Contents lists available atScienceDirect
Journal of Mathematical Analysis and Applications
www.elsevier.com/locate/jmaa
Linear combination of projections in von Neumann factors
Stanisław Goldstein∗, Adam Paszkiewicz∗
FacultyofMathematicsandComputerScience,UniversityofŁódź,Banacha22,90-238Łódź,Poland
a r t i cl e i n f o a b s t r a c t
Articlehistory:
Received31January2019 Availableonline9April2020 SubmittedbyD.Blecher
Keywords:
Projection Factor
Linearcombination
Itisshownthatanyself-adjointoperatorinafinitediscreteorinfinitevonNeumann factor canbewritten as areallinear combinationof 4projections. Ontheother hand,inanytypeII1algebraandinanytypeII∞factorthereexistsaself-adjoint operatorthatisnotalinearcombinationof3projections.
©2020TheAuthors.PublishedbyElsevierInc.Thisisanopenaccessarticle undertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).
1. Introduction
Weproveherefourresults,twopositive andtwonegative:
(1) Any self-adjoint operatorA acting ona finite-dimensional(complex) Hilbertspace canbe writtenas alinearcombination of 4 projections, with two of thecoefficients chosen arbitrarily from theinterval [2A,∞[.
(2) Anyself-adjoint operatorA in aninfinitevon Neumannfactorcanbe writtenas alinearcombination of4 projections,withtwo ofthecoefficientschosenarbitrarilyfromtheinterval]2A,∞[.
(3) Thereisaself-adjointoperatorinanytypeII1 algebrathatcannotbewrittenasalinearcombination of3 projections.
(4) Thereisaself-adjointoperatorinanytypeII∞ factorthatcannotbewrittenas alinearcombination of3 projections.
Anyself-adjointoperatoractingina(complex)finite-dimensionalHilbertspacecanbewrittenasalinear combinationofafinitenumberofprojections—this followseasilyfromthespectraltheorem.Whatisnot immediately obvious is whether we canmake the number of projectionsindependent of the dimension of theHilbertspace. IsthisalsopossibleiftheHilbertspaceisinfinitedimensional?Thefirstpositive results
* Correspondingauthors.
E-mailaddresses:stanislaw.goldstein@wmii.uni.lodz.pl(S. Goldstein),adam.paszkiewicz@wmii.uni.lodz.pl(A. Paszkiewicz).
https://doi.org/10.1016/j.jmaa.2020.124135
0022-247X/©2020TheAuthors.PublishedbyElsevierInc. ThisisanopenaccessarticleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).
in this directionwere obtainedby Fillmore1967 (see [3]), who wasable to get downto 9 projections [4].
Pearcy and Topping[16] reducedit to 8 alreadyin1967, thenthe second-namedauthor to 6 in1980 (see [15]).Using theideasfrom[15], Matsumotoshowedin1984that5 isenough.Alsoin1984Nakamura [12]
provedthatanyHermitian matrixisalinearcombinationof4 projections.
In 1985 the second-named author proved that any self-adjoint operator in B(H) can be written as a linear combination of4 projections, both forfinite and infinite-dimensional HilbertspaceH, with aproof for HermitianmatricesdifferingfromthatofNakamura[12].Thepaper [14] withtheresultswassenttoa journal and gotapositive review, requiringonly minorimprovements, butforsomeinexplicable reasonit hasnotbeenresentto thejournal.Thisfactwentunnoticedbythefirst-named authorwhowascompiling thejointpublicationonlinearcombinationsofprojectionsinvonNeumannalgebras[5] whileonleavefrom his alma mater, and thepaper wasadvertised as acceptedfor publication.Evenworse,some resultsfrom thejointpublication dependontheunpublishedpaper, includingthecriterionforavon Neumannalgebra to beacomplexlinearspanofitsprojections.
In 2016,the first-namedauthor met ViacheslavRabanovichat theconference “Groupsand Operators”
in Gothenburg, Sweden. There he learned that [14] has never been published, and that Rabanovich [17]
himself proved that4 projections are sufficient in the caseof B(H) with aseparable infinite-dimensional HilbertspaceH.
We decided thatthere are good reasonsfor complementingour earlier paper[5] with this one,dealing with linearcombinations of projectionsin vonNeumann factors. BothNakamura’s [12] andRabanovich’s [17] proofsareelegant, andtheyshowclearlythat4projectionsareenoughforself-adjointoperatorsinH both infinite- andinfinite-dimensionalcases.However,thereareseveral reasonsforourpresentation:
(1) Theorem1.1(a)of[5] attributedto[14],andhencedidnotprove,adecompositionofself-adjointoper- atorsinatypeIn factoras linearcombinationsof4projections,withaspecificformofthecoefficients.
Thatform isnotavailablefrom Nakamura’s[12] but isrequiredtoobtaininTheorem2.1(a)of[5] the decompositionofself-adjointoperatorsinafinitediscretevonNeumannalgebraasalinearcombination of4projectionswith coefficientsinthecenterof thealgebra.Aproof basedontheunpublishedpaper [14] is presentedinTheorem 2.1below.
(2) Theorem 1.1 (a) of [5] attributed to [14] a decomposition of self-adjoint operators inan infinite von Neumannfactorasalinearcombinationof4projections.TheproofpresentedbyRabanovich[17] holds for aseparable Hilbertspace and cannot be easily adaptedto thecase of anarbitrary infinitefactor.
Anindependentproofbasedon[14] ispresentedinTheorem3.5.
(3) Given Theorem2.1andTheorem 3.5,alltheresultsfrom [5] are nowproved.Theresultshavealready been usedbyseveralauthorsthroughout theyears.
Theinterestinthisfieldofresearchcontinues(see,forexample,[7]),withmoreattentiondirectedrecently to asituationinC∗-algebras (see[10]).Nevertheless,asseenbelow,there aresomeopenquestionseven in thevon Neumannalgebracase(see‘Openproblems’attheend ofthepaper).
ForavonNeumannalgebraA wedenotebyAhitsself-adjointpartandbyProjA thelatticeofprojections in A. For a projection E, we denote by AE the reduced von Neumann algebra EAE acting in EH. For A∈ Ah,eA(·) denotesthespectralmeasureofA.TheunitofA isdenotedby1Aorsimply1,ifthereisno danger of confusion.For projectionsP,Q∈ A, we write P ∼ Q ifthere is apartial isometry U ∈ A such thatP = U∗U andQ= U U∗,and thenP Q ifP ∼ R ≤ Q foraprojectionR∈ A, andP ≺ Q ifP Q and P Q.
2. Finitediscretefactors
LetA befactorof typeIn,n<∞,andτ the tracialstateonA.Weassume(forconvenience)thatA is representedonaHilbertspaceH insuchawaythatA= B(H).
Theorem2.1. [1.1(a)in[5]] LetA∈ Ahandα = τ (A).Foranyβ,γ≥ 2A,thereareprojectionsP,Q,R,S inA withP ∼ Q andR∼ S suchthat
A = (β + α)P − βQ + (γ + α)R − γS (1)
whenα≥ 0,and
A = βP − (β − α)Q + γR − (γ − α)S (2)
whenα≤ 0.
Proof. (2) followsfrom (1) appliedto−A.Weshow (1).IfA= α1, takeP = Q= 1,R = S = 0. Consider n ≥ 2 and A = α1.Assume thatA = 1, otherwise use (1/A)A. Put A = A˜ − α1. Then A˜ = 0. Let {e1,. . . ,en} bean orthonormal basisin H consistingof eigenvectors of A (andA),˜ and let α1,. . . ,αn be thecorrespondingeigenvaluesofA.˜ Weorder theeigenvalues(andthecorresponding eigenvectors)insuch awaythatforeachj,j = 1,. . . ,n, βj := α1+· · · + αj∈ [0,2].ByLemma1in[12] (orLemma3.4below), forany β,γ≥ 2,there existrankoneprojectionsPj,Qj inA withPj,Qj≤ Ej+ Ej+1,when Ej isarank 1 projectionontothesubspacegeneratedbyej,suchthat
βj(Ej− Ej+1) + αEj =
(β + α)Pj− βQj for j odd, (γ + α)Pj− γQj for j even.
Wehave(notingthatβn= 0 andαn =−βn−1)thatfor1≤ j ≤ n− 1
A = [β1(E1− E2) + αE1] +· · · + [βn−1(En−1− En) + αEn−1] + αEn
= (β + α)(
j odd
Pj)− β(
j odd
Qj) + (γ + α)(
j even
Pj)− γ(
j even
Qj) + αEn.
Ifn isodd,putP =
j oddPj+ En,Q=
j oddQj+ En,R =
j evenPj,S =
j evenQj;ifn iseven, addEn to thesumsovereven indices.(Forn= 2,putR = S = E2.)
3. Infinitefactors
Thefollowinglemmabelongsto mathematicalfolklore.Wegivehereitsproofforcompleteness.
Lemma 3.1. Let A be a von Neumann factor, and let P,Q,R ∈ Proj A with P + Q = R, and R infinite.
ThenR P orR Q.
Proof. Let E,F ∈ Proj A besuch thatE + F = R,E ∼ F ∼ R (see Lemma6.3.3in[8]).Then P∧ E Q∧ F or P ∧ E Q∧ F . AssumeP ∧ E Q∧ F ; then, using Kaplansky’s Parallelogram Law, we get R∼ E = P ∧ E + (E − P ∧ E) Q∧ F + (E ∨ P − P )≤ Q,soR Q.
TheproofinthecasewhenP∧ E Q∧ F isobtainedbyexchangingtherolesofP andQ intheproof above.
Thenextfourlemmas areessential.
Lemma 3.2. Let A bean infinite factor,and let A∈ Ah.There exists a∈ [−A,A] such that eA
]a− ε,a+ ε[
∼ 1 foreach ε> 0.
Proof. Assume there is no such a. Then, for each x ∈ [−A,A], there is εx > 0 such that eA
]x− εx,x+ εx[
≺ 1.Fromcompactnessof[−A,A],thereisafinitenumberofintervals]xn− εxn,xn+ εxn[ coveringtheinterval[−A,A].Hence,thereisafinitenumberofdisjointintervalsIn,eachcontainedin ]xn−εxn,xn+εxn[,withunion[−A,A],sothat1=
pnforpn= eA(In)≤ eA
]xn−εxn,xn+εxn[
≺ 1, whichcontradicts Lemma3.1.
Notethata satisfiestheconditionsoftheLemmaifandonlyifitbelongstotheessentialspectrumofA with respecttothelargest closedidealofA (theonegeneratedbyalltheprojectionsnotequivalent to1).
Lemma 3.3. LetA bean infinite factor and letA∈ Ah.Assumethat for each δ > 0,eA(]− δ,δ[)∼ 1.For any ε> 0, there are:a sequence(εi),i∈ Z of positivenumbers satisfying ε0 =A and
i=0εi ≤ ε, and a sequence(Ei),i∈ Z ofprojections fromA such that Ei ∼ 1 andEiA= AEi foralli,
i∈ZEi = 1 and
−εiEi≤ AEi≤ εiEi fori∈ Z.
Proof. Considerthefollowing cases:
I eA({0}) ∼ 1. Choose Ei,i ∈ Z\{0},E0 so that E0 ∼ Ei ∼ 1 and
i=0Ei + E0 = eA({0}). Put E0= E0 + eA(R\{0}),ε0=A andεi= 0 for i∈ Z\{0}.
II eA {0}
1. First observe that we can replace the set of indices Z by any other countable set.
Define, for a sequence (εn),n = 0,1,. . . with ε0 = A,ε1 = ε/2 and εn+1 ≤ εn/2 for n ≥ 2, sets In = [−εn,−εn+1[∪]εn+1,εn] andDn = eA(In). Since∞
n=0Dn ∼ 1 (byLemma3.1),we canchooseεn so thatDn = 0 foreachn,byassumptionandagainLemma3.1.There areagaintwocases:
(a) Forinfinitelymanyn,sayn∈ M ⊆ N,Dn∼ 1.ThenwecanusethemforformingEi,withE0obtained as 1−
i=0Ei.
(b) IfDn ∼ 1 forfinitelymany(or none)ofindices n,wecanassumethatDn 1 for eachn = 0.
Now,if1 isσ-finite,wearenecessarilyinasemifinitefactor(allnon-zeroprojectionsinaσ-finitetype III factor are equivalent to 1), and we can use a faithful normal semifinite trace τ on A. We have
∞
n=1τ (Dn) = τ eA
]− ε1,ε1[\{0}
= τ (1) = +∞, by assumption and Lemma 3.1. Now we can find infinitely many disjoint infinite sets Mi ⊆ N such that
n∈Miτ (Dn) = +∞. Thus we cantake Ei=
n∈MiDn,i = 0,with(asbefore)E0= 1−
i=0Ei.
If1 isnotσ-finite,wemayassumethatallprojectionsDn(n = 0) areinfinite(and≺ 1).Foraprojection F ∈ A,denote by#F thelargest cardinalnumberofamaximalorthogonalfamilyofnon-zeroσ-finite projectionsmajorizedbyF (thiscorresponds directlytothegeneralizeddimensionfunctionofTomiyama inthefactorialcase(see[19]).ByLemma3.1,eA
]− ε1,ε1[\{0}
∼ 1.SinceA isafactor,wehave,by Tomiyama[19,Theorem 4anditscorollary] (see alsoBlackadar[1,III.1.7.1])
∞ n=1
#Dn= #
∞ n=1
Dn= #eA
]− ε1, ε1[\{0}
= #1.
This means thatthere isno cardinalκ< λ= #1 such that#Dn ≤ κ for n= 1,2,. . . . Hence,λ is a limitcardinalof countablecofinality(see, forexample,[11]),andwecanform asubsequence(Dmn) of
(Dn) suchthat#Dmn λ.Weusethesubsequenceto build countablymanysets Mi ⊆ N suchthat
n∈Mi#Dmn= #1 foreachκ.Wefinishtheproof asintheσ-finite case.
Lemma3.4.LetA beavonNeumannalgebra,E,F ∈ Proj A,E⊥ F ,E = U∗U ,F = U U∗forsomeU ∈ A (so thatE ∼ F ).Forany α,β ≥ 0 and D∈ A+ satisfying D≤ βE, thereexistP,Q∈ Proj A with P ∼ Q suchthat
(α + β)P− βQ = αE + D − UDU∗ andthat P,Q≤ E + F .
Proof. If β = 0,thenD = 0 andonecantakeP = Q= E.Hence,we canassumethatβ > 0.Weuse the standardmaterialontherelativeposition oftwoprojectionsinavon Neumannalgebra(see Halmos[6] or Takesaki[18,pp.306–308]).WeconstructP,Q∈ Proj A insuchawaythatP,Q≤ E + F andP ∼ E ∼ Q (inAE+F).Morespecifically,we representtheoperators as2× 2 matricesoverAE,sothat
A11 A12 A21 A22
= A11+ U A22U∗+ A12U∗+ U A21
forAij ∈ AE (notethat1AE = E).Then
E =
1 0 0 0
, P =
C2 CS CS S2
and Q =
C 2 CS CS S 2
forsuitablychosenC,S,C,S with0≤ C,S≤ 1,0≤ C,S ≤ 1 andC2+ S2= 1,C 2+ S 2 = 1.
Define firstH ∈ AE byH := (2D + α1)−1(D + α1) for α > 0 and H := (1/2)1E forα = 0. Checkthat 0≤ (α + β)−1[D + (α + β)1]H ≤ 1 and 0≤ β−1(β1− D)H ≤ 1.Let C andC be thesquare rootsof the two operators,respectively, and letS = (1− C2)12,S = (1− C 2)12. Note thatC,S,C andS belong to theabelianvonNeumannalgebrageneratedbyD andE.Wecheckthat
(α + β)P− βQ =
α1 + D 0
0 −D
,
whichendstheproof.
Theorem 3.5.[1.3(a) in [5]] Let A be aninfinite factor, andlet A∈ Ah.For any β,γ > 2A,there exist α∈ R with|α|≤ A and projectionsP,Q,R,S∈ A suchthat, depending onα,(1) or(2) ofTheorem2.1 holds.
Proof. ThecaseA= 0 istrivial.WemayassumethatA= 1,otherwiseuse(1/A)A insteadofA. Let α besuch thateA
]α− ε,α + ε[
∼ 1 foranyε> 0.Itisenoughto consider thecaseα≥ 0:ifα < 0, use
−A insteadofA. Put A = A˜ − α1. Takeε> 0 satisfying2(1+ ε)< β,γ.Let(Ei)i∈Z and(εi)i∈Z beas in Lemma3.3,forA and˜ ε.
Let Uj ∈ A be any partial isometry satisfyingUj∗Uj = Ej,UjUj∗ = Ej+1. Put U =
j∈ZUj andlet h denotethemappingX → UXU∗ onA.Note thatU isunitaryandh−1 exists andmapsX toU∗XU .Let Aj = EjAE˜ j and B0 = (1− α + ε)E0. Use therecurrencerelation Aj = Bj − hBj−1 to define Bj for all j = 0.Then
(−1 − α)E0≤ A0≤ (1 − α)E0;
forj ∈ Z,j = 0,
−εjEj≤ Aj ≤ εjEj; finally,
0≤ Bj≤ 2(1 + ε)Ej. ApplyLemma3.4toobtain
αEj+ Bj− hBj =
(β + α)Pj− βQj for j odd, (γ + α)Rj− γSj for j even.
Then put
P =
j odd
Pj, Q =
j odd
Qj, R =
j even
Rj, S =
j even
Sj.
Note that
A = ˜A + α1 =
j∈Z
(Aj+ αEj) =
j∈Z
(Bj− hBj+ αEj)
to endtheproof. 4. TypeII1algebras
LetA beaσ-finitetypeII1algebraandletτ0beafixedfaithfulnormaltracialstateonA.Inthissection τ denotesafixed tracethatisascalarmultipleofτ0.
WewillbeconstructinganoperatorA∈ A+ withthepropertythatwhateverα,β ∈ R+ andγ∈ R,and whatever projectionsP,Q,R∈ A,
A + γR = αP + βQ. (3)
Weshallfirstdescribethepropertiesoftheright-handsideof(3),introducingconvenientterminology.
Definition4.1.Adistributionfunction isanon-decreasing,left-continuousfunctionF : R→ [0,∞[ suchthat limt→−∞F (t) = 0.Foradistribution functionF we denote by F+ itsright-continuousvariant: F+(t):=
limstF (s),andbyF(δ)andF(δ)+,foranyδ∈ R,thefunctionsF(δ)(t):= F (t)+F+(δ−t) andF(δ)+(t):=
F+(t)+ F (δ− t),respectively.
Forany A∈ Ah,thefunctionFA givenbyFA(t):= τ (eA(]− ∞,t[) isadistributionfunction,calledthe distributionfunctionofA w.r.t.thetraceτ .Notethatlimt→+∞FA(t)= τ (1) andFA+(t)= τ (eA(]− ∞,t])).
Definition 4.2.A distribution functionF iscalled (α,β)-symmetric (with 0≤ α ≤ β)if it is constanton eachoftheintervals]− ∞,0],]α,β],]α + β,∞[ and F(α+β)isconstanton]0,α]∪]β,α + β].
F issymmetric ifitis(α,β)-symmetric forsome0≤ α ≤ β.
It is clearthatthenotionsof symmetryjust defineddonotdepend onthechoiceof thetraceτ (being alwaysa scalarmultipleofτ0).Similarly,ifthedistributionfunctionofanoperatorissymmetric,thenthe distributionfunctionofapositivemultipleofthisoperatorissymmetricaswell.
Remark 4.3.If the distribution function FA (w.r.t. τ ) of an operator A ∈ Ah is (α,β)-symmetric, then A∈ A+ and
FA(0) = 0 (4)
FA+(α + β) = τ (1) (5)
FA+(α) = FA(β), (6)
and
FA(α+β)takes on at most three different values. (7) Infact,positivityofA and(4) followfromthedistributionFAbeingconstanton]− ∞,0[.Then(5) and (6) followfrom FA being constanton]α + β,∞[ and]α,β],respectively.Thedefinitionof(α,β)-symmetry alsoimpliesthatF(α+β) isconstantandequaltoτ (1) on]− ∞,0]∪]α + β,∞[,andthatF(α+β)isconstant (andequalto2FA(β))on]α,β],whichyields (7).
We will now investigateproperties of the distribution function of alinear combination A = αP + βQ of two projections P,Q ∈ A with α,β ≥ 0. Thestructure of such an operator A was described (for real coefficients) byNishio [13]. The symmetry of thespectrum of A follows immediately from Corollary 3of [13],asobservedbyRabanovich,whoaddsafewmorefactsonthesymmetrypropertiesofA insection2of [17].Nevertheless,wehavenotspottedany proof(orevenstatement)ofthemoregeneralfactgivenbelow andconcerningsymmetryof thespectral measureofA, aresultcrucialforthesequel.
Lemma 4.4.Let A be anarbitrary vonNeumann algebra and letP,Q∈ A be twoprojections in ageneric position (i.e. P ∧ Q = P ∧ Q⊥ = P⊥ ∧ Q = P⊥ ∧ Q⊥ = 0). Then, for any α,β ≥ 0 and any Borel Z ⊆ [0,α + β],
eαP +βQ(Z)∼ eαP +βQ(α + β− Z).
Proof. Weuse2× 2 matrixrepresentationofA over AP (cf. referencesintheproofofLemma3.4).Then
P =
1 0 0 0
and Q =
C2 CS CS S2
forsome0≤ C,S≤ 1,C2+ S2= 1.LetV =
0 1
−1 0
. Then
V∗(αP + βQ)V = αP⊥+ βQ⊥= (α + β)1− (αP + βQ), hencethespectralprojections
eαP +βQ(Z) and eαP +βQ(α + β− Z) = e(α+β)1−(αP +βQ)(Z) areunitarilyequivalent.
Corollary 4.5. Let A be an arbitrary von Neumann algebra, and let P,Q∈ Proj A,α,β ∈ R,0 ≤ α ≤ β.
Then
eαP +βQ(]0, α[)∼ eαP +βQ(]β, α + β[) and eαP +βQ(]α, β[) = 0. (8)
Moreover, if 0< α≤ β,then
eαP +βQ({0}) = P⊥∧ Q⊥ and eαP +βQ({α + β}) = P ∧ Q.
If, additionally,α < β,then
eαP +βQ({α}) = P ∧ Q⊥ and eαP +βQ({β}) = P⊥∧ Q, while ifα = β,
eαP +βQ({α}) = P ∧ Q⊥+ P⊥∧ Q.
(Trivially, for 0 = α < β we have eαP +βQ({0}) = Q⊥ and eαP +βQ({β}) = Q, and for α = β = 0, eαP +βQ({0})= 1.)
Proof. Let A= αP + βQ.Put(cf. Takesaki[18, pp.306–308)])
P0= P − P ∧ Q − P ∧ Q⊥, Q0= Q− P ∧ Q − P⊥∧ Q, (9) and
A0= αP0+ βQ0. (10)
IfP Q= QP ,thenP0= Q0= 0.IfP andQ donotcommute,thenP0and Q0 arebothnon-zeroand in agenericpositioninAP0∨Q0.Note that
1 = P0∨ Q0+ P∧ Q + P ∧ Q⊥+ P⊥∧ Q + P⊥∧ Q⊥.
As observed byRabanovichinthefirst paragraphof section 2of[17], itfollows from Halmos [6] that the only partsofP andQ that contributeto thepossiblynon-zerovaluesofthespectral measureeA(·) atthe one-point sets {0},{α},{β} and{α + β} are the ‘commuting’ ones, that is P ∧ Q,P ∧ Q⊥,P⊥∧ Q and P⊥∧Q⊥.Bytheabove,eA(]0,α[)= eA0(]0,α[), eA(]α,β[)= eA0(]α,β[) andeA(]β,α+β[)= eA0(]β,α+β[).
Note thateA0(]α,β[)= 0.This follows,for example,fromtheformula inCorollary 3inNishio [13].Hence also eA(]α,β[)= 0.Therest followsfromLemma4.4.
WereturntoourassumptionthatA isaσ-finitealgebraoftypeII1,andτ isascalarmultipleofafinite tracial stateonthealgebra.The followingresult iswell known(see,for example,theproof or Lemma2.5 (iii) in[2]),butwe givehereasimpleproofforcompleteness.
Lemma 4.6.For0≤ B ≤ A and λ∈ R wehave FA(λ)≤ FB(λ).
Proof. Assume that FA(λ) > FB(λ) for someλ. This means that τ
eB([λ,∞[)
> τ
eA([λ,∞[
, so that there isanon-zeroξ∈
eB([λ,∞[)∧ eA(]− ∞,λ[)
(H),forwhichAξ,ξ< λξ2 andAξ,ξ≥ Bξ,ξ≥ λξ2,acontradiction.
Lemma 4.7. For any P,Q ∈ Proj A in generic position and any 0 ≤ α ≤ β, the distribution function F := FαP +βQ (w.r.t. τ )satisfies
F (0) = 0, (11) F (t) = 1
2τ (1) for t∈]α, β], (12)
F+(α + β) = τ (1), (13)
F(α+β)(t) = τ (1) for t∈ R. (14)
In particular,F is(α,β)-symmetric.
Proof. Since P and Q are in a generic position, also P and Q⊥ are in a generic position, which implies Q ∼ P ∼ Q⊥, so thatτ (Q) = (1/2)τ (1).By Lemma4.6, wehave FαP +βQ(0) ≤ F0(0) = 0,which yields (11),andFαP +βQ(β)≤ FβQ(β)= τ (1)− τ(Q)= (1/2)τ (1).ApplyingLemma4.4 withZ =]− ∞,t[ gives
F (t) = τ (eαP +βQ(]− ∞, t[))
= τ (eαP +βQ(]α + β− t, ∞[)) = τ(1) − F+(α + β− t) (15) forall t∈ R.Hence(14) and,by(11), F+(α + β)= τ (1)− F (0)= τ (1),which shows(13). For(12) note thatby(15) wehaveF+(α)= τ (1)− F (β)≥ τ(1)− (1/2)τ(1)= (1/2)τ (1).
Lemma4.8. Forany commuting P,Q∈ Proj A and any 0≤ α ≤ β,thedistributionfunctionF := FαP +βQ
isconstant oneach of theintervals]− ∞,0],]0,α], ]α,β],]β,α + β] and]α + β,∞[. Moreover,
F(α+β)(t) = 2τ (1)− τ(P ) − τ(Q) for t ∈]0, α]∪]β, α + β]. (16) In particular,F is(α,β)-symmetric.
Proof. Notethat for commutingP and Q, αP + βQ= αP ∧ Q⊥+ βP⊥∧ Q+ (α + β)P ∧ Q isa linear combination ofthreemutually orthogonalprojections,whichshowsthatthedistributionfunctionFαP +βQ
hastobe constantoneachoftheintervals]− ∞,0],]0,α],]α,β],]β,α + β] and]α + β,∞[.Inparticular,
τ (P⊥∧ Q⊥) =
F (t) for t∈]0, α], F+(α + β− t) for t ∈]β, α + β]
and
τ ((P∧ Q)⊥) =
F (t) for t∈]β, α + β], F+(α + β− t) for t ∈]0, α].
Asimplecalculationgivesnow(16).
Theworkdoneso farcanbe summedupinthefollowing:
Theorem 4.9. Let P,Q be projections from A, and let 0 ≤ α ≤ β. Then the distribution function F of A= αP + βQ (w.r.t. τ ) is(α,β)-symmetric,and
F(α+β)(t) = 2τ (1)− τ(P ) − τ(Q) for t ∈]0, α]∪]β, α + β]. (17)
Proof. Let usfirstextractcommutingandgenericparts ofP andQ.Weput Pc:= P∧ Q + P ∧ Q⊥, Qc := P∧ Q + P⊥∧ Q;
P0:= P− Pc, Q0:= Q− Qc.
We consider P0 and Q0 in the reduced von Neumann algebra AP0∨Q0. Similarly, we treat Pc and Qc as elements ofA(P0∨Q0)⊥.Put F0:= FαP0+βQ0 andFc := FαPc+βQc.ByLemmas4.7and4.8,bothF0andFc are (α,β)-symmetric,andtheysatisfy,fort∈]0,α]∪]β,α + β],
F0(α+β)(t) = τ (P0∨ Q0) = 2τ (P0∨ Q0)− τ(P0)− τ(Q0), Fc(α+β)(t) = 2τ ((P0∨ Q0)⊥)− τ(Pc)− τ(Qc).
HenceF = F0+ Fc isalso(α,β)-symmetricandF satisfies(17).
Thenextfewlemmas showtheexistenceofapositive operatorA suchthatthedistributionfunctionof A+ γR isnotsymmetric,whateverthechoiceofγ∈ R andprojectionR∈ A.
Lemma 4.10.Choose arbitrary λ > 0 and 0 ≤ Λ < τ (1). If A = B + C,B,C ∈ A+,B < λ and τ (supp C) ≤ Λ, then FA(λ) ≥ τ(1)− Λ. Consequently, if A = B1 +· · · + Bm + C1+· · · + Cn with Bi,Cj ∈ A+,B1+· · · + Bm< λ andτ (supp C1)+· · · + τ(supp Cn)≤ Λ,thenFA(λ)≥ τ(1)− Λ.
Proof. Suppose that FA(λ) < τ (1)− Λ. Then τ (eA(] − ∞,λ[)∨ supp C) < (τ (1)− Λ)+ Λ = τ (1), so that eA([λ,∞[) ∧ (supp C)⊥ = (eA(] − ∞,λ[) ∨ supp C)⊥ = 0, and there is a non-zero vector ξ ∈
eA([λ,∞[)∧ (supp C)⊥
(H). Hence Aξ,ξ =Bξ,ξ ≤ Bξ2 < λξ2, while at the same time
Aξ,ξ≥ λeA([λ,∞[)ξ,ξ= λξ2,acontradiction.
In therestof thissection weassumethat thetrace τ satisfiesτ (1)≥ 1071.
Remark4.11.Belowyouwillfindafewtechnicallemmasthatleadtotheconstructionofanoperatorthat cannot be written as alinearcombination of three projections. To this aim, webuild in 4.19anoperator A such thatthe distribution function F of A+ γR is not symmetric, whatever the values of γ ∈ R and R∈ Proj A.Thisshows,accordingtoTheorem4.9,thatA+γR isnotalinearcombinationoftwoprojections with positivecoefficients.ThecaseofarbitraryrealcoefficientsisthenobtainedeasilyinTheorem4.20.
Note thattheoperatorA:= B + C + 4E1+ 21E2 constructedinLemma4.19isalinearcombination of 9 projections, withB (constructed inLemma4.15and Corollary4.16)alinearcombinationof 4 mutually orthogonal projections,sayEB0,. . . ,EB3, thenC (builtinLemma4.17andCorollary 4.18)of 3 mutually orthogonalprojections,sayEC1,EC2,EC3,andfinallywithE1 andE2mutuallyorthogonal.Theconstruc- tionsofB,C andthepairE1,E2areindependentofeachother.Specificvaluesofτ ontheprojectionswill be usedinprovingCorollary4.22,dealingwithafactoroftypeIn.
Only the values of the trace on the 9 projections (and, of course, the values of the coefficients of the linearcombinations) matterforthevalidityofthelemmas mentionedabove.Thechoiceofallowedvalues, used in Lemmas 4.15and 4.17, wasmade to avoidcomplicated fractions.Theassumption τ (1) ≥ 1701 is needed tomakeroomfortheconstructionoftheseprojections.
Lemma 4.12. LetB,C ∈ A+,B < 1,τ (supp C) < 1, andlet E1,E2,R ∈ Proj A withE1 ⊥ E2. Denote Λi:= τ (Ei),c:= τ (R) andassumethat c> 0,Λ1> Λ2 andthat
A = B + C + λ1E1+ λ2E2+ γR
forsome0< λ1< λ2< γ.Thenthedistribution functionF ofA satisfies:
1. F (t)+ F+(γ + λ1− t)≤ 2τ(1)− c− Λ1− Λ2 fort∈]0,λ1];
2. F (t)+ F+(γ + λ2− t)≤ 2τ(1)− c− Λ2 fort∈]0,λ2];
3. F (γ)≤ τ(1)− c;
4. F (1)> τ (1)− (c+ 1+ Λ1+ Λ2);
5. F (γ + 1)> τ (1)− (1+ Λ1+ Λ2);
6. F (λ1+ 1)> τ (1)− (c+ 1+ Λ2);
7. F (γ + λ1+ 1)> τ (1)− (1+ Λ2);
8. F (λ2+ 1)> τ (1)− (c+ 1);
9. F (γ + λ2+ 1)> τ (1)− 1.
Proof. For3.,useA≥ γR andLemma4.6.For1.and2.,userespectivelyA≥ λ1(E1+ E2)+ γR andA≥ λ2E2+γR,Lemma4.6and(17) inTheorem4.9.TheothersresultfromLemma4.10,andrepresentations:4.
A= B+(C +λ1E1+λ2E2+γR);5.A= (B+γR)+(C +λ1E1+λ2E2);6.A= (B+λ1E1)+(C +λ2E2+γR);7.
A= (B+λ1E1+γR)+(C+λ2E2);8.A= (B+λ1E1+λ2E2)+(C+γR);9.A= (B+λ1E1+λ2E2+γR)+C. Corollary4.13.LetB,C∈ A+, B< 1,τ (supp C)< 1,andletE1,E2,R∈ Proj A,withE1⊥ E2,τ (E1)= 16,τ (E2)= 4 and c:= τ (R)> 0.Let
A = B + C + 4E1+ 21E2+ γR with γ > 21.
ThenthedistributionfunctionF of A satisfies:
1. F (t)+ F+(γ + 4− t)≤ 2τ(1)− c− 20 for t∈]0,4];in particular F (t)≤ τ(1)− c− 10 or F+(γ + 4− t)≤ τ (1)− 10 for anyt∈]0,4];
2. F (t)+ F+(γ + 21− t)≤ 2τ(1)− c− 4 fort∈]0,21];inparticularF (t)≤ τ(1)− c− 2 orF+(γ + 21− t)≤ τ (1)− 2 for any t∈]0,21];
3. F (γ)≤ τ(1)− c;
4. F (1)> τ (1)− c− 21;
5. F (γ + 1)> τ (1)− 21;
6. F (5)> τ (1)− c− 5;
7. F (γ + 5)> τ (1)− 5;
8. F (22)> τ (1)− c− 1;
9. F (γ + 22)> τ (1)− 1.
Proof. Putλ1= 4,λ2= 21, Λ1= 16,Λ2= 4 andγ > 21 inLemma4.12.
Lemma4.14.If F isadistributionfunctionsatisfying 1.− 9.ofCorollary4.13forγ > 44 andc> 42,then F isnot symmetric.
Proof. AssumeF is symmetric. Weshall consider four cases,oneof whichwould necessarilytakeplace if thefunctionF weresymmetricwithrespect tosome(α,β).
Iβ ∈ [γ,γ + 1[).By4.13.8,
F (β)≥ F (γ) ≥ F (22) > τ(1) − c − 1.
Hence,by4.13.2witht= 11+ γ− β,(5) and(6),either
F+(10)≤ F (11 + γ − β) ≤ τ(1) − c − 2 < F (β) = F+(α) or
F+(β + 10)≤ τ(1) − 2 < τ(1) = F+(β + α).
Thus α≥ 10,andby4.13.6and4.13.7,
F (5) + F+(α + β− 5) ≥ F (5) + F (γ + 5) > 2τ(1) − c − 10. (18) Notethatα + β− γ − 2≤ 2β − γ − 2< 2(γ + 1)− γ − 2= γ.Hence,by4.13.1witht= 2,(5) and4.13.3, we haveeither
F (2) + F+(α + β− 2) ≤ (τ(1) − c − 10) + τ(1) = 2τ(1) − c − 10 (19) or
F+(γ + 2) + F (α + β− γ − 2) ≤ (τ(1) − 10) + (τ(1) − c) = 2τ(1) − c − 10. (20) ByDefinition4.2,thelefthandsidesof(18),(19) and(20) areallequaltothevalueofF(α+β)on]0,α]∪]β,α+
β],whichyieldsacontradiction.
II β≥ γ + 1. By(6) and4.13.5,F+(α)= F (β)> τ (1)− 21> τ (1)− c.By4.13.3,α≥ γ.
By4.13.5,(5) andthe definitionofsymmetry 4.2,we haveas well21> τ (1)− F (γ + 1)≥ F (α + β)− F+(β) = F (α)− F+(0); by 4.13.3 and 4.13.5, F (γ + 1)− F+(0) ≥ F (γ + 1)− F (γ) ≥ 21. This implies α < γ + 1.
Sinceα≥ γ,wehaveF (α + β− 22)≥ F (2γ − 22)≥ F (γ + 22)> τ (1)− 1,by4.13.9.Thisimplies,again bysymmetryofF and4.13.8,F (2)+ F+(α + β−2)= F (22)+ F+(α + β−22)> (τ (1)−c−1)+ (τ (1)−1)= 2τ (1)− c− 2,hence
F+(2) > τ (1)− c − 2. (21)
In particular, F (3)> τ (1)− c− 10,and by (6) with 4.13.1, F (β)= F+(α) ≤ F+(γ + 1)≤ τ(1)− 10,so that,by4.13.7,β < γ + 5.
The inequality(21) yieldsalso F (11) > τ (1)− c− 2, so that,by 4.13.2,F+(γ + 10)≤ τ(1)− 2. Since γ≤ α < γ +1 andγ + 1≤ β < γ +5,wehaveγ + 10= α + β−t forsomet∈]0,α] witht< γ.Consequently, F (t)+ F+(γ + 10)= F (22)+ F+(α + β− 22), so thatF+(γ + 10)= F (22)+ F+(α + β− 22)− F (t)>
(τ (1)− c− 1)+ (τ (1)− 1)− (τ(1)− c)= τ (1)− 2,from 4.13.8and4.13.3.Weobtainedacontradiction.
III β∈ [0,γ/2[ Wehaveα≤ β < γ/2,so thatγ > α + β andF (γ)= τ (1),whichcontradicts4.13.3.
IVβ ∈ [γ/2,γ[ Wehaveα + β≥ γ, sinceby(5) F+(α + β)= τ (1) and F (γ)≤ τ(1)− c,by 4.13.3.By 4.13.4, there is a0< < 1 satisfying F () > τ (1)− c− 21.Suppose α + β > γ + 1+ . Choose γ with β < γ < γ. Wehaveγ + 1= α + β− t1,γ = α + β− t2 for some< ti < α,i= 1,2.Usingα≤ β < γ, 4.13.3andthechoiceof,wehaveF (ti)∈]τ(1)− c− 21,τ (1)− c] fori= 1,2.Ontheotherhand,by4.13.3, F+(α + β−t2)= F+(γ)≤ τ(1)−c,andF+(α + β−t1)= F+(γ + 1)> τ (1)−21,by4.13.5.Thiscontradicts c > 42 andthe equalityF (t1)+ F+(α + β− t1)= F (t2)+ F+(α + β− t2),satisfied fort1,t2 ∈]0,α]. We showedthatα + β≤ γ + 1+ < γ + 2,inparticular,by(5),F (γ + 2)= τ (1).
SinceF+(γ + 10)= τ (1),by4.13.2witht= 11 wehaveF (11)≤ τ(1)− c− 2.By4.13.8and(6),wehave F+(α)= F (β)≥ F (γ/2)≥ F (22)≥ τ(1)− c− 1,so thatα≥ 11.Consequently, F (2)+ F+(α + β− 2)= F (5)+ F+(α + β− 5).Since γ≤ α + β < γ + 2,wehaveα + β− 2,α + β− 5∈ [γ − 5,γ[⊆]22,γ[,and it followsfrom4.13.3and4.13.8thatF+(α + β− 2)− F+(α + β− 5)< 1 andF (5)− F (2)< 1.By4.13.6this
impliesF (2)> τ (1)− c− 6,henceF+(γ + 2)≤ τ(1)− 10 by4.13.1with t= 2.Wegotcontradictionwith theequalityF (γ + 2)= τ (1),whichends theproof.
Lemma4.15.Letτ beamultipleofτ0suchthatτ(1)> 17.ThereisanoperatorB∈ A+satisfyingB< 1 andthefollowingcondition: forany operatorC∈ Ah with τ(supp C)≤ 1,thedistribution functionFB+C
(w.r.t. τ) isnotsymmetric.
Proof. IfthedistributionfunctionF ofanoperator(w.r.t.τ)satisfies,forsome0= λ0< λ1< λ2< λ3/2, the inequalitiesF+(λi)− F (λi)≥ 4 for i= 0,1,2,F+(λ2)≥ τ(1)− 3,F (λ3)< τ(1), then F cannotbe symmetric.Infact,supposethatF is(α,β)−symmetricforsome0≤ α ≤ β,andputδ := α + β.Forδ < λ3 onehasF+(δ)≤ F (λ3)< τ(1), contradicting(5).Ontheotherhand,ifδ≥ λ3,then
F(δ)+(λi)− F(δ)+(λi−1) =
F+(λi)− F+(λi−1) +
F (δ− λi)− F (δ − λi−1)
≥
F+(λi)− F (λi) +
F (δ− λi)− F (δ − λi−1)
≥ 4 + F (δ − λ2)− τ(1)
≥ 4 + F+(λ2)− τ(1)≥ 1,
fori= 0,1,2 and anyλ−1< 0.Consequently,F(δ)hasatleastthreejumps,soittakesonmorethanthree values,contradicting(7).ThusF isnot(α,β)-symmetric.
Putnow B :=3
i=0λiEi,where Ei aremutuallyorthogonalprojectionsfrom A withsum1Asuchthat τ(E0),τ(E1),τ(E2) ≥ 5 and τ(E3) = 2, with λi as above, and assume additionally that λ3 < 1, so thatB < 1.The operatorB + C haseigenspaces Ei(H) for eigenvalues λi, with Ei ≥ Ei∧ (supp C)⊥, for i = 0,1,2,3. Hence τ(E0),τ(E1),τ(E2) ≥ 4,τ(E3) ≥ 1 and FB+C+ (λi)− FB+C(λi) ≥ 4 for i = 0,1,2,FB+C+ (λ3)− FB+C(λ3)> 0.Moreover,byLemma4.10appliedtoτ, FB+C+ (λ2)≥ τ(1)− (τ(E3)+ τ(supp C)) ≥ τ(1)− 3, so thatthe first partof the proof applies and, consequently, B has the desired properties.
Corollary 4.16.There existsan operator B ∈ A+ satisfying B< 1 and thefollowing condition: forany operatorC∈ Ah with τ (supp C)≤ 63, thedistributionfunctionof B + C (w.r.t.τ ) isnotsymmetric.
Proof. Putτ = (1/63)τ and build B forτ as inLemma4.15. Forany C with τ (supp C)≤ 63, we have τ(supp C)≤ 1,sothatFB+C is notsymmetric.
Lemma 4.17.There is C ∈ A+ satisfying τ (supp C) < 1 and the following condition: for any operator B ∈ Ah with−· 1A≤ B≤ (1− )· 1A forsome ∈ [0,1],thedistributionfunctionF ofB+ C (w.r.t.
τ ) isnotsymmetric.
Proof. Define C :=3
i=0λiEi, where λ0 = 0,λ1 = 2, λ2 = 4 and λ3 = 11, and where Ei are mutually orthogonal projections from A with sum 1A satisfying τ (E1+ E2+ E3)< 1 and τ (Ei)> τ (E3)> 0 for i = 0,1,2. ByLemma4.10, for any B ∈ A+ with B≤ 1, the distributionfunction F := FB+C of the operatorB + CsatisfiesF+(λi+ 1)≥ τ(1)−τ(Ei+1+· · ·+E3)= τ (E0+· · ·+Ei) fori= 0,1,2,3 (anempty sumisthezerooperator).ByLemma4.6,itsatisfiesF (λi)≤ τ(E0+· · · + Ei−1) fori= 0,1,2,3.Hence,for anoperatorB as inthestatementofourlemma, wecantakeB := B+ · 1A, andthen thedistribution functionFsatisfiesF(t)= F (t+),sothatF +(λi+1−)≥ τ(E0+· · ·+Ei),F(λi−)≤ τ(E0+· · ·+Ei−1) fori= 0,1,2,3.Thuswehave
F(t) =
⎧⎪
⎪⎪
⎪⎪
⎨
⎪⎪
⎪⎪
⎪⎩
0 for t∈] − ∞, −] τ (E0) for t∈]1 − , 2 − ], τ (E0+ E1) for t∈]3 − , 4 − ], τ (1)− τ(E3) for t∈]5 − , 11 − ], τ (1) for t∈]12 − , ∞[.
(22)
SupposethatF is(α,β)-symmetricforsome0≤ α ≤ β,andletδ := α + β.Ifδ < 11− ,thenF +(δ)≤ F(11−)< τ (1).Ifδ≥ 11−,then,foranyt≤ 6,F (δ)(t)= F(t)+F +(δ−t)∈ [F(t)+F +(5−),F(t)+ τ (1)]⊆ [F(t)+ τ (1)− τ(E3),F(t)+ τ (1)]. Wehavegot F (δ)(t0)< F (δ)(t1)< F (δ)(t2)< F (δ)(t3) for arbitrarily chosen t0 ≤ −,1− < t1 ≤ 2− ,3− < t2≤ 4− ,5− < t3 ≤ 6, using τ (E3)< τ (Ei) for i= 0,1,2.Thus,foranyδ∈ R,eitherF +(δ)< τ (1) orF (δ)takesonmorethan3 values,whichcontradicts (7).
Corollary 4.18.ThereisanoperatorC∈ A+ satisfying τ (supp C)< 1 and thefollowingcondition:forany operator B∈ Ah with −1· 1A≤ B ≤ 66· 1A,thedistribution functionofB + C is notsymmetric.
Proof. Take C := 67C, where C is the operator from Lemma 4.17, used with = 1/67. Then, for any B ∈ Ahwith−1· 1A≤ B ≤ 66· 1A,wehaveforB := (1/67)B thatFB+C isnotsymmetric,sothatFB+C is notsymmetric,either.
Lemma 4.19.Put A:= B + C + 4E1+ 21E2,where E1,E2∈ Proj A with E1⊥ E2,τ (E1)= 16,τ (E2)= 4, and where B and C are operators from Corollaries 4.16 and 4.18, respectively. Whatever the choice of projection R∈ A andγ∈ R, thedistributionfunctionof A+ γR is notsymmetric.
Proof. IForR suchthatτ (R)≤ 42 (andanyγ∈ R)wehaveτ (supp(C + 4E1+ 21E2+ γR))≤ τ(supp C)+ τ (E1)+ τ (E2)+ τ (R)< 1+ 16+ 4+ 42= 63,andCorollary 4.16togetherwith thedefinitionofB imply thatthedistributionfunctionof A+ γR isnotsymmetric.
II For −1 < γ≤ 44, we have−1A ≤ B + 4E1+ 21E2+ γR≤ 66· 1A. Corollary 4.18implies thatthe distributionfunctionofA+ γR = C + (B + 4E1+ 21E2+ γR) isnotsymmetric.
IIIForγ≤ −1 andτ (R)> 42 theoperatorA+ γR isnotpositive.Infact,τ (supp(C + 4E1+ 21E2))≤ τ (supp C)+ τ (E1)+ τ (E2)= 1+ 4+ 16< τ (R).Hence,for 0 = ξ ∈ (R ∧ (supp(C + 4E1+ 21E2))⊥)(H), we have(A+ γR)ξ,ξ=Bξ,ξ+ γξ2≤ (B+ γ)ξ2< (1+ γ)ξ2< 0.
ThusFA+γR(0)> 0,whichbyRemark4.3(4) meansthatA+ γR isnotsymmetric.
IVForγ > 44 andτ (R)> 42,thedistributionfunctionoftheoperatorA+ γR isnotsymmetric.Infact, itsatisfies theassumptionsofCorollary4.13,sowe canuseLemma4.14.
It should be noted that if aself-adjoint operator A cannotbe written as a real linear combination of three projections,itcannotbewrittenas acomplexlinearcombinationofthreeprojections,either.Infact, if A = αP + βQ+ γR with α,β,γ ∈ C, then A = (1/2)(A+ A∗) = (α)P + (β)Q+ (γ)R. Hence, Theorem 4.20, Corollaries 4.21 and 4.22, as well as Theorem 5.3 of the next section, hold for arbitrary combinations ofprojections.
Theorem4.20.LetA beaσ-finitetypeII1 algebrawithafaithfulnormaltracialstateτ0.Forany 0< δ ≤ 1 there exists anoperator A∈ A+ with τ0(supp A)≤ δ such that A isnot areal linear combination of three projections from A.
Proof. Put τ := (1071/δ)τ0. We will apply Lemma 4.19 with the trace τ , and show that the operator A defined there cannot be of the form αP + βQ+ γR, whatever α,β,γ ∈ R and choice of projections P,Q,R∈ A. Infact: