Contents lists available atScienceDirect
Journal of Functional Analysis
www.elsevier.com/locate/jfa
L
p-L
qmultipliers on locally compact groups
✩Rauan Akylzhanova, Michael Ruzhanskyb,a,∗
aSchoolofMathematicalSciences,QueenMaryUniversityofLondon, United Kingdom
bDepartmentofMathematics:Analysis,LogicandDiscreteMathematics, Ghent University,Belgium
a r t i c l e i n f o a b s t r a c t
Articlehistory:
Received14June2018 Accepted19September2019 Availableonline25September2019 CommunicatedbyStefaanVaes
MSC:
primary43A85,43A15 secondary35S05
Keywords:
Locallycompactgroups Fouriermultipliers Spectralmultipliers Hörmandertheorem
In this paper we discuss the Lp-Lq boundedness of both spectralandFourier multipliers on general locally compact separableunimodulargroups Gfor therange 1< p ≤2 ≤ q <∞.AsaconsequenceoftheestablishedFouriermultiplier theorem we also derive a spectral multiplier theorem on general locally compact separable unimodular groups. We thenapplyittoobtainembeddingtheoremsaswellastime- asymptotics for the Lp-Lq norms of the heat kernels for general positive unbounded invariant operators on G. We illustratetheobtainedresultsforsub-Laplaciansoncompact Lie groups and on the Heisenberg group, as well as for higherorderoperators.Weshow that ourresults implythe known results for Lp-Lq multipliers such as Hörmander’s Fourier multiplier theorem on Rn or known results for FouriermultipliersoncompactLiegroups.Thenewapproach developedin thispaper relies on advancing theanalysis in thegroup vonNeumann algebra anditsapplication to the derivationofthedesiredmultipliertheorems.
©2019TheAuthors.PublishedbyElsevierInc.Thisisan openaccessarticleundertheCCBYlicense (http://creativecommons.org/licenses/by/4.0/).
✩ The second author was supported by the FWO Odysseus grant and by the Leverhulme Grant RPG-2017-151.TheauthorswerealsosupportedbyEPSRCGrantEP/R003025/1.
* Correspondingauthor.
E-mailaddresses:r.akylzhanov@qmul.ac.uk(R. Akylzhanov),Michael.Ruzhansky@ugent.be (M. Ruzhansky).
https://doi.org/10.1016/j.jfa.2019.108324
0022-1236/©2019TheAuthors.PublishedbyElsevierInc.ThisisanopenaccessarticleundertheCC BYlicense(http://creativecommons.org/licenses/by/4.0/).
Contents
1. Introduction . . . . 2
1.1. Hörmander’stheoremonlocallycompactgroups . . . . 3
1.2. Spectralmultipliersonlocallycompactgroups . . . . 6
2. Notationandpreliminaries . . . . 8
2.1. Fouriermultipliersonlocallycompactgroups . . . 16
3. PaleyandHausdorff-Young-Paleyinequalities . . . 17
4. Nikolskiiinequalityonlocallycompactgroups . . . 22
5. Hörmander’smultipliertheoremonlocallycompactgroups . . . 25
5.1. Thecaseoflocallycompactabeliangroups . . . 28
5.2. ThecaseofcompactLiegroups . . . 29
5.3. Thecaseofq=∞ . . . 31
5.4. Thecaseofnon-invariantoperators . . . 32
6. Spectralmultipliersonlocallycompactgroups . . . 35
7. Heatkernelsandembeddingtheorems . . . 37
7.1. Sub-RiemannianstructuresoncompactLiegroups . . . 40
7.2. Sub-LaplacianontheHeisenberggroup . . . 41
7.3. RocklandoperatorsontheHeisenberggroup . . . 43
7.4. RocklandoperatorongradedLiegroups . . . 44
References . . . 46
1. Introduction
The aim of this paper is to give sufficient conditions for the Lp-Lq boundedness of Fourier and spectral multipliers onlocally compact separable unimodulargroups. It is known thatin this casewe must havep≤q and two classicalresults are available on Rn,namely,Hörmander’smultipliertheorem[37] for1< p≤2≤q <∞,andLizorkin’s multiplier theorem[41] for1< p≤q <∞.There is aphilosophical difference between these results:Hörmander’s theorem doesnot requireany regularity ofthe symbol and appliestopandqseparatedby2,whileLizorkintheoremappliesalsofor1< p≤q≤2 and 2≤p≤q <∞butimposes certainregularityconditionsonthesymbol.
InthispaperweaimatprovingtheHörmandertypetheoremexpressingconditionsin termsofthesharpdecaypropertyofthespectralinformationassociatedtotheoperator, on general locally compact separable unimodular groups based on developing a new approach relying on the analysis in the noncommutative Lorentz spaces on the group vonNeumannalgebra.Thissuggestedapproachseemsveryeffective,implyingasspecial casesknownresultsexpressedintermsofsymbols,insettingswhenthesymboliccalculus is available. The obtainedresults are for generalFourier multipliers, inparticular also implying new results for spectral multipliers. Lizorkin type theorem in the setting of locally compactgroupsrequiresarathersubstantialmodificationoftechniquesandwill appearelsewhere.
The class of groups covered by our analysis is very wide. In particular, it contains abelian,compact,nilpotentgroups,exponential,realalgebraicorsemi-simpleLiegroups, solvable groups(notallofwhicharetypeI,butwedonotneedtoassumethegroupto be oftypeIorII),andmanyothers.Asfarasweareawareourresultsarenewinallof these non-Euclideansettings.
Inthispaper wefocus ontheLp-Lq multipliersasopposed tothe Lp-multipliersfor which theorems of Mihlin-Hörmander or Marcinkiewicz types provide resultsfor both Fourierandspectralmultipliersinsomesettings,basedontheregularityofthemultiplier.
Lp-multipliershavebeenintensivelystudiedondifferentkindsofgroups.However,most oftheresultshavebeenobtainedforLpspectralmultipliers,forwhichawealthofresults isavailable: e.g.[44,45] onHeisenberg typegroups,[13] on solvable Lie groups,[43] on nilpotent and stratified groups, to mention only very very few. Lp Fourier multiplies havealsobeen studiedbutto alesser extentdue tolack ofsymbolic calculusthatwas notavailable until recently,e.g.Coifman andWeiss [14,15] on SU(2),[58,59] and then [26] oncompact Liegroups,or[12,27] ongradedLie groups.A characteristicfeatureof theLp-Lq multipliersisthatlessregularity ofthesymbolisrequired.Therefore,inthis paper we concentrate on the Lp-Lq multiplier theorems aiming at obtaining unifying resultsforgenerallocally compactgroups.Wegiveseveralapplications oftheobtained resultstoquestionssuchas embeddingtheorems anddispersiveestimates forevolution PDEs.
Theapproachto the Lp-Fouriermultipliersis different from thetechniqueproposed in this paper allowing us to avoid making the assumption that the group is compact or nilpotent. In this paper we are interested in both Fourier multipliers and spectral multipliers,forthelattersomeLp-Lqresultsbeingavailableinsomespecialsettings,see e.g.[17],andalso[18],aswellas[1] forthecaseofSU(2),andforthediscussionofsome relationsbetweenthoseinthegroupsettingwecanreferto [59] andreferencestherein.
In the context of harmonic analysis, group von Neumann algebras are used in two differentways.Inthefirstapproach[32],thegroupGcanbeunderstoodasthe‘frequency domain’and itsgroup von Neumannalgebra VNR(G) can be thought of as the ‘space domain’.Inthispaper,weviewGas‘spacedomain’anditsgroupvonNeumannalgebra VNR(G) asthe‘frequencydomain’.
Bythe combinatorialmethod it is possible to establish theLp-Lq estimates for the Poisson-type semigroup Pt on discrete groups G [38]. Finally we note that multiplier estimates on noncommutative groups are in general considerably more delicate than those in the commutative case, recall e.g. the asymmetry problem and its resolution in[21]. A linkbetween Fourier multipliersand Lorentzspaces on groupvon Neumann algebrashasbeenoutlined in[4].
Wecan alsomention afew related works:[19] for generalC∗-algebras presentation, vonNeumann’soriginalworks[68,69],and[23] forrecentworkondirectintegraldecom- positionsforvonNeumannalgebras.
Wenowproceedto makingamorespecificdescriptionoftheconsideredproblems.
1.1. Hörmander’s theoremon locallycompactgroups
Toputthisincontext,werecallthatin[37, Theorem1.11],LarsHörmandershowed thatfor1< p≤2≤q <∞,ifthesymbolσA:Rn→C ofaFouriermultiplierAacting ontheSchwartzspaceS(Rn) satisfiesthecondition
sup
s>0
s
⎛
⎜⎝
ξ∈Rn:|σA(ξ)|≥s
dξ
⎞
⎟⎠
1 p−1q
<+∞, (1.1)
then AisaboundedoperatorfromLp(Rn) toLq(Rn).Here,asusual, theFouriermul- tiplier AonRn actsbymultiplicationontheFouriertransformside,i.e.
Af(ξ) =σA(ξ)f(ξ), ξ∈Rn, f ∈ S(Rn). (1.2) Moreover, itthenfollowsthat
ALp(Rn)→Lq(Rn)sup
s>0
s
⎛
⎜⎜
⎜⎝
ξ∈Rn
|σA(ξ)|≥s
dξ
⎞
⎟⎟
⎟⎠
1 p−1q
, 1< p≤2≤q <+∞. (1.3)
TheLp-Lq boundednessofFouriermultipliershasbeenalsorecentlyinvestigatedinthe contextofcompactLiegroups,andwenowbrieflyrecalltheresult.LetGbeacompact Lie group and G its unitary dual. For π ∈ G, we write dπ for the dimension of the (unitaryirreducible)representation π.In[2] theauthorshaveshownthat,foraFourier multiplierA actingvia
Af(π) =σA(π)f(π) byitsglobalsymbol σA(π)∈Cdπ×dπ wehave
ALp(G)→Lq(G)sup
s>0
s
⎛
⎜⎜
⎝
π∈G σA(π)op≥s
d2π
⎞
⎟⎟
⎠
1 p−1q
, 1< p≤2≤q≤ ∞. (1.4)
Here forπ∈G, theFouriercoefficientsaredefinedas f(π) =
G
f(x)π(x)∗dx,
and σA(π)op is the operatornorm of σA(π) as the linear transformation of the rep- resentation space of π ∈ G identified with Cdπ. For a general development of global symbols and the corresponding global quantization of pseudo-differential operators on compact Liegroupswecanreferto [56,57].
One ofthe resultsofthis paper generalises both multipliertheorems (1.3) and (1.4) to thesettingofgenerallocallycompactseparableunimodulargroupsG.
ByaleftFouriermultiplierinthesettingofgenerallocallycompactunimodulargroups wewill meanleft invariantoperatorsthat aremeasurablewith respecttotheright group vonNeumannalgebra VNR(G),seeSection2.1foradiscussion.
Thus,inTheorem 5.1weprovethefollowing inequality
AfLp(G)→Lq(G)sup
s>0
s
⎡
⎢⎣
t∈R+: μt(A)≥s
dt
⎤
⎥⎦
1 p−1q
fLp(G), 1< p≤2≤q <+∞, (1.5) whereμt(A) arethet-thgeneralisedsingularvaluesofA,seeDefinition2.8fortheprecise definitionandproperties(following[67]).Anextensionof(1.5) to q=∞will beshown inTheorem 5.9.
Remark 1.1.The measurability assumption on A implies that the domain Dom(A) ⊂ L2(G) isdenseinL2(G).WithoutlossofgeneralityweassumethatDom(A) isdensein Lp(G) forevery1< p≤ ∞.
The proof of inequality (1.5) is based on a version of the Hausdorff-Young-Paley inequalityonlocally compactseparablegroupsthatweestablishforthispurpose.
The key idea behind the extension (1.5) is that Hörmander’s theorem (1.3) can be reformulatedas
ALp(Rn)→Lq(Rn)sup
s>0
s
⎛
⎜⎜
⎜⎝
ξ∈Rn
|σA(ξ)|≥s
dξ
⎞
⎟⎟
⎟⎠
1 p−1q
σALr,∞(Rn) ALr,∞(VN(Rn)),
(1.6) where 1r = 1p − 1q, σALr,∞(Rn) is the Lorentz space norm of the symbol σA, and ALr,∞(VN(Rn)) is thenorm of the operatorA inthe Lorentzspace on thegroup von NeumannalgebraVN(Rn) ofRn.Inturn,ourestimate(1.5) isequivalenttotheestimate
ALp(G)→Lq(G)ALr,∞(VNR(G)) sup
s>0
s
⎡
⎢⎣
t∈R+: μt(A)≥s
dt
⎤
⎥⎦
1 r
, (1.7)
whereA isactingontheSchwartz-BruhatspaceS(G) andALr,∞(VNR(G)) isthenon- commutative Lorentz space norm on the right group von Neumann algebra VNR(G) of G. Thus, the Lorentz spaces become a key point for the extension of Hörmander’s theoremto thesetting ofgenerallocally compact(unimodular)groups.
The assumption ofunimodularity ofGensures theexistenceof thePlancherel trace onthegroupvonNeumannalgebra VNR(G) andthuscanbeviewedasnaturalallowing onetousebasicresultsofFourieranalysisovervonNeumannalgebras,suchas,forexam- ple, Plancherel formula(see Segal [61]).Otherwise, thenoncommutative Lorenzspaces cannot be constructedas subsetsofτ-measurable operators.Nevertheless,theunimod- ularity assumption maybe inprinciple avoided,see e.g.[22], butthe expositionwould involvetheTomita-TakesakimodulartheoryandtheHaagerupreductiontechnique.For amoredetaileddiscussionofpseudo-differentialoperatorsinthegeneralsettingoflocally compact groups(possiblynon-unimodular)wereferto[42].
1.2. Spectral multiplierson locallycompactgroups
LetusillustratetheuseoftheFouriermultipliertheorem(1.5) intheimportantcase of spectral multipliers on locally compact groups. Later, in Theorem 6.1 we will give a spectral multiplier result on general semifinite von Neumann algebras, however, we nowformulateitsspecialcaseforthecaseofgroupvonNeumannalgebrasassociatedto locally compactgroups.
Interestingly, this resultasserts thatthe Lp-Lq norms of spectral multipliersϕ(|L|) depend essentially only on the rate of growth of traces of spectral projections of the operator|L|:
Theorem 1.2.Let G be a locally compact separable unimodular group and let L be a left Fourier multiplier on G. Assume that ϕ is a monotonically decreasing continuous function on[0,+∞)suchthat
ϕ(0) = 1,
u→+∞lim ϕ(u) = 0.
Then wehavetheinequality ϕ(|L|)Lp(G)→Lq(G)sup
u>0
ϕ(u)
τ(E(0,u)(|L|))1p−1q, 1< p≤2≤q <∞. (1.8) Here E(0,u)(|L|) are the spectral projections associated to the operator |L| to the interval (0,u), seeSection 2forprecise definitions, andτ is thecanonical traceon the right groupvon NeumannalgebraVNR(G),seealsoSection2foradiscussion.
Also wenote thatmoregeneralstatements,weakening theaboveassumptionson ϕ, are possible,seeCorollary 6.3.
The estimate (1.8) says that if the supremum on the right hand side is finite then the operator ϕ(|L|) is bounded from Lp(G) to Lq(G). Moreover, the estimate for the operator norm can be used for deriving asymptotics for propagators for equations on G. For example, we get the following consequences for the Lp-Lq norm for the heat
kernelofL,applyingTheorem1.2withϕ(u)=e−tu,orembeddingtheoremsforLwith ϕ(u)= (1+u)1 γ.
We note that estimates of the type (1.10) are exactly those leading to subsequent Strichartzestimates.Here,ourmethodisverydifferentfromtheusualonesaswedonot getitbyinterpolationfromtheend-pointcase.
Corollary 1.3. Let G be a locally compact unimodular separable group and let L be a positiveleft Fouriermultiplier suchthat forsomeαwehave
τ(E(0,s)(L))sα, s→ ∞. (1.9)
Thenforany 1< p≤2≤q <∞ thereisaconstant C=Cα,p,q>0suchthat wehave e−tLLp(G)→Lq(G)≤Ct−α
1 p−1q
, t >0. (1.10)
Wealso havetheembeddings
fLq(G)≤C(1 +L)γfLp(G), (1.11) providedthat
γ≥α 1
p−1 q
, 1< p≤2≤q <∞. (1.12) The number α in (1.9) is determined based on the spectral properties of L and is oftencomputable.Forexample,wehave
(a) if L is the sub-Laplacian ona compact Lie groupG then α= Q2, where Q is the HausdorffdimensionofGwithrespectto thecontroldistanceassociated toL; (b) if L is the sub-Laplacian on the Heisenberg group G = Hn then α = Q2, where
Q= 2n+ 2 isthehomogeneousdimensionofHn; (c) More generally, if L = (−1)N
n
k=1
Xk2N+ n k=1
Yk2N
, where the vector fields X1,. . . ,Xn,Y1,. . . ,Yn,T isthe(usual)basisintheLiealgebrahnoftheHeisenberg groupHn suchthat[Xk,Yk]=T andallothercommutatorsarezero,thenα= 2NQ . Consequently,inbothofthesub-Laplaciancases(a)and(b),Corollary1.3impliesthat forany1< p≤2≤q <∞thereisaconstantC=Cp,q>0 such thatwehave
e−tLLp(G)→Lq(G)≤Ct−Q2
1 p−1q
, t >0. (1.13) The embeddings (1.11) under conditions (1.12) show that the statement of The- orem 1.2 is in general sharp. Taking ϕ(s) = (1+s)1a/2 and applying (1.8) to the
sub-Laplacian Δsub in either of examples (a) or (b) above, we get that the operator ϕ(−Δsub)= (I−Δsub)−a/2 isLp(G)-Lq(G) boundedandtheinequality
fLq(G)≤C(1−Δsub)a/2fLp(G) (1.14) holds trueprovidedthat
a≥Q 1
p−1 q
, 1< p≤2≤q <∞. (1.15) However, this yields the Sobolev embedding theorem which is well-known to be sharp at least inthe case(b) ofthe Heisenberg group([29]), showingthe sharpness of Theo- rem1.2andhencealsooftheFouriermultipliertheorem(1.7).Moreexamplesaregiven inSection7.
Throughout thepaper wewill usethenotation ofthetypefX fY ifwe have fX ≤CfY withtheconstantC thatmaydependonthespacesX,Y butnotonf. 2. Notationandpreliminaries
In thissection wefix thenotation andbriefly recall somepreliminarieson vonNeu- mannalgebrastobeusedfordevelopingsubsequentharmonicanalysisonlocallycompact groups. For exposition purposes it seems beneficial to recall several generalnotions in thecontextofgeneralvonNeumannalgebrasM.However,forourapplicationtomulti- pliersonlocallycompactgroupsGwewillbelatersettingM to betherightgroupvon NeumannalgebraVNR(G).Inparticular,wewillbe abletoreadilyapplythenotionof noncommutative Lorentz spaces onM as developed in[51], one of thekey ingredients forouranalysis.
LetM⊂ L(H) beasemifinitevonNeumannalgebraactinginaHilbertspaceHwith a traceτ. The semifiniteassumption simplifies the formulationsand is satisfied inour main exampleM= VNR(G).
Definition 2.1(Affiliated operators).A linearclosedoperatorA (possiblyunboundedin H)issaidtobe affiliatedwith M,symbolicallyAνM,ifitcommuteswiththeelements of thecommutant M! ofM,i.e.
AU=U A, for all U ∈M!. (2.1)
This relation ν is anaturalrelaxation ofthe relation ∈: ifA is abounded operator affiliated withM, then bythedouble commutant theorem A∈M.Oneof theoriginal motivations [46,47] of Johnvon Neumannwas to build amathematical foundation for quantummechanics.Inthisframework,theobservableswithunboundedspectrumcorre- spondtoclosed denselydefinedunboundedoperators.AlthoughthealgebraM consists
of bounded operators, the technique of projections makes it possible to approximate unboundedoperators.
Somepropertiesof tracesshall be usedintheproofs ofour theorems.Therefore, we giveabriefbackgroundontracessummarisingtheresultsthatwillbeusedinthesequel.
For some descriptionof measurable fields of operators and links to the representation theoryandgeneralvonNeumannandC∗-algebraswereferto[28,AppendicesBandC].
Thefollowingdefinitionistakenfrom[20,Definition I.6.1,p. 93]:
Definition 2.2.Let M be avon Neumannalgebra. A traceon the positivepartM+ = {A ∈ M:A∗ = A > 0} of M is a functional τ defined on M+, taking non-negative, possiblyinfinite,real values,possessingthefollowingproperties:
• IfA∈M+ andB∈M+,wehaveτ(A+B)=τ(A)+τ(B);
• IfA∈M+andλ∈R+,wehaveτ(λA)=λτ(A) (withtheconventionthat0·+∞= 0);
• IfA∈M+ andifU isaunitaryoperatorofM,then τ(U AU−1)=τ(A).
Wesaythatτisfaithful(orexact)iftheconditionA∈M+, τ(A)= 0,implythatA= 0.
We say thatτ is finite if τ(A)< +∞ for all A ∈ M+. We say that τ is semifinite if, foreachA∈M+, τ(A) isthesupremum ofthenumbersτ(B) overthoseB ∈M+ such thatB≤Aandτ(B)<+∞.Wesaythatτ isnormalif,foreachincreasingfilteringset S ⊂M+withsupremumS∈M+,τ(S) isthesupremumof{τ(B)}B∈S.AvonNeumann algebraM issaidto besemifinite ifthere existsasemifinitefaithfulnormaltraceτ on M+.
Definition 2.3 (τ-measurable operators S(M)).A closeable operator A (possibly un- bounded) affiliated with M is said to be τ-measurable if for each ε > 0 there exists aprojectionpinM suchthatpH ⊂D(A) andτ(I−p)≤ε. Here D(A) isthedomain ofAinH. Wedenote byS(M) theset ofallτ-measurableoperators.
Werecallthefollowing resultwhichwillbepartially used.
Theorem 2.4 ([62, Theorem 4, p. 412]).If operators A and B are τ-measurable with respecttoavonNeumannalgebra M,thensoare A∗,A+B andAB,i.e.themaps
+ :M×M(A, B)→A+B∈M, (2.2)
·:M×M(A, B)→AB∈M, (2.3)
∗:MA→A∗∈M (2.4)
arewell-defined.
Wenote thatthenotionofτ-measurability doesnotappearintheclassicaltheoryof SchattenclassessinceforM =L(H) wehaveS(L(H))=L(H).
It canbe seenthatA∈M+ ifandonlyifA= (A1/2)∗A1/2.
Example 2.5.Let G be a locally compact unimodular group with VNR(G) the group von Neumannalgebragenerated bythe rightregular representationπRof G.LetA be alinearboundedoperatorcommutingwiththeleft regularrepresentation. Thenbythe double commutanttheoremA∈VNR(G) anditsactionisgivenby
L2(G)h→Ah=h∗KA∈L2(G),
where KAisitsconvolution kernel.Wecandefineatraceτ onVN+R(G) by
τ(A) :=
K
A122L2(G), ifKA1/2∈L2(G),
∞, otherwise. (2.5)
The trace τ on M+ can also be extended to the space S(M) of all τ-measurable positiveoperators.
Proposition2.6. Let(M,τ)beavonNeumannalgebraandletAbeaτ-measurablelinear operator. Assumethat ϕisaBorelfunctionon sp(|A|)⊂[0,+∞).Thenwehave
τ(ϕ(|A|)) =
+∞
0
ϕ(t)dμ(t), (2.6)
where μt=τ(Et)and
|A|=
+∞
0
tdEt(|A|).
Takingϕ(t)=t,wecanalternativelydefineatraceτ as follows τ(A) =
tdμ(t).
Proof of Proposition 2.6. For the spectral measure we can take the family {E[0,t)}t≥0
of spectral projectionsE(0,t)corresponding to theintervals [0,t). Thereadercancheck thatthespectralmeasure axiomshold true.
The traceτ is continuous with respect to the τ-measure. In view of the monotone convergence theorem(see [67,Theorem 3.5])wecanassume,withoutloss ofgenerality, thatAisaboundedτ-measurableoperator.Indeed,foreveryτ-measurableoperator|A| there existsasequence{An}ofτ-measurablebounded operators
An= n 0
tdEt(|A|)≤A
convergingtoAintheτ-measuretopology.Then,takingthelimit
nlim→∞τ(An) =τ(A),
we justify the claim. We notice that every Borel function can be uniformly approxi- matedbyboundedBorelfunctions.Thus,weconcentratetoestablish(2.6) forbounded measurableAandboundedBorelfunctionsϕon[0,AB(H)].Bythespectralmapping theoremwehave
sp(ϕ(|A|)) =ϕ([0,AB(H)]).
Let 0 ≤ λ1 ≤ λ2 ≤ . . . λN be a partition of the interval ϕ([0,AB(H)]). Then the Riemann-likesums
RN = N k=1
λkEϕ−1(λk−1,λk)(|A|)
convergeto ϕ(|A|) intheτ-measure topology.Thetraceτ onRN isgiven by
τ(RN) = N k=1
λkτ(Eϕ−1(λk−1,λk)(|A|)). (2.7) Onecannoticethatthesumin(2.7) isaLebesgueintegralsum
N k=1
λkμ(λk−1,λk)
fortheintegral
A
0
ϕ(t)dμ(t),
whereweset themeasureμ((a,b))=τ(E(a,b)), (a,b)⊂[0,AB(H)]. 2
Example 2.7.Let M = {Mϕ: L2(X,μ) f → Mϕf = ϕf ∈ L2(X,μ)}ϕ∈L∞(X,μ)
and take τ(Mϕ) :=
X
ϕdμ, where (X,μ) is a measure space. Then an operator Mϕ
isτ-measurable ifand onlyifϕisaμ-almosteverywherefinitefunction.
The∗-algebraS(M) isabasicconstruction forthenoncommutativeintegration.Let A=U|A|be thepolardecomposition.Thespectraltheorem yieldsthat
|A|=
Sp(|A|)
λdEλ(|A|), (2.8)
where {Eλ(|A|)}λ∈Sp(|A|) are thespectral projectionsassociated with the operator|A|. Here dEλ(|A|) shouldbeunderstoodastherelativedimensionfunctionfirstconstructed in[46].Since A isaffiliated withM, theprojections satisfyEλ(|A|)∈M. Now, weare ready to‘measurethespeedofdecay’oftheoperatorA.
Definition 2.8 (Generalised t-th singular numbers).For an operator A ∈ S(M), define thedistributionfunctiondλ(A) by
dλ(A) :=τ(E(λ,+∞)(|A|)), λ≥0, (2.9) where E(λ,+∞)(|A|) is the spectral projection of |A| corresponding to the interval (λ,+∞).Forany t>0,wedefinethegeneralisedt-thsingularnumbersby
μt(A) := inf{λ≥0 : dλ(A)≤t}. (2.10) Forthesakeoftheexpositionclaritywenowformulatesomepropertiesofthedistri- butionfunctiondAwhichwewill beusingintheproofs.
Proposition 2.9. LetA∈S(M).Thenwehave
dA(μA(t))≤t; (2.11)
μA(t)> s if and only if t < dA(s); (2.12) sup
t>0
tαμA(t) = sup
s>0
s[dA(s)]α for 0< α <∞. (2.13) Theproofofthispropositionisalmostverbatimtotheproofof[33,Proposition1.4.5 on page46]. Theword ‘almost’standsfor theright-continuity ofthe non-commutative distribution function dA(s) which is discussed after [67, Definition 1.3 on page 272].
Therefore, inthefollowingproof weshallusetheright-continuity ofdA(s) withoutany justification.
Proof of Proposition 2.9. Let sn ∈ {s>0 :dA(s)≤t}be suchthatsn μA(t).Then dA(sn)≤tandtheright-continuityofdAimpliesthatdA(μA(t))≤t.Thisproves(2.11).
Now,weapplythispropertytoderive(2.12).Ifs< μA(t)= inf{s>0 :dA(s)≤t},then sdoesnotbelongtotheset{s>0 : dA(s)≤t} =⇒ dA(s)> t.Conversely,ifforsomet ands,wehadμA(t)< s,thentheapplicationofdAandproperty(2.12) wouldyieldthe contradiction dA(s) ≤ dA(μA(t)) ≤t. Property (2.12) is established. Finally, we show
(2.13).Givens>0,pickεsatisfying0< ε< s.Property(2.12) yieldsμA(dA(s)−ε)> s whichimpliesthat
sup
t>0
tαμA(t)≥(dA(s)−ε)αμA(dA(s)−ε)>(dA(s)−ε)αs. (2.14) Wefirst letε→0 and thentakethe supremumover alls>0 to obtainonedirection.
Conversely,givent>0,pick0< ε< μA(t).Property(2.12) yieldsthatdA(μA(t)−ε)> t.
Thisimplies thatsups>0s(dA(s))α ≥(μA(t)−ε)(dA(μA(t)−ε))α >(μA(t)−ε)tα. We firstletε→0 andthentakethesupremumoverallt>0 toobtaintheoppositedirection of(2.13). 2
Hereweformulatesomepropertiesofμtthatweuseintheproofof Theorem5.1.
Lemma 2.10 ([67,Lemma 2.5, p. 275]).Let A,B be τ-measurable operators. Then the followingpropertiesholdtrue.
(1) The map (0,+∞) t → μt(A) is non-increasing and continuous from the right.
Moreover,
tlim→0μt(A) =A ∈[0,+∞]. (2.15)
(2) μt(A) =μt(A∗). (2.16)
(3)
μt+s(AB)≤μt(A)μs(B). (2.17) (4)
μt(ACB)≤ ABμt(C), for any τ-measurable operatorC. (2.18) (5) Forany continuous increasingfunction f on[0,+∞) wehave
μt(f(|A|)) =f(μt(|A|)). (2.19) InLemma2.10,weformulateonlythepropertiesweuse,whereasin[67,Lemma 2.5, p. 275] thereadercanfindmoredetails.
Example2.11.FortheoperatorMϕinExample2.7,fromDefinition2.8wecanshowits generalisedt-thsingularnumberstobe
μt(Mϕ) =ϕ∗(t),
whereϕ∗(t) istheclassicaldecreasingrearrangement(seee.g.[6]).
Asanoncommutativeextension[51] oftheclassicalLorentzspaces,wedefineLorentz spacesLp,q(M) associatedwithasemifinitevonNeumannalgebraM asfollows:
Definition 2.12 (Noncommutative Lorentz spaces).For 1≤p<∞, 1≤q <∞, denote byLp,q(M) thesetofalloperatorsA∈S(M) satisfying
ALp,q(M):=
⎛
⎝ +∞
0
t1pμt(A)qdt t
⎞
⎠
1 q
<+∞. (2.20)
Forq=∞,wedefineLp,∞(M) asthespaceofalloperatorsA∈S(M) satisfying ALp,∞(M):= sup
t>0
tp1μt(A). (2.21)
Withthis, for1≤p<∞,wecanalso defineLp-spacesonM by
ALp(M):=ALp,p(M)=
⎛
⎝
+∞
0
μt(A)p dt
⎞
⎠
1 p
.
The classical Lorentz spaces Lp,q(X,μ) correspond to the case of commutative von Neumannalgebra.Modulustechnicaldetails[20,p. 132,Theorem 1],anarbitraryabelian von Neumann algebra ina Hilbert space H is isometrically isomorphic to the algebra {Mϕ}ϕ∈L∞(X,μ) fromExample2.7.ThennoncommutativeLorentzspacescoincidewith theclassicalones:
Example 2.13 (Classical Lorentz spaces).Let M be the algebra {Mϕ}ϕ∈L∞(X,μ) from Example 2.7consisting ofallthemultiplicationoperators Mϕ: L2(X,μ)f →Mϕf = ϕf ∈L2(X,μ).ByExample 2.11, wehave
μt(Mϕ) =ϕ∗(t).
Thus, theLorentzspaceLp,q(M) consistsofalloperatorsMϕ suchthat
+∞
0
[t1pϕ∗(t)]qdt
t <+∞, whichgivestheclassicalLorentzspace.
Concerning the structure of semifinite von Neumann algebras, given an arbitrary semifinitevonNeumannalgebraM withatraceτ,thereisanisomorphismofM ontoa certain HilbertalgebraU ([20, p. 99,Theorem 2]).Thus,we constructthetraceonthe
Hilbertalgebra yielding thetrace onM due to isomorphism. We referto [20], [48] for moredetailsonthis.
Letnow G be alocally compact unimodularseparable group.Denote by πL(g) and πR(g) theleft andtherightactionofGonL2(G),respectively:
πL(g)f(x) :=f(g−1x), πR(g)f(x) :=f(xg),
andbyVNL(G) thegroupvonNeumannalgebrageneratedbyalltheπL(g) withg∈G, i.e.
VNL(G) :={πL(g)}!!g∈G, andsimilarly
VNR(G) :={πR(g)}!!g∈G,
where !! is thebicommutant ofthe self-adjoint subalgebras {πL(g)}g∈G,{πR(g)}g∈G ⊂ L(L2(G)).It hasbeen shownin[60] that
VNL(G)!= VNR(G), (2.22)
VNR(G)!= VNL(G). (2.23)
We do notmake assumption that Gis either of type I or type II. The decomposition theory forunitary representationsof locally compact separableunimodulargroupshas beenestablishedin[24,25].
FromnowonwetakeM= VNR(G).
Forf ∈L1(G)∩L2(G), wesay thatf on Ghas a Fouriertransform whenever the convolutionoperator
Rfh(x) := (h∗f)(x) =
G
h(g)f(g−1x)dg (2.24)
is a τ-measurable operator with respect to VNR(G), i.e. Rf ∈ S(VNR(G)). The Plancherelidentitytakes ([61,Theorem3onpage282])theform
RfL2(VNR(G)) =fL2(G). (2.25) Inthissetting,theHausdorff-Younginequalityhasbeenestablishedin[40] intheform
RfLp(VNR(G)) fLp(G), 1< p≤2. (2.26)
The constantin (2.26) has been computed in [55] for simply connected real nilpotent Lie groups with explicitly computable Plancherel measures. It canbe shown that the constantislessthan1 forlocallycompactgroupswithnocompactopensubgroups[31].
In [51], as an application of the technique of the t-th generalised singular values, the Hardy-Littlewood theorem ([34]) has been generalised to an arbitrary locally compact separableunimodulargroupG:
Theorem 2.14([51]). Let1< p≤2andf ∈Lp(G).Thenwehave
RfLp,p(VNR(G))≤ fLp(G). (2.27) Remark 2.15.ThePlancherel equality (2.25) by Segal [61] and Kosaki’s version[51] of Hardy-Littlewood inequality(2.27) have been originally established forthe left convo- lution Lfh=f ∗h. However, the sameline of reasoningyields inequalities (2.27) and (2.25) withtherightconvolution Rf.Weworkwiththerightconvolution operatorsRf
heresinceitnaturallycorrespondstoleft-invariantoperatorswhenanalysingtheFourier multipliersongroups.
Using thetechniqueof thet-thgeneralisedsingular valuesdevelopedin[67], wecan formulateboth theHausdorff-Young(2.26) andHardy-Littlewood(2.27) inequalitiesin thefollowing forms(for1< p≤2):
⎛
⎝
+∞
0
μt(Rf)pdt
⎞
⎠
p1
≡ RfLp(VNR(G)) ≤ fLp(G), (2.28)
⎛
⎝
+∞
0
tp−2μt(Rf)pdt
⎞
⎠
1 p
≡ RfLp,p(VNR(G))≤ fLp(G). (2.29)
Inthesequel,whenweprovePaleyinequalityinTheorem3.1,theHardy-Littlewood inequalities(2.27) and(2.29) (fortherightconvolutionRf)willalsofollowasitsspecial cases.
2.1. Fouriermultiplierson locallycompactgroups
LetGbealocally compactseparableunimodulargroup.Thefirstquestionishowto understandthenotionofFouriermultipliers.Inthefirstinstanceweadoptthefollowing definition:
Definition 2.16.A linearoperatorA is saidto be aleft Fouriermultiplier on G ifA∈ S(VNR(G)).
Ifwe now recallDefinition 2.1we cansee thatA isa leftFourier multiplieron Gif and only if A is affiliated with the right group von Neumann algebra VNR(G) and is τ-measurable.WecanthenclarifyDefinition2.16further:
Remark2.17.For M = VNR(G) the operators affiliated with M areprecisely those A thatareleft-invariant onG,namely,
Ais affiliated with VNR(G) ⇐⇒ AπL(g) =πL(g)A, for allg∈G. (2.30) Summarising this observation with Definition 2.16, left Fourier multipliers on G are preciselytheleft-invariantoperatorsthatare measurable(inthesenseofDefinition2.3).
Forclarityand inviewofitsimportance,wegiveashort justificationofthis.
Proof of Remark2.17. =⇒.ByDefinition2.16,wehave
AU=U A, for all U ∈VNR(G)!. (2.31) Thenby(2.23),and bytaking U =πL(g),g∈G, weseethatA mustbeleft-invariant.
⇐=.Wehave
AπL(g) =πL(g)A, for all g∈G.
Bydefinition,thealgebraVNL(G) istheclosure oftheinvolutivesubalgebra {πL(g)}g∈G⊂ L(L2(G))
inthestrongoperatortopology.Therefore,we obtain
AU=U A, for all U ∈VNR(G)!, (2.32) whereweused (2.23).Thiscompletes theproof ofRemark2.17. 2
3. PaleyandHausdorff-Young-Paleyinequalities
OuranalysisofLp-Lq multiplierswillbe basedonaversionoftheHausdorff-Young- Paleyinequalitythatweestablishinthissectioninthecontextoflocallycompactgroups.
Itwill be obtainedbyinterpolation betweentheHausdorff-Younginequalityand Paley inequalitythatwediscussfirst.
Westartfirstwithaninequalitythatcanbe regardedas aPaleytypeinequality.
Theorem3.1(Paleyinequality). LetGbealocallycompactunimodularseparablegroup.
Let1< p≤2.Supposethat apositivefunctionϕ(t)satisfies thecondition
Mϕ:= sup
s>0
s
t∈R+
ϕ(t)≥s
dt <+∞. (3.1)
Then forallf ∈Lp(G) wehave
⎛
⎝
+∞
0
μt(Rf)pϕ(t)2−pdt
⎞
⎠
1 p
≤M
2−p
ϕp fLp(G). (3.2)
As usual,theintegraloveranemptyset in(3.1) isassumedto bezero.
Wenotethattakingϕ(t)=1t werecoverKosaki’sHardy-Littlewoodinequality(2.29).
Inthissense,thePaleyinequalitycanbeviewedasanextensionof(one of)theHardy- Littlewoodinequalities.AsasmallbyproductofourproofofTheorem3.1wethusgeta simpleproofof Theorem2.14.
Proof of Theorem3.1. Let measure ν be absolute continuous with respect to the Lebesguemeasure onRn+,i.e.i.e.
ν(t)
dt :=ϕ2(t), t∈R+ (3.3)
where ν(t)dt is theRadon-Nikodym derivativeat the point t∈R+. We definethe corre- spondingspaceLp(R+,ν),1≤p<∞,asthespaceofcomplex(orreal)valuedfunctions f =f(t) suchthat
fLp(R+,ν):=
⎛
⎜⎝
R+
|f(t)|pϕ2(t)dt
⎞
⎟⎠
1 p
<∞. (3.4)
Wewill showthatthesub-linearoperator
T:Lp(G)f →T f :=μt(Rf)/ϕ(t)∈Lp(R+, ν)
is well-definedandboundedfromLp(G) toLp(R+,ν) for1< p≤2.Inotherwords,we claimthatwehavetheestimate
T fLp(R+,ν)=
⎛
⎜⎝
R+
μt(Rf) ϕ(t)
p
ϕ2(t)dt
⎞
⎟⎠
1 p
M
2−p
ϕp fLp(G), (3.5)
whichwouldgive(3.2),andwhere wesetMϕ:= supt>0t
t∈R+
ϕ(t)≥s
dt.Wewill showthatT is ofweak-type(2,2)andofweak-type(1,1).Moreprecisely,weshow that
ν{t∈R+: |T f(t)| ≥y} ≤
M2fL2(G)
y
2
with norm M2= 1, (3.6) ν{t∈R+: |T f(t)| ≥y} ≤ M1fL1(G)
y with normM1=Mϕ, (3.7) whereν isdefinedin(3.3).Then(3.5) would followfrom(3.6) and(3.7) bytheMarcin- kiewiczinterpolationtheorem.Now,toshow(3.6),usingPlancherel’sidentity(2.25),we get
y2
t∈R+ μt(Rf)
ϕ(t) ≥y
ϕ2(t)dt≤ T f2L2(R+,ν)=
R+
μt(Rf) ϕ(t)
2
ϕ2(t)dt
=
R+
μ2t(Rf)dt=Rf2L2(V NR(G))=f2L2(G).
Thus,Tisofweak-type(2,2)withnormM2≤1.Further,weshowthatT isofweak-type (1,1)withnormM1=Mϕ;moreprecisely, weshowthat
t∈R+ μt(Rf)
ϕ(t) ≥y
ϕ2(t)dtMϕfL1(G)
y . (3.8)
FromthedefinitionoftheFouriertransformitfollowsthat
μt(Rf)≤ fL1(G). (3.9)
Indeed,fromtheDefinition2.8,wehave
μt(Rf)≤ RfL2(G)→L2(G).
TheYounginequalityforconvolution(e.g. [30,p.52,Proposition2.39])yields RfgL2(G)≤ fL1(G)gL2(G).
Thus
RfL2(G)→L2(G)≤ fL1(G).
Thisproves(3.9).Therefore, wehave y < μt(Rf)
ϕ(t) ≤fL1(G)
ϕ(t) .