoenvironmental Research, Volume 11)

Generalized Lie theory in mathematics, physics and beyond

Current Developments in Mathematics 1998

Numerical Semigroups (Developments in Mathematics)

Communications In Mathematical Physics - Volume 288

Number Theory: Tradition and Modernization (Developments in Mathematics)

Determinantal Ideals (Progress in Mathematics)

Morse Homology (Progress in Mathematics)

Hypercomplex Analysis (Trends in Mathematics)

Lie Theory: Unitary Representations and Compactifications of Symmetric Spaces (Progress in Mathematics)

Tree Lattices (Progress in Mathematics)

Ramsey Theory: Yesterday, Today, and Tomorrow (Progress in Mathematics, 285)

Trends in Optics. Research, Developments, Applications

Analysis and Mathematical Physics (Trends in Mathematics)

Adeles and Algebraic Groups (Progress in Mathematics)

Modules and Comodules (Trends in Mathematics)

Adeles and Algebraic Groups (Progress in Mathematics)

Developments and Trends in Infinite-Dimensional Lie Theory (Progress in Mathematics, Volume 288)

Surveys in Number Theory (Developments in Mathematics)

Surveys in Number Theory (Developments in Mathematics)

Graph Theory in Paris (Trends in Mathematics)

Advances in Ring Theory (Trends in Mathematics)

Modules and Comodules (Trends in Mathematics) (Trends in Mathematics)

Naive Lie Theory: Undergraduate Texts in Mathematics

Naive Lie Theory (Undergraduate Texts in Mathematics)

Naive Lie Theory (Undergraduate Texts in Mathematics)

Positivity (Trends in Mathematics)

Developments and Retrospectives in Lie Theory: Algebraic Methods

Recent Developments in Lie Algebras, Groups and Representation Theory

Floer Memorial Volume (Progress in Mathematics)

New Trends and Developments in Automotive Industry

Dendroclimatology: Progress and Prospects (Developments in Paleoenvironmental Research, Volume 11)

Generalized Lie theory in mathematics, physics and beyond

Current Developments in Mathematics 1998

Numerical Semigroups (Developments in Mathematics)

Communications In Mathematical Physics - Volume 288

Number Theory: Tradition and Modernization (Developments in Mathematics)

Determinantal Ideals (Progress in Mathematics)

Morse Homology (Progress in Mathematics)

Hypercomplex Analysis (Trends in Mathematics)

Lie Theory: Unitary Representations and Compactifications of Symmetric Spaces (Progress in Mathematics)

Tree Lattices (Progress in Mathematics)

Ramsey Theory: Yesterday, Today, and Tomorrow (Progress in Mathematics, 285)

Trends in Optics. Research, Developments, Applications

Analysis and Mathematical Physics (Trends in Mathematics)

Adeles and Algebraic Groups (Progress in Mathematics)

Modules and Comodules (Trends in Mathematics)

Adeles and Algebraic Groups (Progress in Mathematics)

Recommend Documents