ContemporaryMathematics 3 0 0 2 Relative Tensor Products for Modules over von Neumann n Algebras a J David Sherman 8 ] Abstract. We give an overview of relative tensor products (RTPs) for von A Neumann algebra modules. For background, we start with the categorical O definition andgoontoexamineitsalgebraicformulation,whichisappliedto . Morita equivalence and index. Then we consider the analytic construction, h withparticular emphasisonexplainingwhytheRTP isnotgenerallydefined t foreverypairofvectors. Wealsolookatrecentworkjustifyingarepresentation a m ofRTPsascompositionofunboundedoperators,notingthattheseideaswork equallywellforLpmodules. Finally,weprovesomenewresultscharacterizing [ preclosedness ofthemap(ξ,η)7→ξ⊗ϕη. 1 v 1 6 1. Introduction 0 1 The purpose of this article is to summarize and explore some of the various 0 constructions of the relative tensor product (RTP) of von Neumann algebra mod- 3 ules. Alternately knownascompositionorfusion,RTPsarea keytoolinsubfactor 0 / theoryandthestudyofMoritaequivalence. Theideaisthis: givenavonNeumann h algebra M, we want a map which associates a vector space to certain pairs of a t a rightM-module anda left M-module. Ifwe write module actionswith subscripts, m we have v: (XM,MY)7→X⊗MY. i This should be functorial, covariant in both variables, and appropriately normal- X ized. Other than this, we only need to specify which modules and spaces we are r a considering. In spirit, RTPs are algebraic; a ring-theoretic definition can be found in most algebra textbooks. But in the context of operator algebras, the requirement that the output be a certain type of space - typically a Hilbert space - causes an ana- lytic obstruction. As a consequence, there are domain issues in any vector-based construction. Fortunately, von Neumann algebras have a sufficiently simple repre- sentation theory to allow a recasting of RTPs in algebraic terms. TheanalyticstudyofRTPscanberelatednicelytononcommutativeLpspaces. Indeed, examination of the usual (L2) case reveals that the technical difficulties 2000 Mathematics Subject Classification. Primary: 46L10; Secondary: 46M05. Key words and phrases. relativetensor product,vonNeumannalgebra,bimodule. (cid:13)c0000 (copyright holder) 1 2 DAVIDSHERMAN come from a “change of density”. (We say that the density of an Lp-type space is 1/p.) Once this is understood, it is easy to handle Lp modules [JS] as well. Modular algebras ([Y], [S]) provide an elegant framework, so we briefly explain their meaning. The final section of the paper investigates the question, “When is the map (ξ,η) 7→ ξ ⊗ η preclosed?” This may be considered as an extension of Falcone’s ϕ theorem [F], in which he found conditions for the map to be everywhere-defined. We consider a variety of formulations. We have tried to make the paper as accessible as possible to non-operator algebraists,especiallyinthefirsthalf. Ofcourse,evenatthislevelmanyresultsrely on familiarity with the projection theory of von Neumann algebras; basic sources are [T1], [T2], [KR]. Primary references for RTPs are [Sa], [P], [F], [C2]. 2. Notations and background Thebasicobjectsofthispaperarevon Neumann algebras,alwaysdenotedhere by M, N, or P. These can be defined in many equivalent ways: • C*-algebras which are dual spaces. • strongly-closed unital *-subalgebras of B(H). B(H) is the set of bounded linearoperatorsonaHilbertspaceH;the strongtopologyisgeneratedby the seminorms x7→kxξk,ξ ∈H; the * operation is given by the operator adjoint. • *-closed subsets of B(H) which equal their double (iterated) commutant. The commutant of a set S ⊂B(H) is {x∈B(H)|xy =yx,∀y ∈S}. As one might guess from the definitions, the study of von Neumann algebras turns on the interplay between algebraic and analytic techniques. Finite-dimensional von Neumann algebras are direct sums of full matrix alge- bras. Attheotherextreme,commutativevonNeumannalgebrasarealloftheform L∞(X,µ) for some measurespace (X,µ), so the study of generalvonNeumann al- gebrasisconsidered“noncommutativemeasuretheory.”Basedonthis analogy,the (unique) predual M∗ of M is called L1(M); it is the set of normal (= continuous in yet another topology, the σ-weak) linear functionals on M ⊂ B(H), and can be thought of as “noncommutative countably additive measures”. A functional ϕ is positive when x> 0⇒ϕ(x) ≥0; the set of positive normal functionals is denoted M+. The support s(ϕ) of a positive normallinear functional ϕ is the smallest pro- ∗ jection q ∈ M with ϕ(1−q) = 0. So if M is abelian, ϕ corresponds to a measure and q is the (indicator function of the) usual support. Forsimplicity,allmodulesinthispaperareseparableHilbertspaces(exceptin Section6),allalgebrashaveseparablepredual,alllinearfunctionalsarenormal,and all representations are normal and nondegenerate (MH or HM is all of H). Two projections p,q in a von Neumann algebra are said to be (Murray-von Neumann) equivalent if there existsv ∈M with v∗v =p, vv∗ =q. Suchanelementv is called a partial isometry, and we think of p and q as being “the same size”. Subscripts are used to represent actions, so XM indicates that X is a right M-module, i.e. a representation of the opposite algebra Mop. It is implicit in the term “bimodule”, or in the notation MHN, that the two actions commute. The phrase “left (resp. right) action of” is frequently abbreviated to L (resp. R) for operators or entire algebras, so that we speak of L(x) or R(M). Finally, we often write M∞ for the vonNeumann algebra of all bounded operators on a separable infinite-dimensional RELATIVE TENSOR PRODUCTS 3 Hilbert space, and M∞(M) for the von Neumann tensor product M∞⊗¯M. One can think of this as the set of infinite matrices with entries in M; we will denote by e the matrix unit with 1 in the ij position and 0 elsewhere. ij The (left) representationtheory of von Neumann algebras on Hilbert spaces is simple, so we recall it briefly. (Most of this development can be found in Chapters 1 and 2 of [JoS].) First, there is a standard construction, due to Gelfand-Neumark and Segal (abbreviated GNS), for building a representation from ϕ ∈ M+. To ∗ each x ∈ M we formally associate the vector xϕ1/2 (various notations are in use, e.g. η (x)orΛ (x), butthisoneisespeciallyappropriate([C2]V.App.B,[S])).We ϕ ϕ endow this set with the inner product <xϕ1/2,yϕ1/2 >=ϕ(y∗x), and set H to be the closure in the inherited topology, modulo the null space. ϕ The left action of M on H = Mϕ1/2 is bounded and densely defined by left ϕ composition. When ϕ is faithful (meaning x > 0 ⇒ ϕ(x) > 0), the vector ϕ1/2 = 1ϕ1/2 is cyclic (Mϕ1/2 = H ) and separating (x 6= 0 ⇒ xϕ1/2 6= 0). Now all representa- ϕ tions with a cyclic and separating vector are isomorphic - a sort of “left regular representation”;we will denote this by ML2(M). It is a fundamental fact that the commutant of this action is antiisomorphic to M, and when we make this iden- tification we call ML2(M)M the standard form of M. If ϕ is not faithful, the GNS construction produces a vector ϕ1/2 which is cyclic but not separating,and a representationwhich is isomorphic to ML2(M)s(ϕ) ([T2], Ch. VIII, IX). Now let us examine an arbitrary (separable, so countably generated) module MH. Following standard arguments (e.g. [T1] I.9), H decomposes into a direct sum of cyclic representations M(Mξn), each of which is isomorphic to the GNS representationfortheassociatedvectorstateω (=<·ξ ,ξ >). Withq =s(ω ), ξn n n n ξn we have MH≃ MMξn ≃ MHωξn ≃ ML2(M)qn. M M M (Hereandelsewhere,“≃”meansaunitaryequivalenceof(bi)modules.) Sincethisis aleftmodule,itisnaturalto writevectorsasrowswiththenthentryinL2(M)q : n (2.1) H≃(L2(M)q L2(M)q ···)≃(L2(M)L2(M) ···)( q ⊗e ). 1 2 n nn X We will call such a decomposition a row representation of MH. Here enn are diagonalmatrix units in M∞, so( qn⊗enn)is a diagonalprojectioninM∞(M). The left actionofM is, ofcourse,Pmatrix multiplication (by 1×1 matrices)onthe left. Themodule(L2(M)L2(M)···)willbedenotedR2(M)(for“row”). Sincethe standard form behaves naturally with respect to restriction - L2(qNq)≃qL2(N)q asbimodules - itfollowsthat L2(M∞(M))is built asinfinite matricesoverL2(M) (see (3.3)). Thus R2(M) can be realized as e11L2(M∞(M)). Proposition 2.1. Any countably generated left representation of M on a Hilbert space is isomorphic to R2(M)q for some diagonal projection q ∈M∞(M). Any projection in M∞(M), diagonal or not, defines a module in this way, and two such modules are isomorphic exactly when the projections are equivalent. In fact (2.2) Hom(MR2(M)q1,MR2(M)q2)=R(q1M∞(M)q2). 4 DAVIDSHERMAN Soisomorphism classescorrespondtoequivalenceclassesofprojectionsinM∞(M), which is themonoid V(M∞(M))in K-theoreticlanguage [W-O].The direct sum of isomorphismclassesofmodulescorrespondstothesumoforthogonalrepresentatives in V(M∞(M)), giving a monoidal equivalence. We denote the category of separable left M-modules by LeftL2(M). For us, the most important consequence of (2.2) is that (2.3) L(MR2(M)q)=R(qM∞(M)q), where “L” stands for the commutant of the M-action. (In particular, the case q =e11isjustthestandardform.) ThealgebraqM∞(M)qiscalledanamplification of M, being a generalization of a matrix algebra with entries in M. Of course everything above can be done for right modules - the relevant abbreviations are C2(M), for “column,” and RightL2(M). Example. SupposeM=M (C). Inthiscasethestandardformmaybetaken 3 as L2(M ) ; L2(M )≃(M ,<·,·>), where <x,y >=Tr(y∗x). M3 3 M3 3 3 Notethatthisnorm,calledtheHilbert-Schmidtnorm,isjusttheℓ2normofthema- trix entries, and that the left and right multiplicative actions are commutants. (If wehadchosenanontracialpositivelinearfunctional,wewouldhaveobtainedaniso- morphicbimodule witha“twisted”rightaction... this isinchoateTomita-Takesaki theory.) The module R2(M3) is M3×∞, again with the Hilbert-Schmidt norm, and the commutant is M∞(M3) ≃ M∞. According to Proposition 2.1, isomor- phismclassesofleftM -modulesshouldbe parameterizedbyequivalenceclassesof 3 projectionsinM∞. Theseareindexedbytheirrankn∈(Z+∪∞); thecorrespond- ing isomorphism class of modules has representative M3×n. In summary, we have learnedthatanyleft representationofM onaHilbertspaceis isomorphictosome 3 number of copies of C3. The same argument shows that V(M∞(Mk))≃(Z+∪∞) for any k. Propertiesofthe monoidV(M∞(M)) determine the so-calledtype of the alge- bra. For a factor (a von Neumann algebra whose center is just the scalars), there are only three possibilities: (Z ∪∞), (R ∪∞), and {0,+∞}. These are called + + types I, II, III, respectively; a fuller discussion is given in Section 7. 3. Algebraic approaches to RTPs When R is a ring, the algebraic R-relative tensor product is the functor, co- variantin both variables,which maps a right R-module A and left R-module B to the vector space (A⊗ B)/N, where N is the subspace generated algebraically alg by tensors of the form ar ⊗b−a⊗rb. In functional analysis, where spaces are usuallynormedandinfinite-dimensional,oneobviousamendmentisto replacevec- tor spaces with their closures. But in the context of Hilbert modules over a von Neumann algebra M, this is still not enough. Surprisingly, a result of Falcone ([F], Theorem 3.8) shows that if the RTP L2(M)⊗M L2(M) is the closure of a continuous (meaning kI(ξ⊗η)k<Ckξkkηk) nondegenerate image of the algebraic M-relativetensorproduct,Mmustbeatomic,i.e. M≃⊕ B(H ).Wewilldiscuss n n the analytic obstruction further in Section 5. For now, we take Falcone’s theorem as a directive: do not look for a map which is defined for every pair of vectors. If we give up completely on a vector-level construction, we can at least make the functorial RELATIVE TENSOR PRODUCTS 5 Definition 3.1 (Sa). Given a von Neumann algebra M, a relative tensor product is a functor, covariant in both variables, (3.1) RightL2(M)×LeftL2(M)→Hilbert: (H,K)7→H⊗MK, which satisfies (3.2) L2(M)⊗ML2(M)≃L2(M) as bimodules. Although at first glance this definition seems broad, in fact we see in the next proposition that there is exactly one RTP functor (up to equivalence) for each al- gebra. The reader is reminded that functoriality implies a mapping of intertwiner spacesaswell,soitisenoughtospecifythe maponrepresentativesofeachisomor- phism class. In particular we have the bimodule structure L(HM)(H⊗MK)L(MK). Proposition 3.2. Let H≃pC2(M)∈RightL2Mod(M) and K≃R2(M)q ∈ LeftL2Mod(M) for some projections p,q ∈M∞(M). Then H⊗MK≃pL2(M∞(M))q with natural action of the commutants. Proof. Byimplementinganisomorphism,wemayassumethattheprojections arediagonal: p= p ⊗e , q = q ⊗e .Using (3.2) andfunctoriality,we have i ii j jj the bimodule isomPorphisms P H⊗MK≃ ⊕piL2(M) ⊗M ⊕L2(M)qj (cid:0) (cid:1) (cid:0) (cid:1) ≃ piL2(M)⊗ML2(M)qj ≃ piL2(M)qj ≃pL2(M∞(M))q. Mi,j Mi,j (cid:3) Visually, L2(M) L2(M)L2(M) ... (3.3) (p)L2(M), (L2(M)L2(M)...)(q)7→(p)L2(M)L2(M) ... (q), ... ... ... ...       where ofcourse the ℓ2 sums of the norms ofthe entries in these matrices are finite. After making the categorical definition above, Sauvageot immediately noted that it gives us no way to perform computations. We will turn to his analytic construction in Section 5; here we discuss an approach to bimodules and RTPs due to Connes. In his terminology a bimodule is called a correspondence. (The best references known to the author are [C2] and [P], but there was an earlier unpublished manuscript which is truly the source of Connes fusion.) Consider a correspondence MHN. Choosing a row representation R2(M)q for H, we know that the full commutant of L(M) is isomorphic to R(qM∞(M)q). This gives us a unital injective *-homomorphism ρ : N ֒→ qM∞(M)q, and from the map ρ one can reconstruct the original bimodule (up to isomorphism) as M(R2(M)q)ρ(N). 6 DAVIDSHERMAN What if we had chosen a different row representation R2(M)q′ and obtained ρ′ :N →q′M∞(M)q′? By Proposition 2.1, the module isomorphism (3.4) MR2(M)q ≃MR2(M)q′ is necessarily given by the right action of a partial isometry v between q and q′ in M∞(M). Thenρandρ′ differ byaninner perturbation: ρ(x)=v∗ρ′(x)v. We con- clude that the class of M−N correspondences, modulo isomorphism, is equivalent to the class of unital injective *-homomorphisms from N into an amplification of M, modulo inner perturbation. (These last are called sectors in subfactor theory.) Our conventionhere is to use the term “correspondence”to mean a representative *-homomorphismfor the bimodule. But the reader should be warned that the dis- tinction between bimodules, morphisms, and their appropriate equivalence classes is frequently blurred in the literature, sometimes misleadingly. Notice that a unital inclusion N ⊂M gives the bimodule ML2(M)N. The RTP of correspondences is extremely simple. Proposition3.3. ConsiderbimodulesMHN andNKP comingfromcorrespon- dencesρ1 :N ֒→qM∞(M)qandρ2 :P ֒→q′M∞(N)q′. ThebimoduleM(H⊗NK)P is the correspondence ρ ◦ρ , where we amplify ρ appropriately. 1 2 1 We pause to mention that it is also fruitful to realize correspondencesin terms of completely positive maps. We shall have nothing to say about this approach; the reader is referred to [P] for basics or [A2] for a recent investigation. 4. Applications to Morita equivalence and index An invertible *-functor from LeftL2Mod(N) to LeftL2Mod(M) is called a Morita equivalence [R]. Here a *-functor is a functor which commutes with the adjoint operation at the level of morphisms. One way to create *-functors is the following: to the bimodule MHN, we associate (4.1) FH :LeftL2Mod(N)→LeftL2Mod(M); NK7−→(MHN)⊗N (NK). The next theorem is fundamental. Theorem 4.1. When L(M)andR(N)arecommutantson H,theRTPfunctor FH is a Morita equivalence. Moreover, every Morita equivalence is equivalent to an RTP functor. This type of result - the second statement is an operator algebraic analogue of the Eilenberg-Watts theorem - goes back to several sources, primarily the funda- mental paper of Rieffel [R]. His investigation was more general and algebraic, and his bimodules werenotHilbertspacesbutriggedself-dualHilbertC*-modules,fol- lowingPaschke[Pa]. Froma correspondencepoint ofview, riggedself-dualHilbert C*-modules and Hilbert space bimodules give the same theory; the equivalence is discussed nicely in [A1]. (And the former is nothing but an L∞ version of the latter, as explained in [JS].) Our Hilbert space approach here is parallel to that of Sauvageot[Sa], andstreamlinedbyourstanding assumptionofseparablepreduals. We will need Definition 4.2. The contragredient of the bimodule MHN is the bimodule NH¯M, where H¯ is conjugate linearly isomorphic to H (the image of ξ is written ξ¯), and the actions are defined by nξ¯m=m∗ξn∗. RELATIVE TENSOR PRODUCTS 7 Lemma 4.3. Suppose L(M) and R(N) are commutants on H. Then NH¯M⊗MMHN ≃NL2(N)N. Proof. If MH ≃ MR2(M)q, then N ≃ qM∞(M)q by (2.3), and H¯M ≃ qC2(M)M. By Proposition 3.2 and the comment preceding Proposition 2.1, NH¯M⊗MMHN ≃N(qL2(M∞(M))q)N ≃NL2(qM∞(M)q)N ≃NL2(N)N. (cid:3) Lemma 4.3 was firstprovenby Sauvageot(in another way). In our situation it means FH¯ ◦FH(NK)≃L2(N)⊗N NK≃NK. (Here we have used the associativity of the RTP, which is most easily seen from the explicit construction in Section 5.) We conclude that FH¯ ◦FH is equivalent to the identity functor on LeftL2Mod(N), and by a symmetric argument FH ◦FH¯ is equivalent to the identity functor on LeftL2Mod(M). Thus FH is a Morita equivalence, and the first implication of Theorem 4.1 is proved. Now letF be aMorita equivalencefromLeftL2Mod(N) to LeftL2Mod(M). Then F must take R2(N) to a module isomorphic to R2(M), because each is in the unique isomorphism class which absorbs all other modules. (This is meant in the sense that NR2(N)⊕NH ≃ NR2(N); R2(N) is the “infinite element” in the monoidV(M∞(N)).) Beinganinvertible*-functor,F implementsa*-isomorphism of commutants - call it σ, not F, to ease the notation: ∼ (4.2) σ :M∞(N)→M∞(M). Apparently we have (4.3) F(R2(N)q)≃R2(M)σ(q). Before continuing the argument, we need an observation: isomorphic algebras have isomorphic standard forms. Specifically, we may write L2(M∞(N)) as the GNS construction for ϕ∈M∞(N)+∗ and obtain the isomorphism (σ−1)t :L2(M∞(N))→∼ L2(M∞(M)), (σ−1)t :xϕ1/2 7→σ(x)(ϕ◦σ−1)1/2. Note that (σ−1)t(xξy)=σ(x)[(σ−1)t(ξ)]σ(y). Now consider the RTP functor for the bimodule MHN =σ−1(M)σ−1(eM11)C2(N)N. Its action is R2(N)q 7→σ−1(M)σ−1(eM11)L2(M∞(N))q (σ≃−1)t MeM11L2(M∞(M))σ(q) ≃MR2(M)σ(q)≃F(R2(N)q). We conclude that F is equivalent to FH, which finishes the proof of Theorem 4.1. Notice that (4.2) and (4.3) can also be used to define a functor; this gives us Corollary 4.4. For two von Neumann algebras M and N, the following are equivalent: 8 DAVIDSHERMAN (1) M and N are Morita equivalent; (2) M∞(N)≃M∞(M); (3) thereis abimoduleMHN wheretheactions arecommutantsof eachother; (4) there is a projection q ∈M∞(M) with central support 1 such that qM∞(M)q ≃N. (The central support of x ∈ M is the least projection z in the center of M satisfying x=zx.) Examplecontinued. M andM areMoritaequivalent. Thiscanbededuced 3 5 easily from condition (2) or (4) of the corollary above. It also follows from the (Hilbert)equivalencebimoduleM3HS(M3×5)M5,where“HS”indicatestheHilbert- Schmidtnorm;thisbimodulegivesusanRTPfunctorwhichisaMoritaequivalence. Regardless of the construction, the equivalence will send (functorially) n copies of C5 to n copies of C3. Apparently Morita equivalence is a coarse relation on type I algebras - it only classifies the center of the algebra (up to isomorphism). At the other extreme, Morita equivalence for type III algebras is the same as algebraic isomorphism; Morita equivalence for type II algebras is somewhere in the middle ([R], Sec. 8). For a bimodule MHN where the algebras are not necessarily commutants, the functor (4.1) still makes sense. To get a more tractable object, we may consider the domain and range to be isomorphism classes of modules: (4.4) πH :V(M∞(N))→V(M∞(M)); FH(R2(N)q)=MHN ⊗N R2(N)q ≃R2(M)πH([q]). We call this the bimodule morphism associated to H, a sort of “skeleton” for thecorrespondence. ItfollowsfromProposition3.3thatifthebimoduleisρ:N ֒→ qM∞(M)q,thenπH isnothingbutρ∞,theamplificationofρtoM∞(N),restricted to equivalence classes of projections. This has an important application to inclusions of algebras. We have seen in ρ Section 3 that a unital inclusion N ⊂ M is equivalent to a bimodule ML2(M)N. Whenthecorrespondenceρissurjective,themodulegeneratesaMoritaequivalence via its RTP functor, and the induced bimodule morphism is an isomorphism of monoids. When N 6=M, itis naturalto expect thatthe bimodule morphismgives usa wayto measurethe relativesize,orindex,ofN in M. (For readersunfamiliar with this concept, the index of an inclusion N ⊂ M is denoted [M : N] and is analogousto the index of a subgroup. The kernel of this idea goes back to Murray andvonNeumann, but the startlingresultsofJones[J]intheearly1980’stouched offanewwaveofinvestigation. Werecommendtheexposition[K]asanicestarting point.) For algebras of type I or II, the index can be calculated in terms of bimodule morphisms. (Therearealsobroaderdefinitionsofindexwhichrequireaconditional expectation(=norm-decreasingprojection)fromMontoN.) Thisamountslargely torephrasingandextensionofthepaper[Jol],andwedonotgivedetailshere. Very briefly, let π :V(M∞(M))→V(M∞(M)) be the bimodule morphism for (4.5) (ML2(M)N)⊗N (NL2(M)M). RELATIVE TENSOR PRODUCTS 9 When M is a factor, π is a monoidal morphism on the extended nonnegative integers (type I) or extended nonnegative reals (type II). It must be multiplication byascalar,andthisscalaristheindex. IfMisnotafactor,theindexisthespectral radius of π, provided that V(M∞(M)) is endowed with some extra structure (at present it is not even a vector space). Example. Consider the correspondence L2(M ) ; M ≃M ⊗I ⊂M ⊗M ≃M . M6 6 M3 3 3 3 2 6 The image of L2(M ) under the RTP functor for (4.5) is M6 6 L2(M ) ⊗ L2(M ) ⊗ L2(M ) M6 6 M3 M3 M3 6 M6 M6 M6 6 (cid:0) (cid:1) (cid:0) (cid:1) (cid:0) (cid:1) ≃ L2(M ) ⊗ L2(M ) M6 6 M3 M3 M3 6 (now counting the dimensions of the Hilbert spaces) ≃M6HS(M12×3)M3 ⊗M3 M3HS(M3×12)≃M6HS(M12×12)≃M6HS(M6×24). We have gone from 6 copies of C6 to 24 copies; that is, π :(Z ∪∞)→(Z ∪∞), 67→24. + + Apparentlythe index is 4,whichis alsothe ratioofthe dimensions ofthe algebras. 5. Analytic approaches to RTPs As mentioned in Section 3, we cannot expect the expression ξ⊗Mη to make sense for every pair of vectors ξ,η. In essence, the problem is that the product of two L2 vectors is L1, and an L1 space does not lie inside its corresponding L2 space unless the underlying measure is atomic. Densities add, even in the noncommutativesetting,andsotheproductin(3.3)“should”be anL1 matrix. To make this work at the vector level, we need to decreasethe density by 1/2 without affectingthe“outside”actionofthecommutants... andthesolutionbyConnesand Sauvageot ([Sa], [C2]) is almost obvious: choose a faithful ϕ∈M+ and put ϕ−1/2 ∗ in the middle of the product. That is, (5.1) ξ⊗ η =(ξϕ−1/2)η. ϕ This requires some explanation. Fix faithful ϕ ∈ M+ and row and column representations of H and K as in ∗ (2.1). We define x1ϕ1/2 D(H,ϕ)=x2ϕ..1/2∈H : Xx∗nxn exists in M. .     D(H,ϕ) is dense in H, and its elements are called ϕ-left bounded vectors [C1]. 10 DAVIDSHERMAN Now by (5.1) we mean the following: for ξ ∈ D(H,ϕ), we simply erase the symbol ϕ1/2 from the right of each entry, then carry out the multiplication. The natural domain is D(H,ϕ)×K. Visually, x1ϕ1/2 x1ϕ1/2 (5.2) x2ϕ1/2,(η1 η2 ···)7→x2ϕ1/2(ϕ−1/2)(η1 η2 ···)=(xiηj). . . . . . .      For ϕ 6= ψ ∈ M+, we cannot expect ξ ⊗ η = ξ ⊗ η even if both are defined, ∗ ϕ ψ although the reader familiar with modular theory will see that (5.3) ξ⊗ η =(ξϕ−1/2)η =(ξϕ−1/2ψ1/2ψ−1/2)η =ξ(Dϕ:Dψ) ⊗ η. ϕ i/2 ψ (An interpretation of the symbols ϕ1/2,ϕ−1/2 as unbounded operators will be dis- cussed in the next section.) Now we define H⊗ K to be the closed linear span of the vectors ξ⊗ η inside ϕ ϕ L2(M∞(M)). Uptoisomorphism,thisisindependentofϕ. (Weknowthisbecause of functoriality; the “change of weight” isomorphism is densely defined by (5.3).) The given definition for D(H,ϕ) ⊂ H makes it seem dependent on the choice of column representation. That this is not so can be seen by noting (as in (3.4)) that the intertwining isomorphism is given by L(v) for some partial isometry v ∈ M∞(M). Butlet us alsomentiona methodofdefining the sameRTP construction without representingH andK. First notice that D(H,ϕ) canalso be defined as the set of vectors ξ for which πℓϕ(ξ):L2(M)M →HM, ϕ1/2x7→ξx, isbounded. (Amoresuggestive(andrigorous)notationwouldbeL(ξϕ−1/2).) Now we consider an inner product on the algebraic tensor product D(H,ϕ)⊗K, defined on simple tensors by (5.4) <ξ ⊗ η |ξ ⊗ η >=<πϕ(ξ )∗πϕ(ξ )η |η >. 1 ϕ 1 2 ϕ 2 ℓ 2 ℓ 1 1 2 The important point here is that πℓϕ(ξ2)∗πℓϕ(ξ1) ∈L(L2(M)M)=M. The closure of D(H,ϕ)⊗K in this inner product, modulo the null space, is once againH⊗ K. ϕ (If we do choose a row representation as in (5.2), we have x1ϕ1/2 x1 πℓϕx2ϕ.1/2=L(cid:18)(cid:18)x..2(cid:19)(cid:19).) . . .   The paper [F] containsmore expositionof this approach,including some alter- nate constructions. 6. Realization of the relative tensor product as composition of unbounded operators In this section we briefly indicate how (5.1) can be rigorously justified and extended. Readers are referred to the sources for all details. In his pioneering theory of noncommutative Lp spaces, Haagerup [H] estab- lished a linear isomorphism between M+ and a class of positive unbounded oper- ∗ ators affiliated with the core of M. (The core, well-defined up to isomorphism, is