of Algebras
SpringerVerlag
New York Berlin Heidelberg
London Paris Tokyo
i,. . PREFACE
Frederick M. Goodman Pierre de la Harpe
Department of Mathematics Section de Mathimatiques The paper [Jol] on subfactors of von Neumam factors has stimulated much research in
University of Iowa Universiti de Genkve
von Neumann algebras. Quite unexpectedly, it was discovered soon after the appearance of
Iowa City, Iowa 52240 CH 1211 Genkve 24
[Jol] that certain algebras which are used there for the analysis of subfactors could also be
USA Switzerland
used to define a new polynomial invariant for links [Jo~].The period of activity following
Vaughan F.R. Jones this discovery saw the creation of a number of related invariants as well as the successful
Department d Mathematics use of these invariants in knot theory. Furthermore, rece11t effort to understand the
University of California  Berkeley
fundamental nature of the the new link invariants has l& to connections with invariant
Berkeley, Cal~fomia94720
theory, statistical mechanics and quantum field theory. In turn the link invariants, the
USA
notion of a quantum group, and the quantum YangBaxter equation have had great impact
Mathematical Sciences Research Institute on the study of subfactors.
1000 Centennial Drive It is not yet the time to give a comprehe~siveaccount of these developments, and we
Berkeley, California 94720
make no attempt to do so here. Our subject is certain algebraic and von Neurnam
USA
algebraic topics closely related to the original paper [Jol]. Hbwever, in order to promote,
in a modest way, the contact between diverse fields of mathematics, we have tried to make
this work accessible to the broadest audience. Consequently, this book contains much
Mathematical Subject Classification (1980): 46L10, DSC50, 16A40 , elementary expositoG material.
We give here a brief preview of the book. Each of the four chapters has its own
Library of Congress CataloginginPublication Data
introduction, with a more thorough description of the contents.
Goodman, Frederick M.
Coxeter graphs and towers of algebras. Chapter 1 begins with a (slightly new guise of) a familiar combinatorial problem: to
(MathematicaI,Sciences Research Institute publications ; 14) classify finite matrices over the nonnegative integers which have Euclidean norm no
Bibliography: p. greater than 2. These are classified by the ubiquitous Cozeter graphs of type A, D, or E
1. Class field towers. 2. Coxeter graphs. (see [HHSV]for other occurrences of these graphs) and the set of possible nomu is
I. La Harpe, Pierre de. 11. Jones, Vaughan F.R.,
1952 . III. Title. IV. Series. (2) u {2ws7r/q : q 2 2).
QA247.G68 1989 512'.55 895991 The central theme of the book  the discussion of which begins in Chapter 2 is the
.
tower of algebras MOc M1 c  ,c Mk c ... determined by a pair MO C Ml of algebras
Printed on acidfree paper.
(with the same identity). The tower can be used to define various invariants of the pair,
including the index [M1:MO]
O 1989by SpringerVerlag New York Inc. . In Chapters 2 and 3, we study two cases of the tower construction in detail.
All rights reserved. This work may not be translated or copied in whde or in part withoutthe written
permission of the publisher (SpringerVerlag, 175 Fifth Avenue, New York, bJY 10010, USA), In Chapter 2, the algebras are finite direct sums of full matrix algebras over some field.
except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with
'
987654321
In Chapter 3, the algebras are finite von Neumann algebras with finite dimensional
centers. Somewhat surprisingly, the results of Chapter 2 essentially extend to this setting. There are several appendices.
But now a pair Eho c M1 is (partially) described by an inclusion matrix A with entries in Appendix I extends the cornputations of Chapter 1.
{2cosn/q : q 2 2) U {I : r 2 21, and pairs with index no greater than 4 are associated to Appendix ILa relates complex semi~imple algebras and finite dimensional
Coxeter graphs of arbitrary type, including types B,F,G,H,I. c*algebras: Appendix X b explains one appearance of the algebras AD,k in statistid
Finally,Chapter 4 is a M h e r analysis of pairs N c M of finite factors of finite index. mw. Appendix 1I.c is a further discussion of ADik for special values of jl.
There are two main themes. The first is the notion of a commuting square, due to Popa
Appendix 111is an exposition of Hecke subgroups in PSL2(R), and thus another famous
[Popl], and its use in approximating pairs of hyperfinite 111 factors simultaneously by
occurrence of the sequence (2cos(~/k))~,~.
finite dimensional von Neumann algebras. The second theme is the derived t o w s of a pair 
of II, factors, which is the c h i n of (necessarily finite dimensional) rdative cornmutants
It is a pleasure to record our gratitude to numerous fiends and colleagues for their
Mo' 17Mk in the tower. All the information in the derived tower can be encoded in a
generous help, including: ,
(possibly infinite) graph, the principal graph of the pair. When the index is less than 4, the R. Baldi, D. Bichsel, H. Dherete, M. Kewaire, A. Ocneanu, M. Pimsner, S. Popa, G.
graph is a Coxeter graph of type A, D, or E. Skandalis, C. Skau, R. Steinberg, V. Sunder, A. Valette, and H. Wenzl.
In Chapter 2, we also describe how a certain quotients of the Hecke algebra of type A
appear in the tower construction associated to a pair Mo c M1 of, say, finite dimensional Wegatefully acknowledgesupport from the MSRI in Berkeley, the IHES in Bures, the
semisimple algebras over the field C. For each choice of a positive faithful trace tr on / United States NSF, ttie IMA in Minneapolis, and our home institutions during our work on
MI, there is a unique trace preserving conditional expectation El : M1 ,MO, and it this project.
t w m out that Ma is naturally generated by M1 and El. Now if the trace tr satisfies
the socalled Markov condition, then the situation propagates up the tower, and each
algebra Mk+l is naturally generated by Mk and a conditional expectation
Ek : Mk IMkl, for aU k 2 1. Moreover the Ekl s are idempotents which satisfy the
"braiding" relations
pEiEjEi=Ei iflij(=l,and
_,
CONTENTS
Preface
<
J
Chapter 2. Towers of multimatrix algebras.
2.1. Introduction.
2.2. Cornmutant and bicommutant
2.3. Inclusion matrix and Bratteli diagram for inclusions of
multimatrix algebras.
2.4. The fundamental construction and towers for multimatrix algebras.
2.5. Traces.
2.6. Conditional expectations.
2.7. Markov traces on pairs of multimatrix algebras.
2.8. The algebras for generic P.
References. 281 This theorem is. proved in section 1.2. Let us now describe the classification of matrices
Index. 287
2 (' ,
'!?".,.
Chapter 1: Matrices over the natural numbers !j 1.1. Introduction
X E Matfin@) with llXll i 2. Obviously, we need only consider X E Matfin({0,1,2)). This theorem is proved in section 1.4, which also contains pictures of the graphs and
The first stgp is to encode the matrix as a bicolored labelled graph, some edges being tabla of matrices. The theorem is also essentidly proved in numerous sources, including,
marked with a sign CO. (The choice of co fits with the marking conventions for Coxeter for example, [ S d and [CGSS]. One point should be stressed: the oombinatorics
graphs, as will be explained below. We distinguish labelled graphs, in which each vertex underlying Theorem 1.1.2 is the same as that which enters into the classification of simple
has a welldefined name, from marked graph, in which some edges are decorated with Lie algebrai or of reflection groups of the comespondix types, but the theory of Lie
various signs.) More generally, a matrix X E Mat (R) is encoded in a labelled bicolored algebras is not needed, nor that of reflection groups. On the other hand, these theories
m,n immediately suggest both statement and proof of the next result.
graph f(X) which has m black v a t i m bl,...,bm and n white vertices wl,...,wn;
We set
there is no edge joining either two b's or two w's; there is an edge joining bi and w.
J
if and only if X i j # 0; W y an edge corresponding to XiSjf {0,1} is marked with some
sign conveying information about the value of X. In particular, an edge corresponding
1,j' so that K = [0,2] n K ( H ) by Theorem 1.1.1. To encode X E Mat (I(), we mark an edge
to X. . = 2 is marked with co and thus looks like o a . For example: m,n
14 , 1 with m if it corresponds to Xi . = 2, with q if it corresponds to X. . = 2 COSTfor some
J 1J q
integer q 2 4, and not at all if X. . = 2 cos $ = 1. For example:
W
As one may expect for any classification, the next step is to establish appropriate
notions of indecomposability and equivalence. We say that X is indecom~osableif f (x) Such a marked graph is known as a Coxetec a.
Theorem 1.1.2 may now be generalized
is connected; two matrices X and X' are pseudoeauivalent if f (x) and f (x') are as
isomorphic as (unlabeled) marked bicolored graphs. These notions will be discussed in
greater detail in section 1.3, which also contains a proposition reducing the classification to Theorem 1.1.3. The encoding above sets up a bijection between:
the indecomposable case. Now comes the solution to the initial problem. [i) Indecomposable matrices in Math(K) 4 nonn smaUer fhan 2, up to
pseudoequivalence, and
Theorem 1.1.2. The encoding described above set% q a bijection between:
(ii) Irreducible Cozeter graphs jom the' following list, together with a bicoloration, wp to
(i). Indecomposabk
 .matrices in Matfin({O,l)) of nonns smaller than 2, up to
isomorphism of bicolored graphs. The list is
pseudoequivalence, and
(ii) Irreducibk Cozeter graphs from the following list, together with a ~bicoloration,up to
n \.
isomorphism ofbicolored graphs. The list is
I
1
A! (1 2 21, DL(1 t 4), El (k6,7,8).
Also it sets up a bfjection between:
Also it sets up a bijection between (iii) Indecomposable matrices in Matfin@) of norm equal to 2, up to
(iii) Indecomposable matrices in Matfin({0,1,2)) of norm equal to 2, up to
pseudoequivalence, and
pseudoequivalence, and (iv) Irreducible Cozeter graphg from the following list, together with a bicoloration, t ~ p
(iv) Irreducible Cozeter graphs from the following list, together with a bicoloration, up to isomorphism of bicolored graphs. The list is
I
to isomorphism of bicolored graphs. The list is
A , e2 1 , BY) (e t 21, C P ) (I t 31,
4 Chapter 1: Matrices over the natural numbers 5 1.2.  Kronecker's theorem .. 5
I
result of J. Shearer [She]. For (vi), see Appendix I, after 1.3.6. What is known of E
. ,
 I
makes it look somehow similar to another set recently studied by [Smy]; see also Lehmer'e
problem, referred to below in remark (3) following Proposition 1.3.4, in Appendix I. I
I
I In other words, X is a (kmjm)th root of unity. #
;
1.2. Proof of Kronecker'a themem.
I ProDosition 1.2.2 Let be a n o w e m real number which' is a root of a manic
polyomial P E H[T]. If all the roots of P are real and in [2,2], then X = 2 cos(2n r)
We repeat firat the two results of [Kro], with their original proof.
for some rational number r.
6 *, ' . Chapter 1: Matrices over the natural numbers +"
w.Let L be the degree of P and set Lemma 1.2.4. Consider integers 4m,n 2 1 vrifi = m+n and a matriz
X E Matman(R). Set
Denote' by X = 2 cos el, 2 cos 02,' , 2 cos el the roots of P. Then Then
llxll = llxtll = IlYll = ~lx~xlll/~
= llxxtl11/2.
P(T) = n$T2 cos Oj)
lsjs w.For any 6 E kRn one has
Q(T) = n j T 2  2 T cos
li js llxt112= ( x t x t l o i llxtxlllld12s llxtllllxllll~12. (*I
(42)
max{IAdI : 1 i j s ~ = 2 c o9 s ~ Proof of Theorem 1.1.1. Let X be a finite matrix over Z with llXll < 2 and let
J Y E Mat@ be as above. As Y is symmetric, its eigenvalues A1,. , ,Ae are real and .
for some natural number q 2 3. llXll = IlYll = max{ I Al I ,,* .,1 All}. These eigenvalues being the roots of the characteristic
polynomial of Y, one has Y = 0 or IlYll = 2 cos for some q > 3 by corollary 1.2.3.
*f. By the previous proposition, there exists pl,. ..,pe, ql, ,qe E H, with
9
The last daim in the theorem follows from the example below. #
(p., q.) = 1, such that A. = 2 cos(2rp./q.). But 2 cos(2rp./q.) and 2 cos(27r/q.) are
J J J J J J J J
conjugate in the field Q(exp(iPrp./q.)), and as P has coefficients in H, the number l3xam~le1.2.5, Given an integer l 2 2, consider >
J J
2 cos(27r/qj) is also a root of P. In case q  2k.+l is odd, 2 cos(27rk.lq.) is also a root
j J J J
 of P, and 12 rm(27rkj/qj)l = 12 w ( r  7r/qj)l = 12 cos(dqj)l. Thus ,
9 = ={q' A')
For j = 1, ..,e, one checks that Y[

 A.t. with
q' = max{ 1 qj 1 : q is odd) j J J
1
q' = max{ I qj I : qj is even}.
Finally q i 3 because 2lcos 7rI = 2 and 2lcos;l =O. # As Y is symmetric one has
The last ingredient in the proof of Theorem 1.1.1 is a well known exercise in matrix
algebra:
8 Chapter 1: Matrices over the natural numbers 1.3. Decomposability and pseud~quivalence 9
One may rewrite the rows and columns of Y in the order E 6, is identified with a matrix in Mat,((O,l)). Matrices X1 . and X2 in Matm,,@)
2,4,...,1,1,3,...,11 if 1 iseven
are defined to be &w i v a l e s t if there exist CY E emand P E 6, with $= d l P ,
2,4,...,I1,1,3,,..,1 if I i s o d d , namely, if appropriate exchanges of rows and columns convert X1 to X2. A matrix is
irredundar& 'if none of its rows and columns is zero. An irredundant matrix
0 X
and obtain a matrix of the form [xt 0] with X E Matm,,(R) is if there exist integers m' ,m',nl ,mn 2 1 with
m = ml+m' and n = nl+n', as well as X' E Matm, ,+, (R) and X' E Matm.,n.(tR)
such that X is pseudmuivalent to
081P'. A matrix is indecom~osabla if it is
irredundant and not decomposable. For example, it is not difficult to check that
Remarks.
(1) The ideas of Kronecker in [Kro] have motivated much further research; see, for
example, [Boy] and [Rob].
(2) The matrix in the example above is "well known". Indeed, it could well be the These notion$ should be distinguished from those defined by Frobenius [Fro] for square
first matrix in the history of mathematics of which the eigenvalues have ever been matrices. Recall that Z E Matm@) is reducible if there q s t s permutation y € em
computed! This occmed in 1759, when Lagrange was studying a system of ordinary
differential equations which "approa~h~~ the equation for a vibrating string ([Lag], nos
with 7Z7' of the form
["'
(I 7,
Z. where ZZ. and Z'
m' ,m' > 1 with m' +m' = m. Matrices such as Z and $37I
are square blocs of siaes
are called eauivalent.
1420). We owe this observation to G.Wanner.
(3) The set of positive algebraic integers which arise as spectral radii of (not
or example, the matrix
necessarily.symmetric) integral matrices is considerably larger than the set qn). Indeed,
the former is dense in [I,@[, even if one considers only aperiodic matrices with entries in IN
[Lin]; but the latter is not, as shown by Theorem 1.1.1.
already considered is irreducible. (Frobenius uses "zerfallend" and llzerlegbar" for
'4 f
(4) It is h o r n that the degree of the algebraic integer 2 ms is d q ) , w h m p
"reducible," and no specific word for llequivalent.l')
is Euler's function [Le2]. It is easy to see that the estimate p 5 q/2 in Theorem 1.1.1is l e t X € Matm,n(R). Denote by xC
E Matm,({O,l}) the matrix defined by x:,~= 1
not sharp.
if X i j # 0 and X! . = 0 for' Xi J.= (I; the w p e m i p t c stands for Ucombinatorial". We
1J
1.3. Decomposabiity and paeudoquidenoe. have explained in section 1.1 how xC caabe encoded in a bicolored labelled graph f (xC)
.
containing m black vertices bl,. .,bm and n white vertices wl,. .,wn. Now set .
We consider first matrices in Mat (IS) for some m,n 2 1. Any permutation CY of 4 = m+n and wnsider the square matrices
m,n
{1,2,. .+,m} will be identified with the mbym matrix of the linear automorphism of
LRm mapping e, to edi1; .with this wnvention at = oil. Similarly any permutation
7 ',Fh
10 Chapter 1: Matrices ove~the natural numbers 5 1.3. ~&omposabilityand. pseudwquivalence.
(, 
and is Caned its *cv matrix in graph theory. Here is a list of easily established m i = mr , n'r = nr, and
observations.
X i pseudoequivalent to 5
(a) The following are equivalent:
(i) one of the matrices X, xC,Y, yCis irredundant;
for r = 1,. ..,a.
(ii) all these matrices are irredundant;
b f . For the existence part, consider the connected components f (xC), of f (xC),
(iii) the graph ~ ( x ~hasJ no isolated vertex.
.
where r = 1,. .,a. Denote by m, [respectively, q]the number of black [respectively,
(b) The following are equivalent:
(i) one of (hence both) the matrices X, xCis fndecomposable;
white] points in f (x')~. Permute the rows of xCand obtain a matrix X' in such a way
(ii) the graph I'(xC) is connected; that the black point bi belongs to f XI)^ if
(iii) one of (hence both) the matrices Y, yC is irreducible.
(c) The following are equivalent:
(i) X: and X: are pseudoequivalent;
(ii) f ($1 and f (xi)are isomorphic aa bicolored graphs. Similarly, permute the columns of X' and obtain a matrix X' E Mat ((0,l)) in such
m,n
(d) The hllowing are equivalent: a way that the white point w. belongs to f (x'), if
J
(i) Y: and Y; are equivalent; 7
(ii) I ' ( x ~ ) and r($) are isomorphic as graphs. nl+..,+nrl<jsnl+...Snr (r=l,...,a) .
..
Pro~osltlon1.3.1. For any irredundant matrix X E Mat (R), there azist an bteger The matrix X' has the desired form.
m,n
a z 1, partitions Uniqueness follows from the fact that f(xC) has well defined connected components.
#
Remarks.
(1) If X corresponds to the bicolored labelled graph f (x), then xt corresponds to
and irredundant indecomposable matrices Xr E Matm (R), for r = 1,. .,a, such that . '
a graph with colors exchanged. The example at the end of section 1.2 is a Jordan block,
r' r say
X is pseudoequivalent to the mat*
for m = 4. As f (x) is clearly isomorphic (as unlabeled bicolored graph) to the graph
with colors exchanged, a Jordan block is pseudwuivalent to its transpose. But
12 Chapter 1: Matrices over the natural numbers 3 1.4. Graphs with norm no larger than 2 13
black vertices bl,..,bm and m whitevertices bi,.w.,b&. There exists a line between
bi and bi, if and only if there exist a line in f ( ~ )between bi and w. as well as
J
are not peudoequivalent to each other, because the graph corresponding to the former has J
.
between w. and %, for some j E (1,. .,n). This shows equivalence of (i) and (ii); the
a white vertex of degree 3. same works'for (i) and (iv).
(2) Consider X,X1 E Matmln(R) and the cohesponding Y,Y1 E Mate (R), with Now a matrix such as X ~ X , which is positive both in the sense of PerronFrobenius
theory and in the sense of operator theory, is irreducible if and only if it is aperiodic. This
e= m+n. If X' is pseudoequivalent to X, say X' = &/then
I. Y' is equivalent to
r11.
follows from Theorem 8 in 3 XIII.5 of [Gan]. #
Y, because Yf = 7 ~ 7 " with , 7 = (' But one may have Yf equivalent to Y 
0
without X' being pseudoequivalent to X. The reader can check this with X and X'
the two 3by3 matrices of the previous remark, or with X E Mat4({0,1)) and
About (a), obserke that X = [::] is redundqt, with XXt imdundant and xtx
redundant.
X' E Matg,2({0,1)) corresponding to
Let r be a finite &aph with e vertices. The 8 of I' is (the
equivalence class of) a symmetric matrix Y = Yr E Mate ({0,1)), with Y. . = 1 if there
is an edge between vertices i and j and with Y. 0
1lJ
..
if not. The ~haracterlstlq
llj 
plvnomial of I' is that of Yr, and the s w t r a l @&Q or of I' is
For future reference, we also note the following fact. A square matrix X E Matm@+) From Proposition 1.3.1, it is clear that the following program will solve the problem of
.
. with nonnegative entries is irreducible if ad only if, for each ilj E (1,. ,m), there exists
classifying matrices X E Matfin({Oll)) satisfying ((X((< 2:
an integer p (depending on i and j) such that the (ij)@ entry of xP is strictly (1) classify finite connected graphs satisfying JJr)J
s 2; ,
positive (see [Gan], 3 XIII.1). It is a~eriodic("primitive" is also used) if there exists an (2) classify bicolorations on these.
integer p such that all entries of xP are strictly positive. Consider now an indecomposable matrix X E Matfin({0,1,2)) and assume that 2 is an
entry of X. Then one has either X = 2 (as 1by1 matrix) or llXll 2 51t2: this is clear
Lemma 1.3.2. from the definition of IlXll, and follows also from Lemma 2 in XII1.2 of [Gan].
(a) For X E Matmln(R), the following are eqdvalent: Consequently the program above will also solve our initial problem of classifying
(i) X is irredundant; X E Matfin({0,1,2)) with IlXll s 2.
(ii) X ~ XE Matrn@) and XXt E Matn(R) are irredundant.
(b) For X e"Matmln([R+), the following are equivalent: 1.4. Graphs with norma no larger than 2.
(i) X is indecomposable;
The elementary classification of the present section is based on two fundamental facts
(ii) XXt is irreducible;
of PerronFrobenius theory, for the proof of which we refer $0 [Gan].
(iii) XXt is aperiodic;
The first one is about ah irreducible square matrix Y E Matk([R+) with nonnegative
(iv) xtx is irreducible; j.
(v) X ~ Xis aperiodic. entries. A PenonFrob yBEtpI for Y is an eigenvector (ER: of Y with
nonnegative entries. The fact is that such a vector
m. Claim (a) is immediate from multiplication rules. (i) always exist;
For (b), if bl,. . ,bm,wl,. ..,wn are the vertices of f(x), then f (XXt) has m (ii) is unique up to multiplication by a positive scalar;
14 jrc"' Chapter 1: Matrim over the natural numbers  #+.:,
$1.4. Graphs with norm no larger than 2
I F~'\,
15
(iii) corresponds to an eigenvalue which is simple and which is the spectral radius of ..
i,j = 1, ,k) and Y' # Y. Then the fact is that the spectral radius of Y' is strictly'
Y. smaller than that of Y. In particular, this implies the following lemma.
In partic&ar, if I' is a finite connected graph, and Y € Matk({O,l)) is its adjacency
matrix, then Y is irreducible, although in general not aperiodic; a Perron Frobenius Lemma 1.4.2. Let I' be a frnite connected graph and let r: be a proper subgraph of
vector C for I' (by which we mean for Y) satisfies r; then llr' II < Ilrll.
Y t = IlI'll C.
As the vector space where Y acts has its natural basis indexed by the vertices of 1', it is
(By a subgaph of I', we mean one obtained by erasing some edges and some vertices
together with all edges emanating from these vertices; for example,
subgraph of ).
o is a 
8, I[
convenient to represent such a ( on the graph I', with the value of each coordinate of
wlitteh near the corresponding vertex of r. For example, if Y = then E !U2 The graphs list& in the next theorem are the ones above together with the following
(Now L is the number of vertices!):
1 1
is represented by .
We consider the following graphs, each given with a PerronFrobenius vector. (Note
that the subscript !. means that the graph has kl vertices (sic).)
p1

Given integers p,q,r with 2 < p 5 q 5 r, we consider also

q1
4 ' 0
.
0"'00
(p+q+r2 vertices). 
r1
The second fact from PerronFrobenius theory concern two matrices (b) IlI'll = 2 if and onlg if I' is one of
Y' ,Y E Matk(nt+) with Y irreducible, Y' s Y (this means (YY')ilj 2 0 for
AP) (e 2 21, D P ) (e I 41, EY) (e = 6,7,8).
16 . Chapter 1: Matrices over the natural numbers $ 1.4. Graphs &th norm no larger than 2 17
Although the theory above is independent of that of semisimple Lie algebras, root
M. Assume first that llI'll < 2. By Lemma 1.4.2, the graph I' contains no cycle systems are very close at hand. In fact, consider X E Mat ((0,l)) with llXll s 2, and
m*
(because ll~j'$ = 2 for l r 2), no point of degee 4 or more (because 1 1 ~ ~ =~ 2), 1 1 set Y = [Ot E MatL((0,l)) as above. Then 2Y is a real symmetric matrix with/
X 0
and at most one point of degree 3 (because llDf 111= 2 for L 2 5). Consequently I' is
positive eigenvalues; as &ch it can be written uniquely as 2Y = Z2, with Z E Mat@) a
either a segment, namely Ae for some L 2 1, or one of T2 e2 = DL for some l 2 4, or
1 , .
symmetric matrix with nonnegative eigenvalues. Let orl,, .,ort denote the oolumns (or
some graph T with q > 3. In this last case one cannot have p 2 3 (because
p,q,r rows) of Z. Then (Z2)i,j is the scalar product (a.10.); since it is 2 if i = j and
= T3,3,3 is of norm 2), nor r ? 6 (because E&') = TZIt6 is of norm 2). Thus J
Yij E {O,1) if i # j, the vectors ..,orl
5,. are of equal length 8 and their mutual
one is left with the list of (a).
angles are lr/2 or 243.
Assume now IlI'll s 2. The same argument shows that I' does not properly contain a
Assume moreover IIXII < 2, so that Y and Z are invertible, and thus 4 , .. ,?
cycle, a vertex of degee 4 together with its nearat neighbors, or one of the DV) for
linearly independent. Then, {al,.,.,eye) is what is called a reduced system of roots. It is
L 2 5. If, moreover, llrll = 2, it follows readily that I' is in the list of (b).
If I' = Al then llrll = 2 cos(lr/(kl)) by the example of section 1.2. For the other clearly irreducible if and only if X is indecomposable; in this case it is of one of the types
A,D,E because all roots have the same length. See [CGSS] for more along these lines.
value8 of h , see the literature,or Proposition 1.2.5 of Appendix I. For various meanings /
of the number hr, see [Stel]. # For the remainder of this section, we assume that the reader has some familiarity with
Coxeter graphs. Recall from section 1.1that
Define the bicoloratio~number of a graph I' to be the number of possible partitions of
its vertices in two nonempty sets, with no pair of adjacent vertices in one of the parts. It is
both elementary and well known that this number is 0 if and only if I' is a point or
contains an odd cycle. For the graphs of the previous theorem, one has and that a matrix X E Mat (K) is encoded as a Coxeter graph, say I', plus a
m,n
0 X
Table 1.4.4. The bicoloration numbers of the graphs listed in 1.4.3 are as follows: bicoloration. In fact, 2  [xt 0]
is twice the matrix of the cmonical bilin* form
A~: o A P ) with I 2 2 even: o associated to r (the form denoted by BM in [BLie], chap. V, $4, no 1). Any Coxeter
A! with l even: 1 A?) with ( 2 3 odd: 1
graph r corresponds to such a matrix X, and llrll = llXll < 2 [respectively [lXll= 21 if
A~ with l r 3 odd: 0 X
2 D P ) with t 4 even: 2
and only if 2  [ ]
xt 0
is positive ddnite [rap., positive semiddnite]. Consequently the
Dewith Lr 4: 2 Df) with l >5 odd: 1
classification of Theorem 1.1.3 is nothing but another phrasing of the classification of
El with l = 6,7,8: 2 E ~ I )with != 69,s: 2
irreducible Cozeter systems which are of fhe finite type [resp., the afbne type]. We refer to
[BLie] for further details. It is again true that an indecomposable matrix X E Mat (K)
This completes the program set up at the end of Section 1.3, and consequently the m,n
corresponding to a Coxeter graph I' has norm 2 cos(r/hr), where hr is either as in
proof of Theorem 1.1.2. The matrices listed in the following tables (Tables 1.4.5 and 1.4.6)
correspond to the graphs of Theorem 1.4.3; the one additional matrix [2] E Matl@) is Theorem 1.4.3 or as in the following list:
~a;:I},
,
More about Coxeter exponents can be found in section 2 of [Cox] and in exercise V.6.4 of
[BLie] .
Table 1.4.5. w
00
Type of Y .size . matrix Y bicolored graph Coxeter number exponents *Yfigures in table
,*
(i i i r)
w
E8
1 0 0 0
7 30 1,7,11,13,
17,19,23,29 yes
B
er.
B
E
N.B.: type Al doesnot appear; it would correspond to the empty matrix, identified to the (linear!) map Ro+ Ro. i
2
Table 1.4.6.
1byl 2 tcU no
P £2'
yes
Be o. ..o 4
1 2 2 L vertices h = 2t Type Ae (1 L 2). . Eigenvalue: 2 cos[lr/(t+ I)] ,
F~
G2
H3
o
6
60
0
c400
5
4 vertices h = 12
2 vertices h = 6
3 vertices h = 10
sin[./(!
*~*
+I)] ?in[2lr/(l +l)]
... 
sin[(!l)r/(t
. .
+I)] , sin[Llr/(t +I)]
H4 5
0000 4 vertices h = 30 Type Be (1L 2). Eigenvalue: 2cos[lr/24
0o
03  4
L2 3 kl vatices
J z 2 2 2 2 8
L2 2 t+l vertices Type EEg' Eigenvalue: 2cos[lr/12] = (13 +1)/p
,
5 vertices sin[lr/l2]

sin[2lr/l2] sin[3lr/12]
i s in[2~/12] sin[lr/l2]
3 vertices
b
I
111 Lemma 1.4.1 and in Table 1.4.7, we have indicated a PerronFrobenius eigenvector ;
for the Coxeter graphs of affine type. For Section 4.5 we will also need to compute the
PermnFrobenius vectors for the connected Coxeter graphs of finite type in the classes A,
D, and E. For completeness, we repeat the case of the gaphs At, already dealt with in
I
Section 1.2, and also give the results for the classes B, F, G, H, and I. Recall that in the ,
standard notation for Coxeter graphs of finite type, the subscript gives the number of
vertices. The details of the verifications are straightforward and are left to the reader. 1
I
,
!
F\
Chapter 1: Matrices over the natural numbers 5 1.4. Graphs with norm no larger than 2
Type F4. Eigenvalue: 2cos[lr/12] Pro~osition1.4.9, With the notation as above, the values of the square 11x1(2 of the
norna of the PerronFrobenius eigenvector for X ~ Xare as follows:
P
00
J3+1
=
C3+1
..
P
Type At: (e + 1)/4 Type Dl : (e 1)/4
Type E6: (3  $3)/2 Type E7: ca. 0,57999
Type Ga. Eigenvalue: 2cos[lr/6] = /
Type E8: ' Ca. 0,38502
101
Remarks.
(1) Let be a graph with .t vertices and let X1 5 ,.s Xe be the ordered sequence
of its eigenvalues. The s~ectrals ~ r e ds(r) of r is XeXI. Of course llrll < 2 implies
s(r) s 4; the converse happens to hold with finitely many exceptions which have been
classified by Petrovit [Pet].
Type Hq. Eigenvalue: 2cos(lr/30]
(2) It has been pointed out to us by D. Cetkovit and C. Godsil that it may also be
possible to classify'indecomposablematrices X E Matfin({l,O,l)) with [[XI[i 2. One can
0 X
mite 2  t [ ]
X 0
= ((41 7)IlriIl as before, so that the problem is equivalent to the
classification of irreducible sets of vectors {al,. ',ae} in Re, all of the same length and
.
Type 12(p). Eigenvalue: 2cos[lr/p] with mutual angles in {lr/3,x/2,2lr/3}. The possible sets of lines spanned by such sets of
(PI vectors are classified in [CGSS]. Once this is worked out, the next cases would be
1 1
X E Matfin(Ku(K)) with l\Xll < 2 as well as X E Matfin@) with IIXII > 2 but IlXll close
to 2.
(3) The subject of spectra of graphs has been extensively investigated. We refer to the
excellent early monograph by Biggs [Big], to more recent books by "CGetkovi~,Doob,
Gutman, Sachs, and TorgaSev [CDS] and [CDGT], as well as to the reviews [CD], [GHM]
and [Schl].
24 Chapter 1: Matrices over the natural numbers 3 1.5. Norms of graphs and integral matrices 25
1.5. The set E of norms of graphs and integral matrim. Ilrpef. It is obvious from considering 1by1 matrices that 4 s ) r U(K(S)). To show
the theorem, set T = 4 s ) and let X E Matm,,(T); one has to find some Z E Matfin(S)
The assignment to a subset S of !It of the set K(S) of norms of matrices on S has
interesting properties. Observe that, obviously, with IIZII = IIXII.
. For any pair (i,j) with 1 i i i m and 1 i j j n, we choose integer p.
1J
.2 1 and a
s na+ c 4 s ) symmetric matrix Y. of size p. over S with IIY. .I1 = Xi,f Let p be the product of
l,j 1,j 1 , ~
..
tion Y = Yl++ +Yk with Y. symmetric matrices in Mate (So) for j = l,, ,k. Set
J
... ,
M. For any Y E Matn(B), set
Let t E W+e be a ParonFrobenius vector for Y and define 6' = (t,t,. ..,() E R+kt.
Then X' t' = I(YIIt', so that t' is a PenonFrobenius vector for X' . Consequently
Ilx' 11 = IlXll. Ib particular one has
Now we particuladze to Y = ztz E Matn(B). For any integer k t 1, the entries of Yk Remarks about E = 1/(f0,1)1
are positive sums of products of entries of zt and Z; it follows that (1) Given s,t E E, one may look for an explicit graph with spectral radius s+t
!
(respectively at, and hs+(s2+4kt 2)1121). Some solutions can be found in [Schl]
r III(X~X)~II.
III(Z~Z)'III (respectively [We], and Theorem 2.13 of [CDS]).
Consequently,
(2) Say s E E is irreducible if a # 0,l and if s = sls2 with sl,s2 E E implies sl = 1
or s2 = 1. Any number in E can be factored as a product of finitely many irreducible.
Are there only finitely many factorizations? (The answer is obviously yes for s r 4.) If
r n 1 4 ~ ( ~ t ~ ) q r~ nf l// k l l ( ~ t ~ ) k l l l l=k nl/k11x112
1 yes, does the number of factorizations relate simply to the minimal t E I for which there
for all k 2 1. The lemma follows. #
I
exists X E Matl (I) 4 t h s = llXll?
(3) Is it true that K(M) = flu)?
One may deduce from Theorem 1.1.4 numerous properties of the set K(S). The
following is a sample.
Corollarv 1.5.2. Let s,t E 4 s ) . The following numbers are also in K(S):
f
a+t, (s2+t2)'l2, $s+(s2+at2)1/2) for evety L E I.
s+sl, s2(s21)1/2
are in the derived set of 4 s ) .
&&. The first claim is a consequence of the following equalities for Perron
Frobenius eigenvectors:
1i i] 1
( = ( ~ ~ + t ~ ) ' with
/ ~ < (=
restrict our attention to semhimple algebras for which the simple components are central
CHAPTER 2 ,
(have center equal to K). In fact we can even restrict attention to multimatrig
TOW of multiniatrix algebra5 I
1 over K, semisimple algebras whose simple components are isomorphic to matrix algebras
over the ground field K Note that if K is algebraically closed, then every semi%imple
2.1. Introduction. I
I
Kalgebra is a multimatrix algebra, since K has no proper finite dimensional division
algebras. We will call an algebra which is isomorphic to Mat (K) for some p > 1 a
The first purpose of this chapter is to study inclusions of one finite dimensional
semisimple algebra in another. Following [Jo~],we introduce a realvalued invariant,
called w,for a pair 16 N c M of axbitrary not necessarily semisimple or even
i
I
I
P
factor. Some authors refer to multimatrix algebras as "split semisimple algebras".
The reason why it sufGces to study multimatrix algebras is that index is stable under
finite dimensional  algebras over a field K, as follows: ! &ggg Qf &f m d field, Let K be a perfect field and E an extension field; for any
First, the fundamental construction associates to N c M the pair M c L where [ Kalgebra M, let ME denote M % E, an algebra over E. If 1E N c M is a pair of
L = ~ n d i ( is ~ the
) algebra of endomorphisms of M viewed as a right Nmodule; M is
/I finite dimensional Kalgebras, then [M:N] = [ M ~ : N ~(Proposition
] 2.4.4). If M and N
are in addition semisimple over K, then it is possible to choose E so that ME and NE
identified with a subalgebra of L, each x E M being identified with the left multiplication
operator (y w xy) E L. Second, the
sequence
induced by the pair N c M is the nested
I E M ~ = N C M ~ = M CCMkCMkSIC"
s e e
i i
\
are multimatrix algebras ov& E. Taking E to be an algebraic closure of K will do, but
one can also accomplish this with a finite dimensional field extension.
We now come to the definition of the index m a t r i ~ A : for a pair of semisimpb
/' Ii Kalgebras 1E N c M. First for a pair of 1E N c M, the reader can easily check
of Kalgebras, where Mk c Mk+l is obtained from Mkl Mk by the fundamental (after looking at Section 2.2) that the index [M:N] is just the ratio of dimensions,
C
II
construction (k 2 1). Third, the & rk(MklMO) of Mk over Mo is defined to be the
smallest possible number of generators of Mk viewed as a right Momodule (this rank lies
I
.
. ;
I
in DI U {a)). And finally the index of NhM is the growth rate i'; In this case A: is the lby1 matrix withaole entry [M:N]'/~. Next mnsider a pair of
Two comments on this definition: First, we could exchange the words left and right and
i' that M = @ Mpi and each Mpi is a factor. Similarly let
i=l
{q. : 1 i j b n) be the
J
minimal central idempotents in N. For each pair i j (1 5 i 5 m, 1 5 j 5 n) set
obtain a rank and an index "from the other side"; but we shall not study this variation I
'here. Secondly, a more interesting variation comes from using tensor products M.lrjp.q.Mp.q.
1J 1 J and N.1,J.=p.q.Npiqj.
1 J
M % M aN .
(a Torlike idea) instead of endomorphisms .
,~ n d ; ( E n d i ( ~ ) ) (an
Extlike idea); we refer to [Jo4] for this. One could check that these variations give the , Since pi is central in M, the product p.q. is an idempotent in Mi. If piqj # 0 then
1 J
same index for semiaimple pairs, but more general examples may have several indices. M. . is a factor (Proposition 2.2.3);
1J
and, since the map x n pixpi from Nq. to N. is a '
J 1,j
The conination of this subject with Chapter 1 is this: For inclusions 1E N c M of
nonzero homomorphism, N. is a h a factor with identity element p.q.. Define A: to
semiaimple algebras over a perfect field, the index turns out to be the square of the norm lj 1J
of a certain matrix of natural numbers AN M associated with the pair of algebras. Thu's the be the mbyn matrix with entries
LI
results of Chapter 1 yield restrictions on the possible values of the index. 1 if piqj = 0, and
Recall that a semisimple algebra over a field K is the direct sum of its minimal two Xij = 0
sided ideals, and each of these is isomorphic to a matrix algebra Mat (A) for some p 1 .= [M. .:N. .l1I2 if piqj j 0.
a 1 +,I 191 U
and some Kdivision algebra A. However, for studying index, we can for the most part
[Mi,fNi,j] is a square integer, so A , is a natural number; in fact
restrict our attention to semisimple algebras for which the simple components are central
CHAPTER 2 (have center equal to K). In fact we can even restrict attention to multimatrix
Towers of multimatrix algebras over K, semisimple algebras whose simple components are isomorphic to matrix algebras
o m the ground field K. Note that if K is algebraically closed, then every semisimple
2.1. Introduction. Kalgebra is a multimatrix algebra, since K has no proper finite dimensional division
f
algebras. We will call an algebra which is isomorphic to Matu(K) for some p 1 1 a
The first purpose of this chapter is to study inclusions of one finite dimensional
semisimple algebra in another. Following [Jo~],we introduce a realvalued invariant,
m. Some authors refer to multimatrix algebras as "split semisimple algebras".
The reason why it suffices to study multimatrix algebras is that index is stable under
calledm, for a pair 1 E N c M of arbitrary not necessarily semisimple or even
pound f i e . Let K be a perfect field and E an extension field; for any
finite dimensional algebras over a field K, as follows: &ggg pf
Kalgebra M, let M' denote M % E, an algebra over E. If 1E N c M is a pair of
First, the fundamental construction associates to N c M the pair M c L where
L = End;(~) is the algebra of endomorphisms of M viewed as a right Nmodule; M is finite dimensional Kalgebras, then [M:N] = [ME:N E] (Proposition 2.4.4). If M and N
are in addition semisimple over K, then it is possible to choose E so that ME and NE
identified with a subalgebra of L, each x E M being identified with the left mdtiplication
are multimatrix algebras over E. Taking E to be an algebraic closure of K will do, but
operator (y n xy) E L. Second, the tower induced by the pair N c M is the nested
one can also accomplish this with a finite dimensional field extension.
sequence
We now come to the definition of the matrix A: for a pair of semisimple
~EM~=NCM~=MC cMkCMkSIC
.. *" '
Kalgebras 1E N c M, First for a pair of a 1E N c M, the reader can easily check
(after looking at Section 2.2) that the index [M:N] is just the ratio of dimensions,
of Kalgebra, where Mk c Mk+l is obtained from MkVl C Mk by the $damental
construction (k 1 1). Third, the & rk(MklMo) of Mk over Mo is defined to be the
smallest possible number of generators of Mk viewed as a right Momodule (this rank lies
in M U {w}). And finally the index of N&M is the growth rate In this ease A: is the 1by1 matrix with sole entry [M:N]'~. Next consider a pair of
multimatrix &g&ug. Let {pi :1i i i m} be the minimal central idempotents in M, so
that M = @ Mq and each Mpi is a factor. Simila~lylet {qj : 1i j i 4 be the
i=l
minimal central i d e m p o t ~ t in
s N. For each pair i,j (1 i i 5 m, 1 i j i n) set
Two comments on this definition: First, we could exchange the words left and right and
. obtain a rank and an index "from the other side"; but we shall not study this variation
here. Secondly, a more interesting variation comes from using tensor products Me  p.q.Mp.q. and Ni,j = p.q.Npiqj.
1, J 1 3 1 J
M M eN .. (a Torlike idea) instead of endomorphisma ..
E n d L ( ~ n d i ( ~(an
))
Since pi is central in M, the product p.q. is an idempotent in Mi. If piqj # 0 then
Extlike idea); we refer to [Jo4] for this. One could check that these variations give the 1 J
same index for semisimple pairs, but more general examples may have several indices. Mi,j is a factor (Proposition 2.2.3); and, since the map x n p.xp. from Nq. to N. . is a
1 1 J 1,J
The connection of this subject with Chapter 1 is this: For inclusions 1E N C M of nonzuo homomorphism, N. is also a factor with identity element piqj. Define A: to
semisimple algebras over a perfect field, the index turns out to be the square of the norm , 1,j
M be the mbyn matrix with entries
of a certain matrix of natural numbers AN associated with the pair of algebras. Thus the
results of Chapter 1 yield restrictions on the possible values of the index.
Xilj = 0 if p.q = 0, and
$Recallthat a semisimple algebra over a field K is the direct sum of its minimal two 1 J
1.. = [ M .:N.
~ .11/2 if p.q. / 0.
sided ideals, and each of these is isomorphic to a matrix algebra Matp(A) for some p 1 1 W J 171 I J
and some Kdivision algebra A. However, for studying index, we can for the most part
[Mi .:Ni .] is a square integer, so Xij is a natural number; in fact
>J J
28 M. p Nij @ MatX. .(K), by Proposition 2.2.2. As the set of pi's is well defined by M
14
1,J
Chapter 2: Towers of multimatrix algebras 3 2.1. Introduction
and the set of q.'s by N, the inclusion matrix is well aefined by the pair N C M up to
J
pseudoequivalence. Obviously one has A pseudc+equivalent to A: for any
9 (N)
automorphism 9 of M. (See also the discussion following Corollary 2.3.2.)
Next consider an arbitrary semi~implepair 1E N c M. The obvious thing would be are immediately determined by that of N c M. These towers have a rich structure, the
to define ANM just as before, using the simp19 components of M and N, but t p s does further study of which requires the introductionof traces.
A on M is a linear map tr: M ,K such that tr(xy) = tr(yx) for x,y E M. It
not suit our purpose. Instead let E be an algebraic closure of K and set
is faithful if the bilinear form (x,y) H tr(xy) is nondegenerate. If K is given as an
extension of the real field R, a trace tr is positive if tr(e) > 0 for any idempotent e in
m
M. A trace tr an M = e Mat (K) is completely descfibed by the rn =
i=1 r%
(tr(fi))lSiSm where fi is a minimal idempotent in the factor Mat (K).
Again A M
N is well defined up to pseudoequivalence by the pair N c M, as the set of 4
factors in NE u d ME are determined by N c M. (In case the simple components of M Consider a multimatrix pair N C M and assume that there exists a faithful trace tr
and N are central, we would in fact obtain the same result without extending the ground on M with faithful restriction to N. (This is always possible if K is of characteristic
field: thus an alternative definition of !A is this: let E be any field extension of K such zero.) Then there exiats a unique Klinear map E :M + N, called a conditional
expectation from M onto N, such that .
that the simple components of NE and ME are central. Then :A = A I.)
N ,
E(Y)= Y for all y E N
For pairs of multimatrix algebras, I E N c M, the index matrix, togethei with the
dimensions of the minimal ideals of N and M, determine the inclusion up to an inner E(y1xy2) = y1E(x)y2 for all x E M and ~ 1 E N~ ~ 2
automorphism of M (Proposition 2.3.3). This is not true for arbitrary semiBimplepairs. tr(E(x)) = tr(x) for all x E M.
We will also refer to the index matrix A! as the inclusion a, when N c M is a pair
of multimatrix algebras. , [ One has of course E E L = ~nd&(M),and we will show that L is generated as a vector
For a semi~implepair, the index is related to the i&l&ion matrix in the following ,space by the elements xlEx2 for xl+ E M; in short: L = (M,E).
simple fashion (see Section 2.4):

Although any multimatrix pair N c,M generates a tower by iterating the
fundamental constiction, traces and conditional expectations in general do not propagate
Theorem 2.1.1. Let N c M be a semisimple pair. Then the indez of N in M is up the tower. That is why we single out a special class of traces as follows. Given 8 E K*
given by and N c M as ahve, define a Markov pf modulus 8 on N c M to be a faithful
trace tr on M with faithful restriction to N for which there exists a (necessarily unique)
trace Tr on L = ~ n d & ( such
~ ) that
To know whether or not such traces exists on a pair N c M, one has again to consider the
Corollary 2.1.2. Let N c M be a pair of semisimple algebras. Then either
inclusion matrix A, and its natural action (here from the right) on vectors with .
[M:N] = 4 cos2(~r/q) for some integer q 2 3, or [M:w > 4. coordinates in K (after reduction of A modulo the characteristicof K).
Theorem 2.1.3. Let N c M be a multimatriz pair with inclusion mat& A and let conditional expectation Ek :Mk + Mkl, which can also be viewed as an element of
P E K* Mk+l. We denote by Atr,k(MocM1) the subalgebra of Mk generated by the unit and
(I) ~ h e de&ts a Markov trace tr of modolw B on N C M if and only if there
ezists a row vector $ E K~ with
El,. ..,Ek1.
I = &,
a 1 1 coordinates of
a I 1 coordinates of
k distinct from 0, and
k~ distinct from 0.
I
Theorem 2.1.7. (i) For any
We now return to the tower , and the natural morphism AP,k + AP,k+l is an injection.
M0=NCM1=Mc cM~cM~+~C ... (ii) If /3 is generic then ABBk is a multimatriz algebra isomorphic to
(K), where $1 = 1 and
c.
where IA is a factor for each X E lk; we refer to Section 2.10 and 2.11 for a more precise
Theorem 2.1.8. Let Mo c M1 and let tr and /3 be as in Theorem 2.1.6,
description.' That there is a relationship between Hecke algebras and the algebras
(4 If /3 is generic, the map AAk 4
J
+.
Mk defined by e. * Ej for j = 1,. ,k1 ,is an
appearing in towers is clear from the presentation of Ablk with eifs and from that of
isomorphism onto Atrlk(M0cM1). Moreover this isomorphism is compatible with the trace
Hqlk with eil 8. More precisely:
trk on and the Markov trace trk on Mk.
(ig Assume K is given as an eztension of W, and that the Markov trace tr is positive Theorem 2.1.9. Let q E K* be a number which is not a root of 1 (with q = 1 allowed
(so that p = [Ml:MO]). Then Atr,k(Mo~M1) is isomorphic to a certain quotient B
Plk when char(K) = 0), and set /3 = 2 + q + ql. Consider an integer k 2 1. Then Am is
AAk which is ezplieitly described in Section 2.9.
isomorphic to the quotient of the Hecke algebra H by the relation
q1k
The braiding relations, which appear above first in Theorem 2.1.6, suggest a strong
connection with Artin's braid groups and with Iwahori' s Hecke algebras. This observation
haa constituted a breakthrough in the study of knots in !R3; see [Jo~],[Jo~],[Fre], and
[HKwl. and one has
For any q E @ and for any integer k 2 1, define the HBELB
HqYk to be
Kalgebra with unit presented by generators gl, ,gk4 and by relations
2
where 4 the subset of lkconsisting of those partitions with Young diagrams having at
gi = (Q114 + q i = l,...,kl
most 2 rows.
gigi+l& = gi+lgi&+l i = 1,. .,k2
gigj = gjq if (ij ( 2 2 i,j= l,...,kl The relation (S) was pointed out to us by Steinberg. In terms of the generators
el,. . ,ekl of Hqlk introduced above, it may also be written as
gi+l
q+l and /3 = 2 + q + ql, and check that
Assuming q # 1, one may also set ei =
Hqlk has a presentation with generators el, .,ekql and with relations
Our exposition is organized as follows.
e? = e.
1 1
.
i = 1,. .,k1 Section 2.2 collects preliminary material on pmmutantt. Section 2.3 shows how to
! ,
qei+,q
1
 /3 ei = ei+leiei+l  /3'ei+, i = 1,.. ,k2 define the index matrix or inclusion matrix of a multimatrix pair N c M, and how to
encode the relevant information in a Bratteli diagram. Concerning chains of multimatrix
algebras (a natural generalization of pairs), we have added an exposition of the m t h m o a
due to Ocneanu and Sunder. Theorem 2.1.1 and Corollary 2.1.2 about the of a pair
Then HQsk is an algebra of dimension k! over K, and the natural morphism N C M are proved in Section 2.4.
Hq,k + Hq,k+l is an injection. Section 2.5 concerns in general, Section 2.6 pnditiond mctationg, and Section
2.7 in particular. Theorem 2.1.3 to 2.1.6 are proved in Section 2.7.
If q is not a root of 1 (with q = 1 allowed if the characteristic of K is zero), then
Section 2.8 is about the algebras Ap,k, with emphasis on generic /3, and Section 2.9
Hqlk is a multimatrix algebra with factors in bijection with the set 'Pk of partitions of
k. We may thus write
discusses nongeneric /3. Section 2.10 is a leisurely digression introducing Recke m,
and the final section ahows how Ahk is a quotient of the appropriate Hecke algebra.
Chapter 2: Towers of multimatrix algebras 3 2.2? Commutaht i d bicommutant
40
T <, , ,'
If K is our favorite field of p m ~ l e xnumb=, multi matrix (or semiaimple) algebras Matd(En%(w)), and En%(W) e Matd(K) via these iaomorphiams. For any
"are the same" as &$.& gimensiona Q . This is of course wellknown, and subalgebra A of EndK(W), the algebra A 4( Matd(K) is thus identified with Matd(A),
made precise in Appendix II.a. If follows that the present Chapter 2 may be looked at as a
study of certain c*aJgebras which are AF, namely approximately finite; see [Eq. In aqd A @ 1 with the set of diagonal matrices
Appendix II.b we report briefly on the appearance of the algebras lo in statistical a
mechanics. Appendix I1.c contains additional material on the algebras Ap,k
nongeneric p.
or with the set of diagonal endomorphisms
Let F be an dgebra over a field K. (By an algebra we will always mean a;n algebra
The following unassuming lemma is the basis of both the Jacobson density theorem
with an identity element I.) The ~ommutantCF(S) of a subset S of F is the algebra of
([BA8], p.39) and of von Neumann's bicommutant theorem ({Talc], p.74), two cornerstones
all elements of F which commute with each element of S. One obviously has of noncommutative algebra.
S c CFCF(S) for any subset S, and the reader can check that CFCF(CF(S)) = CF(S)
for any S. If A is a subalgebra of F, then CFCF(A) is an algebra containing A, a ~e& 2.2.1. Let W be a vector space over # and let A be a subalgebra o f
sort of closure of A in F. EndK(W) with 1E A. Write A' for CEnd (A).
", K(
Let W be any Kvector space (not necessarily finite dimensional). We remark that
for each d E OI there are canonical dgebra isomorphisms
(i) T ~ commvtant
C of A B I in ~nd,,p,vd) ir A' % Matd(K).
(b) The commvtant of A t Matd(K) in ~ n d ~ is( A'~ @ ~1. )

' ~ r m (a)
t Since ( a o l ) ( X x i j e eij) = I ? , j @ eij, and ( C x i j e i , j ) ( l )
@
(For atly algebra A, Matd(A) denotes the algebra of dbyd matrices over A.) For = Z I , j a e eij, it follows that C x i j @ e i j lies in the cornmutant of A @ 1 if and O ~ Y
1 r i I d let ri :wd .W be the projection on the ith component, and for 1 I j I d let if x. . E A' for all (ij).
1,J
e. : W ,Wd. be the injection such that ri o e. = 41 on W for all i. Given (b) If x = C x i j @ e i j commutes. with A % Matd@), then in plutidar it
J J
x E ~ n q w d )d h r matrix [ 3 , j E Matd(En%w)) by 3,j= ri 0 x 0 ei. One can 1,J
commutes with the matrix units 1 @ e for all (p,~),
check that the map x n [xi .] is an algebra homomorphism, and that it ha. an inverse, A"
,J
which t a k a a matrix [xi,j to x r i o 3,jo 7 E ~ n d l ( ( d ) . This establishes the first
i,j
isomorphism. G t {ei,j 1I i,j r d) denote the standard system of matrix units in
Matd@); the matrix s . has a 1 in the (i,j) position and zeros elsewhere. Then the .
It follows that x = 0 for j # u and x. = 0 for i # p, and x = x
1J ",J "," P,P' Thus
x= Cy @ ei,i = y @ 1 for some y E EnddW). Since in addition x commutes with
i
A @ 1, it follows that y E A'. On the other hand any element of the form y @ 1,y E A',
COmm~te~ with 4 % Matd(K). #
basic facts abput the representation theory of F are: This index is the square of an integer by Lemma 2.2.2. In case M "itsel'f is a subfactor of a
factor F, the same leinma implies
(i) Any finite dimensional Fmodule is completely reducible, and
(ii) any irreducible Fmodule is equivalent to the standard module V.
(By a module over a Kalgebra A, we will always mean a module; that is the
. which is a preview of Proposition 2.3.5 below.
identity of A acts aa the identity on the module. Thus an Amodule is also a Kvector
space.) If W is any finite dimensional Fmodule, it follows from (i) and (ii) that Pro~osition2.2.3, Let M be a multimatrix subalgebra of a factor F with 1 E M C F.
dimK(V) divides %(W), say dimK(W) t dim@) = dl and W is equivalent to the
d
Fmodule V with the diagonal action
I (a) CF(M) is a multimatrix algebra, and has the same minimal central idempotents
..
a : En$(W) 4 M c En%(V)
Remark@.
(1) If M is a subalgebra of a factor F and if M is not semi%imple, one may have
M is
M C (C (M)). An example is given by the algebin M of matrim of the form
# F F ]:
[M :N]= (dimK~)(diXilK~)'.
I in the factor F = Mat2(K), for which CF(CF(M)) = F.
Chapter 2: Towers of multimatrix algebras,p8% 3 2.2. Commutant and bicommutant ,"""4
41
1 i
:" .d
(2) The biccommutant theorem is valid for all finite dimensional s&%imple algebras;
see [Weyl], Theorem 3.5 B. Brauer [Br] and Weyl (op.cit.) stressed the importance of the using CF(CF(M)) = M. Now take cornmutants and apply Lemma 2.2.2 in qFq to obtain
theorem for invariant theory. The ultimate form of the theorem is Jacobson's density
the conclusion. #
theorem ([Jacl], [Jac2]). It was noted by Bourbaki ([BA8], p. 39) that Jacobson's
theorem follows fkom von Neumann's proof of his bicornmutant theorem for operator .
algebras ([vN] or [Tak]). Prowsition 2.2.5 Let M be a multimath subalgebra of a factor F with 1E M C F
(3) Let V be a vector space such that F = EnUV). The usual notation in and let q E M U CF(M) be a nonzero idempotent. Then
functional analysis for CF(M) is M'. In some books of algebra, V viewed as a (a) qMq is a multimatriz algebra.
(b) CqFq(~Md'qCF(M)q.
CF(M)module is called the countermodule of the Mmodule V.
EIQQf. The notation being as in the proof of Proposition 2.2.3, set % = piq and
2.2.4. Let M be a subfactor of a factor F with 1 E M c F and let
q E M U CF(M) be a nonzero idempotent. Then
(a) q ~ isq a factos;
(b) CqFq(~Md= qCF(M)q
t
I
observe that q =
E
i=l
qi. One obviously has
4CF(M)q =
m
@
i=l
qiCF(M)$ =
m
@
i=l
~ ~ ~ ~ ~ ~ ~ p ~
t E CF(M). As tq = qt one has iI by Lemma 2.2.4. We leave details of (a) to the reader. #
.
is identified with N. @ 1. We will see that, for multimatrix algebras, the matrix AN
1J
M
r:
n
@
j=1
Maty.(K)
J
,N c M with r(Mat,.(K)) =.Nq. for each j is then a representation of
J J
i
I
n
@ Mat,(g) on V. Each Vi is a nontrivial Nsubmodule and we let ~r~ denote the
j=l J
corresponding subrepresentation. If q.p. # 0 then q.p.V. = q.V. is again a nontrivial M is irredundant.
(c) AN
J 1 J11 J 1
Nsubmodule (although an Msubmodule), and the corresponding subrepresentation
(d) The repmentation IndE(/3.) obtained by induction is equivalent to
J
x4,j4 for
Iri,j is zero on Mat ul(K) (l # j), while the restriction of r.1,j to Mat,.(K)
J
is a unital i
isomorphism onto N. . c M. = EndK(qjVi). The multiplicitx of this representation is j = I,.,n.
w 1,j (e) The repressntatiqn ailN obtained by restriction is equivalent to x + , j b j for
A.
1,j3
so there is a basis of q.V. such that the matrix of qj(xl,, +,xn) is
J 1
. j
J
.
diag(x.,. .,x.), with x. being repeated A. times; cf. the proof of lemma 2.2.2. Hence
J J 1,j
i = l,...,m .
there is a basis of Vi such that the matrix of ri(xl,. ,xn) is .. (f) A: is indecomposable.if and only if zMn ZN = K.
',
.
diag(xl,. .,xli3,. ~ ~ 3 .;xI1,.
; . e , ~ ) , u f . (a) We have to compute the image . [fj]M € KO(M) of the basic element

m ' m
,\ I
I
with x. being repeated A. . times. [fjN E K ~ ( N ) ; where j E (1,. p}. As f. = x p . f . one bas [fdM =
J 1J
b{lM. But
J W i=1 i=l
Next we introduce some terminology in order to give the Ktheoretic 'interpretation of
f.p  f.q.p. is the sum of A minimal idempotents in Mij, as one sees from the proof of
the inclusion matrix. The KOgroup of a ring R is an abelian group constructed from the J ~  J J ~ i,j
2.2.2 or the analysis of representations above. Furthermore, a minimal projection in M i j
representation theory of R; the set of equivalence classes of finitely generated pr?j&ive m
<
RmodFes forms an abelian semigroup, the operation being direct sum, and KO(R) is the
1
'quotient ,goup of this semigroup. For a multimatrix algebra M g @ piM g @ EndK(Vi),
remains minimal in Mpi, and is thus equivalent to ei. Thus [fdM = ZAij[eiIM, aa was
i=l
Ko(M) is the free aabean group generated by the classes of the standard representations
1 to be shown.
(b) This follows because $ = [1IM and 8=
ai(x) = xpilV.. This can also be expressed in terms of idempotents. For any idempotent
1
e E M, let [elM denote the class of the left, ideal Me regarded as an Mmodule. If ei is (c) Given i E (1,.
I
..,h),there exists j with p.q.
1 J
# 0 because x
n
i=l
qj = 1, hence
any minimal idempotent in piM, then Mei = Mpiei is a minimal left, ideal of M, and as
the ith row of A
: is not zero. Similarly, no column of :A can be zero because
a left Mmodule it is equivalent to the standard module {ai, Vi); thus {[?IM} is a basis
of KO(M). Similarly, choose minimal idempotents f. of q.N and denote by P. the
J J J
..
corresponding'irreducible representation of N, for j = 1,. ,n. There is a canonical map
i=l
(d) Observe that Ind;(fli) acts on M % Nf. which is canonically isomorphic to
KO(N)+ KO(M) induced by the inclusion N c M, which is the unique morphism of J
Mfj; thus indK(4) is left multiplication by elements of M on Mf..J The proof of (a)
abelian groups taking [elN to [elM for any idempotent e in N.
n
shows that lnd$(pj) = ,Aijq.
..
n 2.3.1. Let N c M be a pair of multimatrix algebras and let the notation
Pro~Osltlo
be as above. (e) This follows directly from the analysis of representations above, i.e. the
(a) A is the matriz of the.map Ko(N) KO(M) induced by the inclusion diagonalization of the representation r.
N 3 M, with mpect to the basis ([?]M)lrism 01 Ko(Mly Zm and ([fjN)lijln of (f) If di%(ZM, il ZN) > 1, there exists an idempotent r E ZM n ZN with r # 0 and
r # 1. With an appropriate ind 'n of the pi's and q.'s, one has
rJi"g J
Chapter 2: Towers of multimatrix algebras
I
i 5 2.3. Inc1usion matrix and Bratteli diagram 47
I
&&. We show this for M a factor, and the general case will follow; we may thus
identify M with En%(V) for some vector space V over K Set
!A
R'
.
= AM = (Al,. .,An) E Ml,n(W).
for some m' ,n' with 0 < m' < m and 0 < n' < n. It follows that M. . = 0 and Xi,j = 0 Let
llJ
unless
H = Matg(K)
but M and
by (x,y) 
are not isomorphic.
. Then AM
N and A M are pseudoequivalent to (2 1)
m
by the particular case observed in the remark following 2.2.2, #
EEQPf. The proposition is obvious if M and N are factors (see the Remark following 3v
2.2.2). In general; write Exam~le2.3.6. If N = Matv(C) @ 1C M = Mat,(C) gMat3(C), then B(NcM) is 111
S
and A! = [3].
irreducible representations of
p
,
e3.We denote them by
'W.
3 2 3 1
As in the examples, we always draw Bratteli diagrams on two levels, with the upper
arresonding to the 2dimensional irreducible representation level representing the larger algebra; the coloring of the vertices is actually superfluous,
since the two types of vertices are labelled by their level. The equation A? = has the
 following interpretation: For a given black vertex v, consider the set of edges entering v,
Similarly for and for each edge take the weight of the white vertex incident to that edge. The sum of
these weights, over all such edges, is the weight of v.
4, corresponding to the identity representation Ir, of e2
w.As (a) is nothing but a restatement of Proposition 2.3.3, we are left with the
...
Let y, ,lr, be the weights of the black points in B and let vl, ,vn be those of
c.
the white points and suppose = A. .v.. set
j
1,J J
$
of respectivb dimensions
=lm3
e3 are r e s ~ e c t i v d ~
B.
(xl,.
Define a map
.J
N 
of lines joining the i& black point with the $
th white point in
M by associating to (yl,. ,yn) E N the element
~ E) M with xi the blockdiagonal matrix
..
*a $ IF IF@!
I .
Xi = diag(yl,, .'Yli' ' .;Yn,' ",yn)
f
52 , Chapter 2: Towers of multimatrix algebras 
I
3 2.3. Inclusion matrix and Bratteli diagram 53
t
%
'. ..
Chains of multimatrix algebras. Now consider an increasing chain (finite or infinite)
II on the kth floor. That is, the m(k)bym(k+l) matrix ) (A! )
A ( ~=
,J
is
of multimatrix algebras over K, irredundant.
(4) Each vector !v has a weight pik E {l,2,ee  ) called its dimension. The
k and the "multiplicities" {A! .) satisfy
dimensions {pi)
9J
k
Let pl,... k
,pm(,) denote the minimal m t r a l idempotents in Mk, let = k be
k
the inclusion matrix for Mk c Mk+l, and let g be the vector of dimensions of Mk, so
~ k(K). (Thua $ = A(~')A(~~) 1(O)p0.) We associate with the
that p ! ~g Mat
pi \ Conversely, given a weighted multigraph B with properties (1)(4) above, we can, by
chain of algebras a (finite or infinite) Bratteli diagram B, which is the union of the iterating the procedure of Proposition 2.3.9.b, construct a chain of multimatrix algebras
diagrams B(MkcMkS1), the upper (black) vertex of B(MkcMk+l) drresponding to with Bratteli diagram B.
p;+l being identified with the lower (white) vertex of B(Mk+1~Mk+2) corresponding to
Pro~osition2.3.10. Suppose
the same idempotent. For example the diagram for CG1 C CG2 C CG3 C CG4 is
~ E M ~ c ..,
i
M ~and
c
~ E A ~ c A ~ c
are two chains of multimath algebras with the same Bratteli diagram. Then there is an
isomorphism gl' of Ma = UMk onto Am= UAk such that #(Mk) = Ak for all k.
k k
I
(Seeexamples 2.3.7 and 2.3.8.) We say that the v&ca
central idempotents p! in Mk belong to the kth f l s of the diagram. The vertices of
= &. ..
Let gl'o :'MO' AO be any isomorphism. Suppose gl'o,s ,% have been 
'+lIMk
the kth and k+lSt floors together with the edges joining them  i.e., the image of defined. Then by Proposition 2.3.9.a, there is an isomorphism %+1 : Mk+l ' Ak+l
B(MkcMk+l) in B  constitute the kth story of B. The Bratelli diagram B is thus a such that %+l(Mk) = Ak, and by Proposition 2.3.3 there is an inner automorphism
weighted multigraph with the following features:
(1) There is a function cp from the set of vertices of B to I = {0,1,2,• *), which
h+l of extending .
~ uo
~ U Swe
Q L Y T~ ~set q+l= #
assigns to each vertex the floor which it occupies.
(a) There are only finitely many vertices on each floor; that is pl(k) is finite 2.3.11. A ~ a t hmodel. Let B be a Bratteli diagram; we use paths on the diagram to
foi ail k. n C'B) t 0, we write ~  l ( k )= (4,...
, vk~ ( ~ ) ) . construct a natural model for the chain of multimatrix algebras associated to the diagram.
(b) The range of cp is either an internal [O,p] in I , if B is finite, or all of I, if We will suppose that B is infinite; it will be obvious how the construction must be
B is infinite. modified for a finite diagram. First we produce an augmented diagram B by adding a
(2) Two vertices v and w are adjacent only if Idv)y(w)l = 1. There are Xk (l)st story corresponding to the inclusion K l c Mo; that is we append one vertex * with
i ,j
,edges joining $and v:+l. '.
d*)= 1 and dim(*) = 1, and we connect *to v? by
J J
edges (1 s j s m(0)).
,
(3) If both the kth and k+lst floors are occupied (i.e., if cp'(k) # 0 and An oriented edge on any graph is an edge together with an ordering of its two vertices;
cpl(k+l) # 0) then each vertex on the kth floor is adjacent to at least one vertex on we will call the first vertex of an oriented edge its start and the second its g&. A path is
the k+lst floor, and each vertex on the k+lst floor is adjacent to at least one vertex
a (possibly infinite) sequence (6)of oriented edges such that end(ti) = tart((^+^) for
Chapter 2: Towers of multimatrix algebras $2.3. Inclusion matrix and Bratteli diagram
[ 55
.
all i. A path (. ,,tk) has end equal to end(&); a path ..)
(G,. has start equal to
start(@. If ( and p are paths such that end(0 = start (p) we define @q to be the
path "fi18t (, then f l A path t on B is monotone increasing if q(end(tk))
= + 1 for d k.
We let n denote the set of infinite monotone increasing paths on B starting at *;
n tr the set of infinite monotone increasing paths starting on the Pfloor of 8; nrl the
so that
set of monotone increasing paths starting at * and ending on the rth floor; and [rd
the set of monotone increasing paths stutiog on the rth floor and ending on the sth floor
\
(I < a). Given ( = ((o,(l, E n, set:a )
trl =(&,".,tpr] *
(0 r), Note that the minimal cent14 projections pi in Ar have the form
tt,s] = (kl,.,.,tJ E nb,s]
(1 i < 4 9
Also let be the vertex end($) = tart((^+^). Suniluly if ( = (to,. .~0E and
. x Q:],.
so that :
nrl = nil (disjoint union), and
It is an easy exercise to check that the factors of ArlS are in bijection with pairs of
It follows that AA .= (Xi,$, and the Brattpli diagram for the chain 1 FA,, C Al C ... vertices (v,w), with v in floor r and w in floor s. The factor corresponding to a pair
I
1
is B. (v,w) is the algebra of endomorphisms of the free vector space over the set of paths from v
As an example of the utility of the path model, let us i t CA!Ar) for 1< s. k t to w.
Remarks. (1) The path model presented here is due to V.S. Sunder [Sun] and A.
Ocneanu [Ocn]. Compare however [SV], in which a maximal abelian subalgebra of
Am = UAk is identified with W1.
k
(2) In case K = C, the action of the "path algebrastt A, on Cfl extends to a
representation on the Hilbert space 8(fl) with orthonormal basis fl. It is evident that
Ttln ' is then a rankone partial isometry with adjoint T*(,, = T,,S' So A, is a
and let Arg = spa%{Tfln : (&q) E Rrls}. Then Arls is an algebra, since again c*subalgebra of ~(F(fl)).
T ~ l $ ~,v ', = 6 ( a t f ) T t ,.
(((.n),(tl .nl)E RrlS), and 1 = C { T U : f E n[rl~l).
2.4. The fundamental construction and towaa for multimatrix algebras.
as operators on WL. Clmly Arc c CA/Ar). obtained by iterating the fundamental construction, as described in the chapter
introduction. It turns out that all the Mk are then multimatrix algebras:
Pro~osition 2.4.1. Let N c M be a pair of multimatriz algebras and let
M c Endi(M) be the pair obtained by the fundamental construction. Then
(a) End;(M) is a multimatrix algebra and its minimal central idempotents are of the
ElPPfL For x E As define P(x) =
form p(q), where q is a minimal central idempotent in N, and p(q) is& r
: ( M I ) E Rr}. multiplication by q.
(b) The inclusion matrix for M c end;(^) M
is the transpose of AN.
One verifies that P is a linear projection of A, onto CA8(A,). But for (&q) E Rsr
u f . Set F = En%(M) and dehne maps Alp : M  F by X(x)(y) = xy and
p(x)(y) = yx for x,y E M. The homomorphism X is the composition of the inclusions
M c ~nd&(M) and E n d i ( ~ c) F; the map p is an algebra isomorphism from MOPP
into F. As the pair N c M is isomorphic to the pair NOPP c MOP* by Corollary 2.3.4, it
is also isomorphic to p(N) c p(M). But ~ n d ; ( ~ )= CF(p(N)) and M = X(M) =
CF(p(M)). Consequently (a) follows from 2.2.3.a and (b) from 2.3.5. #
Chapter 2: Towers of multimatrix algebras
>
5 2.4. The fundamental construction
and zi > 0 for all i, and t.J > 0 for some j. It follows that
a :En$(M) %E I EndE(ME)
by
cu(Pa)(*b) = d x ) @ ba (v a En$(M), a,b e E, x E M).
The inclusion b t r i x for MOC M2k is (At A)k , and that for MOC M2k+1 is ( A A ~ ) ~ ~
Define also
Thus dink MU = ll(At A)kull + 2 and di% M2k+1 = ll(AAt )kA S I ~Therefore
~.
aa follows. Let {ai} be a basis of E over I( For each @ a EndE(ME)and each i, there
Lemma 2.4.3. Let N c M be a pair of finite dimensional algebras over a field K, and is a unique cpi E En$(M) such that
let (Mk)k20 be the associated tower. Then
1
@(*I) = cpi(x) @ ai (X E M)..
[M:N] = 1i m sup {di% Mk}Ilk.
k+ m Only finitely many cpi(x) are nonzero for any particular x, and since M is finite
dimensional over K, mly knitely many cpi are nonzero altogether. Then 0 can be

Proof. As Mo and Mk are finite dimensional Kalgebras, one has
defined by
and therefore It is easy to check that a and P are isomorphisms of Ealgebras which are inverse to
each other.
[M : N] = 1i m sup {rk(Mk I M&}'/~ = 1i m sup {% M~}'/~. # Next observe that
k+ m k+m .
cu(X(m)@a)= X(m@a), and
Prowsition 2.4.4. Let . I E N c M be a p&r of finite dimensional algebras over a field a(p(n)@a) = p(n@a) (m E M, n a N, a E E).
K, let E be any extension field of K and set
It follows from this that
M ~ = M % E and N ~ = N % E .
,
Then
(b) Let (Mk)k>O be the tower of extensions generated by N c M and let (Ak&>,,
(a) ~ n d i ( %  
~ E) Y Endr E ( ~ E ) .
E E
N : be the tower generated by N~ E ME. We produce a sequence of isomorphisms
(b) [M:N] = [M : N 1.
I
k
$ : Mk % E 4 Ak such that ak+l = ak for all k. Take %,al to be the
$ I ~ k v
[ identity and 02 to be the isomorphism defined in part(a); we have o2
&QQ$. (a) This is an example of a theorem on "change of rings in Hom"; see for
example [R], p.24. We give a simple proof appropriate to the special case at hand.
Define
1 .
(2.4.4.1). Suppose orl,. .,% have been defined. Let
I
C
62 Chapter 2: Towers of multimatrix algebras
I 3 24. The fundamental construction
Exam~le2.4.5 Consider two integas m' ,me L 1 and set m = m' + m'.
b3
Let M be
I Mk+l . E = E ~ ~( M
%+l: ~ ~ ) . E  . E ~ ~ ' ~M
( )! the factor Matm(C), let P be its "parabolic" subalgebra
Mk1 Mkl
A B
be the isomorphism defined as in part (a), and let
= {[o c] E M : A E Matm. (c), B E Matm. ,m,(C), C F at,.(^)}
and let L be the "Levi" subalgebra {[t;]} of P. hen L and M are semisimple,
a+1:EndMk1 (Mk)E 'Ak+l = Ak1 (Ak) and [M :L] = 2 as above, but P is of course not semisimple.
We claim that [M : PI = 1. Indeed, from left multiplication
be induced by the pair of i S 0 ~ 0 r p h i ~ ~ \ one has the inclusion
E
Set ak+l = 7k+1 o $+,;this extends 4 becaw extends the identity on Mk,
and &+l extends at where AA is leftmultiplication by A (and pA below is right multiplication). As the
the cominutant of P in M is reduced to the center C of M, the cornmutant of XfP) in
C n q e t l , we b e d ) = ( M ) = ( A ) for dl k,
E E # EndC(M) is isomorphic to M; moreover the natural morphism from M to
equality [M :N] = [M : N ] follows from this and Lemma 2.4.3.
CEnd(M)(X(P)) is an isomorphism. Consequently the tower generated by P c M is
Proof of Theorem 2.1.1 and Coro l l m 2.1.2 .. Because of 2.4.4 and the definition of P c Mc Mc ...
and the index is 1.
M for arbitruy semisimple algebras (given in the chapter introduction), it suffices to
AN We also claim that [P : L] = 1. From left multiplication
consider the case that K is algebraically closed, so M and N are multimatrix algebras. one has the inclusion
But then
L End ( P )
{M : N] = l i m {diml( M ~ } =
k
' ~ ~IAE~~, I
by 2.4.2 and 2.4.3. The corolluy follows fmm Kronecker's Theorem 1.1.1. #
R e . The norm of a product of two matrim is not, in generd, the product of their Thus CEnd(P)(L) isthe subalgebra
norms. It follows that, given a nested sequence 1E L c P c M of semisimple algebras,
I
the inequality
,
is in general stria. ~0, even this inequality failsto hold for *bras with radicals, as
we now show. Endc(p), isomorphic to (Mat,, (C) @ Mat2([)) @ Matm.(C). As right mdtiplicaion
9 4[t g] is r e p r ~ t e in
d Endt(P) by the matrix
Chapter 2: Towers of mult~matmitlgt:uryn Y 4.3. UJUW,~UCI,~"~
LUG I U U ~ ~ ~ . L U ~ U L~LL
rcX%? 00
Thus * is the reflection through the first floor. (Nevertheless w@regard the reflection of
an upward oriented edge to be upward oriented.)
n
Let vj = Knj e l m j and V = @ V.. ~ e f i n ea linear i p U from M to V by
21 01 j=1 J
[ P 
the canonical morphism P 4 N is given by
The argument used to show [M :PI = 1 shows also that the canonical construction
applied to P c N gives an algebra isomorphic to N. Finally, the tower generated by F('e'~) =T(~>(i),(C,ti*)
L c P is L c P c N c N C . . . andtheindexisalso 1. (C E n 2j1 and qo E n j ; i s j s n).
01
2.4.6. A reprise of Proposition 2.4.1. Let 1E N cM be a pair of multimatrix
Note that U breaks and unfolds the round trip path toq', while F folds and joins the
algebras with inclusion matrix A. Write {qj '. 1< j < n} and {pi : 1< i < m) for the pair ( t ' ,%). For example:
..
minimal central idempotents of N and M respectively. Let B be the twmtory Bratteli
diagram whose 0" dory is B(NcM) and whose lStstory is the reflection of B(NcM);
that is A(') = A and A(') = At. Let B be the augmented diagram, as in 2.3.11. For
example for CS3 c CG4, B is .
V carries both a right action of N and a left action of A2, arising from the right action
of N on Kf2 and the left action of A2 on Kn2]:
01
We identify the pair N c M with the pair A. c At, of path algebras d a t e d with 8.
(See 2.3.11.) j ' 1 i j < n) for the minimal central idempotents of the path
Write {"q
algebra A2. According to 2.4.1 and 2.3.9, there is an isomorphism of ~ n d h ( onto
~ ) A2 It is easy to check that A2 is in fact the cornmutant of p(N) in EndK(V), and that U
which takes X(M) onto M. Our purpose here is to use the path model to provide an intertwines the right actions of N on M and V. Hence
explicit isomorphism. Except as noted above, our notation is as in 2.3.11.
An edge on f3 is specified by the data q = (k;i,j,e), where k is the story on which q a: QUopF
lies, v.k and "!v are the two vertiegl of q, and the index L distinguishes among the
J is an isomorphism from ~nd;(M) = CEnqM)(p(N)) to CEn
k
Ai J. edges joining vkj and vf'.' Define an involution * of n[o,llU "1,21 by (PO)) = A2.
)
Let (4.d E % and ( 0 , ~E)RI (so TLq E $ and E M). One checks that
66 Chapter 2: Towers of multimatm rtlgebras 5 2.5. Traces
"'(~~,)('u,r) = ~(~~~or;)~(Gytl),(ro,t~).
i  ..a
where ei is a minimal idempotent in Mpi For example, the trace defined by
It follows that al(x) = X(x) for x E M c AT Also = 1 and = 0 for k # i mmaponds to the ith vector of the canonical
1 pkM
basis of K ~ Any
. row vector E K~ determines a unique trace
P(4,= a  l ( C TL$ = p(qj),
'4
as required by Proposition 2.4.1. \
Remark. Later we will want to modify the definition of U somewhat. If with associated vector g.
A trace tr on M is faithful if and only if the associated vector 5 has no zero entries.
c : fl[O,ll 4 @ is any function and we instead d&ne U by When the characteristic if K is zero, we say that tr is positive if si 2 0 for all i. (There
is an ambiguity here; if K is given as an extension of the reds, the meaning of si 2 0 is
U(TgS = c ( ~ ~ ) ( t ( , . ~ ~ { )70.
.,,sm) in C
@
I
Chapter 2: Towers of multimatnx rugeorau,  OY
not divisible by characteristic, then the weights t. We nOn*rO, and the restricted trace is
J
faithful. #
Remarks
(1) With the notation of the proposition, one has, when 0 extends 7, (Note that On any Bratteli it is actually superfluous to record the dimemiom
except on the first floor. Similarly on a finite Bratteli diagram it is superfluous to record
m the weights of a trace except On the top floor, but on an infinite Bratt& diagram it is not
(5, $1 = En = u(1) = ~ ( 1=) ( 6 3. (in general) su~erfluowto record the traces, since the tr on the higher floors are not
i=l determined by those on the floors below.)
(5) s( 3 R, and N c M is a pair of multimatrix algebra over s( with
By Propositions 2.3.1.b and 2.5.l.a, this implies ZN mM= and with inclusion matrix A. ~ e t( M ~ be) the ~ tower
~ ~ obtained by
t M + iterating the fundUWXltal c0IlStrllction. Then it follows from PerronFrobenius theory that
(5, )A
?! = (sAN, v),
there is a unique positive normalized (tr(]) = 1) trace on M~ = UM
k k'
which is, of course, obvious! In fact, let t(O) be the PerronFrobeniw eigenvector for i \ t ~ , nor&zed by
(2) A faithful trace on M may have zero restrictionto N. Consider for example Xtfo)vj = 1. Define t(2k) = l l ~ 1 1  @I ~~ t(2k1) = $ W A t (k l). Then
2 I
as r Eple 2.3.7, and the trace on M associated to the vector (1,1,1) E 8.Or consider
the two elemat field F2 and the pair F2 C Mat2(F2) (with inclusion matrix (21); trace An m ~ m m similar
t to one given in the proof of 2.4.2. shows that n ( ~ t ~ ) r ( ~ ; )
rLO
on M ~ ~ ~ has ~ )reatdction to the center F$ One may thus say about tram
( Fzero On consists of PerronFrobeniu~ eigenvectors for A ~ A . Suppose tr is m y positive normalized
traceon Mm and f(li) is the vector determining tr on MI Then for k r,
multitrix algebras, that pdikvitY is hereditary, but faithfulness is not.
(3) The assignment of a vector f E Km to a trace tr : M + K has been defined above
via the values of tr on (classes of) minimal idanpotents of M. In Chapter 37 we f(2k+2r)(zA)r = :(2k),
consider a new situation, where M is a finite direct Sum of COntihuous (type 111)
N~~~~~~ factors; since no minimal idempotents are present in this situation, we shall whence f(2k) is a PerronFmbenius eigenvector for n t ~ .since t(ik)(AtA)k = ~ ( ~ 1 ,
we have $(2k) = l l ~ l l f(O).
~~
describe a trace tr by the vector z = ( tr(p1),' .,tr(pm) ) of values of tr On
projectiom of M. In principle, the description of tr via is possible for
2.6. Conditional expectations.
present chapter.
(4) Given a Bratteli diagram representing a sequence of iXlClUSi0ns of multimatrix We are primarily interested in the following situation:
(3) E :M + N is the orthogonal projection of M onto N with respect to the inner But if V is onedimensional, spanned by v, then the functional (a,bv) n a +b is
product determined by the tram. faithful on A.
However, to clarify somewhat the roles played by semi simplicity, the pair of faithful The next proposition concern the existence of faithful conditional expectations.
traces, and the conditional expectation E, we begin in a more general setting.
A conditioa wectation from a Kalgebra M onto a subalgebra N is an Pro~osition2 . Q . Let N c M be a pair of Kalgebras with N finite dimensional, and
(N,N)Iinear map whose restriction to N is the identity. Recall that such a map E is let tr :M I K be a jaithjW trace with faith@ restriction to N. Then there ezists a anique
faithful if for each nonzero x E M there is a y E M such that E(xy) # 0. For example, if Klinear map E : M N such that
+
M is a factor, M = Matp(K), where K has charaxtersitic 0 or ,u is relatively prime to (i) tr(E(x)) = tr(x) xEM
(ii) E(Y)= Y YEN
char(#), then the trace on M normalized by trm(1) = 1 is a faithful conditihnal
(iii) E(xy) = E(x)y x E M, y E N. !
expectation of M onto K. L
Moreover E is a faithful conditional expectation from M to N, namely
Consider H O ~ ~ ( M , Nthe
) , set of right Nlinear maps from M to N, with its left r
(iv) E(~x)= yE(x) x E M, y E N Ii
z E N, so that E(xy) = 0 by faithfulness of T and x = 0 by that of E. Thus u is tr(E(t)y) = tr(E(ty)) = tr(ty) = 0 'd
, faithful. It follows that any Klineax map M 4 K is of the f o m x I+ u(ax) for some
so zy 'E N". Now xy  E(xy) and x  E(x) are in N", and hence also xy  E(x)y E N".
Rema&. (1) If N is a multimatrix algebra, then N has a faithful Klinear The difference
\
functional. 2
(2) Let V be a Kvector space and define a multiplication on A = KP V by F (xyE(xy))  (xyE(x)~) = E(x)yE(x~)
'
3
(X,v)(Xf ,vl) = (AX' ,XV' +X1v). The result
0 P V is an ideal. Suppose dim V 2 2. If
is a Kalgebra for which any subspace of
(p :A + K is any Klinear functional, then is in N' n N = (0), which proves (iii). One obtains (iv) similarly.
ker((p) n V is a nonzero,ideal in ker(~).So A has no faithful linear functional. i?
PY""
5
Chapter 2: Towers of multimatrix algebrr? 1 ) 2.6. Conditional expectations
*'u%A31.p
73
.ar
mf.Hypothesis (ii) says that E~ : M + M* = HO~;(M,N) is an isomorphism. As Sorollarv 2.6.4. Consider a pair of multimatriz algebras
projective modules of finite type are flat (see [BAC 11, page 28), the Klinear map n m
1E N = .@ q.N c M = @ p.M
~ = 1J i=l 1
as well as
is an isomorphism. Let
u f . (a) Condition (i) of Proposition 2.6.3 is fulfilled because any module over a
is an isomorphism. Routine computations show that 1 semisimple algebra is projective, and condition (ii) is fulfilled by Lemma 2.6.1.
To prove (b), first note that Q is a morphism because E is an idempotent which
commutes with A(x) for all x E N. If x E N and cp(x) = 0, then also x = cp(x)(l) = 0,
so 9 is injective. Finally Q is surjective by part (a).
,Z bnirpter a: lowers or mgimatnx a.tgeDraa $2.6.. Conditional expectations 75 '
.&mark: The following instructive proof of 2.6.4.a was given by Wenzl [We&]. First
note that the map Q of 2.6.4.b is an injective homomorphism. Now consider the while
subalgebra A of L generated by X(M) and E, and note that
A = { I(yo) + Z A ( ~ ) E X ( Y:~4
) Yj E MI, and
i
EAE = gP(N) g N Hence
' 2.6.5 Reprise of 2.6.4 using the path mod$. Let N, M, and L be as in 2.6.4.
Suppose tr is a faithful trace on M with faithful restriction to N, and let E : M ,N
be the conditonal expectation determined by tr, as in 2.6.2. Let B be the Bratteli
diagram for N c M c L, and let f! be the augmented diagram, as in 2.3.11 and 2.4.6. We
identify N c M with the pair of path algebras AOc A1, but we distFguish between
L = ~ n d i ( ~and
) the isomorphic path algebra A2. Let 5 and t be the vectors
determin the trace tr on M and N. We dso regard 5 and t as functions of vertices
B 1 = si. R e d l the *operation
on the Ot and lstfloors respctively: t ( 4 ) = t. and s(vi)
J J
Hence for (a,@)E R2,
It follows that ..
Definition 2.6.6. Let N c M be a pair of algebras over a field K, and E : M + N a ( 3 ) If N, M and L are *algebras, E = E*, and f = f*, then F is self adjoint,
faithfizl conditional expectation. j k c t e n s i ~ gpf M is a pair (L,f), where L is an because $ is a *morphism.
algebra containing M, f E L, and (4) If N and M are c*algebras, L is a *subalgebra of a algebra, E = E*
L is generated aa an algebra by M and f. and f = f*, then is positive. Indeed x H M is positive and is positive.
(i)
(ii) f2 = f.
Proposition 2.6.9. Assume that M L projective of jnite tgpe as a right Nmodule and
(iii) fyf = E(y)f = fE(y) for all y E M.
x H xf is injective. E is very faithfil. Let (L,f) be an eeztension of M. Then
(iv) The morphism
for all x,y E M, and hence TI is determined by its values on elements of the form X(x)E q.N 2 MatJK) piM p Mat (K) o(qj)L 2 Mat,.(K),
J J '5 ' J
for x E N.
Let tr be a faithful trace on M with faithful restriction to N and let E :M+ N be
the conditional expectation defined in Proposition 2.6.2. Let P E K Define tr to be a
Markov trace of modulus 4 if there exists a trace TI : L 4K such that
Tr(X(x)) = tr(x)
PTr(X(x)E) = tr(x)
] for all x E M.
PI.oof. We use the notation of Corollary 2.64. If such a trace TI exists, then for
j, ,
(ii) 4 (i) Set ? = rlt and let TI : L + K be the corresponding trace. Then Tr
p rj = flr(X(fj)E) (by 2.6.4.c) exten& tr because
iilt = p  l t ~ t= / r l . f ~ ~=t f.
82 Chapter 2: Towers of multimatrix algebras
t
s4At is not a scalar multiple o f t . This shows that one cannot delete the first equality in
by Proposition 2.7.2. From t,he proof of 2.7.1, we know that TI is described by ? = flit.
condition (ii). Although tAtA = t p follows from tAAt = gp (because t = $A), we
Consequently
prefer to state (i) in a symmetric form. cA t A ~AA~=P$
(2) We stress that ,D> 0 holds without any positivity assumption on tr, in case
char()() = 0.
and (a) now follows from 2.7.2.
. (b) The b i i e a r form (u,v) H Tr(uv). is nondegenerate on L and its restriction to
Theorem 2.7.5 Let K be a Peld eztension of IR. Let N C M be a pair of multimatriz
X(M) is nondegenerate; thus L = X(M) e x(M)', where orthogonality is meant with
algebras over K with inclusion matriz A, and with ZM n ZN = K. Let /3 E K*.
respect to this biinear form. For all x E M one has
A necessary and'sufficient condition for the ezistence of a positive Markov trace of
modulus on M is that 3/ = [M:N] = llA)12. Any two positive Markqv traces on M are
proportional.
so that PE1 E x(M)'. Aa D is theorthogonal projection of L onto X(M), this implies

Proof. Since ZM n ZN = K, it follqws that A is indecomposable and AAt is D(@) = 1.
(c) By Mlinearity of D one has DX(E)D = X(D(E))D, so (c) follows from (b).
irreducible (2.3.lf and 1.3.2b). Recall that [M:N] = llA112 by Theorem 2.1.1.
(d) Choose x,y E M and set u = X(x)EX(y) E L. The maps from M to M,
If tr is a positive Markov trace of modulus P on M, then P = llhAt[l = [M:N] by
the previous proposition and PmonFrobenius theory.
Conversely, set P = [M:N]. Let k be a PerronFrobenius vector such that
s AAt = fi. Let t = $A; it follows as in remark (1) above that tAtA = t. Hence if tr
t
is the (positive) trace corresponding to the vector 'ZT, then tr is a Markov trace of
modulus @ by 2.7.2.
The final statement follows from the uniqueness of the PerronFrobenius eigenvector are equal by (N,N)linearity of E. By (M,M)linearity of D one has
for Aht. #
....
A crucial property of a Markov trace tr on a pair N c M is that the trace TI on L
= ~ n d i ( ~entering
) the definition of the Markov property is again a Markov trace on
M c L. More precisely:
Consequently, using (a),
Chapter 2: Towers ot mut1ma,tm ageuraa,, 9 2.7. Markov traces / x'T 00
tr = trl on M1, which is a Markov trace of modulus P on the pair Mo C M1. We denote
U f . Statements (a) and (c) follow from (a), (c) and (d) of Proposition 2.7.4. If
by tr2 the extension of the trace to M2 denoted previously by Tr, and by j 2 i+2, then Ei E Mkl, and (b) follows because E. is Mkllinear. #
J
E1=E:M1+Mo, E1EM2
E2=D:M2+M1, E2€M3 Observe that this Proposition contains Theorem 2.1.6. .
the associated conditional expectations. According to Proposition 2.7.4, the process of
extending a Markov trace on Mk to MkS1 iterates; namely, if 2.7.6. The path model for M and the idempotents Ei. Let Mo c M1 be a pair of
multimatrix algebras and let
%: Mk+ Mkl is the conditional expectation associated to trk and trk,l, and
trk+l : Mk+l ' K is the unique extension of trk satisfying
P trk+1(xEk) = trk(x) for all x E Mk (see 2.7.1),
be the tower generated by iterating the fundamental construction. Let B be the
then trk+l is also a Markov trace, aqd the process can continue. Note that Mk+l is the augmented Bratteli diagram of the tower and
algebra generated by Mk and Ek, for short Mk+l = (Mk,Ek). Denote by Moo the A0cAl C *' c A ~ c A ~C +* ' ~
inductive limit (union) of the nested sequence
the chain of path algebras associated to B as in 2.3.11. Having identified Mo c M1 with
MoCMIC ..CMkCMk+lC.... AOC A1, we can obtain an explicit sequence of isomorphisms ak : Mk 4 Ak with
= for all k, by iterating the procedure of 2.4.6.
This is a Kalgebra with unit which is the union of its finite dimensional semi~imple
subalgebras, and which has a finite dimensional center isomorphic to ZM tl ZN. The union . If tr is a Markov trace of modulus P on MI, then tr extends uniquely to a trace
of the trkl s constitutes a trace t r : M(9+ K which is nondegenerate (namely, tr(xy) = 0 on Moo which is faithful on each Mk and which has the Markov property: if
I and tr = trl is positive, then t r is also
for all y E Moo implies x = 0). If K 3 R
Ek :Mk ' Mkl is the conditional expectation determined by the trace, then
positive in the sense that tr(r) > 0 for any non zero idempotent r in Moo. If this holds, Btr(%x) = tr(x) for dl x E Mr If t p ) denotes the weights of the trace on the kth
and if moreover ZM n ZN g K, then tr is the unique positive trace on Moo, up to floor of B, then tIk) = ,K1tjk2) for all k and j. We also write tr for the
J J
normalization; see Remark (5) at the end of Section 2.5. \ correeponding trace on Am= UAk.
I k
Assuming (just for the sake of having definite formulae) that # .is quadratically closed,
we can choose the isomorphisms {ali) so that $ = %(%) =
Chapter 2: Towers.of multimatrix algebras 5 2.8. The algebras for generic b .87
where Sk denotes reflection of an edge through the kth floor of B. In fact we h o w that
(Observe k indexes the algebra generated by idempotents up to k1; this agrees with the
this choice determines (4 completely because of the decomposition 2.6.4.(a). Then
usual convention for Artin's braid groups, but is not as in [Jol] or [Jo~].)
{%} is a sequence of idempotents (selfadjoint projections on ?(n) in cas? 'K= C and .
A monomial in AP,k is a product c. c. ,.ti where each ci is one of el,. , rkl;
'1 '2 q j
.
tr is positive) satisfying (a)(c) of 2.7.5.
in Ak as a
Iterating the decomposition (2.6.5.6), we can write any matrix unit T the unit 1 of Aplk is a monomial (the empty product).
a,P
monomial in the matrix units of A1 and the idempotents el, e,ekl. For example for
Proposition 2.8.1. Any monomial w E A may be written in one reduced form
(a,@)E R3 (Ta,p E A3) one finds P,k
'
goal is to understand the structure of these algebras. We shall see that, when the modulus
p of the Markov trace tr lies in a certain generic set, these algebras depend only on P
and k, and not on any other data pertaining to the inclusion Mo c M1 or the trace tr.
Moreover dur4( AP,k s
I"[ .

Proof. Consider an integer m with 0 < m < k1; we prove the first part of the lemma
For p in this generic set, Atrlk(M0cM1) is isomorphic to an abstractly defined algebra
by induction on m for a monomial w in {el,. .,cm}. Asthis is obvious for monomials
whose structure we describe in detail in the next section. For nongeneric P, new
with m 5 1, we may assume that m 2 2 and that the claim holds for m1.
phenomena can occur, and our knowledge is much less satisfactory in this case; see Section Suppose w is a monomial in which cm appears at least twice. Then w has one of
2.9. The following two sections borrow heavily from [Jo 11.
the forms
W=W
l Emaemw2
2.8  The algebras Afl$ for generic fl or
\ w =~ ~ ~ ~ a ~ ~  ~ b ~ ~ w ~
For any integer k 2 1 and for any number /3 f 0 in the basic field K, let AP,k be
the algebra abstractly defined (as an associative algebra over K) by where a,b aremonomials in { E ~ ,   . , E ~  ~ )As
. cm commutes vhth these, w equals
Chapter 2: Towers of multimatrix algebras + y LhU. 5SWrPr P,k "' 6'"
rr
I .* >

either Wlrmaw2 with paths from (a+l,a) to (a+n+l,a+n) which do not touch the main diagonal, and
their number is
1
or . wlaF fmbw2,
and the number of em's has been reduced. Consequently we may assume that w
involves exactly one rm. Consider finally a sequence (il,jl,, ,
,iP"P' ) corresponding to a reduced monomial in
Let w = wlrmw2 with wl and w2 monomials in {el,. .,rml). Using first the We may associate to this sequence the following path from (0,O) to (k,k), and any
induction hypothesis on w2 and then the commutation r r. = e.e for j < m2, we caa path from (0,O) to (k,k) which remains on or below the diagonal can be obtained in this
mJ Jm way.
reduce to the case that w = wlcmrml... cn, with wl a reduced monomial finishing,
say, with el. If 1 n one has
with all desired relations for the i' s aqd the j's. This ends the induction argument.
We now count the number of reduced monomials, following Chapter 111in [Fell. By a
path in the lattice z2, we mean here an oriented connected polygonal line with vertim at
integral points and with edges being either horizontal and directed to the right or vertical
and directed upwards. A path starting at (a,b) and ending at (c,d) has ca + db unit
edges, ca horizontal ones and db vertical ones. The number of these paths is
consequently the binomial coefficient
N[$] = ca 1.
[ca+db
it follows that the number of reduced monoinials is
["4. #
IS
Chapter 2: Towers of multimatrix algebras
90
5 2.8. The algebras Ai,k for generic 3
/
I
Lemma 2.8.2. Let k ;! 1 be an integer and set m = [k/2], the greatest integer less than
or epudl to k/2. Then

proof. By comparison of the coefficients of tC on both sides of
( l + t ~ ~ ( l +=
t )(l+t)a+b,
~ one has
so that in particular

.Z'["I
k +Z
'[k12,
m and
ik1l2=;KI :[;12.
j =O (Observe Pk here is as in [Wenl], but as Pkl in [Jol].)
Setting a = b = k and c = k + 1 in (*), one obtains Pm~osition2.8.3. Consider an integer k ;! 0 and set m = b]. Then
k (i) The polynomial Pk is of degree m. Its leading coefficient Is (l)m if k = 2m
2 i
j=O
k]El] C =
j=O
[il] = EL]
k k
[j] is even and (l)m(m+l) i f k = 2m+l b odd.
(ii) Pk has m ' distinct' roots which are given by for j = 1,2,..;,m.
A< q 1
\
For k odd (k = 2m+l), one obtains similarly
\
Then P1(A)> 0, P2(A)>0,. ..,Pk(A)>0, Pk+l(A)t 0.
(iv) Set Qk(A) = P~(x(x+I)~). Then
/CS'*
92 Chapter 2: Towers of multimatrix algebras ~7 'JJ. 9 InelLlgeDrw generic P ak
YJ
For ,b E K*, let q be a number distinct from 0 and 1;in K or possibly in some
quadratic extension of K, such that 8, = q'(q+l)'. Claim (iv) of Proposition 2.8.3 shows

Proof. Claims (i) and (iv) are easily checked by induction.
m .
For (ii), we compute in the ring Q[A,] and proceed as in the proof of 1.2.2. The that ,b is not generic if and only if z q l = 0 for some integer m 2 2. In particular, if K
+
difference equation for the Pkl s has an indicial equation ?j  p A = 0 with roots j=O
is a finite field, no ,b is generic.
For generic /3 € K*, we shall define inductively a neated sequence (BAk)k21 of
associative Kalgebras with unit, and a normalized trace on each of these.
k + D&. By adjustment of the constants C,D to fit PO,P1 k e find Set Bp,l = K and denote by trl the tautological trace on BPll. Set
so that Pk = Cpl
1 k+l k+l BA2 = Kel @ K(1el) where el is an idempotent, not zero. Define tr2 : BPY2 K by J
Pk = (plp2) (pi p2 ) for each k 2 0. Consider now a real number 0 with
1 ei 0 ei 0 tr2(el) = 81 and tr2(1el) = 1 8'. Identify B,b,l with the multiples of the identity
O < 8 < 7r/2 andset A=, so that pl = and h=pco~f Then
4cos 0 in B c Bp,2 is
P,2' The Bratteli diagram of the pair Bp,l
pk(A) =,*
.2 cos ( 0) s m 0
k
the (unique) nonzero idempotent in QO. 
Proof (see $5.1 in [Jol]). During the proof, we write Bk for Bp,k.
c Bak Both trn and it restriction trnl to Bnl are nondegenerate by (vii), since P is
. (iv)
I
The inclusion BWkl is desc~bedby the Bratteli diagram:
tional expectation associated to tr, a
q, in
Then EnenlEn = ?En. Indeed, for all y E Bn and all z E BW1, one has, first by
2.6.2(i) and then by (vi)
enlEn(enl~) = e p l x (*I
The j of the subfactors increases from right to left, So the white [ ~ M Pblack]
.
k Obviously (*) holds for x = 1 because En(enl) = 7 1 by the previous claim.
vertez on the eztnme right reprrcsnts 9il [ r a p . Qd. Next we check that (*) holds if x = yenl for some y E BPI, First, if y = ylen2y2
1

with y1,y2 E Bn2, then enlx = ry1enlY2 = 7en1Y1Y27 and
2
enlEn(enlx) = enn1En(en1)~1~2= 7 enlY1Y2 =
Thus (*) holds when x = yenl, for any y € Bn2 + Bn2en2Bn2, namely for all
Now using the Bnl.linearity of En, we see finally that (*) holds for all
Suppose in oddition that K = c, that each BOyk (k 13 has a c*&ebw stmdure
x E Bn1 + BnlenlBnl, namely for all x E B,.
mating the idempotents ei selfadjoint projectiom, that 0 > 0, and that p k ( r l ) > 0 for
Define BhS1 to be the algebra obtained from the pair Bnl c Bn by the
I
'+1
trn,+l(q41en) = 7J Pn+l2(j+l)(r)
and
(all n and jj. = = fi~n~($l)
This shows (iv), and consequently also (iii). Now (v) follows from L e y 2.8.2. and (**) follows because trn+l extends t r n l .
Define the trace trn+l :BnS1 7K by assigning the weight ~JP,+~~.(T)to the We now set w = xenly for some x,y E Bnq1. Then enwen = xenenleny = rxe,y
factor Qn", as desired for (vii). Let fk denote a minima( idempotent in Q\. When n by (ii) and, using the case of (**) already checked
j J
is even and j = n/2 we have
trn+1(fi/2) = n+' 2) 
PI~P~=
( ~p12,
)
while On the other hand, by the induction hypothesis
) PI2.
trn(fit2) = P / ~ P ~ ( T=
Theorem 2.8.5. Consider an integer k 2 1 and a number P E K* such that Suppose that m > j+l and that the result is verified for elements of
P )#0 r j <  1 w e e (Pj)j>l are the polynomials ofProposition 2.8.9. .
d g { l , ~ ~ + ~,eml).
,. It suffices then to deal with a reduced word w = xtmy where x
i
ci = (r1r2.. .rhl). ..(7172)~1
whenever 1 < j < k1 and w is in the subalgebra generated B$I l,tl,. ~ , e ~  ~Moreover
. in AS+ % K(q) for 1< i 5 k1. Thae are invertible, with $ = (qI + l)ri  I, and one
tr, is faithful i f pk(/T1) # 0. verifiea by induction that' C.~.C;~ = yhl and c.e.c;' = e for i 5 j1. In particular,
J 1J J 1 J hl
(c) The natural map AP,kl I is injective and trk extends trkl. ak: x ++ c p c i l is the automorphism of part (0, This automorphism is trace presewing,
(d) If Bp,k is as in Lemma 28.4, the assignment 6. I+ e. (1 j s k1) eztends to an because the trace trk extends uniquely to Ap,k % K(q).
J J #
isomorphism from ApIk onto Bp,k.
(e) The trace trk on also satisfies Corollary 2.8.6. Consider an integer k 2 1 and an grbitraa number @ E K*. Let cp
be Qe homomorphism ' Ap,k+l which, for j 5 k1, maps 6. viewed as a generator
J
of Ap,k to e (sic) viewed as a generator of A
j P,k+l
(a) is of dimension
whenever 1 5 j< k2 and w is a word in {ejSl, ..., c ~  ~ )More
. generaUy we have
(b) cp is an injection and any element x E A can be written as x =
P,k+l
d u ) + Zdui)%q(wi), whae u, vi, and wi are elements of A
P,k'
(c) There is a sequence of traces tre : 4 K (1 < l ( k) such that
.
whenever u is a word in {cl,. .,4.) and w is a word in { c ~ + ~ , ,. c ~  ~ ) .
J
.. tre (1) = 1, and trl+l(p(u) + xq(ui)ee "(i)) = trl (u) +~ ' Z t (viwi)
r ~ for all U,
the one nonzem c$,, is a power of r1 depending on u and r. In particular there are (b) In particular, with the notation above, the map ,y : E. I+ E. eztends to an
J J
monomi* ~ $ , ~ ( Et )Kit] such that c$,, as above is just cp0,r(8') for any u,r,p E S. isomolphbm of onto Atr,$MOcM1), and the restriction to Atr,k(Mo~M1) of the
Define now the "generic" algebra A@n,k over the polynomial ring K[t] as the free K is nowdegenerate.
Markov trace tr : Mk I
I
K[t]module over S, with canonical basis denoted again by (c,),,~~, and with b&&(a) It follows from 2.8.5(b) that tr 0 x = trk' Hence if x E ker(x), then for
multiplication defined by all y E one has trk(xy) = tr(x(x)x(y)) = 0, so that x = 0, by the nondegeneracy
of trk. Thus x is an isomorphism and tr is nondegenerate.
(b) By 2.7.5, the map x extends to an homomorphism of AD,k onto Atr,k(MO~M1),
and 2.7.5 together with 2.8.5(b) imply that tr o x = trk. Thus (b) foHows from (a).
#
The relations which express that this multiplication is associative are polynomial, and they
hold when t is specialized at r1
for any generic P EK*, by Theorem 2.8.5. Hence they Suppose ,8E s(* is generic. The following picture sums up the structure of the traced
hold identically, and Agen,k is a welld&ned associative algebra. Indeed, it is the algebra algebras introduced in this.section (with r = F1).
with unit over K[t] abstractly deGned by generators el,. . , E ~ and
~ relations
&mark: In general the traces tr! of claim (c) are not faithful; see Theorem 2.9.6.d.
Consider now the situation at the end of Section 2.7: One has a multimatrix pair
M,,c MI and a Markov trace tr : M1 3 K of modulus B; these ge&rate a tower, and the
.
conditional expectations E.J : M.J I Mkl for j = 1,. .,kl generate (together with 1) a
subalgebra Atryk(M0cM1) of Mk'
2.9. An appmach to the mngeneric case. Let Mo C M1 C M2... be the tower generated by Mo t M1. Since Mk is generated
as an algebra by M1 and Atrjk(M0cM1), if for some no the algebras A and
If 1,3 E is nongeneric, then tr,no
(1) The algebra AP,k defined by generators and relations as in Section 2.8 need not be AtrTE0+, were equal, then M4 = Mno+l as well, and therefore Mk = M for all
no
semiflimple. k 2 no. But dimc Mk increases as [Mi:Mo]k = 9k, by Proposition 2.4.2, Hen&
(2) Given a multimatrix pair MOc Ml and a Markov trace tr of modulus on
MI, the restriction of tr to Atr,k(MocM1), the algebra generated by {],El,. .,Ek_l)
in Mk, need not be faithful.
(3) Given a second such pair M0 c and a Markov trace & of modulus /3 on for all k. On the other hand h;ll = Mn for all k and A ~ , ~ ( M 2~ Cc for
M all
~ ~k.
(1,1). Then tr is a Markov trace of modulus 1 on M1. Consider also a pair MO= The algebras BP,k, For the rest of this section we fix a P E s(* which is ngeneric
with any faithful trace ?r on then ?r is evidently a Markov trace of modulus 1 on but not (n+l)g~eric for some n 1 1. That is pk(/3l) #0 for k < n, but
(3) A t r , k ( M o ~ ~ land
) A ~ , ~ ( M ~ care
Mnonisomorphic
~) for all k 2 2.
'
For k i n, define Bp,k and trk exactly as in Lemma 2.8.4; since P is ngeneric there
is no problem i n doing so. For k 2 n define Bp,k+l to be the algebra obtained by
&&. The matrix =A = (:t] brs eigenvectors (1,1) a+ (1,l) with : applying the fundamental construction to the pair BSkl C BPyF Observe that B
P,n+l
eigenvalues 1 and 9 respectively. The PerronFrobenius eigenvalue 9 is also the index c isthesameas B'
P,n+l
in 2.8.4. For k < n+l, define trk as in Lemma 2.8.4; then
[M1:MO]. But the other eigenvector (1,1) also defines a Markov trace tr on M1 with trn+l is also faithful because Pn+l does not appear in the computation of the weights of
modulus = 1. I.
IVY
,4, V"*pYYL .. &"I."*"
"1 A
'
."
. YI".". lo""
".>
,* ZVO
\
Bp,n+l = extends that on Bp,,; it thus follows from 2.8.4(vi) (with k = n+1)
that trn is a Markov trace of modulus @ on Bp,nl c BPp For k 2 n+l, we define
trk as in Proposition 2.7.4. Thus trk is a Markov trace on Bp,kl C Bp,k for k 2 n,
but not for k < n. Note that B is a multimatrix algebra generated by the identity
P,k
..
and idempotents {el, ,ekl) satisfying the relations 2.8.4(ii); in fact these relations
hold for {el, .,en) by 2.8.4 and for {en,en+l,. ..) by 2.7.5. For k 2 n+1 the identity
is contained in the algebra generated by {el, .,%l), in contrast to the case of generic
fl this follows from 2.6.4.
Note that if K = C and P = 4ms2(d(n+2)), then the algebras Bp,k can be given a
structure such that the generators {ei) are selfadjoint projections, and the trace is
faithful and positive. This is shown in 2.8.4 for k 5 n. The assertion for k 2 n+l follows, B2,l 1,l
because the tower construction for a pair of finite dimensional c*algebras with a positive
Figure 2.9.4.
Markov trace produces a chain of c*aIgebras with a positive trace, and selfadjoint
projections ei; see the discussion in Appendix IIa.
Example 2.9.4. Assume that the characteristic of K is not 2 and let P = 2, so that
n = 2. The structure of the algebras Bp,k and of the traces trk is shown in figure 2.9.4
below.
Figure 2.9.5.
In general, the picture for the B 's is obtained fmm that of the .4P,k's at the end
P,k
of Section 2.8 by deleting the factor":Q (represented as the extreme right point in the
(nt1)st row) as well as all factors above and to the right.
vnapter z: lowers or I I I U I G I  L ~ ~itlgaula 5 2.9. An approach to the nongeneric case
106 I~
t
Theorem 2.8.5 gives a complete description of AP,k when P is (k1)generic. The trk 1  trkI, and
following theorem indicates how part of the picture changes when P is not generic. Recall Ckl
that we may (and do) always identify with a subalgebra of (see Corollary
Theorem 2.9.6. Comider an integer n 2 2. Let /3 E I(. be such that p j ( r l ) # 0 for (f) The trace trk on also satispes
j 3 n and P ~ + ~ ( B ' )= 0, where (Pj)j21 are the polynomiak of Proposition 2.8.3. Then
P trk(Ejw) = tr(w)
one has for all k 2 1,
(a) BP,k is a m d t h a t r i z algebra, and there ezists a homomorphism rk of APYk
onto B mapping each generator c. onto e. (1 < j < k1).
whenever 1 5 j 5 k2 and w is an element of alg {l,ej+l,. .~ , e ~  ~More
) . generally, we
P,k J J have
(b) There ezists a normalized trace trk : Bak 4 K such that, for any j E 11,. .,k1) . trk(uw) = trk(u) trk(w)
ptrk(wej) = trk(w) .
whenever u E alg {&el,. .,e.) and w E alg {P,ejSl,. ,ekl).
J
..
whenever w is in the subalgebra Bp,j
restriction of trk to B is tr. for j s k.
P,j J
B~,k.
M O T ~ Otrk
U ~ ~is faithfir and the
e.
J

(g) The map 4 +I % extends to a trace preserving automorphkrn I?k of
ekj eztends to a trace preserving automorphism Fk of Bp,k. These automorphisms
are inner in case K contains an element q satishing q1(q+l) 2 = P.
and

Proof. Claims (a) to (c) follow from the construction of the Bp,k above. The traces
tr o rk on APIk satisfy (*). The uniqueness statements in (b) and (d) are proved
as in 2.8.5(b). We have trk(xy) = trk(rk(x)rk(y)), so that if x E ker(rk), then x E 1
P,k'
Conversely if x E I then rk(x) = 0 by faithfulness of trk on B This proves (d),
P,k' P,k'
and (e) follows similarly. Statement (f) ia proved as 2.8.5(e), and statement (g) as
(d) There is a unique family of normalized traces trk : + K S U C ~that
2.8.5(f). #
I
a""
F vuapm A, ruwc~u
ur L l r u i u l u m e i r A olgcuroa An approach to the
$is&',, nongeneric case
$
that the idempotents Ei (1 5 i 5 k1) are positive linear combinations of oertain minimal
idempotents T B be the Blinear span of the
see especially 2.6.5.2 and 2.6.5.4. Let Mk
6, 6
mat* units geneiating Mr Thus 4 is a m u l t i  m a t e b a o r I, and
Mk = M
! %K Let A! 4 generated by {&El,. ..,Ekl}.
be the Csubalgebra of
The trace tr restricts to a positive Rvalued trace on 4. Note that Ak
B is closed under In order to accomodate vectors and matrices of different sizes, we adopt the convention
I defined by ~ that Id imbeds in Id+' via I'
the M n e a r involution r of Mk f= TlYf.
, ~ Positivity of the trace implies
I
that tr(x*x) > 0 for all nonzero x E Mk, and as this holds in particular for x E Ak, we
conclude that t r
1A!W is faithful. It follows by linear algebra that tr is also faithful on
With this convention we have for n&l
Atr,k(MO~Ml)= Ak % K, and therefore 2.9.6(e) implies the conclusion. #
(2.9.8.1) 0 (n odd).
t$ = 11((n3k)l12= ((A~A:)~~(I
I
where the P: are the polynomials of Proposition 2.8.8. (A;A,) (k1)/2( for k odd
J (hk) =
A ~ ( A ~ A ~for) k~ even,
/ ~  ~ ~
m.Set An = ABpvn
Bp,nl
and = di+BAk). Also let ((n(n.k) be the vector of ' Hence
dimensions of the multimatrix algebra Note that the Bratteli diagram for  (2.9.8.2) bf: = ( ( A ; A ~ ) ~  ~ ~(nI even).
Bp,k.
Bp,nl c BP,, is the Coxeter graph An+1, with a particular bicoloration and labelling of
the vertices. (See 2.8.4(iv) for the picture, substituting n for k.) Thus for n odd An is One can visualize these results quite easily by adding to the Bratteli diagram of the
the by Jordan block : chain (Bp,k)k2 some "phantom" vertices with zero dimension. The picture for n = 5,
for example, is
>
Recall also that our labelling of the vertices on each floor increases from right to left,
<
Since An< = (n odd) and AC: = 4 (n even), (2.9.8.1) and (2.9.8.2) give
The difference fn+l(z)  [zfn(z)+l] is computed using 2.9.8.5, and the resolvent identity:
(2.9.8.3) b i = ( ( A : A ~ ) ~tO) ~ ~(n ~ d d ) ,and
(2.9.8.4) b: = ((A,A:)~( I 4) (n even).
P1(z) 1 where the direct sum is over the i ' s with epi # 0. The Bratteli diagram for the pair
while s2(z) = = #
C c H(G,Go) is &at part ofthe Bratteli diagram /or the pair CIGO]c C[G] which lies above
the wertez corresponding to e.
will no longer denote an idempotent below.) Identifying the double w e t s is a special case
114 Chapter 2: Towers.of multimatrix algebras 9. z.ru. . necKe algebras 115
n h ~ h  l ), ( B ~ B ) / B
where W is the I1Weyl group1', namely here the symmetric group Gn embedded in
GLn(q) as permutation matrices (see 5IV.2 in [BLie]). Thus to each permutation w E 6,
I B/(B
c l a s s of b class ofbh
there is associated an element aw of the Hecke algebra H(G,GO), which is the is well defined (if b,bl are in the same class modulo B n hBhl, there exists b' E B
with b' = bhb'hl, and blhB = bhB) and bijective. Then the number of left classes
characteristic function of BwB divided by the order of B. For i = 1,2,. .,n1, let si .  modulo B in BhB is the index [B : B n hBh'1. It follows that
be the element of W given by the matrix
1.
1
'1 Let us compute (aw*aw,)(g) when w = w' = s. As C(s)C(s) = B u C(s) this is
0 1 zero unless g E B U C(s). For g . B~ one has by (w)
1 0
1,
where the first diagonal 0 is the (i,i)th entry, and set 4 = asi. As al is a convolution unit in C[B\G/B], this implies
Furthermore the elements 4 (1 6 i < n1) generate the Hecke algebra H(GLn(q),B).
PEeefL (see [BLie] as well as Propositions 11.30 and 11.34 in [CR]). For each
permutation w E 6, set C(w) = BwB. Let a, E C[B\G/B] be the quotient by I B I of
the characteristic function of C(w); then is a Cbasis of the Hecke algebra. and consequently X = q  1. This shows (a).
Introduce the length function l : 6, ,{0,1,2,. ..} with respect to the generators S.
For w1w1,w' E 6, and for g E C(we), one has
Then
Indeed, if 4sw) > 4w)> then C(s)C(w) = C(sw) by no IV.2.4 in [BLie], so that as * a,
is a scalar multiple of as, by (*). Let g,h E C(s) and u,v E C(w) with gu = hv; then
If C(w) n gC(wl)I is not empty, there &a bl,. ..,b4 E B with blwb2 = gbjwf1b41
vuI = h'g E C(s)C(s) = B u C(s); but vuI E C(s) would imply v E C(S)C(W) =
so that g E C(w)C(wl).
r
8 fP*wnewt:ageoras 117 ,,
,
.Y
;.
:(sw), which. is incompatible with v E C(IV); hence g E hB, and thus any element in
' ' (i) to reduce the length of a word in the {gi) (i.e., to write it as a linear combination
:(sw) can be written in exactly ( BI ways as a ~roductof one element in C(s) by one in' of shorter words), and
:(w). This shows that as * a, = a,. It follows in particular that {as) generates (ii) to reduce the number of occurences of the largest gi in a word, and to move it to
[(GLn(q),B). the right,
Consider finally s,t E S with (st)3 = 1. Then 4s) = 1,qst) = 2,4sts) = 3 and thus whenever the corresponding operation can be performed on the corresponding word in the
* at * as = asts by (**). Similarly at * as * at = at,, and (b) holds. Claim (c) . .
{si). It follows that d i ~ &H
q,n
is at most n!. On the other hand, we will exhibit below a
.way.. #
>llowsin the saine
.
sufficient family of inequivalent irreducible representation of H to obtain the other
. . . . q,n
 inequality. See [HKW,$4] for a more explicit proof. #
Now remember that 'the symmetric group in n letters has a presentatioh with
enerators the transpositions .si = (i,i+l) for 1s i s n1 and relations For convenience we take 9( = C in the following discussion. For q a prime power,
Hq,+ is the same as H(GLn(q),B) in 2.10.a, and is in particular semisimple. But we
si2 = 1 . s si+lsi = si+lsisi+l 1 j .
s.s  S.B.
J I
if, Jijl L 2. have no reason a pn'ori to believe that there is any relationship between these algebras for
different values of q. Also, the decomposition of any H as a direct sum of matrix
q,n
'here is an easy proof of this which shows at the same time that the abstract algebra algebras is not obvious, each summand corresponding to some irreducible representation of
enerated by n1 generators subjected to the relations of 2.10.3 is of dimension at most
GLn(q).
!. (See the beginning of $4 in [HKW].) For q a prime power, it follows then that the
Ohewe however that, if we put q = 1, we recognize H as the algebra C[Gn] of
gations of 2.10.3 give a presentation of the Hecke algebra H(GLn(q),B). But we shall see l,n
lat it is important to consider a more general family of algebras, defined for all q f 0.
relations: as in 2.10.3. H ~ , + and polynomia~structure constants PP (q) such that g$T=Cp$,T(q) gII ~t
('97
 P
Proposition 2.10.4. One has di H = n! for all q E K and for aU n L 1. follows that degeneracy is determined by a polynomial equation in q, so for all but a finite
q,n set of q E C (n fixed), H is semi4mple of dimension n!. Also Hq,nl embeds in
q1n

Proof. We take for granted the presentation of 6, in terms of the transpositions Hq,n via the obvious identification of the generators 4 for 1 s i i n2.
3i). Each of the n! dements r of Gn can be written uniquely as a reduced word w in We now argue intuitively, though Mremely plausibly. For the values of q for which
, and Hqp are semi simple, the inclusion Hq,nl c Hq,n is completely deecribed
le isi) with
by a vector of integers (for the dimensions of the factors in H ) and an integer valued
(i) minimum length among all words representing Ir, q,n
(ii) the largest si in w appearing only once, and moved as far to the right as matrix (the inclusion matrix). As these should vary continuously with q, they should be
independent of q for these values. In particular they can be determined by examining the
possible, and
case q = 1. But then they are determined entirely by the dimensions of the different
(iii) all subwords of w reduced according to criteria (i) and (ii)..
1
representations of C5nl and Gn and the restriction rule from en to FOI this
he corresponding n! words in the generators {gi) of H span H linearly, because reason we shall now describe this structure. In 2.10d we will identify a certain singular set
q1n q,n
le Hecke algebra relations 2.10.3(a)(c) can be used
110 bnapcer a: lowers or mulr1maFnx ageoras 9 2.11~. neme ageoras
Example 2.10.5. To the partition X = [X1,X2,XpX4,X5] = [5,3,2,2,1] of 13, one figure 2.10.6
associates the Young diagram The dimensions of the corresponding representations are given by the number of ascending
paths on 2.10.6 beginning with o and ending at the Young diagram in question. The
above facts will actually follow from the construction of irreducible representations for the
Hecke algebras Hq,n, to which we now return.
The most important rule is the restriction rule: if one restricts the representation of
6, corresponding to a Young diagram Y to it is isomorphic to the direct sum of 2.lO.d  Irreducible representations of H for q $ St.
q1n
all representations corresponding to all Young diagrams Y' obtained by removing one The material that follows is taken fiom Wewl's thesis [Wen2]. The Kalgebra H
box from Y, all occuring with multipIicity one. q,n
is that defined at the beginning of 2.10.b; in particular, the field K is arbitrary.
Thus the irreducible representations of Gn (and hence the Bratteli diagram for
By our intuitive argument, we expect that figure 2.10.6 should also represent the
CG1 c CG2 C CG3 c ...) arewnveniently pictured by the following important diagram: structure of H
q,n
for all but a countable number of values of q. While this could be
proved in an elegant manner due to Tits (see exercise 26 in 3IV.2 of [BLie], or Lemma 85 in
[Ste2]), two important pieces of information would be missing: there would be no
indication of which values of q are."badl' (and that would be particularly frustrating for
K countable!), and there would be no construction of concrete representations of H
q,n'
We shall now show how to cons an irreducible representation of HaSn for each
partition of n, provided q is not in the set St defined below. It is first convenient to
dispose of another presentation of Hqln than that of 2.10.b.
(a) e,g=
i+l
m i=l,...,n1
(a) By adding two boxes to the same column of Xi.' In this u s e fivp = vp.
These generate H and constitute with the relations
q9" (b) By adding two boxes to the same row of Xi.' In this u s e fivn = 0.
(b) e; = ei .
i = 1,. .,n1
r
(c) By adding boxes in differeht rows and columns of Xi'; mort precisely there is
(c) eiei+lei  q(q+l) 2 ei = ei+leiei+l  q(q+l) 2 ei+l i = 1, ..,n2 pair of integem (1,s) with r # s and A':
:'li A': such that = A': + 1 and A;
(d) e.e . e.e. when lij1 2 2 i,j = l,...,nl A:+' = Xi' + 1. In this case there is precisely one asceading path from Q to A which
1j 3 1
differs from p in its ith vertex only; we call this path p' . For example:
a presentation of H
4,n'

Proof. A straightforward computation.' #
Set
l + q + . ..+q
d and 0 b S e ~ ethat d # 0. Define d' in the same way for the path p' and note that
(Remark: When d > 0, then ad(q) = Note also that
 (l+q)( l + q + . +qdl)'.. d' = 4. Finally, define
9
1s $k
ad(q) = Qk(q) where Qk is as in Proposition 2.8.3.iv.) Suppose given a partition
J to be zero. We think of
X as a Young diagram. Let VA be the free Kvector space on the set of ascending paths Observe that fi leaves invariant the subspace Kv @ Kv of VX as well as its canonical
: P P'
p fcom to X on figure 2.10.6, and denote by {vP) its canonical basis. We define now , complement; on Kv @ Kv ,, it is described by the matrix
I P P
.
endomorphisms f l, ,fnl of V
Let i E (1, ,n1). For each ascending path
\
p = ( X 1=u,X2,..,Xn=X)
i We have taken advantage of the equaIity ad(q) + ad,(q) = 1, which follows from the
we have to define fivp. The partition Xi+' is obtained from in one of three ways definition of ad and from d + d' = 0.
Chapter 2: Towers of multimatrix algebras 3 2.11. AD,, and Hecke algebras
The veiification that fl,. s,fnl satisfy the relations (b) and (d) of Proposition with ad written for ad(q). It follows that Hqp has a c*algebra structure for q E IR
2.10.7 is trivial. They also satisfy (c), but this is more tedious to check and we refer to and q > 0, for which ei is an orthogonal projection: In fact, the main interest of [Weh2]
[Wen2]. We conclude that, for each partition A of n there is a representation rX of
is in the values q = e*2d/n E R, for which Wenzl has constructed c*+dgebras which are
Hq,n in VA defined by rA(ei) = fi' A remarkably easy inductive argument shows that quotients of the corresponding Hecke algebras.
the rA's aie irreducible and mutualljr inequivalent when X runs over the set .'P, of all
partitions of n (for q E K \ R). Indeed,, these representations are absolutely irreducible,
2.11. The relationship between Am and Hecke algebras.
because the same argument applies to any extension of K. By theorem of Burnside and
Frobeni~lsSchur,this implies that Hq,, has a quotient isomorphic to the multimatrix
It was first pointed out to V. Jones by R. Steinberg that the defining relations of AD,,
algebra e En%(VX), of dimension n!. But we have already reported that the
A€? u (see section 2.8) actually imply the Hecke relations. This is obvious from the definition of
dimension of H is no more than n!. (See the end of 2.10.a above, and $4 in [HKW.) Ap,, and Proposition 2.10.7, but we would rather state this in terms of the generators gi
q,n
Consequently the dimension is precisely n!, we have a complete set of irreducible again.
representations of H for q 6 K \ R, and Hq,, ie isomorphic to e EndK(VX). (In
q,n A€%
B Proposition 2.11.1. Consider q E K \ {1,O) and ,O = 2 + q + qI E #*, an integer
particular, setting q = 1, this gives for K = OJ the usual complete set of irreducible 2 1, and the algebra Ab,n. Set
representations of the symmetric group an.)
7
Another trivial consequence of the construction is that the restriction of a 7i+1
7i = (q+l)fi  1 so that ci = 7 i = 1,. .,n1.
representation rAof Hq,n to Hq,+ is a direct sum te XX,, where A' runs over all q+
partitions of n1 obtained from the partition of X of n by removing one box from the
and constitute with
Young diagram. We reformulate this as follows.
Theorem 2.10.9. Let K be a field and let R c K be the union of {0), ofthe non trivial
roots o f 1, and of 1 in case char(K) # 0. Consider q E K \ R, an integer n 2 1, and the
Hecke algebra H generated by gl,. q,gnl with Me relations of 2.10.b, or equivalently
q,n
.
by el,. *,enl and the relations ofProposition 2.10.7. Then
(a) Hqp is o f dimension n!.
(b) Hq,, is a multimatrix algebra. a presentation of A
P,n'
(c) The natural mapping Hq,, ' Hq,n+l is an imbedding.
(d) The structure of the chain Kq,l c Hq,2 c ... c Hq,., c q . is gyen by figure 
Proof. A straightforward computation. # '
2.10.6.
Corollary 2.11.2. There D a surjective algebra morphism
We make one further comment on Wenzl's paper, and for this we assume K = C. His
exposition does not involve the matrix 2.10.8, but rather the related one a Hq,n: ' AD,,

defined by Iln(gi) = 7i for i = 1,. ,n1. If n = 1 or n = 2 it is an isomorphism. If
n 2 3 ikr kernel In is the twosided idea1 o j H generated by
q,n
Chapter 2: Towas of multimatrix algebras p.
A@,, and Hecke algebras
'1
.g1gg1+ glg2 + g2g1+ g1+ g2 + 1. Lemma 2.11.3. Let q E K \ 0 , let. n L 3 be an integer, k t A be Young diagram with
f
n boxes and with at most two rows, and let
Moreover the diagrdm
r .H
A . 9,n
4 End(VA)
1 1 i = 1,s. + , n d
(g162. 'h1)9(~!1" 'g2 g i ) =
Thus In is generated by xl. The last claim is obvious. # where A' has one less box than A.. In particular A' has at most two rows, and
5,(x1) = 0 by the induction hypothesis. Consequently rA(x1) = 0. #
Thus, if q is a value for which H is semisimple, it must be possible to identify
q,n
AD,, with a certain ideal of Hq,n, given by some subset of the set of partitions of n. We Consider q E K \ 0 and an integer n 1 3. Let 'P, be the set of partitions of n. We
shall show that this subset is precisely the set of a3l partitions with at most two rows; this know from section 2.10.d that H is a direct sum @ IA of simple two sided ideals,
qP AE.~,
was also explained to us by R. Steinberg.
We recall that R has been defined in 2.10.d, and that the relations which constitute the notationbeing such that, for each AO E Tn, the representation r~ of Hqp restricts
with gl,. ,gngnl a presentation of H are written in Proposition 2.10.3 (see.also 0
q>n to an isomorphism IA + End(VA ) and maps IA to (0) when A # A,,. We denote by
2.10.b). 0 0
f
the subset of 7, of partitions with at most two rows.
126 Chapter 2: Towers of multimatrix algebras 5 2.11. ApIn and Hecke algebras
of Theorem 2.8.5.
be as in Corollary 2.11.2. Then
To sum up, we have shown that, for generic pand corresponding q, the algebra A
P,n
is isomorphic to the quotient of by the twosided ideal generated by
Hq,n 
+ + + +
gl%gl glg2 + g2g1 gl g2 1. This ideal corresponds precisely to the direct
A Ef,
mnmands IA of H given by all Young diagrams A E 'P, having at least 3 rows.

Proof. By the previous lemma for one has ( 1=0 so that (4,''
IA n In = (0). Consequently In c e IA.
Ann\f
Let A = [r,s] E f . From the definition of VA (see 2.10.d) one has
Thus
I: AEf
(dim
2 1 2n
= rn[
by Lemma 2.8.2.
where L ~ ( M )is the Hilbert space obtained by completion of M for the scalar product
(xi y) = t r ( ~ * ~ )It. was shown in [Jo4] that this definition of index a g r w with the purely
In this chapter we study pairs of finite von Neumann algebras with finite dimensional
centers, and the index of such pairs. ringtheoretic definition of Chapter 2.
Sections 2 to 4 are purely expository, and may be taken as an encouragement to the If [M:N] < W, the pair N c M generates a tower of IIlfactors
reader having essentially no previous experience with von Neurnann algebras. Sections 5 to
7 present a generalization to the present setting of some of the ideas of [Jol] for pairs of ...
I E M ~ = N C M ~ = M Cc M ~  ~ c M ~ c .  .
factors. Though this chapter c a ~ o be t so selfcontained as the previous ones, we have
tried to minimize the technical background in operator algebras assumed from the reader. by a fundamental constructi~gwhich is defined as follows. The natural conditional
expectation from Mk onto Mkl can be seen as an orthogonal projection
Let us first describe Sections 2 to 4. Let M be a von Neumann algebra which is a
f&@of type II1. (The definition is given in Section 3.2.) We denote by tr : M + C the % :L'(M~) + L ~ ( M ~  ~and
) , Mk+l is the von Neumam algebra of oprnntors on
normalized trace on M, For every Hilbert space H on which M acts, M m a y and von L2(Mk) generated by Mk and ek. This MkS1 is again a 111 factor. It is aparticular
Neumann have defined a positive number (possibly a) called the constant case of Proposition 3.1.4 below that this way to define the fundamental construction agrees
between M and its commutant; we denote this number by dimM(H). Two with that of Chapter 2. Moreover the Markov relation holds:
representations of M by operators on two separable Hilbert spaces H and H' are I
[M:N]trk+l(xek) = trk(x) for all x E Mk,
equivalent if and only if dimM(H) = dimM(H1). Section 3.2 is as exposition of the
definition and the basic properties of these coupling constants. Except for the
presentation, all this material comes from the original papers by Murray and von where trk and trk+l denote the normalized traces on Mk and Mk+l respectively.
Neumann. I The sequence (ek)k21 of projections in U Mk satisfy the relations
In Section 3.3, we present some geometric ezamples of coupling constants arising in the k2O
theory of discrete series representations of Lie groups; they are borrowed from
IM:N] e.e.e. = ei if 1 ij 1 = 1
AtiyahSchmid [AS]. In particular, we show: 1 J 1
e.e. = e.e. if (ijl 2. 2
1J J 1
Theorem 3.1.1. Let G be a connected real semisimple, noncompact Lie group
without center. Let r be a lattice in G, and let M be the uon Neumann algebra of the and provide consequently a representation of the algebras with P = [M:N]. (SW
discrete group I'. Then M is a 111 factor. If n : G ,U(H) is an irreducible discrete
Section 2.8 and Theorem II.16,) From this follows
series representation of G, then T &ends to a representation of M on H, and
1r 2
Theorem 3.1.2. If N c M is a pair of 111factors, either [M:N] = 4 cos (r/q) for
dimM@) = C O V O ~d,,( ~ ) some integer q 2 3 or [M:N] E [4,m].
where d, is the formal dimension of 7. There is substantial overlap between Sections 3.2 to 3.5 and Sections I to I11 of
Comes' report [Con].
b
In Section 3.4, we consider a pair N c M of finite factors and we recall some aspects of F Let us now describe Sections 5 to 7, where we consider a pair N C M of finite von
the original work [Jol] on this subject. First the of N in M is now defined to be g
Neumann algebras with pnite dimensional centers. There are projections pl,' .,p, which
!li.
128 are centrd in M and projections ql,. ..,qn which are central in N such that
plM, + .,pmM, qlN,. .,qnN
unapter 3: Finite von Neurnann algebras
M and T =. !T
Then there exists a pair N c M as above with A = AN
131
Moreover, M and
n m N may be chosen hyperfinite.
N .=. @ q.N C M= @ piM. ,
j=1 J i=l
The SkolemNoether theorem does not hold for 111factors and Proposition 2.3.3 does
If dimC(M) < m, this is the situation of Chapter 2. At this stage, let us assume that each not w r y over to the present setting: the matrices !A and T
: do not characterim N
)f the factors piM, qjN is of type Ill (see the comment after 3.5.4). as a subalgebra of M.
M Once a faithful trace is given on M, <thefundamental construction gives a new algebra
As in Section 2.3, we define an index & AN = (Ai,j) E Matm,,(R+ U {a}) by
(M,eN), just as described above in the case of factors.
 [ O ifpiqj=0
1,j  Proposition 3.1.4. Let N be offinite indez in M.
[ M : N~ ~j]112
,~ ifnot (a) The algebras (M,eN) and ~ n d i ( ~
are) isomorphic,
(b) The algebra (M,eN) is again a finite sum of 111 factors. There is a natural
1J
.
vhere N. = p.q.N is a subfactor of the factor M. . = p.q.Mp.q.. We say that
1 J IJ 1 3 1 J
N is of
een the minimal central idempotents of N and those of (M,eN),
inite index in M if A
: doe8 not have any infinite entry.
For an analysis,of traces on M and N (see Section 2.4 when dimc(M) is finite), we A convenient isomorphism is described in Corollary 3.6.5, and the bijection of (b)
Lefine also the &racematrix TM N = (ci,$ E Matm,,(R+) by appears in Proposition 3.6.l.i~.
The partid description of N c M by !A and !T is useful because, if N is of
here tr denotes the normalized trace on the factor piM. 4 trace on M is described
piM
L Mt
y the vector E ( R + ) ~with si = tr(pi), and its restriction to N by the vector A,=(A,) T ~ =N F
N ~ T ~
!T E ( R + ) ~ .Traca are always assumed to be positive in this chapter, so that si 2 0 for
E Mat,,,@+) is defined by
= 1,. .,m.
If dimE(M) < M and !T
the matrices AN are simply related by c i j = ~ ~ , ~ u ~ & ~ ,
itb py = dimc(piM) and u? = dimC(qjN). This relation has no analogue when the
J
,M's and the q.N1s are factors of type 111:
J
and where !F is a diagonal matrix e n s u i q that L
(TM)j,i)jd=l for j=l,..+,n.
Proposition 3.1.3. Consider two irredundant matrices 1s i i m
See Propositions 3.6.6 and 3.6.8 for the details.
.
l
} ~ ~ ~and T = (e. E Matqn(lR+)
A =(A. .) E Matmp({2 C O S ( % / ~U) [2,m]) As in Chapter 2, a trace on M is said to be a Markov t r a of p o d u l ~B for the pair
IJ
N c M if it extends to a trace tr on (M,eN) for which
tisfging:
= 0 ~ c . =. 0 and
1,J
c i j = 1 for i E (l,...,m) . P tr(xeN) = tr(x) x E M.
li j ~ n
There exists at most one such extension. As traces are positive in this chapter, 8, has to
be a positive number. The analogues of Theorem 2.1.3 and 2.1.4 hold as follows. Recall
that a pair N c M is s@g&$& if the intersection Z(M) n Z(N) of the centers is reduced
to C1.
Moreover [M:N] = 1 1 ~ 1 =1 ~4 cos2(lr/h), where h is the Cozeter number. (See tables
m n 1.4.5 and 1.4.7.)
Theorem 3.1.5; Let M = @ piM and N = @ q.N be finite direct sums of 111
(b) If [MN = 4, then :A corresponds to one of the graphs
i=l j=1 J
pctors, let N be a subalgebra of M offinite indez, and write T, T for TM M
N , TN.
(a) Let t r :M +C be a trace, let 5E R? be defined by q = tr(pi), and let @ ER;.
Then tr is a.Markov tl'ace of moddw P for the pair N c M if and only if
so that [M:N] = 1 1 ~ 1 1 ~ (See
. tables 1.4.6 and'l.4.7.).
(b) Ifthe conditions of (a) hold, then the Markov dension (My%) +C of tr is a The index range described by Theorem 3.1.2 appears also in the remarkable family of
Markov trace of modulus P for the pair M c (M,eN). W , which are discrete subgroups of PSL(2,R) generated by two parabolic
\ 
transformations. We have included an Appendix I11 on these groups. Its purpose is to
(c) If N c M is connected, there ezists a unique nonalized Markov trace on N c MI expose the spectacular comparison with Theorem 3.1.2 as well as to illustrate Section 3.3.
and its modulw P is the spectral radiw of TT.
Comparing Theorems 2.1.4 and 3.1.5, we may define the index of N in M as 3.2. The coupling constant: definition.
[M:N] = p ( ~ T ) , Let H be a (complex) Hilbert space. We denote by B(H) the *algebra of all
where p denotes spectral radius. bounded operators on H, with x* the adjoint of the operator x E B(H). :Besides the
I topology associated to the norm
Corollary 3.1.6. Theorem 3.1.2 holds for finite direct s u m of 111 factors.
We note that the definition of [M:N] given above is not the same as that of Chapter
2. However, P. Jollissaint has shown, in unpublished work, that the two definitions of the algebra has also the ultraweak $@x&gy or wtopology which is defined by the
seminorms
index coincide.
If N C M is a connected pair of finite dimensional multimatrix algebras with
[M:N] s 4, we have shown in Theorems 2.1.1 and 1.1.2 that the corresponding graph is a
Coxeter graph of one of the types A,D,E. The chief result of Section 3.7 is that connected
pairs N c M of'finite direct sums of 111factors with [M:N] 5 4 give rise to possible
Coxeter graphs associated with finite and affine groups.
that p is continuous with respect to the wtopology on both M1 and M2 (corollary theorem, there is a net (x$ in M with Ilx,ll s 1 for all cu such that Ir(xcu) converges
5.13 in [SZ] or section 1.4.3 in [DvN]). A ~ Q QNeumann &g&& is a *algebra M which is strongly to a; that is, n(xcu)[ converges to a< for all 4 E H. Taking < = 1, this means
*isomorphic to a von Neumann subalgebra of B(H) for some B; by the result just that (r(x&) is a Cauchy net for the )I.))2distance, so converges with respect to this
recalled, such an algebra has a welldefined wtopology.
A f a is a von Neumann algebra M with center ZM reduced to the scalar distance to some element r(xo) by the assumed completeness of the ball of M. One can
multiples of the identity. Von Ne algebras axe known to be principal in the sense check that the strong topology and the 11. 112topology coincide on the unit ball of Ir(M)',
that any wclosed twosided ideal is generated by a central projection (see section 1.3.4 in so a = r(xO) E r(M). #
[DvN]). Thus a von Neumann algebra M is a factor if and only if m y twosided ideal
J # 0 in M is wdense. There is not any continuity problem for representations of a
Let M be a finite factor acting on some Hilbert space H. We are going to define the
factor M in the following sense: any *homomorphism M + B(H) is wcontinuous.
Coupling constant dimM@) which is a measure of the size of H aa an Mmodule, the
(See theorem V.5.1 in [Tak]; the separability of H is crucial here.)
A HI&g&g is an infinite dimensional factor M which admits a normalized finite definition b n g made so that the standard Mmodule L ~ ( M= ) ~ ~ ( ~ , has
t r size
) 1.
Before comparing other Mmodules to that one, we recall the following facts.
trace tr : M 4 C such that
(i)
(ii)
(iii)
tr(1) = 1
tr(xy)=tr(yx) x , y ~ M
tr(x*x) 2 0 x E M.
It is known that, on a 111factor, such a trace is unique in two senses. First, in the usual
den+ r3
Lemma 3.2.2. (a) Let J : L 2 ( ~ ) L 2 ( ~ ) be the conjugate linear isomety which
x w x
+
B(H) for some H of finite dimension. Such a factor is simple as a cbmplex algebra by EUMf, (a) Let x,y,z E M. By definitionof J
[DvN], 111.5.2. Here is a characterization of finite factors; for more on this, see [KvN].
JxJy = (xy*)* = yx* = yJx.
m o s i t i o n 3.2.1. Let M be a c*algebra with unit and with center reduced to the
scalar multiplies of 1. Let tr : M ,C be a faithful normalized trace (namely b linear form
Applying this twice we get .
satisbing (i), (ii), (iii) above and tr(x*x) > 0 for x # 0). Assume that the unit ball of M
is complete with respect to the metric d(x,y) = IIx~(~,where = (tr(x*x))lJ2. Then
JxJyz = yzJx = yJxJz,
M is a finite factor. and setting z = 1,
(JxJ)y = y(JxJ).
&g& Let H = L 2 ( ~ , t r ) be the Hilbert space obtained by completion of M with
respect to the scalar product defined by cxly> = t r ( ~ * ~ )for x,y E M. Let Thus JMJ c MI where M' = E ~ ~ ~ ( L ~ ( M ) ) .
r :M 4 B(H) be the *representation of M on H, with r(x) being the extension to H
of the left multiplication by x on M. Then r is injective because tr is faithful,. Let
.
Let moreover a E M' By definition of the adjoint
4M)' denote the double commutant of M in B(H), which is, by von Neumann's
(Y*x* la) = (x* lya) =(x* lay) = {a*x*l y) = (x*a*l y) = (a*lxY).
bicommutant theorem, the wclosure of r(M) in B(H).
To show that M is a von Neumann algebra, it is enough to show that the inclusion of
M in n(M)' is surjective. Let a € 4M)' with JlalJ= 1. By Kaplansky's density
 i
so that J a = a*. Thus the first computation shows also that JM'J C M' and, taking
adjoints, M' c JMRJ.
By von Neurnann' s bicommutant theorem, one has M' = JMJ.
(b) Let x E B(L~(M)@ K). Choose an orthonormal basis (q)iQ of K, and . For example,
repreah x by a matrix (3,I.)i,jEI over B(L~(M)).If x commutes to the aWm of M, TrMt (JxJ@p) = trM(x) dimC(pK)
this matrix must have entries in EndM(L2(M)), and thus x E EndM(L2(M)) @ B(K).
Conversely any bounded matrix ("ij)with entries in E ~ ~ ~ ( L ~ (commutes
M ) ) with the if x E M+ and if p E B(K) is a projection.
~ : H  + L ~ ( M ) ~ K
which intertwines the actions. # for x E EndM(H)+, and thus * E EndM(L2(~)@ K)+. If ul,% are two possible
ChoiC€!~for U, then u;ul = ugUz = idH and urd%
= for X M; as TIM,
As there will be many traces with various normalizatiob in the sequel, we introduce is a trace,
the following convention. If M is a finite factor, trM will denote its normalized t r m .
So if TI is any other trace on M, then TI = Tr(l)trM, a formula which we will use
often. Occasionally, we will have to consider a trace TI on a factor P which is not finite
(for example B(H) or M@B(H), with H of infinite dimension). Let P+ denote the and TIM, (x) does not depend on the choice of u.
positive cone of P, consisting of those element of the form z*z with z E P. Then a trace The word "natural" is justified by the following property (which again shows the
TI is a map P+ , [O,m] such that independence just observed).
(i) Tr(x+y) = Tr(x) + Tr(y) x,y E P+
Lemma 3.2.3. Let HI,H2 be two Mmodules; let a : HI I H2 and b : H2 ,H1 be
(ii) Tr(Ax) = XTr(x) X E R+, x E P+ (with 0.m = 0)
two Mlinear bounded operators. Denote bg T. l e natural trace defined on EndM(Hj)
(iii) TI(*) = Tr(x) x E P+, u a unitary in P. J
as above, for j = 1,2. Then
Given a finite factor M acting in a Hilbert space H as in Lemma 3.2.2, we define
now the paturd t r a o TIM' on its commutant. It is crucial for what follows that TIM,
is not necessarily normalized.
First, if H = L2(M) as in (a), we define TrM, (JxJ) = trM(x) for all x E M; in this
ease, TIM, is notmalied. Secondly, if H = L ~ ( M@
) K as in (b), consider an
tj 3.2. Coupling constant: definition 139
TI : EndM(H)+ + [O,W] is defined as above. If u is as in 3.2.2.~~one has also trace. In case EndM(H) is infinite, they define cM = +w.
dimM(H) = TrM, (uu*) by 3.2.3. The Mmodule K gives rise to othm modules as follows. Let e E B(H) be a
projection (e # 0), with range denoted by eH. If e E EndM(H) then eH is naturally a
Proposition 3.2.4. Let M be a finite factor and let H,Hf ,H1,H2,. .. be Mmodules Mmodule (a submodule of H); if moreover EndM(H) is finite, the value D(e) of the
which are separable as Hilbert spaces. Then normalized trace of EndM(H) on e is called the dimension of e. On the other hand, if
(a) dimM(H) = dimM(H1) if and only if H and H' are isomorphic as Mmodules, e € M, then eH is a eMemodule; the algebra eMe is a finite factor (because it is
(b) dimM(?Hi) = x d i m M ( ~ i ) ,
x.,
simple, a fact easy to check) which is called the reduction of M by e. Following common
1
i practice, we also write Me for eMe.
(c) dimM(L2(M)) = 1,
(d) dimM(H) < m if and only ifthe factor EndM(H) is finite. Proposition 3.2.5. Let M be a finite factor and let H be a Mmodule. Assume that
the factor EndM@) is jnite (namely that dimM(H) < w). Then

Proof. Claim (a) follows from the comparison theorem for projections in the factor. (e) dimM(eH) = D(e) dimM(H) for any nonzero projection e E EndM@).
E ~ ~ ~ ( L @~K),
( Mclaim
) (b) fmm the oadditivity of the trace T r M on the same (f) dimM(H) = cM, the coupling constant ofMurray and von Neumann.
I 
1
factor, and (c) is obvious.
In all cases, EndM(H) is a semifinite factor, and thus admits a nonzero trace which
is unique up to a multiplicative constant. Claim (d) holds because EndM(H) is finite if (h) dimeMe(eH) = 1 dimM(H) for any nonzero projection e € M, where
The coupling constant of Murray and von Neurnann for M and EndM(H) is
= trM(et)/trEndM(H)(ei)*
Let C E H with ( # 0 and let q E L2(M) @ K with q # 0. As earlier, denote by e E M
C
and e' E EndM(H) the projectiom of H onto
C Endlyl(H)i and m. Likewise denote by
since we are assuming that EndM@) is finite. By uniqueness of the normalized trace on
f RE M and f' E EndM(L2(M) @ K) the projections of L2(M) @ K onto
11 EndM(H), one has
EndM(L2(M)@K)qand fi. With respect to the orthogonal decomposition L2(M) @ K =
H 4 HI, the algebra M acts by operators of the form 6 9, the algebra d
(3.2.5.5) trEnd H (PXP) = Tr
M( )
(x) + Tr (PI
E.~~(L~(M)~X) E ~ ~ ~ ( L ~ ( M ) @ K )
fC= [:(:1, so that it is the same element in M which acts as f on L2(M) @ K and
C
= {tr (e )
6 pEndM(L2(M)@Kfi)' TrEndM(L2fM)sK)
(PI
as e on H. Consequently
C
(3.2.5.2) trM(ft) = trM(et).
1 and claim (f) is proved.
Observe also that, more simply Claim (g) now follows trivially from (f). As for (h), using (e) and (g) as well as
EndeMe(eH) = e(EndM(B))e, we have
3 I The f o n d dimension dr depends on r and on the choice of the Haar measure for
If M = Mat (c) for some integer p 2 1, then dimM@) =
d
P
I'
of H in the naive sense. For all this, see section 16 in [Rbt] or Chapter 14 in [DC*].
Given an arbitrary (unimodular) group G, its discrete series may be empty. This
happens for G infinite abelian, or infinite discrete, or G = SL(2,C), or G = SL(n,lR)
with n 1 3, to quote but a few examples. When G is a semisimple connected real Lie
3.3. The coupling constant: examples. group with maximal compact subgroup K, then G has discrete series representations if
and only if G and K have the same rank. In particular SL(2,lR) has a discrete series, as
The situation for which the coupling constant is computed in this section is of the well as ~ 0 ( n , l ) Ofor n even.
following kind: G is a noncomp&t semisimple connected real Lie group which has the
same rank as its maximal compact subgroups, r : G ,U(H) is an. irreducible 3.3.b. Factors defined bv icc vrouns.
representation of G in the discrete series, and M = w*(I') is the von Neumann algebra
of an appropriate discrete subgroup I' of G. Then H is naturally an Mmodule. On a discrete group I', we consider always the counting measure; the space of square
Theorem 3.3.2 below is a computation of dimM(H), due to AtiyWchmidt [AS,(3.3)]. summable functions from 'I to E is denoted by ?(I?). The von Neumann algebra
w*(I') of 'I is the (u1tra)weak closure of the linear span of +(I?) in B ( ~ ~ ( I ' ) )by
; von
First we discuss some background; the knowledgeable reader should jump to Theorem
3.3.2. Neumann's theorem, it is also the bicommutant of Xr(r) in B ( ~ ~ ( I ' ) )and
, w*(I') is
thus also denoted by A#)".
3.3.a. Discrete series,
Let 6, c.t2(r) be the function which takes the value 1 at the identity e of I' and 0
Let G be a locally compact group. We assume that G is unimodular,.we choose a elsewhere. It is easy to check that x x(b,) is a linear injection of w*(I') in ?(I?),
Haar measure dg on G, and we denote by lG : G u ( L ~ ( G , ~ ~the
) ) left regular ~ ) is a normalized finite faithful trace on w*(T); see
and that the map tr(x) = ~ x ( 6(6,>
representation of G. \ the end of 4.2 in [Sak . It follows that the von Neumann algebra W*(I') is finite, and that
For an irreducible unitary representation r : G 4 U(H) of G, the following 1
the Hilbert s p w L (W*(r),tr) defined before Lemma 3.2.2 is canonically isomorphic to
properties are equivalent: e2g).
r is a subrepresentation of XG; more precisely, there exists a projection p in
8
(i) Moreover w*(I') is a factor (and thus a factor of type 111) if and only if I' is an
the commutant of XG(G) such that the restriction of XG to the range of p is equivalent infinite coniueacy group, or for short an icc m o u ~(Lemma 4.2.18 in [Sq). The
to Ir; following lemma exhibits a rich class of icc groups. Before this, we recall that the quotient
(ii) There exist b q E H  {O) such that g I+ <li(g)gl q> is in L ~ ( G , ~ ~ ) ; G/I' of a unimodular locally compact group G by a discrete subgroup I' has always a
For all S,q E H the function g I+ <n(g)( ( q> is in L ~ ( G , ~ ~ ) . Ginvariant measure, which is unique up to a scalar factor; by definition, I' is a &t&g in
(iii)
I, G if the meaure of G/I' is finite.
If these hold, r is said to belong to the (unitary) discrete seria. On may then attach to
r a real number d r > 0, called its formal dimension, such that Schur's orthogonality Lemma 3.3.1. A lattice I' in a connected semisimple r e d Lie group G vithout
relations formally hold. In particular, for any r : G + U(H) in the discrete series center and without a compact factor is an icc group, and W*(I') is consequently a
111factor.
IYY ,y'
r
~  Y cnaprer 5: r lnlre von lveumann ageoras /+*.a. coup~ingconstant: examples ,JAE,

Proof. The main point is Borel's density theorem, which we quote without proof (see L'(D,W ~ e n c ethe von ~ ~ u m malgebra
n XG(I')' is isomorphic to w*(r) o c g
[Bar] or [Zim]): I' is Zariskidense in G.
Consider h E I' and its conjugacy class Ch in I'. The m&p
i7;;= Ch and
w*(r).
More generdy, let p E B ( L ~ ( G , ~ ~be) )a projection which commutes with AG(I').
Denote by H the range of p, by r .'I + U(H ) the corresponding subrepresentation
P P' P
of XGlr, and by r (I')' the von Neumann algebra generated by r (I?) in B(H ).
P P P
{g E Glgh = hg} is a closed subgroup of finite index in G. But the algebraic group Then the ~*mor~h.ism (b(I')' pxr ~ ( r ) '
r ia obviously surjective. 1f I' is moreover
corresponding to G is Zariski connected, and it follows that {g E Glgh = hg) = G.
an icc group, then XG(I')' g w*(r) is a factor of type I I ~and is in particular a simple
Thus h is central in G, so that h = e. This shows that I' is an icc group. #
ring, so that the map XG(I')' t r
P
(r)*is an isomorphism.
A final remark about this: let rl c and r2c G2 be two examples of the We shall particularize below to the case in which the projection p commutes with all
situation in the previous lemma. Assume moreover that G1 and G2 have real rank at of XG(G), and definw an irreducible representation of G in the discrete series.
least two. It is a conjecture, due to A. Connes and "beyond Mostow and Margulis", that
w*(I'~) is isomorphic to w*(r2) if andonly if rl and r2 are isomorphic.
3.3.d The formula dimM(H) = covol(I')g,
3.3.c. w*(I')modules associated to subre~resentationsof A . Now the relevant background has been established, and we demonstrate the main
result of this section.
Let G be a unimohlar Lie group with Haar measure dg and let I' be a discrete Theorem 3.3.2. Let G be a connected semisimple real Lie group with Haar measure
subgroup of G. In the present context, it is convenient to define a fundamentalhmaiq for 4
r in G to be a subset D of G which is measurable and satisfies dg, let r be a discrete subgroup in I', k t M denote w*(I') and let r : G + U(H) be
an irreducible representation in the discrete series. Assume that is an icc grou$. Then
7 , n~ r 2 has
~ nu^ measure for 71.'y2 E r vfth 71 # 72 and dimM(H) = covol(I')dr.
By Lemma 3.2.2.b, 'this cornmutant M' is generated by finite sums of the form orthonormal basis (67@en } F r l n a of t2(l') @K is more conveniently viewed as the
X = C p 7 @ a r For each 7 E I?, the symbol p 7 stands for JXl'(dJ E E ~ ~ ~ ( L ~ ( M ) ) basis {AG(7)J%r,na of L ~ ( G , ~ Let
~ ) . rl be a unit vector in L2(G,dg); assume that
YE^ :
E H, namely that pq = q. For any g E G one has (writing X instead of XG)
and a is a finite rank operator in B(K). Let (en)na be an orthonormal basis of K.
7
Let @ 5, denote the operator [ I+ (eml ()en on K. One may write
a =
7
a7,m,n <@
en, where the a
7,m,n
are complex numbers. By definition of
m,nEiN
TrMt one has Consequently, as p commutes with X(G):
where trM is the normalized trace on M and where TK is the trace on B(K)
normalized by T K G @ f m )= 1 for all m E I. With x as above, one has consequently
By Schur' s relations
Let q : L2(G,dg) ,K be the orthogonal projection given by restricting functions from
G to D, and let T denote the trace on B(L~(G,&)) taking value 1 on projections of
rank one. Then TK(y) = T(qyq) for y E B(K)+ or y E ?(K). In particular, for x of
the form x = p @ a we have
C 7 7 \ and the proof is complete.
7 a #
I
irreducible discrete series representation. (These rk consitute the holomorphic discrete
Recall that 6 denotes the characteristic function of (7) in,'l and that en, which I series, and the full discrete series contains a second "half", the antiholomorphic part.)
7
is a function on D, is also naturally a function on G (vanishing outside D). Thus the Ig Define the Haar measure dg on G as follows: let T = S0(2)/{*1) be the maximal
compact subgroup of G, such that {! induces a diffeomorphim G/T g 1 then
1
!
* *" ,"*\
r
"YY,,"". V. L *IYYW .VY *."UYIUYY U ~ W V I I V
ddz) = ~  ~ d x d is
y a Ginvariant measure on 'P, if cp is a continuous function G t C
with compact support, set f(z) = (cz+drP f [a],[E11 z E 7, E SL(2,Z).
' JGdg)dg = @(z)Jydldgt) Z = b(i) The second one is a growth condition: observe that f(z) = f(z+l), so that f can be
defined on the punctured unit disc A* by ~ ( e = f(z);~ ~ the~ second
) defining condition
is that the Laurent expansion of f in A* is of the form ?(w) = x a n w n for w E A*.
where dt is the Haar measure on T of total measure 1.
Then the virtual dimension of rk is known to be given by dk = $$ see theorem ni 1
It is a result of Hecke that a cusp form f of weight p satisfies If(x+iy) 1 d B ~ for ~ / ~
+
all x iy E 7 and for some constant B; see page 1.24 in [Ogg].
17.8 in [Rbt]. (Warning: under the Cayley transform Let M = w*(I'). Consider an integer k 1 2 and the Mmodule Hk of example
3.3.4. Given a cusp form f of weight p, the growth condition implies that f induces a
multiplication operator Af : Hk 4 HkSp, defined by (Afcp)(z) = f(z)dz), which is
chosen in the present section is that which is defined by the Riemannian structure for bounded (in fact IIAfll 5 B with B as above). The invariance condition implies that Af
which 'P has constant curvature 1; the computation may be found, for example, in is Mlinear. Consequently, given two cusp fonns f,g of weight p, the operator
Section 5.10 of [Car].) t EndM(Hk). Let Tk denote the natural trace on
Now consider an integer q 2 3, set X = 2cos(lr/q) and let rXbe the Hecke subgroup
EndM(Hk). Then the space of cusp forms of weight p has a natural hermitian form
of PSL(2,lR) generated by the classes modulo *l of the matrices
A computation in the same spirit as that presented in the proof of Theorem 3.3.2 shows
Then cov01(l'~) = 4
1): by the GaussBonet formula, because has a triangular.
fundamental domain with angles 0,z,z
9q
(see Appendix 111). (f ldl,= + k m g ( z , ~ p 2 d x d ~
Altogether, we have shown:
. with D a fundamental domain for I' in P' . Up to a constant factor k;i,
1 this is known
Example 3.3.4. Given integers q>3 and k 2 2, consider the 111factor
as the Peterson scalar product for cusp forms.
M = w*(rX) defined by the Hecke group rAwith X = 2cos(lr/q) and the Hilbert space
Hk of the holornorphic discrete series of PSL(2,lR). Then Hk is a Mmodule of coupling This suggests a natural project, which could be interesting for the study of cusp forms:
constant evaluate ~ L n o r m sdefined by
Ilflkq= { ~ ~ l (d ~2 ; l/q.
~ ~ ) >
We particularize further, and set q = 3 in example 3.3.4. That is, we consider 3.4. Index for subfactors of 111 factors.
r = PSL(2,Z) as a discrete subgroup of PSL(2,lR).
Given an integer p 2 1, recall that a cusp form of weight p is (in this situation) a
There were two main motivations for the introduction in [Jol] of the concept of index
holomorphic function f : 'P 4 C on the Poincar6 halfplane satisfying two conditions. The for subfactors. The first was that, if rl < I'2 are two icc discrete groups, the 111 factor
first one is an invariance:
IDU u a p t e r s: rlIllte von NeUnUmU algebras . 3 3.4. Index for subfactors
N = A(rl)' acts in an obvious way on t2(r2) and dimN(t2(r2)) = [r2:r1]. Theorem 3.4.3. Let N be a subfactor of a 111 factor M.
Furthermore t2(r2) is the same as L ~ ( M )where M is A(r2)'. This suggested the (i) Either [M:N] = 4cos2r/q for some integ~rq 2 3, or [M:N] 2 4.
following definition: (ii) If [M:NJ < 4, then N is automatically irreducible in M .
(iii) There ezist subfactors of the hyperjinite 111 factor R with any of the indec values
Definition 3.4.1, The index pf pubfactor N ef a fulite f w &$ is allowed by (i).
(iv) There are ezamples,of factors M for which the set of all possible values [M.N] is
countable.
This was the original definition of index; it was ahown in [J&] that this definition agrees Remarks: Statements ji) to (iii) are from [Jol]. We prove (i) below. A generalization
with the ringtheoretic one which we have given in Chapter 2, when M and N are finite to finite direct sums of 111 factors is shown in Corollary 3.7.6. A second proof of (i) occm
factors. The index C a n also be computed as IM:N] = dimN(H)/dimM(H), where H is in Corollary 4.6.6, as a byproduct of the analysis of "derived towers".
any Mmodule of finite dimension over M; see Proposition3.4.6. Statement (ii) is proved as Corollary 3.6.2(c).
The second motivation was a result of M. Goldman ~Gol],who showed that, if N c M We will verify Oii) by giving 'several constructions of subfactors of R. The first
are 111 factors (always with the same identity R) then, if % ( L 2 ( ~ ) ) = 2, there is a construction, in this section, works for all allowed index values. Another construction,
valid for the index values 4cos2r/q is given in Theorem 4.4.2. A third construction, in
crossed product decomposition M = N r 2/22. Consequently if one defines [ M a ] as
Section 4.5, produces irreducible pairs; the index values 4cos2r/q are obtained once more,
above, Goldman's result is seen to be a beautiful analogue of the fact that a subgroup of
as well as sporadic values greater than 4. In Section 4.7.d, we give examples of
index 2 of a group is normal.
honeonjugate irreducible subfactors of R of index 4. We would also like to mention the
It would also have been nice to have been able to call a subfactor N c M, normal
work of Wenzl [Wen2], in which a family of irreducible subfactors of R of index greater
when its (unitary) normalizer generates M. But unfortunately standard terminology
r than 4 is produced by a construction involving the Hecke algebras Ha(q) for q . a
reserves "normal" for subfactors N such that (N'nM)' n M = N, and the term a is .
used for subfactors with the normalizer property described above. We take this primitive root of unity.
opportunity to introduce some more terminology. Statement (iv) is from [PP2], md will not be proved here.
For arbitrary 111 factors, the question of existence of subfactors of index 4 eos2rJq
k
Definition 3.4.2. If N c M are factors we say that N is breducibl~ N' fl M = C remains open, more precisely we know of no example of a full 111 factor M having a
subfactor of index 4 cos2a/q, q # 3,4,6. (A 111 factor is called "full" if the group of inner
It is not hard to see that a regular irreducible subfactor has integer index (or w which J
,
we shall treat as an integer) see [Jo~].A more refined analysis based on [,To61 shows that automorphisms is closed in the topology of pointaise strong convergence in the whole
all regular subfactors have integer index. On the other hand dimM@) can be any automorphism group  an example of such a factor is X(PSLf2,2))'.)
positive real number so the question naturally arose:
Proof of 3.43 {i). As for finitk dimensional algebras (2.6.2) there is always a (unique)
faithful tracepreserving conditional expectation from M onto N, which, viewed as an
(a) What are the possible values of [M:N]?
operator on L2(M) is the orthogonal projection % onto L 2 ( ~ ) . The fundamental
{b) What are the possible values of [MN] for an irreducible pair N c M ?
construction again yields a 111factor
Question (a) was settled completely in [Jol] for M = R, the hyperfinite 111 factor.
Question (b) remains open even for M = R, and question (a) is open for arbitrary 111 end;(~~gn,tr)) = E~~;{MI= (M,eN).
Indeed, the linear fprm defined on N by x w tr(eNx) is a trace (3.6.l.iii). As 1 = satisfying the relations 3.4.3.1, and a positive faithful normalized trace t r on B
2 P,k
tr(eN) [M:N] by Proposition 3.2.5.e applied to the Nmodule L (M), the property is satisfying the relation 3.4.3.2 for 1 5 j 5 k. Since tr is faithful, the trace representation
valid for x E N, by uniqueness of the normalized trace on N. But then for x E M, we
rtr is faithful as well, and we can take R to be u ~ ~ ( u B ~ , ~ ) ' .
have [M:N]tr(eNx) = [M:N]tr(eNxeN) = [M:N]tr(eNEN(x))= tr(EN(x)) = tr(x), wing A simpler procedure is available when P is the square of the norm of a nonnegative
3.6.1.i. integer valued matrix (i.e. @ E ,&(IN).
dimens&
In this case there is a connected pair of finite
c*algebras I3 c A with [A:B] = P, and the tower construction for this pair
Now the tower construction of Chapter 2 works and yields an increasing sequence of
yields a sequence of projections (ei)i,l satisfying 3.4.3.1, and a positive faithful trace on
111factors
Mo=NcM1=Mc .cMk~Mk+lC;.., alg {&el,. ,.) satisfying 3.4.3.2. Cf. 2.7.5 and the discussion at the end of Appendix IIa.
Lemma 3.4.4. [Jol] With the notatioa above, R i s the hyperfinite 111 factar.
and a sequence of selfadjoint projections (ei)i2 satisfying
llEgPfr It is clear that R is a finite, hyperfinite von Neumann algebra. We claim that
(3.4.3.1) P 9eiklei = ei
if z is in the center of R, then
e.e  e.e.
I j J 1
if 1 ij 1 2 2,

= ltr((z Y)%+! (XI)I + I(tr(y)  tr(z))tr(x) l
5 2 6 llxl12. It follows from the relations eele = T1e and eleel= /Tiel that e and el are
Since c is arbitrari, this finishes the proof. #
b
equivale t projections in (R,e). Since e is finite in (R,e) by 3.6.l(v), the projection el
is finite in (Ke). But 1 is the sum of finitely many projections each equivalent in R to
Lemma 3.4.5, [Jol] Set R
P {P,e2,e3,. ..)' . Then [R:RP] = P. a subpr~jectionof el, so (R,e) is finite. Hence [RRP] = tr(e)" = tr(e1)" = 8.
This completes the proof of the lemma, and also of 3.4.3(iii). #
We know (by 2.8.5 and 2.8.7 or by 2.9.6) that for each k > 2 , the
relation P tr(elx) = tr(x) holds when x E alg {P,e2,, ,.ek), and, taking limits, we have It is tempting to guess that the pair R 3 R
P is irreducible, also for P > 4, since on a
the same relation also for x E Rg Therefore ER (el) = 8'1. purely algebraic level it is easy to see that there is no element of the algebra generated by
P e l } which commutes with {e2,e3,. ' v } .V. Jones confesses to spending
Similarly EN(eZ) = B11, where N = {l,e3,e4,. :I*. For k ? 3, any
considerable &fort to prove this, but it turned out that RP has nontrivial relative
.
x E alg {l,e2,. .ek} is of the form x = a + bigci, with a,bilci E dg { I q , .ek}.. commutant in R when p, 4. A laborious proof of this nonobvious fact was given in
i
[Jol] and a simpler proof in [PPl]; we will give a proof due to Popa in 4.7.5. The
Consequently, EN(x) = s +flCbici and elxel = EN(x)el. Taking limits again, we
'
difficulty is that m e cannot write down an explicit form for an element in RP' tl R
i
have without invoking a beautiful representation of {el,e2,. '1' discovered by Pimsner and
Popa:
(*) elxel = EN(x)el for all x E RP. We have seen that one way to obtain a sequence of projections (ei)iL1 satisfying the
' relations 3.4.3.1 is to form the tower from an indecomposable pair B c A of finite
One next' verifies that xel = fi ERJxel)el, for all x E R, by first checking this for dimensional c*algebras. Then, as we have observed in Chapter 2, the restrictions on
P index are related to restrictions on the type of inclusions B c A which yield a modulus
x E alg {l,el, .q) (that is, for x of the form x = r + xbielci, \with a,bi,ci E p < 4. This is where the Coxeter graphs of types A, D, and E enter the picture. But to
i meate the sequence (ei)i21 one can also use a pair N c M of finite direct sums of
.
alg {keg,. .ek)) and then by taking limits. Consequently , ,bl= RP el, and RelR
111factors. In the following sections we will see how, if one allows this extra freedom, the
=RP elRP. Observe also that R = RelR, because finite factors are algebraically simple
remaining Coxeter graphs appear!
([DvN], Cor. 111.5.3).
2 2
Let e be the orthogonal projection of L (R) onto L ( R d . One has exe = ER (x)e We finish this section by recording one useful fact on indev of subfactors from [Jol].
D
for all x E R, by 3.6.1.i. below, so that in particular, eele = T1e. We claim that also
Pronosition 3.4.6. Let N c M be and let H be aay Mmodule such that
eleel= riel. Sine! R=R e R
P 1 P' it suffica to check this equality on vectors xelyfi,
dimM(H) is finite. Then [M:N] = (, )i(In paPtimlar, dimN(H) 1 dimM@).)
where x,y E R and fl is the trace vector for R But m~
P
EreQfc If HI and Ha are any two Mmodules such that dimM(Hi) is finite for
eleel(xel~n) = eleelEN(x)~n (by (*)I
i = 1,2, then there is a finite dimensional Hilbert space K and an Minvariant
= e1ER$elEN(x)y)fi (by definitions of e and E
: projection q such that H1gq(H2@K) as Mmodules. Then
= e1ER$4)EN(x)y" (by R p e a r i t y of ER. ) dimN(H1) < dimN(H2 @ K) = dimN(H2) dim@), by 3.2.5(i), so dirnN(H1) is finite if
P
= 8' elEN(x)yfi and only if dimN(H2) is. In particular, [M:N] is finite if and only if dim@) is.
(by (*I).
1
156 t
p~'p''p'
I
vnapcer 3: r m c e von ne iugeulaa R%3 J.O. UCIWIOIIS 01nmte von lveumann agebras (57
Assuming that [M:N] is finite and choosing an Mmodule isomorphism Recall that if P is a finite factor, trp denotes its unique normalized trace, and if TI
H 2 q ( L 2 ( ~@) K), as above, we have is any other trace on P, then Tr = Tr(l)trp.
n
dimN(H) = d i m N ( q ( ~ 2 (@~K))
) Let N = @ N. be another direct sum of finite factors, contained in M and having
j=1 J
= trNJ(q) d i m N ( ~ 2 (@~K)
) (by 3.2.5(e)) ..
the same identity. Let ql19 ,qn be the minimal central projections of N.
= tr,. (q) d i q ( K ) dimN(L2(M)) (by 3.2.5(i)),
Definition 3.5.1. If N c M are as above, we d&ne the mbyn matrix T M
N = (cij)
while dimM(H) = trM, (q)d i ~ ( K ) . # by
4~ = trPiM(piqj).
c..
Proposition 3.5.2.
3.5. Inclusi0118 of finite von Neumann algebras with finite dimensional centers.
(i) The nzatriz T
: k rowstochastic; i.e., c.1,J. 2 0 and
We saw in Chapter 2 that a unital inclusion B c A of finite dimensional c*algebras J
can be specified by the inclusion matrix A E Matfin@) and a vector d E ~i~ for some n, (ii) If k apecipes a trace on MI then 5 !T spwijes its ndriction to N.
specifying the algebra B up to isomorphism. It is impossible to specify an inclusion so (iii) If N c M C L are fnite direct sums of fnite factors, then Tfi = T&T.!
precisely in the 111case since, for example, it is possible to find infinitely many
nonanjugate subfactors of index 4 in R, even irreducible ones, as we shall see in Chapter
4. What we will do is specify enough information to be able to calculate all the needed
coupling constants, which will enable us to find the Markov traces as in Section 2.7. (ii) A s ~ ~ ~ = l ,
The situation will differ in two ways from the finite dimensional case. The first is that i
there are no minimal projections around, so integers do not appear inThis way. The second
is that the subfactors can have indices different from squares of integers. This extra
freedom allows the appearance of new Coxeter graphs.
m
First some notation. Let M = @ Mi be a direct sum of finite factors with
i=l L is the
(iii) Let {rk} denote the minimal central projections of L, so that TN
..
corresponding minimal central projections pl,, ,pm. Since the trace on a finite factor iB matrix whose (kj) entry is tr, L(rkqj). Since q  p.q., one has
unique up to a scalar multiple, a trace on M is completely specified by a row vector 1 k iC
i 11
!
..
s = (slle ,sm), with si = tr(pi), ( Warning: This is not the same vector which was used
in Chapter 2 to specify a trace on a direct sum of finite dimensional factors; there we used
the vector whose ith component is the trace of a minimal projection in Mi.) A trace is
positive (i.e., trace (a*a) 2 0) if and only g has nonnegative components. We adopt t& But in the finite factor A = rkL, if esf are two projections, then
convention that ('trace" means 'Ipositive trace". A trace is faithful (i.e., trace (a*a) = 0
rn trA(e) = trA(f)trfAf(e). Thus
implies a = 0) if none of the components of 5 are zero, and normalized .if
T.
si = 1. A
1=
trace is automatically n o d , i.e., if {fi} is a family of mutually orthogonal projections,
m
m
then trace ( V fi) = ztrace(fi).
i=l i=l
158 . Chapter 3: Finite von Neumann algebras 3 3.5. Lnclusions of finite von Neumam algebras 159
If rkpi # 0, then. x w trr p.G. p.(rkx) is." trace on piM whose value at pi is 1, so in (iii) For any faithfil trace tr on M , the regular representation of M on
k l k l L 2 ( ~ , t r )is b finite representation of the pair N c M.
fact tr Lr p.(rkx) = trpiM(x). Hence (iv) For any faithjkl representation {lr, 3) of M such that 4M)' is finite, the
'kpi k I
. . . L
algebra r(N)' is also finite.

Proof. (iv) E=, (iii) * (i) is evident.
(i) 4 (ii). If T is a faithful finite representation of the pair N c M on H, then the
oommutant of lr(N. .) on x(p.q.)H is ir(p.q.)lr(N)'~(p.q.), which is finite. It follows
I rJ 1 J 1 J 1 J
as desired. # that dimN. .(lr(p.q.)H) < m (Proposition 3.2.4.d), and
1 J
1,J
A second piece of data needed is the matrix of indices of the "partial embeddingso'.
Note that Nij = Np.q. = {p.q.x : x E N) is a fidite factor, a subfactor of
1 1 1 J
M. . = p.q.Mpiqj.
l,J 1 J
.
!
Lemma 3.5.4. Suppose N c M are finite direct s u m of finite factors. The following N = { ~ ' E R y: = x + rp(x)forsomex€R ),and
P
are equivalent: M = R @ RIep.
P
(i) N is of finite index in M. Then
(ii) The matriz !A has only finite entries.
1.I A ~ A $= (1 I)[:] = 2,
,
%
a
,

+
, 9 a.0. l n e ~ u n a a m e n construction
t~ +=yl
and with [qlj: P.1 J.]'I2 = (A).1,J. (possible by [Jol], Theorem 4.3.2). For each i and j such
that (T). . # 0, choose an isomorphism 0. . :R + P. (possible since all the factors are
1,~ IJ 1,j
! n
111 and hyperfinite]. Set q. = z q . ., put Nj = {XdiJ(x) :x E R}, and N = e Nj.
by Corollary 2.2.5 of [Jol] or 4.7.2. These are not equal, unless tr(p) = 112. IJ
Of course, if N c M c L is a triple of finite m,then [L:N] = [L:M] [MN] by i i j=l
Proposition 3.4.6. Then q.N = N., and N is the required subalgebra. #
J J
If N and M are as in 3.5.4, and the inclusion N c M is connected, then all factors of 3.6. The fundamental mnetructioa
N and M are of type 111, or d i v ( M ) < CO. It is also known that all factors of N and M
The discussion of the fundamental construction in Chapter 2 was purely ring theoretic.
share (or do not share) the property of being hyperfinite (Lemma2.1.8 in [Jol]) or the
In the von Neumann algebra framework, where the preferred modules are Hilbert spaces, it
property T (see [Ana] and [PP2]).
is natural to make a construction which, apparently, depends on the choice of a trace on
M. We begin by showing that in fact the ring theoretic construction is exactly the same.
If r is a finite faithful representation of the pair N c M on H, then the centers of
First we recall some notions from [Jol] which work for arbitrary finite von Neumann
r(M) ' r(;)]',are the same as those of M and N respectively, and the rows and
M algebras exactly as for factors. Let N c M be finite von Neumann algebras with the same
columns of A, , are naturally indexed by the columns and rows of AN. The
identity. Given a faithful normalized trace on M, there is a unique faithful normal
generalization of Proposition 2.3.5 to this setting is the following. conditional expectation EN : M ,N determined by tr(xy) = tr(EN(x)y) for x E M and
y E N. In fact EN is .the restriction to M of the orthogonal projection
Lemma 3.5.5. Let N C M be a pair of finite direct sums of finite factors, as above, as
% :~ ~ ( ~ 3 , t ~r ~) ( ~ , tWe
r ) denote
. by (M,eN) the von Neumann algebra on L2(M,tr)
suppose r is a faith'fil finite representation of the pair. Then
I
generated by M and %.
We let J denote the conjugate linear isometry of ~ ~ ( ~ , extending
t r ) the map
x w x* on M.

Proof. If M and N factors, the equality holds because [x(N)?n(M)'] = [MN]
Proposition 3.6.1.
by Propositions 3.4.6 and 3.2.5.g. To extend the equality to the general case, one proceeds
exactly as in the finite dimensional case (Proposition 2.3.5), with Proposition 2.2.5b being (i) eNxeN = EN(x)eN for XE M
replaced by [DvN], Proposition 1 of $1.2, which says: if Q is a von Neumann algebra on (ii) JeNJ = eN
H and p is a projection in Q or in Q' , then Endpgp(pH)) equals pEndQ (H)p. #
(iii) For x E M, x commutes with eN if and only if x x N.
(iv) (M,eN) = JN'J
Also note that r(M)' is of finite index in 4N)' by Lemmas 3.5.4. and 3.5.5.
N ,(M ,eN)
(v) The map $J is an injective morphism onto eN(M,eN)eN.
Proposition 3.5.6. Given an irredundant mbyn math A over Y Y ~ N
(0) U (2 cos r/q :q 2 3) U [~,co],and an mbyn row stochastic mat& T having the (vi) The central support of eN in (M,eN) is 1.
same pattern of zero entries as A, there ezists a pair N C M (both hyperfnite) with (vii) The space MeNM, which denotes the linear span of {x' eNx' : x' ,x' E M}, is
!A = A and !T = T.
a strongly dense *subalgebra of (M,eN).
(ii) Follows from ~ ~ ( x =* E) ~ ( x ) * . (b) If N c M are 111 factors, then dimC(N1nM)5 [M:N].
(iii) Note that x commutes with eN if and only if left multiplication by x (c) If N c M are 111 factors with [M:N] < 4, then N' fl M = C1.
commutes with EN.' This is clearly so for x E N. On the other hand, if x E M and x
commutes with EN, then x = xEN(B) = EN(x) E N. 
Proof. We first consider the case that N and M are factors. Let H = L ~ ( M )and
n
x = {xo + CxieNYi: n E N, xy.
1 lEM)
i=l
4
x = x v i y i , with yi E N.
which implies x = 0 by 3.6.l.v and the faithfulness of EN. This proves the claim that *
In fact viyi = viEN(vix).
F(eN) is invertible.
Since the centrd support of eN in (M,eN) is 1 and since (M,eN) is finite with
Now we may obtain a formula for xeN. Suppose first that x is in MeNM, namely
that x is a finite sum x a j % b j with a.,b. E M. Then F(x%) = x a . ~(b.)F(eN) and
finite dimensional center by 3.6.1(iv), there exists a finite set wl,. .,wn. of partial
J J I N J
isometries in (M,eN) with
*
l eN and z w j w ; = 1; in particular the w. have
W.W.
J J J
mutually orthogonal range projections. (See [Tak], V.1.34.) As wjeN = wj, there are, by
3.6.3, elements vl, .,vn EM with w. = v e for all j. We verify that the vi have the
J jN
This formula holds for any x E (M,eN) because both sides are strongly continuous in x properties (a)(c). For i # j
and because M%M is strongly dense in (M,eN) by Proposition 3.6.l.vii. Thus
*
, so EN(v.v.) = o by 3.6.l(v). Sirnilarb, since wiwi is a projection in (M,eN) and
Theorem 3.6.4. Let N C M be type 111 uon Neumann algebras with finite dimensional I * I J * *
centers and k t tr be a jaithjil normal trace on M for which N' is finite on L 2 ( ~ , t r ) .
i wiwi = EN(vivi)eN, 3.6.l(v) implies that fi := EN(vivi) is a projection in N.
9 3.6. 'me Wdarnental construction 167
1
Furthermore (ii) Let o : M + N be a right Nlinear map and set a =za(vj)v;. b a l l from
* *
vitieN = vieNvivieN = W.W. W.
1 1 1 j
*=w.=ve Section 2.6 that E;(a) : M ,N is defined by E:(a)(x) = EN(ax) for x E M. We have
I i N7
fi~N(v;x) = ~ ~ ( f ~ v 
; xE )N(v x) for X E M .
by Nlinearity of EN
(iii) It follows from 3.6.3 that M%M is a twosided ideal in (M,eN). But MeNM
= E VjEN (v*x)eN,
j
mnttbh z v j e N < = ~ W JI *j  I, so h e N =~ (M,eN).
j
j j
and hence x = v.E (v.x), by 3.6.3. To show uniqueness of the expansion, suppose (iv) If a : M 4 M is right Nlinear, , then for XEM,
~ J N ; 4x1 = a ~ v ~ E ~ ( v = ; xZa(vj)EN({x);
) thus o = ZA(a(vj))oENoA(v;), where
J
using Nlinearity of EN and properties (a) and (b) of {vi}. We will refer to a family Corollary 3.6.5. Let N c M be a pair of von Neurnanla algebras of type 111 having
{vi} having properties (a)(c) as a PimsnerPopa basis of M over N; see [PPl]. finite dimensional centers, and suppose that N is of finite index in M . Let tr be any
Now consider the Nlinear map faithjul nonnal trace on M and define eN and EN via tr. Then
M4 @ f.N
1s jsn M aN M g(M,eN) as Nbirnodules, and
~ n d i ( ~(M,eN)
) as Calgebras,
isomorphism M eNM 2 (M,eN) by using a PimsnerPopa basis. # Lemma 3.6.7. If p.q. # 0, .then
1 J
C. .
(i) dimNt. .(piqjH) = A, and
The next proposition determines one part of the spatial data for the inclusion 1,~ j,!'
M c (M,eN). (ii) d. . dimN, .(piqjH) = dimqeNt(qjH)
J,' i,~ J
Proposition 3.6.6. Let N c M be finite direct sums offinite factors such that N is of
(weN) Mt 
Proof. By 3.4.6,
jnite index in M, and let tr be any faith@ trace on M. Then AM = (AN) . dimN. .(PiqjH)
j,!' =.

Proof. This follows from 3.5.4, 3.5.5, and the formulas JN'J = (M,eN),
and by 3.2.5(h),
JM'J = M. #
dimM. JpiqjH) = trpiM(qjpi)ddimp.M(~i~).
IrJ I
To describe M c (M,eN) more precisely, we also have to compute the matrix of traces
But since M is in standard form on H, so is piM on piH, and dimpVM(piH)= 1.
T F y e N ) . This is the part of the theory which differs most from the finite dimensional 1
=0 p.q. = 0,
1 J
r.l,j = if piqi # O,
the sum being over those i such that p.q. # 0, and let F be the diagonal matrix
I J
( M 7 e ~=):T
Our present goal is to compute the entries of TM : , namely +,
F = diag((ol,. ..,cp,). Furthermore, let be the nbym matrix
wlapbrn a: r~lucevon neumaonageoras 5 3.6. The fundamental construction. 171
x2
When p.q.
I J
1 0, we have JJ. = $
(9. i,j = A i j 2;
J
a d when piqj = 0, (T). . = 0 =
J71
"i
J
Thus
(weN)
Proposition 3.6.8. TM = FT.
n
Set L = (M,eN), L = @ L.; then L. g Mat,(C) whae rj = ( ~ ' 2 )=&Aij.
~ Note

Proof. Combining 3.6.7(i) and (ii) we get j=1 J J J i
that
Thus
if p.q. # 0, and d.. = 0 otherwise. To eliminate dimq:N,(qjH) we use the fact that
1 J J71
( M,%)
T~ is row stochastic,
which is in accord with the relation observed above between the inclusion matrix and the
index matrix.
. ..
As the minimal central projection in (M,eN) = JN'J are precisely {Jq.J : 1 i j 5 n),
J
any trace Tr on (M,eN) is specified by a row vector f, with r. = Tr(Jq.J). It will turn
J J
out to be useful to calculate the q u a n t i t i Tr(eNJqjJ). h c a l l that JeN = eNJ. Also
A?
Putting this back in (3.6.8.a) gives d. . =
1.1 vjci,j if p.q.
!.ti
I J
.
# 0 and d.J,] = 0 otherwise, as observe that
desired. #
( M,%)
Let us check what that formula TM = F'i? means for finite In fact, let 51 denote the trace vector in H = L2(hf,tr), i.e. the identity 1 of M
dimensional algebras. Suppose that piM g Mat (C) nod qjN g MatUJC). As noted reqarded as an element of H. The l i n e i space {x51 :x E M) is dense in H and we have
Y
M determines the trace matrix T = T N via
before, the inclusion matrix A = AN d e Jq.J(x51) = e J .x*n = e x 51
N J
= ~N ~q~( x=qEN~ qj
( )x ~) Q ~ ~
= qjEN(x)n = EN(qjx)fi
= eflj(xfi).
since q.p. is the sum of Aijuj orthogonal minimal projections in piM Setting
J 1
= diag(pl,. ..llr,) .
and "v diag(ull. .,un), this can be written
T = jtl~"v.
Lemma 3.6.10. Let Tr be any trace on (M,eN) and let r. = Tr(Jq.J). Then We restrict our attention to pairs of finite direct sums of finite factors and continue to
J J
use the notation of the previous section.
Hence
/j $ =b;TLMleN) = /3 :F'f (by 3.6.8)
=if (by 3.7.3.1)
= t TT
:.
3.7. Markov traces on EndN(M), a generalization of index.
'i= /T14 M 1
(P) Given a trace tr on M satisfying tT ~ =T dpline s TNF
pefinition 3.7.1. Let N c M be finite von Neumann ,algebras with N of finite index (motivated by 3.7.3.1), and define a trace Tr on (M,eN) by Tr(Jq.J) = Then
in M. We say that a faithful trace tr on M is a Markov trace of ~ o d u l u sp for the pair J 'j.
N c M if it extends to a trace, also called tr, on (M,eN) for which
J N) 5 r j q = (FF)~
Tr(q.e
Hence f defines a Markov trace on (M,eN) byeTheorem3.7.3. #
= TFF~F).3
= T1tj, Remark. Before going on, let us see how the analysis above agrees with that in
Chapter 2 for finite dimensional algebras. Assume that Mpi g Mat (C) and
as desired. # Y
Nq. E Mat,JC). We noted in the remark following 3.6.8 that
_Corollarv 3.7.4. Suppose N c M are jnite direct sums of finite factors, with N of J J
M
Pnite i n d a in M . Set T = TN.
(i) If N c M is a connected inclusion, then there is a unique normalized Markov
trace on N c M; it is faithhl and has modulw eqwl to the spectral radius of TT.
(ii) If tr is a Markov trace of modulw p on N c M, then the unique eztension of where i; = diag(pl,. ..,pm) and "v diag(ul,. ..,un). Thus
the trace to (M,eN) satishing (3.7.2) is a Markov trace of modulw /3 (for M C (M,eN)).
trace to (M,eN) satisfying the Markov condition (3.7.2) is specified by the vector ~ ( T T=
) (fl;)(~l~~t;)
( MaN> \
= F"I', with entries = ftAAt;=pf$= pf. #
? = ~ l fTFl.
i Let R denote thematrix TM
Definition 3.7.5, Let N c M be finite sums of 111 factors with the same identity and
with N of finite index in M. Let A =!A = (A. .) be the matrix of indicea and
IrJ
T = TM
N = ( c ~ , ~be) the row stochastic matrix of t r m as above. Form T = '?(A,T), the
A? .
matrix whose (j,i) entry is 0 if c. . = 0 and
191
3 otherwise. Then the index of N in M,
( M,%) i ,j
Since AM = (A!)~, the m a t h , R (which is to R as T is to T) has entries [M:N], is the largest eigenvalue of the matrix TT.
&mark. It is easy to see that this definition agrees with that of Section 3,4 when N
and M are factors. We mention again that P. Jolissaint has recently shown that this
definition always coincides with the ring theoretic definition given in Section 2.1 and [Jo~].
That is 'fi = TFl. But then Corollarv If NcM are as above and [M:N] < 4, then
$7'6t
[M:N] E (4 cos r/q : q 2 3).
;RR = ( r l f TF~)(FT)(TF~)
= ,rlf TZTFI &g& The index is the largest of the numbers [Mz:Nz], where z is a minimal
'
= f TFI (by 3.7.3) projection in Z(M) f l Z(N), so we can assume that M :,N is connected.
= p;.
By 3.7.4(i), there is a Markov trace tr on M of modulus [M:N]. Examole 3.7.10. Let M be a 111factor, p a projection of trace t in M and
Then 3.7.4(ii) al1ows:us. to iterate the fundamental construction in the usual way to
+
N = pMp Q, where Q is a subfactor of index X in 11p). Then
obtain a tower
!A = (1 X1l2) and TT#
=.(t 14). So f = and [$ift] TT = 1 + A.
This is 1 4
2
when X = 1 , 2 , 4 cos ir/5, or 3.
a sequence of selfadjoint projections (ek)k,l with Mk+l = (Mk,ek) for all k, and a
Remark%The index matrices in example 3.7.10 correspond to A3, B3,H3, and GP),
respectively, under the corespondence of Theorem 1.1.3, when X = 1, 2, 4cos2ir/5; and'3.
trace tr on UMk satisfying the Markov property
k This is no accident, as we will see.
[M:N]tr(ekx) = tr(x) for x E Mk. Pronosition 3.7.11. Let A = (A. .) be an irredundant m a t h over ( 2 cos(ir/q) : q 2 2)
 1,J
and T = (c. .) is a row stochastic matrix wilh the same pattern ofzero .entries as A. Let
1,J
The projections ek then satisfy the usual relations and therefore the restriction on [M:N]
t=~(A,T) be the mat& whose (j,i)entW is zero i j c. . = 0 and equal to
A?
$
follows from [Jol]; see the argument given in Section 3.4. # Id i ,j
othekise.
Next we provide some examples. Note that by 3.5.6, to. construct examples it suffices If the spectral radius of TT is less than 4, then it equals 4 cos2ir/q for some
to give the +trices A and T. q E {3,4,5. .).
+
I
Examnle 3.7.7. The simplest new example is where M is a 111factor, p is a We can suppose that A and T are indecomposable. By 3.5.6, there is a
M
projection in M and N = pMp + (Ip)M(Ip). Here the matrix :A is (I I), and T N connected inclusion M J N of finite direct sums of Illfactors with A = A
: and
c i j = [ Z + $  ~ A ~ , ~or, T = XA, where X is the mbym diagonal matrix whose Let 3 be the set of k+ltuples with il = ik+l; thus a r a ' is bijection of Wk
J
(i,i)sntry is
J
[EAiVj]l.
Then the (j,i) entry of
J
f is A i j [ Z A i j ] , i.e., f = d t r l . tr(Ak) =
a€Wk
a, =
acWk
$x
(aa + aa
Thus TT = XAAtXl, which has the same spectrum as Adt.
(b) By 3.5.6 there is a pair N c M with A = AN M and M Then
T =TN.
tr(Ak) t ( a g 1)112 = go = tr(Gk),
(M:N = p(TT) = llA112. acVk 0w
' k
(c) It suffices to consider A irredundant, so the result follows from (b) and
3.7.6. #
for all k E PI. When k is even, we have
Of course, 3.7.12(c) was dready known as a consequence of Theorem 1.1.3. Theorem
1.1.3 suggest8 (but does not immediately imply) the following, which is the main result of np(Alk = n p ( ~ z~tr(Ak)
) t tr(Gk) 2 p ( ~ ) k ,
this section.
where the first equality and last inequality result from considering canonical forms for A
Theorem 3.7.13. Let N cM be a connected inclusion of jinite direct sums of and G, noting that the eigenvalues of G~ are positive. Taking kt h roots and then the

II, factors.
limit as k + w gives the result. #
(a) If [MN] < 4, e n A is the matriz associated (in Theorem 1.1.9) fo a Lemma 3.7.15, Let A= be an mbyn irredzmdant mat& over
bicoloration of one of the following Cozeter graphs: : {I E B :r = 0 or r 2 1). Let T = (G1 J.) be a row stochastic matriz PuiIh the same pattern of
zero entries as A. Let 2 he the nbym matriz whose (j,i)entry b 0 if c. . = 0 and
At (L 2 2)7 Be (L 2 3), DL (L t 4), (L = 6,7381, 1J
Moreover [M:N] = Al($: = 4 cos2~/hYwhere h i s the Cozeter number. (See tables If p(T"i') i 4 then (lA112 $ p ( e ) .
A t ) (L odd, k L I), 861) (L L 2), C P ) (L t 31, p,v E (1,2) are all nonzero; tha&is, the graph f (A) contains a subgraph of the form I
/I
D P ) (L L 4), ~ 1(L =
~ 6,7,8),
) ~ f )G
, F).
 ;I
Lemma 3.7.14.
mat& and kt
(Schwenck, [Sch2]) Let A = (a. .)
G = (%,J be the motri. with entries
I>J
be a nonnegative
q , j = (a. .a. .)'I2.
1 J 131
nbyn
Then
i4.
o I
I
I
I
Rearranging the rows and columns, we can suppose il = jl = 1 and i2 = j2 = 2. Denote
IlGll = p(G) < p(A), where p denotes spectral radius.
&& For any k+lt,uple o = (il,i2, .,iW1) with 1 I i.J $ n, let al denote the
reversed tuple al = (ik+l,. ..J2,il), and let
\
the (i,j)entry of TT by %,i Then .liJ =xv,
k
, c. x2
j,k
the sum being over those k I
I
.
71 9 1 z
k
'2 with equality everywhere if and only if BU
~,k
a =ai ai ..
a 2, 4k1ik+l. I
1UU
c"'i ctnapar a: riulrc: vuu r~eulnauul u ~ v u ~ s a m . 7 . Markov traces
",I %
.o,.
By monotonicity of the PerrbnFrobenius eigenvalue, we have [ 0 if the ith and j rows of A are orthogonal
with equality if and only if A is 2by*. This in turn is no smaller than Observe that ko depends on (ij), and ko(i,j) ='ko(j,i). On the other hand,
7 =.
191
ZXt,. Note that for all pairs (ij), the (ij) entry of AAt is the geometric
k
mean of the (i,j) and (j,i)entries of TT; i.e.,
by the observation above, with equality if and only if all the nonzero Xjlk with j j 2 are
equal to one. Truncating the sums defining the entries of the last matrix we see that the
spectral radius is at least = Io if the , ith and jth rows of A are orthogonal
1 li,koAj,ko otherwise.
Hence by Lemma 3.7.14
1 1 ~ 1 =1 ~p(hAt) 6 P(Tf) #
with equality if and only if ,Ijfk= 0 for j = 1.2 and k > 2. If we replace the offdiagonal
entries by their geometric mean, we do not alter the spectrum, so the las quantity is equal Prwf of Theorem 3.7.13. (a) Let T = !T and f = T(A,T). By hypothesis
to 1 [M:N] = p(Tf) < 4, so by 3.7.15 we have 1 1 ~ 1 s1 p(TT).
~ Let S be the (convex) set of all
row stochastic matrices of the same dimension and zeropattern as A and T. For each
S a S define S = f (A$), the matrix whose (j,i)entry is 0 if A. . = 0 and ,IT ./(S)i,j
1,J >J
. .
otherwise, and cp(S) = p(SS). Then cp is a continuous function of S by elementary
, .
c 'c
.. . perturbation theory, and (P assumes the value [M:N] = ~ ( T T )as well as the value 1111112,
where o = 2'2. Finally, this is at least by 3.7.12(a). But by 3.7.11, the set of values of cp less than 4 is discrete, so by convexity
.: C2,1C1,2 . .
of S, cp is constant and
CHAPTER 4 In Section 4.2, we study the notion of commuting squares and give a number of
~ ~ m u tsqam,
@ subfactors, and the derived tower examples of constructiowhich produce commuting squares. In particular we coqsider the
behavior of commuting squares under the fundamental construction. "
4.1. Introduction.
C1 B1
There are two main themes in this chapter. The first is the approximation of a pair P r o ~ o stiion 4.1.2. Consider a commuting square U u with respect to a trace
N M of hypefinite I I ~factors by pairs cnc Bn of finite din~ensionalvan Neumann
Co Bo
algebras, with tr which is a klarkov trace for the pair Bo c B1 offinite von Neumann algebras &th finite
N c M
u U d i m ~ ~ ' o n acenters.
l Let B2 = (Bl,el) be the von Neumann algebra qbtained via the
Bn+l' kndamental comtruction for Bo c B1, and let C2 = {Cl,el)*. Then
u U
Cn Bn c 2 B2
U U
a,ld M = (UB,)", N = (U c,)". 1n order for the approximating "ladder1' of finite
C1 B1
dilnensiond algebras to behave well with respect to the fundamental construction and the is ako a commuting square.
index, it should behave well with respect to the conditional expectations:
Therefore iterating the fundamental construction will produce an infinite ladder of
be tile conditional expectation of Bn onto Cn. We are thus led to the following defillition
commuting squares. Now suppose that we have a pair N C M of finite von Neumann
wllicll was first introduced by Popa (Lemma 1.2.2 in [PoplI; see also [Pop2]): . algebras with a Markov trace tr of modulus 0 and a ladder of commuting squares
E ~ l M c (M, e)
C1t B1 algebras (An)n20 generate (M, e), and u U is a commuting square with
u

u Bn An
respect to the Markov extension of the trace to (M, e).
Co Bo In Section 4.3 we prove a theorem of H. Wenzl on pairs N c M generated by a ladder
E ~ a of commuting squares satisfying a periodicity assumption. (See Section 4.3, Hypothesis
commutes. (B), for the exact assumption.)
182
184 Chapter 4: Commuting squares and subfactors 5 4.1. Introduction 185
/
Theorem 4.1.3. Suppose N c M is a hyperfinite pair (with a finite faithful trace tr on sequence of projections in the tower construction. We already know another construction
M) generated by a ladder of commuting squares  
of an irreducible pair with index /3, namely {el,e2, '1' 3 {e2,e3,,  } " (Theorem 3.4.3).
4 An argument due to C. Skau shows that {el,e2,.  9 ) ' n (UMk)" = N (Theorem 4.4.3).
In Section 4.5 we present a construction which yields irreducible pairs of hyperfinite 111
factors, starting with a Coxeter graph of type A, D, or E and a choice of a distinguished
vertex wl on r. In particular for r = E6 and wl an end vertex on one of the long
+
arms of r , we obtain the index value 3 p,which is at present the smallest known value
of finite dimensional von Neumann algebras. Suppose that the inclusion data for the ladder larger than 4 of the index of an irreducible pair. The construction goes a8 follows.
is periodic, in the sense of Hypothesis B ofSection 4.3. Then Give r the bicoloration with wl white and with r white vertices altogether. Let
(i) N and M are factors and [M:N] < m.
Let e and (An)n20 be as above and let zn be the central support of e i n An. Mo denote the abelian von Neumann algebra C' and M1 the finite dimensional von
2
For large n, zn = 1. Equivalently, An is isomorphic to the result of the Neumann algebra containing Mo such that 'I is the Bratteli diagram of the inclusion
(ii)
fundamental constrnction for Cn c Bn. I Mo C MI. Form the tower (Mj)j20, starting with the pair Mo c M1 and the Markov
trace tr on M1, and let be the usual sequence of projections. Let M be the
(iii) For large n, [M:N] = [Bn:Cn] = lit(n))2/11~(n)112,where t(n) and g(n) are
the vectors of the trace on Cn and Bn respectively. factor ( U M.)" and let N be the subfactor of M generated by P and the e.'s. Set
j20 J J
2
/3 = [M1:MO] = llrll ; since ,/3 < 4, Skau's Lemma 4.4.3 applies and M n N' = Mo. Let
Section 4.4 contains a contruction of (necessarily irreducible) pairs of hyperfinite
factors with index less than 4, as follows: Start with a connected pair Co c Bo of finite & p be the minimal projection of Mo corresponding to the vertex wl and set C = pN and
dimensional von Neumann algebras with index /3 < 4, and let B1 = (BO, eC ) be the
0
p B = pMp. Then C c B is a pair of factors with C' n B = p(N1nM)p = pMOp= Cp;
is C c B is irreducible. The index of this pair can be computed as follows:
that
result of the fundamental construction for Co c Bo, with respect to the Markov trace of Let r also denote the matrix of the bicolored graph I', and let tE
denote the unique
modulus 0 on BO. Define q E T by /3 = 2 + q + q", set g = qeC  (leCd, a PerronFrobenius row vector for rtI', normalized so that its first coordinate
0 (corresponding to the distinguished white vertex wl) is 1. Let A be the Coxeter graph of
C1 B1 type A with the same Coxeter number as I', and with a bicoIoration having at least one
unitary element in B1, and set Cl = g~ogl. Then U U is a commuting square, white end vertex, which is labelled as the first white vertex. Denote also by A the matrix
Co Bo of the graph A, and let 5be the PerronFrobenius row vector of A t A, normalized so
that its first coordinate (corresponding to the chosen white end vertex) is 1. then
with
. respect
. to.the Markov extension of the trace to B1. Let (Bn)n20 be the tower [B:C] = 115112/11t112.
The proof uses Wenzl' s index formula from Section 4.3.
obtained by iterating the fundamental construction, with Bn+l = (Bn, en) and set The second main topic of Chapter 4, presented in Sections 4.6 and 4.7, is the
and princi~al& of a pair of finite factors N c M of finite index. The derived
Bn+l
is a ladder of commuting squares,
 the chain of relative commutanis (N' n Mk)k>O, where (Mk)k>O is the tower
tower is
 
Cn+l = alg{Cn, en} for n >_ 1. Then U U
for the pair N c M. It follows from 3.6.2 that dim (Nr n MI) 5 [ M : N ] ~for k 2 0.
n' Bn
with respect to the Markov trace on U Bn. It turns out that the inclusion data is periodic Let (ek)k>0 be the projections in the tower construction, let Yk denote N'
 n Mk
and that B = (U Bn)" 3 C = (U Cn)' is a pair of factors with index p. and Ak the inclusion matrix for Yk c YkS1. The following summarizes the structure of
Let (Mk)k,O be the tower obtained from a pair N c M of finite von Neumann

algebras with finite dimensional centers, with index /3 c 4 and let (ek)k>l
 be the usual
186 Chapter 4: Commuting squares and subfactors 5 4.1. Introduction 187
Theorem 4A.4. (b) If N c M is ofinfinite depth, then 'l is one of the following:
(i) The inclwion Yk c Yk+l is connected.
YkekYk is an ideal in Yk+l, and if zk+l = z(ek) is the corresponding

(ii)
t
mat& Akl.
I
yk
k' e
xzkf 1
kk' has inclusion
/\/\/ "' (end uertez at distance n from *)
on Yk1.
t
fc) Ak = Ak1.
II~~=
~ I [M:Nl'
I~ (end vertex at distance n fiorn *
(vi) If the equivalent conditions of (v) hold for k, then they also hold for k+l.
Section 4.7 is devoted to computing the derived tower for a number of examples
We call the ideal YkekYk "the old stuff", since it is determined by Ykl c Yk; the Crossedproducts and fixed point algebras for outer actions of finite groups give example
with depth 2. The pairs R c R (of Proposition 3.4.4) when P < 4 have principal graph
complementary ideal is called "the new stuff1. Then (iv) says that "the new stuff coma P
) 0. The princi~al& of the pair
only from the old new stuff", or ( x ~ ~ ) ( l  + ~ = of type An; for 0 = 4 the principal graph is Am. In 4.7.c we give a general hethod whicf
N c M is obtained as follows: on the Bratteli diagram of the derived tower, delete on each allows the computation of the derived tower in many examples coming from group actions
level the vertices corresponding to the old stuff, and the edges emanating from them; the In 4.7.d we use this method to obtain the derived towers for the index 4 subfactoc
result is a connected bipartite graph with a distinguished vertex *, the unique vertex on R~ c (R @ at^(^))^,
where the hypefinite 111 factor R is realized as the weak closur~
level 0. The Bratteli diagram of the derived tower can be reconstructed from the principal m
of the CAR algebra @ Mat2(C) in the trace representation, and G is a closed subgroup o
graph. The pair N c M is said to be of finite d e ~ t hif the principal graph is finite; the
is the maximum distance from any vertex to * . SU(2) acting by the infinite tensor product of its action by conjugation on Mat2(C). 11
This analysis, together with the work of Chapter 1,yields a new proof of the restriction this way one obtains as principal graphs all the affine Coxeter graphs of type A, D, and E
on index values: as well as the infinite graphs Am, and Dm listed above. Finally we compute thc
Corollarv. (i) Suppose N c M is a pair of 111 factors with [M:N] < 4. Then derived tower for the pair RP c R when /3> 4. This is the most difficult result of tht
(a) [MN] = 4 cos2 nth for some integer h 3. chapter, involving a representation of the sequence (ei)iL1 in the CAR algebra due tc
.'
(b) The depth of N c M is no greater than h2. Pimsner and Popa and a theorem of Popa on the tunnel construction (a mirror image oj
(c) The principal graph of N c M is a Coxeter graph oftype A, D, or E, whose the tower construction). Ultimately one identifies the pair R c R with the pail
P
norm is [ M : N ~ / ~ . N ~ (N
C@ at^(^))' where N is the completion of the CAR algebra with respect to s
(ii) Suppose N c M is a pair of Ill factors with [M:N] = 4.
certain Powers state. The principal graph is therefore Am,&
(a) If N c M is of finite depth, then the principal graph 'l is a completed Cozeter
graph of type A, D, or E, i.e., one of the graphs in Table 1:4.6.
188 Chapter 4: Commuting squares and subfactors 3 4.2. Commuting squares 189
4.2. Commuting squares. Assume (i) holds and let bo E Bo. For all co E Co one has
We begin with a proposition, inspired by Lemma 2.1 of [Pop2], which gives a number tr(Ec (b0)co) = tr(bocO) = tr(Ec (b0)co).
of equivalent conditions for a commuting square. 0 1
Proaosition 4.2.1. Consider a diagram As Ecl(bO) E Co, this implies EC (bo) = EC (bo), and (v) follows. As (v) implies (iii),
0 1
conditions (i) to (v) are equivalent.
The equivalence of (vi) and (vii) follows from the formula
of finite von Neumann algebras and a finite faithful normal trace tr on B1. All conditional
for bo E Bo and cl E C1.
expectations being with respect to tr, the following are equivalent. The next step is to show that (ii) and (vii) are equivalent. Observe first that one has
6) ECl(BO) C GO.
(iii) EC EB = EC EB .
1 0 0 0 for any x E B1. Thus (ii) can be reformulated as
(iv) EC EB = EBOECl and Bo ilC1 = C,.
1 0
EB (x) EC (x) I C1 for all x E B1.
Ec 0 0
c1ClB1
(v) The diagram U U commutes. Suppose (ii) holds. Then, in particular, bo I C1 for bo E Bo with EC (bo) = 0.
0
Consequently, for all cl E C1 and co E Co, one has
Proof. Let p,q,r be three projections acting on some Hilbert space. The following are
clearly equivalent:
(a) P9 = r
(b) pq = rq and r d q
which is zero, since the conditional expectations are trace preserving. Consequently (ii)
(c) pq = qp = r.
holds.
As we may view the conditional expectations as projections on L2(Bl,tr), this shows the
equivalence of (ii), (iii) and (iv). Obviously (ii) implies (i).
Chapter 4: Commuting squares and subfactors 3 4.2. Commuting squares 191
Finally, as (iv) is symmetric with respect to Bo and C1, we may exchange Bo and
'j+kc Bj+k
C1 in any of the conditions (i) to (vii). # Proof. (i) For each j 2 0 and k L 1, the diagram U U is a commuting
Cj C Bj
It follows for example from (v) that in diagrams like
N c M
square, by induction on k. It follows that the limit diagr& U U is also a
Cjc B
j
commuting square, and thus for any b E B one h~ eNbeN = EN(b)eN = EC.(b)eN.
j J
Since elements of N, and in particular EC (b), commute with eN, this shows (i).
j 3,
Pro~osition4.2.2. Consider a pair N c M of finite von Neumann algebras, a finite for all x E Bj+l. Consider now yo,yh,y& E B..J Then
faithful nonnal trace tr on M, and the algebra (M,eN) obtained by the fundamental
construction. Assume that M [respectively N] is generated as a von Neumann algebra by
a nested sequence (Bj)j20 [resp. (C.). ] o f von Neumann subalgebras in such a way that
J J20
one has for each j 2 0 a commuting square
Thus EB (A.) C B. for a dense *subalgebra A. of A
j+l J J J j*
Let x E A.. By the density theorem of Kaplansky, there exists a sequence ( x ~ ) ~ , ~ ,
J
with xk E A. and llxkll s llxll for all k 2 1, such that x = 1i m xk in the topology
J k+ m
defined by the norm 11. 112.
It follows that EB (x) = 1i m EB (xk) c Bj. Thus
j+l k+m j+1
andset A. = {Bj,eN)'. Then E (Aj) c B. and this proves (iii). #
J Bj+l J
(1 e ~ b e N= ECj@)eN = eNEb:(b) for b c Bj, j 2 0
J C1 B1
(ii) T h e algebras (Aj)j20 generate (M,eN) as a von Neumann algebra. Corollarv 4.2.3. Consider a commuting square U U with respect to a trace tr
Co Bo
Suppose moreover that tr is a Markov trace of modulw /3 for the pair N c M, and
which is a Markov trace for the pair Bo c B1. Let B2 = (Bl,el) be the von Neumann
denote the Markov edension of tr to (M,eN) by tr again. (See Definition 3.7.1.) Then
algebra obtained via the findamental constkction for Bo c B1, and let C2 = {Cl,el)'.
C2 B2
(iii) U U is a commuting square with respect to tr . j Then U U is also a commuting square.
c A.J I~j+l
B~
192 Chapter 4: Commuting squares and subfactors 5.4.2. Commuting squares 193

Proof. This is the special case of 4.2.2 applied to U
Co C
U
C1 C C2
. # 
Proof. (i) Let f be a minimal projection in (Co)q and let pf = x g i be a
i=l
decomposition of pf into orthogonal minimal projections in (BO)p (so
Remark. Suppose moreover that Bo and B1 have finite dimensional centers. Then
n = [(Bo)pq : (C ) ]It2). Then (by 2.6.4) £el = u(fec ) is a rninimml projmtion in C$
0 Pq 0
the fundamental construction iterates to give the tower (B ) with BjS1 = (Bj,ej) for
j j>O
and
all j. Define inductively CjS1 = {Cj,ej}' for j 2 1. Then we obtain a ladder of
f e l b felp (by 3.6.9)
'j+l Bj+l = fpel (because p E Bo)
commuting squares U U . We are going to use this idea to construct
= pfel (because p E Z(Bo))
Cj C Bj 1
/ n
examples of subfactors below, starting with a commuting square of finite dimensional
algebras. The next two lemmas concern conditions which cause the inclusion matrices for
the resulting ladder of finite dimensional algebras to be repeated with period 2.
Thus (fel); is a sum of n orthogonal minimal projections in (B2):?
C1 B1 (ii) We are now supposing that tr is a Markov trace. The statement Cj+l = C.e.C.
J J J
Lemma 4.2.4. Consider a commuting square U U of finite dimensional
is valid for j = 1 by hypothesis. Suppose it is valid for some j. Then Cj+lej+lCj+l is
Co Bo
von Neumann algebras, with respect to some trace tr on B1. Let B2 = (Bl,el) be the
an ideal in CjS2 containing pejej+lej = ej' where P is the modulus of the Markov
J J J 3 1, so Cj+lej+lCj+l = 'j+2'
trace. Then Cj+lej+lCj+l 3 C.e.C.
finite dimensional von Neumann algebra obtained via the findamental construction for
Bo c B1 and let C2 = alg{Cl,el}. It follows that for all j, the tower C.J1 c C.J c Cj+l is isomorphic to
 
Suppose that C2 = ClelCl (or equivalently, by 2.6.9, that xxiec0yi xxiem C. C C. C End (Cj), so the inclusion matrices ACj are alternately AC1 and
J1 J Cj1 'j1 C~
is an isomorphism from the algebra (C1,eCo) obtained by the findame6tal c o ~ t r u c t i o nfor B.
Finally the statement regarding hCJ follows from (i) and induction. #
Co c C1 onto C2). Then j
Co and AC 1 = (AC0) .
for Cr1 c C.J are alternately AC1 C2 C1 t
194 Chapter 4: Commuting squares and subfactors 4.2. Commuting squares
Suppose in addition ffiat tr is a Markov trace with respect to Bo c B1. Let (Bj)j20
be the tower obtained by iterating ffiehndamental construction, with BjS1 = (Bj,ej), and
and the claim follows. #
set CjS1 = alg{C.,e.) for all j 2 1. Then
J J
Remark. A similar result holds for reduction by projections in Cg.
(ii) The chain Cjl c C. c Cj+l is isomorphic to C.
3
c C. c EndC. (Cj) for all j.
J1 J 31
Next we give some examples of commuting squares involving relative commutants,
The inclwion matrices ACj+' are alternately At and A (j 2 0), and the inclusion
CJi fixedpoint algebras of groups, and crossedproducts.
a n alternately At and A (j 2 0).
Pro~osition4.2.7. Let N c M be a pair of von Neumann algebras, let tr be a jnite
faithfil normal trace on M, and let S be a selfadjoint subset of N. Then

Proof. (i) We have C2 = for some matrix ill, by
B C B B
+
02. Therefore A% = A ~ A ; = ntA 02nl. On the other hand A: =A ~ A ;
is a commuting square.
2 1 1 1
= AtA. This is only possible if K = (0), because otherwise i12ill # 0. The remainder of
(i) and (ii) now follows from the previous lemma. # 
Proof. We may suppose that S is a von Neumann subalgebra of N. Choose x E M.
Denote by C the 11. l12closure of the convex hull of {mu* : u is unitary and u E S) in
The next result is that commuting squares are preserved under reduction by certain L2(M,tr), and denote by y the projection of the origin onto C. Then y E M because the
projections. ball of radius JJxJ) in M is a )I.)12closed subset of L2(M,tr). Moreover, by the
tr(pEB (y)pu) = tr(pypu) = tr(yu) for all u E pBop. Consider z E pC1, say z = pc,
0
with c E C1. Then
196 Chapter 4: Commuting squares and subfactors 5 4.2. Commuting squares 197

Proof. Let us show that EB (x) E Co for any x E C1. This is obvious when x E Co.
0

Proof. For each x E M, one has
By Proposition 2.8.1, one may then assume without loss of generality that x = yenz with
y,z E Co. As EB (en) = B1 (see the proof of 4.2.2.iii), one has EB (yenz) = y r1z E Co.
0 0
g = qeN 
\
where n indicates now a crossed product.
and observe that g is a unitary which commutes with N. Then
We next describe three examples which are interesting in light of the connections
between the theory of subfactors and that of the braid groups.
&%'
c (M,eN)
.
Examnle 4.2.9. Let el,. ,en be a sequence of projections acting on some Hilbert U U
N c M
space such that
is a commuting square.
P eiejei = e.1 if lijl = 1
e.e. = e.e. if lijl 2 2
1 3 J 1 
Proof. Let x E g ~ g  l . If y = glxg E M, one has
for some real number 0 1 (see the last remark of Appendix IIc). Let tr be a normalized
faithful trace on the algebra generated by the identity and the e.'s,
J and assume that the
Markov relation Since EM(eN) = we have
Remarks.
(1) Up to scalars, g and 6' are the only unitaries in alg{l,eN} for which the
above construction works. Observe that g is precisely the element involved in the braid
roup representation of [Jo~].
is a commuting square. (2) This example is the basis for the examples of Section 4.4 below.
198 Chapter 4: Commuting square and subfactors 4.3. Wenzl's index formula
Example 4.2.11. Let N c M be a pair of factors, of finite index /I, and let tr denote of finite dimensional von Neumann algebras.
the normalized trace on M. Assume that there exists a projection eo E M such that
all y e N. Because of the density theorem of Kaplansky (see the proof of Proposition j
.
4.2.2.iii), we may check this for all x E alg{P,eo,. .,en) and for all n > 0. If n = 0, this
Then C = (UCj)" is a von Neumann subalgebra of B. In the periodic case which we
follows from the hypothesis on eo. To end the proof, we may assume that n 2 1 and that j
the claim holds up to n  1. consider below, tr is the unique tracial state on UC. and UB so that C and B are
j~ jj'
..
For aOb, E alg{l,eo,. ,enl) and x =zaaenba, factors.
n If E : B + C and E. : B. ,C. denote the conditional expectations with respect to
J J J
C c B
Ej; that is U u is a commuting square for each j. Let A = (B,e)
C. c B.
J J
which is by induction be the result of the fundamental construction for C c B with respect to tr, and let
, Aj = {Bj,e)' for each j. Then A. is an E.extension of B in the terminology of
J J j'
Section 2.6. Hence if (B.,f.) is the result of the fundamental construction for C. c B
J J 3 j'
then the formula o.(xa.f.b.) = x 5 e b i (ai,bi E B.) defines an isomorphism from
J 1J 1 J
This shows that the claim holds up to n. i i
#
(Bj,fj) onto the two sided itieal B.eB. generated by e in A., by 2.6.9.
J J 3
Remark. It would be interesting to have a systematic classification of commuting
squares Lemma 4.3.1.
c1 c i~,,tr) : (i) The central support z. of e in A. is a.(P); this is also the central support of
J 3 J
U U 7 the ideal B eB
j j'
Co Bo t
(ii) !im z. = 1 in the strong operator topology.
j+w J
200 Chapter 4: Commuting squares and subfactors 5 4.3. Wenzl's index formula 201

Proof. (i) This is straightforward, since the central support of fj in (Bj,fj) is 1, by Proof. (i) That B and C are factors follows from the uniqueness of the trace on
3.6.l(vi). ) is a finite factor. In any case A is
semifinite, so has a faithful normal semifinite trace Tr; we have to show that Tr(1) < m.
= [AeXd = 1,. That is, z. increases to P. # Now eAe = Ce is isomorphic to C, which is a finite factor, so e is a finite projection
J
and Tr(e) < W. Fix some j 2 jo and let qi be a minimal central projection in C and
j'
Next we introduce a very strong periodicity assumption on the inclusion data for the ng minimal central projections in (B f )
j' j
Cj+~CBj+l
ladder of inclusions U u .
Cj c B
j
Tr(eCi) Tr(eqi)
= (using 3.6.9)
Hypothesis (B). We assume there is a jo 2 0 and a p 2 1 and a suitable ordering of Tr(ti Tr(Ci
the factors in the B.'s and C.'s such that for all j 2 jo:
J J
(i) The inclusion matrix for B. c Bj+l is the same as that for B
Similarly for C. c Cj+l and Cj+p tions in <.A. (by 2.6.4) while Ci is the
J Cj+p+~. 1 J
Let d. = min{v~)/((A!~.dj))~).Then
J i J J
primitive.
(iii) The iiiclusion matrix A. for C. c B. is the same as that for Cj+p C Bj+p.
J J J
J = x ~ r ( e < 2~ d) x ~ r ( ? &=) d.Tr(z.).
Tr(e) = Tr(ez.)
j J J
i i
eigenvector for @ and similarly for the vectors ~ ( j ) . at Tr(e) > d Tr(P), and Tr is finite.
j'
Since p = [B:C] < mi the normalized trace tr on A has the Markov property:
Lemma 4.3.2. Assuming hypotheses (A) and (B), tr(ex) = ~ l t r ( x )for x E B. It follows from this and 2.6.4(c) that the weight vector of tr
(i) B and C are factors and [B:C] < CO. on B.eB. = z.A. is r1 t(j).
(ii) A . < B : ( 1 for all j, the inequality holding component J J J J
bycomponent.
(iii)
J J
for some
J J
j' A! J being the inclusion matrix of B. c z.A By the remark
J J j.
above, the weight vector of tr on A. has the form (/T1 f(j),?(jl), so that
tr(nzj+tp) = ( 4)
t  [B:Cllt(j)AtA $LT',(~+~P)), J
j S' j
202 Chapter 4: Commuting aquares and subfactors § 4.4. Examples of irreducible pairs 203
4.4. Examplea of irreducible pairs of factors of index less than 4, and a lemma of C. Skau.
Remarks.
(1) If it is known a pn'ori that B and C are factors with [B:C] < W, Wenzl can
obtain the index formula assuming only periodicity of the inclusion data for (C.).
J JLO'
(2) We used periodicity of the inclusion data for (Bjj10 only to obtain that B is a for all x E Bm= B
6 J>O j.
factor. g
204 Chapter 4: Commuting squares and subfactors 5 4.4. Examples of irreducible pairs 205
Theorem 4.4.2. With the notation above, if P( 4 then the pair C c B of factors of
type 111 is irreducible, and its index satisfies
and consequently, using the formula for @(gk),
&%& For any complex number w of absolute value 1 and for j 1 0, set Remark. The construction of the pair of factors C c B with [B:C] = P still makes
 (1  e.), and for x E U B., define
gj(w) = J J j20 J sense if p = 4, but the pair need not be irreducible. For example take M = N @ Matr
Then
g 0 ~ s i 1C (M , eo)
u u
The same argument as above shows that this defines an endomorphism \ of B, and the N c M
map is isomorphic to
206 Chapter 4: Commuting squares and subfactors 8 4.5. More examples of irreducible pairs 207
M0=NcM1=Mc ... c M ~ c M ~ +..~ c Let E denote the conditional expectation from R onto R'
P n R, and recall that Fk+l
is the conditional expectation from Mw onto RbnMm. As ekeR, one has
the projections (ek)k21 with el = eN and the Markov extension of the trace on the finite
von Neumann algebra Mooobtained by GNScompletion of U M
Now, if p < 4, the relative commutant R'
P n R is trivial and E is just the trace. (by
j20 j'
.6.2(~)when P< 4 and by [Jol], 55.3 when P = 4.) Thus Fk+l(ek) € C, and the proof
Theorem 4.4.3. (Skau's lemma). If /3 s 4 then {el,e2,. # ) ' n Mw = N.

Proof. Set = {el,e2,. .)' n M,. As N C fl is obvious, we have to show the
Remark. The condition P s 4 is necessary for Skau's lemma. For any pair N c M of
opposite inclusion. For each k 2 I, let Fk be the conditional expectation of Mw onto & 111factors with [M:N] > 4, we claim that
{ek,ek+l,, .1' Observe that FkFe = Fhn(k,e). We have to show that
F1(Mw) C N.
It is enough to show that F2(MJ c M, because this and Proposition 3.6.15 show
that FIF2(M,) c N. 1b We will see in Section 4.7.f that there is a projection eo and a subfactor P of N such
that M = (N,eo) and M is obtained by the fundamental construction for P c N. Then
Suppose we know that Fk+1(ek) E C for each k 2 1. One has then for t2 1 and for k
i
a,b E Me
k by 4.7.5,
Consider a Coxeter graph 'I of finite type in one of the classes A,D,E, with a
bicoloration involving m black vertices and n white vertices, and with a distinguished
white vertex wl. We shall associate to these data an irreducible subfactor C of the
hyperfinite factor B,of type 111 and we shall compute the index [B:C].
208 Chapter 4: Commuting squares and subfactors 5 4.5. More examples of irreducible pairs
In particular, if l? = E6 with the vertex wl chosen as vector of A ~ A ,normalized so that its first coordinate (corresponding to the chosen white
end vertex) is 1.
Let A be the Coxeter graph of type Ahl, with a bicoloration having at feast one dimension 0 on the Oth floor), and again form alternate stories by and rt. Compute
the dimension vectors on each floor by $(2j) = (rtr)j,dO) and = r(rtr)j$O).
white end vertex; choose one white end vertex and label it as the first white vertex.
Denote also by A the matrix of the graph A, and let 5
be the PerronFrobenius row
21d Chapter 4: Commuting squares and subfactors 5 4.5. More examples of irreducible pairs 2'11
Finally erase all vertices of dimension 0 and all edges emanating from such a vertex. If to the chain (B )
j heo
is in fact (isomorphic to) a tower obtained by iterating the
is the maximum distance from wl to any vertex of I', then for j 2 to1 fundamental construction.
The Bratteli diagram for (C.). is the same as that for (Nj)j20, and is obtained
I' if j is even J ~20
from' A, the Coxeter graph of type Ahl, in exactly the same way as that of (Bj)j20 is
obtained from I'; see section 2.9. In particular if j > h2 then
For example if I' is E6 with the distinguished white vertex at the end then the diagrams
are:
It now follows from Lemma 4.2.4 that for j 2 jo = max{to,h2) the "horizontal"
B. B
inchion matrices are also periodic, A Jl = A j+'. Therefore by Wenzl's
Cj1 Cj+l
Theorem 4.3.3, [B:C] = ilt(j)l12/llf(j)l12. for j > jo, where t(j) and are respectively
the w&ht vectors of Tr on C. and B Now for 2j 2 jO, d 2 j ) [resp. ~ ( ~ j is ) ]a
J j'
t
PerronFrobenius eigenvector for A ~ A[resp. I' I'] and so is proportiopal to 5 [rap. 8.
It remains only to establish the correct normalization of ~ ( ~ and
j ) t ( 2 ~ ) .We have
so the first component of k(2j) is ~j and $(2j) = ~ j t ,Similarly t(2j) = pljfi, .and
If tr is the normalized trace on M, note that Tr(+)= tr(.)/tr(p) is the unique, thus [B:C] = ))t(2j)1)2/))g(2j)))2
= )lfi)12/))t)12.
#
normalized trace on yBj, and if E. : Mj + MjVl is the trpreserving conditional
J
Pro~osition4.5.2. The possible values of the indez in Theorem 4.5.1 are as follows:
expectation, then E.(B.) = Bjl (since p E Mo) and E.
J J J Bj
I
is Trpreserving. Finally
for a E Bj, e.pae.p = E.(a)e.p, so {B.,e.p}" is an E.extension of B in the For I' of type At ( l 2 2): sin2[kr/(e+l)]/sin2[r/(l+1)], k = 1,2, *' ,[(l+l)/2].
J J J J J J J j'
terminology of Section 2.6. We have For I' of type Dl ( l ? 4): 2 sin2[klrj(2e2)]/sin2[rr/(2e2)], .
k = 1,2,. ,,l2,
Remark: The only one of the values between 4 and 5 is 3 + 8 g 4,732. The values Definition 4.6.1. The derived tower aM/dN is the chain of algebras
between 5 and 10 are:
[sin2 3r/lj/[sin2 r/lj for l 7, > C=N' n N c N ' flMcN' nM2c . * ..
8cos2 r / l for e > 6,
3+fig55,36 (D6)'
Lemma 4.6.2.
6 + 243 2 9,464 (E6), (i) 3'n Mk is Pnite dimensional for all k.
k
[sin2 4=/10]/[sin2 r/lO] g 9,472 (A3), (ii) If N and M are factors dim(N1nMk) i [M:N] .
ca. 7,759 (E7). (iii) With respect to tr, ,
310 9. ii3 L
4.6. The derived towcr and the Coxctcr invariant.
The results of this section, with the exception of 4.6.3(vi), were all known to V. Jones
before the inception of A. Ocncanu's work on subfactors. Nonetheless, our exposition has
been strongly influenced by conversations with Ocneanu, to whom we would like to record
is a commuting square for all k.
our gratitude.
The proof given in [Jo 11 for the restrictions on the possible values of the index of a
pair N C M of 111factors proceeded by constructing the tower associated to N C M and 
Proof. (i) It follows by induction on k, using condition 3.5.4(iv) that N is of finite
index in Mk for all k; hence by 3.6.2(a) N' n Mk
is finite dimensional.
then examining the algebra generated by the projections ei in the construction, as in
k
(ii) If N and M are factors, then [Mk:N] = [M:N] , so the inequality follows from
.
Chapters 2 and 3. It was of great importance that {ellea,. .,en}' is finite dimensional
3.6.2(b).
for each n > 1. Here we will describe another chain (Yk)k>O of finite dimensional
(iii) Follows from 4.2.7. #
algebras associated to a pair N c M, such that Yk contains {el,e2,+. .,ekl}', but is in
general strictly larger. It turns out that the chain (Yk)k,O is determined by a certain We assume henceforth that N, and M are factors and we denote N' 0 Mk by Yk,
(possibly infinite) graph, called the princi~al& of the pair N c M, which is a , the inclusion matrix for yk c Y ~ by
by E ~ and + A~ ~ Since
. ekxek = Ek(x)ek
E~kl
conjugacy invariant of the pair. In case [M:N] < 4, the principal graph is a Coxeter graph
=E (x)ek for X E Yk (by 4.6.2(iii)), {Yk,ek)" is an E extension of Yk, in
of type A, D, or E whose norm is the square root of index; this provides another proof of Ykl ykl
the restrictions on the index values, as well as a conjugacy invariant fiaer than index itself; the terminology of Section 2.6. Let Xk+l = (Yk,ey ) be the algebra obthined by the
these results were announced in [Jo~]. k1
Consider a connected inclusion N c M of finite von Neumann algebras with finite fundamental construction for Ykl c Yk, with respect to tr, for k 2 1. Give Xk+l the
dimensional centers, with N of finite index in M. Let not necessarily normalized trace Tr defincd by the weight vector ~ l ; ( ~  l ) , where
B(") is the weight vector of tr on Ykl, and the minimal central projections of Ykl
correspond to those of Xk+l as in 2.4.1. By 2.6.9, ukSl(xey x ' ) = xekxl defines an
k1
be the tower obtained by iterating the fundamental construction, as in Chapter 3, with injective algebra homomorphism from Xk+l into {Yk,ek)' c YkS1 with image
Mk+l = (Mk,ek) for all k. Write P = [M:N], and let tr be the unique trace on UMk YkekYk; it is evident that uk+l is a *homomorphism. We will refer to U ~ + ~ ( X ~ + ~ )
k
with the PMarkov property with respect to each inclusion Mkl C Mk. as "the old stuff" in Yk+l, because it is determined by Ykl t Yk. It turns out that the
old stuff is an ideal, whose complementary ideal we call "the new stuff".
214 Chapter 4: Commuting squares and subfactors 5 4.6. The derived tower 215
Theorem 4.6.3. (v) If $1 denotes the dimension vector of Yj, then for fued i and for ?i > 0,
(i) For k 1 and a E Xk+l, tr(%+l(a)) = Tr(a).
(ii) For k 2 1, U ~ + ~ ( X =~ Yk%Yk
+ ~ ) is an ideal in Yk+l.
(iii) For k 2 1, uk+l(lX ) is the central support o f % in Yk+l.
k+l
(iv) For all k, the inclusion Yk c Yk+l is connected.
(v) For all k, l l ~ ~ r1#[M:N].
1~
(vi) ("The new stuff comes only from the old new stufi") For k 2 2, i f x IE Yk+l dim Yi+2p = ll$(i+2p)l12 2 (~(A~A~)$~)l(i)ll~.
and X U ~ + ~ ( X=~ 0,+ ~then
) xuk(Xk) = 0.
2
(vii) For all k 2 1, the following are equivalent: Suppose that for some i, [M:N] r llAill , and choose c > 0 such that (~A~ll'(lc)> [M:N].
(a) flk+l(xk+l) = Yk+1. Let 3 be a PerronFrobenius vector for h;hi (which exists due to (iv)), normalized so
(b) Tr is normalized on Xk+l. that $(i) > 3 (componentbycomponent inequality). Then for p >1
(c)
(dl Ak =
is an eigenvector o f A ~  ~ A with
~  eigenvalue
~ b.
dim Yi+2p , =
ll(~:~~)~:11~ 11$114P11a12,
whence
I I ~ ~  ~=I iMZN].
I~ (dim Y )l/i+2p 2 [f$]1ii+2p.
(viii) If the equivalent conditions of ( 6 ) hold for k, then they also hold for k+l.
11~~11~
(i) We have to show that Tr(eyklx) = tr(ekx) for x IE Yk Since
Since the right side converges to 1 as p increases, it follows that for some k,
= Tr(ey
Weyk k1 Ek(x)) and tr(ekx) = tr(ekEk(x)), it is enough to prove the
(dim yk)lIk 2 (1r)llhi(12 > [M:N]
equality for x E Ykl, and since both x I+ Tr(ey x) and x I+ tr(ekx) are traces on
k1
Ykl, it suffices to prove it for x a minimal projection in Ykl. But if x is a minimal which contradicts 4.6.2(ii).
projection in some direct summand of Ykl, then by 2.6.4(c), e x is a minimal (vi) Because of (ii) we can suppose x is a central projection in YkS1; then xek = 0
Ykl
impliea xekl = eklxekekl = 0, so that also xaeklb = 0 for all a,b E Yk.
projection in the corresponding direct summand of Xk+l, whose trace Tr(ey x) is by
k1 (vii) Conditions (a) and (d) are equivalent since the inclusion matrix for Yk C XkS1
definition /T1tr(x) = tf(ekx).
is A&, and (a) is equivalent to (b) by (i). If (d) (and thus (b)) hold, then the vector of
(ii) We must show that if a,b E Yk and x IE Yk+l, then xae b E U ~ + ~ ( X ~ As
+~).
k tr on Y ~ is+ / ~ l ; ( ~  l )and
in the proof of 3.6.3, xaek = &+l(xq)ek, and by the NN bilinearity of Ek+l,
EkSl(xwk) E Yke SO =ekb E YkekYk.
(iii) This follows from (ii) and the fact that e has central support 1 in Xk+l.
ykl
(iv) An equivalent statement is that Ak is indecomposable for all k. This is evident Thus (d) implies (c). I£ (c) holds, then l l ~ ~ > fix~ and
) ~since tbe opposite inequality is
; true by (v), we have that (c) implies (el. If (e) holds, but not (d), than hk is a non
for k = 1 since Yo = C. By (ii) and 2.6.9, the matrix Ak+l has the form r:
r has no row of zeroes; hence if Ak is indecomposable, so is Ak+l. I negative matrix of the form , with , # 0, so , l ~ ~ >1 11 1 ~~ ~  ~=1 [M:N],
1~
?, contradicting (v).
Chapter 4: Commuting squares and subfactors $4.6. The derived tower 217
(viii) This follows from (d) and (e) of (vii) together with IlAll = 1 1 ~ ~ 1 1 . # The pair N c M would have depth 4 according to our conventions. The Bratteli
diagram of aM/aN for a finite depth pair N c M will always exhibit a growth in
Theorem 4.6.3 gives an interesting qualitative picture.of the Bratteli diagram for the complexity up to a certain level, after which the remaining structure is obtained by
derived tower aM/BN. One is led to the following concepts which have been emphasized reflecting. Note that if N c M has depth k, then the Bratteli diagram for YkWl C Yk is
by Ocneanu: isomorphic to the principal graph.
We can now record the following consequence of Theorem 4.6.3 and the work of
Definition 4.6.4. A pair N c M is said to have finite d e d if there is a k for which Chapter 1. ,
the conditions of 4.6.3(viii) hold. In this case the smallest such k is called the &p& of
N c M. Corollarv 4.6.6. Suppose N c M is a pair of 111factors with [M:N] < 4. Then
(a) [M:N] = 4 cos2 s / h for some h ? 3.
Definition 4.6.5. The princi~al & of N c M is the bipartite multigraph (b) The depth of N c M is no greater than h2.
constructed as follows: On the Bratteli diagram of the derived tower dM/ON delete on (c) The principal graph of N c M is a Cozeter graph of type A, D, or E, whose
each floor the vertices belonging to the old stuff, and the edges emanating from these norm is [ M : N ] ~ / ~ .
vertices.
Since the new stuff is connected only to the old new stuff, the resulting graph r is Proof. If N c M were not of finite depth, then one would have a sequence Ai of non
connected. The principal graph r has a distinguished vertex *, the unique vertex on
floor 0, and the distance of any vertex from * is the number of its floor. The pair N c M
negative integer matrices with 11$112 L l l ~ ~ + r~ [M:N]
l l ~ < 4, which is impossible by the
has finite depth if, and only if, I? is finite, in which case the depth of N c M is the classification of Chapter 1. If k is the depth of N c M, then for j 2 k1 the graph of Aj
maximum distance of any vertex from *. The Bratteli diagram of aM/aN can be 2
is isomorphic to the principal graph r, and 1 1 ~J . 1 1=~ llrll = [M:N] < 4. Therefore, by
reconstructed from r (given together with the distinguished vertex *). 2
Chapter 1, is a Coxeter graph of type A, D, or E and [M:N] = 4 cos s/h, where h
For purposes of illustration, let us give an example of what might be the Bratteli
is the Coxeter number. Furthermore k r diam(r) s h2. #
diagram of aM/aN and the principal graph for a pair N c M of finite depth. (We are
not claiming that this example actually occurs; this is a much more delicate question!)
This completes the proof of the restriction on index values outlined in [JoS].
4.7.a. Finite group actions. We shall analyze the derived tower for a pair N c M,
(doubly injnite linear graph) when N is the fixed point algebra M~ for an outer action of a finite group G on a 111
factor M, and also when M is the crossed product N r G of N by an outer action of a
finite group G on N.
(i) N = M ~ In
. this case we know from [Jol], that (M,eN) = M2 can be identified
with the crossed product M r G, so that Y2 = M2 n N' contains the group (von

Neumann) algebra CG. The inclusion matrix for C C CG is [no = l,nl,. ',nk], where
the ni are the degrees of the irreducible representations of G. Thus
llAlll 2
2
1 2
ni = I GI. On the other hand, by [Jol], [M:N] = IG I. Hence by 4.6.3,
(end vertez at distance n t o r n *) Y2 = CG and N c M is of depth 2. Note that the derived tower is independent of M or
the action of G; for example, in case G = S3 the Bratteli diagram for OM/aN is

Proof. (a) Follows from 4.6.3 and 1.4.3.
rk denote the subgaph of I' containing vertices of distance no greater than
(b) Let
k from *; then rk is also isomorphic to the Bratteli diagram for Ykl C Yk, 80
llrk12< 4 for all k. Thus each rk is a Coxeter graph of type A, D, or E; furthermore
is obtained from rk by addition of one or more vertices at distance k+l from *.
It is easy to verify that the only possibilities are those listed. #
We will see in Section 4.7 that subfactors of finite depth and of infinite depth do occur.
In particular all possibilities allowed by 4.6.7 do actually occur except possibly A,, and
D,,; we will also see why A is also labelled U.
m,m
Finally, we cannot resist saying a few words about the truly exciting results of
Ocneanu, who has added to the structure described here something we have completely Remark. According to Ocneanu, depth = 2 ,and N' n M = C characterizes fixed
neglected in our treatment, namely the involutions Ji coming from each basic point algebras of outer actions of finite dimensional Kac algebras.
construction in the tower Mi. He shows that they combine to define an endomorphism of (ii) M = N r G. In this case (M,eN) is known to be the crossed product of M by
aM/aN and, together with the eils and the principal graph, seem to complstely determine the "dual action" of G (see [NT]). To be more concrete, denote by u the canonical
g
N c M in many cases. In particular, he can show that there are osly finitely many unitaries of the crossed product. On L2(M), each of the projections e = u e u* onto
g gNg
subfactors of the hypefinite 111factor (of index < 4) for each Coxeter graph, up to the closure of Nu = u N is evidently in (M,eN) n N' = Y2. By the same reasoning as
g g
conjugation by automorphisms, and he determines which Coxeter graphs are allowed, in case (i), the (commutative) algebra which they generate is equal to Y2, and one always
has the following Bratteli diagram for aM/aN .
Chapter 4: Commuting squares and subfactors 3 4.7. Examples of,derived towers 221
unless i n3.)
Since Skau's lemma also is valid for P = 4, the same argument shows that the
principal graph for the pair R c R when P = 4 is
P
4.7.c. A eeneral method. The following result is useful as it allows the computation of
the derived tower in many examples coming from group aictions. A more powerful result
has been discovered and exploited independently by A. Wassermann [Wa].
Remarks. (1) Ocneanu's endomorphism allows one to reconstruct the multiplication
Lemma 4.7.1. Let N c M c P be von Neumann algebras, cp a faithjil normal state of
table for G, once the elements have been put into bijection with the factors on the third
P, e a projection in N' n P, and G a group of automorphisms of P preserving N, MI
line!
Q, and e. Suppose:
(2) By choosing G = 2/22 one obtains the Coxeter graph A3 as the principal graph,
(i) eMe = Ne,
and G = 2/32 gives D4 (for either the fixed point algebra or crossed product case). One (ii) dxe) = cp(e)cp(x) V x E M,
can also check that the pair N . .
Fi2 c N Fi3 or MG3 c Mb2 has principal graph As.
(iii) tr = cp
PI is a trace,
(6, denotes the symmetric group.) (iv) NG c MG c p G are 111 factors with [ M ~ : N =
~ [] P ~ : M=~ dB)'.
]
G
Then there is an isomorphism @ of ( M , e ) onto pG such that @(eNG)= e
NG
4.7.b. The An Coxeter graphs. Let be a sequence of selfadjoint
{eiIB0 and @(x)= x for all x E MG. ..
projections satisfying the relations
m. Let us first show that w e = E G(x)e for x E MG. By hypothesis, exe = ae
N
eieialei = ~ l e ~
for some a E N, and for all g E G we have ae = exe = g(we) = g ( x ) = g(a)e. But then
e I. eJ . = ej ei for lijl22,
de)cp((ag(a))*(ag(a))) = rp((ag(a))*(ag(a))e) = 0, so a E NG since cp is faithful.
Also if y E N ~ then, cp(e)tr(yx) = tr(yxe) = tr(eyxe) = tr(yexe) = tr(yae) = cp(e)tr(ya),
for some P > 0. Realize the hyperfinite 111 factor R as R = {l,eo,el, ...)' and write so a = E G ( ~ ) .
N
RP for {P,el,e2,...) " . We have computed that [R:R ] = P (Proposition 3.4.4).
P
When P = 4 cos2 ir/n for some n 1 3, it follows easily from the proof of Skau's
Now by 2.6.9 and 3.6.4 there is a *algebra isomorphism @ of (M
G
.
such that @(xe Gy) = xey for x,y E M G . The map @ is trace preserving (since
eNG) Into
G
lemma (4.4.3) that the principal graph of R c R is the Coxeter graph Anql. In fact,
P N
write el,e2,... for the projections occurring in the tower construction for R c R, so tr(e G) = 1 ~ ~ : ~ = tr(e)),
~ l  so l normal and unital. We need only show that is
P N
that Mi = {l,ei+,l ,...,el,e2 ,...)" (i 1 0). The proof of Skau's lemma shows that
surjective. But the image of @ is a 111 factor containing MG (since @ is unital) as'a
N' r l Mi = {l,ei+l,...,el)', and we saw in Section 2.9 that the chain (N' n Mi)i2l has
subfactor of index tr(e)' = [PG:M G1, so the image is P G. #
the appropriate Bratteli diagram. (The statement of Skau's lemma does not apply since
{ei+l,...,eo,el)' is not isomorphic to the result of the fundamental construction for
222 Chapter 4: Commuting squares and subfactors 3 4.7. Examples of derived towers 223
This lemma will be used repeatedly in 4.7.d and 4.7.f to calculate derived towers by ((R @ ~ a t ~ ( C ) ) ~ onto
, e ~ (R
~@) Mat2(C)) @ ~ a t ~ ( C taking
) ) ~ eRG to el. If G is
calculating it in a simple situation a q i then passing to the fixed point algebra of some
group action. See also [PP3]. Wassermann calls the lemma the invariance principle, since infinite, we have to do a little more work to reach the same conclusion.
in cases where [M:N] makes sense it should also be rp(e)'. Identify R (Gequivariantly) with Rg @ Mat2(C), where Ro g R. Set
1
eo = lRO @ {ell @ en  e12 @ eZl  eZ1 @ el, +e22 @ ell). Thus (R @ Mat2(C))eo
4.7.d. Some examales of derived towers for index 4 subfactors. We realize the
hyperfinite 111 factor R as the completion, with respect to the unique tracial state tr, = Reg (by 3.6.3 or by direct computation). If x E (R @ Mat2(C))G, then there is an
m xOE R such that xeo = xoeo, and
of the infinite tensor product of Mat2(C), R = (@ Mat2(C)). Any closed subgroup G of
m
SU(2) acts on @ Mat2(C) by the infinite tensor product of its action by conjugation on
Mat2@). The action preserves the trace, so extends to an action on R. The group G
acts in the same way on R @ Mat2(C), so one has a commuting square
so (R @ ~ a t ~ ( C )=) RGeO.
~ e ~ Therefore,
1
el = 21R {ell
@ @ 22  el2 e21 e21@el2 + e22 ell)
@ @
Thus we have identified the derived tower for R~ c (Re at^(^))^ with the sequence of and JMJ = MIl = (N1,e0), SO M = JMLIJ = (N,eo). That is, the pair N c M is the
[Mat2(C) @ at,(^)]^, etc. Iterating this construction, one obtains a decreasing "tunnel1'of 111 factors
It is fairly evident that these algebras are just the cornmutants of the tensor powers of
the fundamental representation of G on C2 determined by its inclusion in SU(2). We
can now use the McKay correspondence between finite subgroups of SU(2) and affine
Coxeter graphs (see [Slo] or [Jo 41) to calculate the Bratteli diagrams or principal graphs along with projections {eo,el,e2,' a ' ) such that (Mk,ek) = Mk+l. The projections
when G is finite. The correspondence is as follows. Let I? be the matrix with rows and
{ei : i 5 0) satisfy the usual relations with ,O = [M:N].
columns indexed by irreducible representations of G, whose (i,j)entry is the multiplicity
of j in the tensor product of i with the fundamental representation of G on C2. Then r The tunnel construction has been exploited systematically by Ocneanu in his
is the adjacency matrix of an afSne Coxeter graph of type A,): DA1), EA1), EP), or classification of subfactors.
Theorem 4.7.3 (Popa). Suppose M is a 111 factor containing a projection f, with <21 Thus ER(f) E Rb n R\C1.
which is impossible since t
tr(f) < $, such that then is an isomorphism 0: f 11f). Let Now let fo be any projection in R' n R with 0 < tr(fo) 5 $. Ultimately we will
P
N = {x+B(x) : x E N f ) . (By 4.7.2, [M:N] = tr(f)' +
tr(1f)' := P; note that P > 4.)' show that f = fo. Since

Let {eo,el,e2,+ ') be projections in the tunnel construction for the pair N c M, set
R = { e o , e  l , ~ ~ ~ )and
' RP= {el,e2,.,.)'. Then 1
f
1
+ 1t = P = [R:R$
1
+ 1 (by the proof of 3.6.2),
0 '0
(1) f E R b n R ,
(2) fRf = R f and (1f)R(lf) = R (1 f). it follows by calculus that tr(fo) ) t. Assume tr(fo) > t. If g := fo A (1f) = 0, then
P P 
fo = fo  fo "1f) N fo V (If)(lf) 5 f, so tr(fo) < t, contradicting our assumption.

Proof. Write t = tr(f), so 8 = tI + (1t) 1. Since EN(eo) = P1,
Since EN(fo) E N n R n Rb = RP n Rb = El, it follows that EN(fo) = tr(fo). Thus
tr(p) =
1
= t(lt), and set q = 1f8(p), so tr(q) = r1
as well. Let v be a partial Define fl = f  f A fo and f2 = s((1f)fo(lf)), where s() denote support. Remark
isometry in M such that v*v = q and vv* = p and define that
s((f0fAf0) (1f) (fofAf0)) = fofAfo
e = (It)p + tq + ATJTfJ(v+v*).
because (fofAfo) E; f = 0. Hence
Then e is a projection. One can check that for any x E M,
= t(fAfo+h) + (lt)(ol(lf)fo(lf) + (If)fo(lf)) 4.7.3. Then f E {eo,el,e2,. .}' n {el,e2,. .)'. But the the pair R 3 Rp is
> th + (1t)(Pf)fo(nf). isomorphic the the pair {eo,el,e2,. .}' 2 {el,e2,. .'1'. #
Hence
th 2 th + (1t)h(lf)fo(Pf)h, We now describe a representation of the ei's due to Pirnsner and Popa [PP 11. Let A
be the infinite tensor product of Mat2(C), that is, the inductive limit of the algebras
so h(1f)fo(lf)h = 0. It follows that n
@ Mat2(C), each imbedded into the next via x x @ 1. It is well known that A has a
I,
1
unique c*norm and that K, the c*completion of A, is asimple c*algebra.
Fix /3 > 4 and let X E ]0,1[ satisfy 2+X+Xe1 = P. Define a state r] on Matp(C) by
so B(fAfo) 5 Iff2. On the other hand, multiplying the equation q(x) = tr [[i !]XI a n d let 9 be the corresponding Powers state [PI on ?
s
i
defined
.
,I
>
which implies that B(fohf) (Iff2) Ilence O(fohf) = PffZ
which preserves 9 and so extends to N. Define el E Mat2(C) @ Mat2(C) by
Finally we compute
Solving gives (12t) tr(fohf) = (12t)t, or, since 12t > 0, tr(fohf) = t = tr(f) = tr(fo).
It follows that f = fAfO= fo and f E R n Rb , as was to be shown.
under an appropriate identification of hIat2(C) @ Mat2(C) with n;l&t4(C). Let
If x E R, then tlEN(xfx) = fxf + O(fxf) lies in R fl N = R p , and
ei = &'(el) for i 2. It is a bit tedious but straightforward to check that {ei : i 2 1)
f d = tlEN(cfc)f E Rp f. Thus fRf = R f. Similarly (If)R(lf) = RP (If). #
P satisfy
p eiei+lei = ei,
Corollarv 4.7.4. [Jol] Let P > 4, and let R 3 R be the pair of factors in Proposition
P 1ij 1 > 2,
3.4.4. Then R n Rb C l . ; eiej = ejei,
P.
&& Let M be the hyperfinite 111 factor, and let f E M be a projection of trace
and that cp al
I g{1,e1,e2,...l
is the Markov trace of modulus The circle group acts on
t= $ ,
(If)M(lf).
so that t'(1t)" = /3. Then there is an isomorphism 8 : fMf
Define the factor N and the projections {eo,el,e_2,e. .) u in Theorem
 A by the infinite tensor product of its action z I+ Ad [" ,IO I
preserves 9 , and extends to an action on N. One checks that the projections ei are
on Mat2(C); this action
Tinvariant. Equivalently (since the Taction is actually the modular automorphism group
230 Chapter 4: Commuting squares and subfact~rs 5 4.7. Examples of derived towers 231
for (o) the eils are in the normalizer of (o, (o(eiy) = dyei) for all y E N. We shall T T
.
projections {el,e2,. .) axe those of a tunnel construction for a(N ) c N Therefore the
apply Lemma 4.7.1 with M = N @ Mat2(C), P = N @ Mat2(C) @ Mat2(C), G = T, , and conclusion f E {el,e2,. .'1' follows from Popa's theorem 4.7.3. #
the state (o @ q @ q on P, in order to compute the derived tower for R c R. The key
P
result is the following, from [PPl]. There is no further difficulty in computing the entire derived tower for RP c R. Using
4.7.1 and 4.7.5, the tower for the pair R c R can be identified with
P
Theorem 4.7.5. With the notation above, {el,e2,. .)' k the $zedpoint algebra NT
for the T action on N.
/
A proof of this result using the ConnesStormer relative entropy is in [PPl]. It is well known that (NT)' n N = C1. In fact, the infinite symmetric group em acts
We shall give a simplified proof due to Popa which is based on his Theorem 4.7.3.
.
First we observe that {l,el,e2,. .) together with the projection f = ell E Mat2(C) ergodically on N (by permutation of the tensorands), and this action is implemented by
T in N' is the group generated by { ok (u) : k?O ).
unitaries in N ; the image of
generate NT. In fact felf = 1 ell @ e22 and (If)el(lf) = X e22 @ ell, so that k k k
T T
2 Therefore (N')' n (N @ (@ Mat2(C)) = 1 @ (@ MatZ(C)) and (N )' rl (N @ (@ Mat2(C)) =
alg{l,f,el) contains the selfadjoint unitary u = as well as the diagonal k
e i j @e.
i , j=l
J ,i 1@ (@ at^(^))^. That is, the derived tower is precisely the sequence of fixedpoint,
k
algebra {dl @ d2 : di is a diagonal 2by2 matrix), and contains in particular 4 f ) = algebras for €he tensor product action of T on @ Mat2(C) (k>O). The Bratteli diagram is
.
1@ e l l By induction, alg{l,f,el,e2,. ,) contains the infinite tensor product of the 2by2
Pascal's triangle [Bra]:
diagonal algebra, as well as {crk (u): k?O). It is a well known fact of invariant theory that,
m r
on the Fock space $ (ic2) over C2, the mmmutant of the direct sum of the tensor
k=O
powers of the representation a I+
[" I
Zl of T on t2is the algebra generated by the
permutations {$(a) : k?O ) together with the infinite diagonal algebra. This implies
that 5 . .
~ , and NT = {l,f,el,e2,. .)'. Therefore it suffices to prove
= ~ * { l , f , e ~ , e)
that f E {el,e2,. .)". +
..)'
The equality NT = {l,f,el,e2,.
+
implies that fNTf = a(NT)f and (1f)N T(If) =
so that 8: xf x(lf) (x E or(NT)) is an isomorphism of fNTf onto
and o(.NT) = {x B(x) : x E fNTf ). Thus the pair 4 ~ ' ) c N' is of the
, and the principal graph is
APPENDIX I Pro~osition1.1.1. Let r be a jnite connected graph which is not a cycle and which
Classi£ication of Coxeter graphs with spectral radius contains a cycle. Then llrll > A,. Moreover Am is the largest constant for which this
just beyond the Kronecker range
holds.
1.1. The reaulta.
One important result about the set E of norms of graphs (see section 1.5) is also
The first purpose of this appendix is to complement section 1.4 and to classify finite essentially due to Hoffman [Hofj, and can be stated in terms of a sequence (A ) of
q22
connected graphs satisfying 2 < llrll < Am where we set
numbers defined as follows. Let v be the largest real root of the polynomial
9
We set
The results are those of Cvetkovi6, Doob and Gutman [CDG]. Then we also classify
Coxeter e r a ~ h with
s norms satisfving the same ineaualitv.
The following graphs enter the classification. First, the T 's already introduced
P,W If vm = $(5lI2+1) is the golden mean, namely, the positive root of &I, it follows by
in Section 1.4. Next, for an integer m 2 3, we set induction on q from the identity L (v)  L (v) = vq'(2v1) that
~,1 q+l q
+ 
p5  9p4 27p3 31p2 12p 1. + 
(ii) For q > 3 and p with 2 5 p 5 q:
llT2,q,r 11 increases strictly with r and converges to X q'
with p+q+r3 vertices. llT3,3,rII increases strictly with r and converges to Am,
The constant Am makes its first appearance in a result of Hoffman (Proposition 3.7 of llHp,q,r 11 decreases strictly with r and converges to X Q'
[HofI):
234 Appendix I: Classification of Coxeter graphs 5 1.2. Characteristic polynomials
To state the corresponding classification of Coxeter graphs, we set 1.2. Computations of characteristic polynomials for ordinary graphs.
F$+ 0
 2 404, Most computations below are by induction, based on the following.
Fk 4
0000. .
4
k vertices
Lemma 1.2.1. Let I' be a graph and let v be a vertez of I?. Let I' be the graph
Hk 5   0 ~  .e
0 .  0 4
k vertices
obtained from I? by deleting v and all edges ending at v; let I'+ be the graph obtained
Theorem 1.1.3. Let I' be a jnite connected Coxeter graph with 2 < IlI'll i Am. Then I' from I? by adding a vertez v+ and an edge between v and v+. Denote by P, P and
is one of the following. P+ the characteristic polynomials of I', I' and I'+. Then
F
+: with norm approximately 2,053; P+(X) = XP(X) P(A).
Fk with k 1 6; the norms llFkll increase and tend to Am;
Hk with k 5; the nonns JIHkJJincrease and tend to Am; w.By standard expansion ofdeterminants; more similar lemmas in [Schl]. #
T2,q,r with 3 i q 5 r, and r t 7 i f q = 3, and r 2 5 if q = $ the norms increase
One example is provided by
and tend to X .
9'
T2,q,r(4) with 3 < q 5 r and r large enough; the n o m s decrease and tend to X .
4'
T3,3,r with r 1 4; the n o r m increase and tend to Am;
T3,4,4 with norm apf;roximately 2,053;
Let P E U[A] be the characteristic polynomial of some graph I'. For our
Hp,q,r(m,m') with 2 5 P i q 2 3 and r large enough, with m,ml E {3,4} and computations, it helps to consider larger rings
m = 3 i f p = 2; the n o m s decrease and tend to X
q'
Remarks.
(1) The number lF:+112 = ]IT3 4112 is the largest raot of p3  6p2 + 8p  2.
(2) The polynomials
checked whether
L2 ( 4
e,
L v
9
L3(u),
I
L4 ( "1
are irreducible in Ulu]. We have not
3 3
For example, we write P(A) = p + pw32 to mean P(A) = X 3X2. The function
&&. Expansion of the determinant defining pf 'l along the first column gives
Promsition 1.2.2. One h m P;(A) = (ypl)l(pl'+lpcckkl) = AP ~ ( A
P~~'(A)A as)a sum of two kbyk determinants; expansions of these along their
~ ( X  ~ C O S ( T ~ / ( ~and ) ) ) = 2m(d(k+l)).
+ ~BAkll first lines give
li jik
and the formula for P The largest root of pl\'l is given by p = 1, namely by
P:+~(A)  AP:(A) + PEl(A) = 0
X = 2 (see also lemma 1.4.1).
~ ~by ~lemma 1.2.1. Set pm = (251/2+1))1/2 = v1l2
The formula for P : ~ follows m
which holds also for k = 1 if p i ( \ ) = 1. The indicial equation being $  Ap 1 = 0, + 4 2
(with vm as in 1.1). If p = pm one has p p 1 = 0 and
+
the general solution of the difference equation is Pk(A) = Cpk ~ p  ~where
, C and D
are constants in ~[p,pl] independent of k. Adjustment of C and D to fit P; and
P: provide the formula with p's. Roots of P;(A) = fk(p2l)1(p2k+2~) are given
by p2 = exp(2ilrj/(k+l)), namely by Consequently the lefthand side has a root larger than pm, and 1 1 ~ P ~ l )> l Am
l = pm+ pi1
for all k 2 2. Let p > y then 4p21 > 0 and ~ : ~ ~ ' ~ ( p + p> 0~ for
) k large enough.
Consequently l i m l l ~ ( ~ ~ =~ )Am.
..
for j = 1,. ,k. The largest of these is 2cos(s/(k+l)). # k+m k
1 1 It is clear that I ~ A ~ is~ strictly
~ ) I decreasing
I when k
and by P:,~
'' its characteristic polynomial.
Say that L is internal if there &st an integer n 2 1 and a sequence of vertices
Lemma 1.2.3. One has P:ll(X) = ,$+I+ pkl 2 and I I A ~ ~=)2.I I One has .
xO,. .,xj = U, xjS1 = w,. .,xn such that all xi are distinct (except possibly xo = x,),
where the degrees d(xi) satisfy
238 Appendix I: Classification of Coxeter graphs 8 1.2. Characteristic polynomials
&f. .
Let xo,, ,x, be as above. We shall assume that xo # xn (the case xo = 5,
substantially easier, is left as an exercise, and is also done with details in [HS]). Let 
..
x  ~ , . 'x1 be vertices adjacent to xo other than xl; by hypothesis m 2 2. Let [ be with at least one ingquality being strict by (**). The conclusion follows.
a PerronFrobenius vector for the adjacency matrix Y of r; one has Y[ = IlI'll[; also
Ilrli > 2 because there exists an integer k such that DL1) is a proper subgraph of I'. Assume finally that
m
i=i
xf,
< k. One has toI IlI'llC, for i = l,...,m. since
adjacency matrix. Suppose one can find a positive vector 2 E llii with $2 $ (II'/I&it will
m 2 2, this implies 2t0 m g i III'IIC~~
and consequently
i=i
then follow from PerronFrobenius theory that llf 11 < IlI'll; see lemma 2 and remark 4 in
5 XIII.2 of [Gan]. We distinguish three cases.
Assume first that there exists t E (1, . ,n1) with [+, 5 4
for i = 0,. .,n. There is .
no loss of generality in assuming xt = u and x ~ =+ W.~ One has
rn
Define 2 by $ = x4 and 4 = to, the other coordinates being as those of [. One
i=i
has
Define i to be the vector with coordinate corresponding to v given by iv= f t and with
other coordinates as those of [. One has
One may also apply lemma 1.2.4 to Coxeter graphs such that, notations being as above,
with at least one inequality being strict by (*). Hence s llr~~i
and Pi # IlI'lli. for~ )i = 1,. . ,n are marked with m = 3.
edges ( X ~  ~ , X
For the two next cases, to< ei .
for i = 0,. .,n, and we assume xo = u, xl = w
without loss of generality.
= 'f,3,k3.
240 Appendix I: Classification of Coxeter graphs 5 1.2. Characteristic polynomials 241
a
P;,,,(\) = (pP1)1{prq+1[vyvq2+. * .
.+I)]  pr+q1[uyu~+2+. .+I)]). A6  6A 4 + 9A2 3 = 8cos3(28)  6 ~ 0 ~ ( 21
8 ) = 2cos(68) 1;
In particular,
1
the roots of P; are given by A = 0 and cos(66) = Z, namely by A = 0 and
P;(U = (P+Pl )(Pkl +P"1) = n(wcos(m.,h))
1s jdk J A = 2cos(b + ji) for j = O,1,2f ,4,5. Finally, set Q(1) = A4 7A3 +14A2 8A +l; a
with h = 2k2 and ml,* ,mk = 1,3, . .. ,2k3,k1;
straightforward computation shows that
Much more than documents.
Discover everything Scribd has to offer, including books and audiobooks from major publishers.
Cancel anytime.