ebook img

The minus order and range additivity PDF

0.25 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview The minus order and range additivity

The minus order and range additivity MarkoS.Djikic´a,1,GuillerminaFongib,2,AlejandraMaestripierib,c,3,∗ aDepartmentofMathematics,FacultyofSciencesandMathematics,UniversityofNiš,Višegradska33,18000Niš, Serbia. bInstitutoArgentinodeMatemática“AlbertoP.Calderón"Saavedra15,Piso3(1083),BuenosAires,Argentina. cDepartamentodeMatemática,FacultaddeIngeniería,UniversidaddeBuenosAires. 7 1 0 2 n a Abstract J WestudytheminusorderonthealgebraofboundedlinearoperatorsonaHilbertspace. By 2 givingacharacterizationintermsofrangeadditivity,weshowthattheintrinsicnatureofthe ] A minusorderisalgebraic. Applicationstogeneralizedinversesofthesumoftwooperators,to F systemsofoperatorequationsandtooptimizationproblemsarealsopresented. . h Keywords: minusorder,rangeadditivity,generalizedinverses,leastsquaresproblems t a 2000MSC:06A06,47A05 m [ 1 1. Introduction v 6 TheminusorderwasintroducedbyHartwig[25]andindependentlybyNambooripad[32], 7 4 inbothcasesonsemigroups,withtheideaofgeneralizingsomeclassicalpartialorders.Itwas 0 extendedtooperatorsininfinitedimensionalspacesindependentlybyAntezana,Corachand 0 Stojanoff[2]andbySˇemrl[36]. Thereisnowanextensiveliteraturedevotedtothisorderand . 1 other related partial orders on matrices, operators and elements of various algebraic struc- 0 7 tures. Seeforexample,[8,30,31]. 1 Themaingoalofthisworkistoobtainanewcharacterizationoftheminusorderforop- : v eratorsactingonHilbertspacesintermsofthesocalledrangeadditivityproperty. Giventwo i X linearboundedoperatorsA and B actingonaHilbertspace(cid:72),wesaythatA and B havethe r rangeadditivitypropertyifR(A+B)=R(A)+R(B),whereR(T)standsfortherangeofanop- a eratorT. Operatorswiththispropertyhavebeenstudiedin[4]and[5](seealso[10]). Recall thatifA andB aretwoboundedlinearHilbertspaceoperatorsthenA≤− B (wherethesymbol “≤−”standsfortheminusorderofoperators)ifandonlyifthereareobliqueprojectionsP and Q suchthatA =PB andA∗=QB∗. Inthispaper,weprovethatthisisequivalenttotherange ∗Correspondingauthor Emailaddresses:[email protected](MarkoS.Djikic´),[email protected] (GuillerminaFongi),[email protected](AlejandraMaestripieri) 1SupportedbyGrantNo.174007oftheMinistryofEducation,ScienceandTechnologicalDevelopmentofthe RepublicofSerbia 2PartiallysupportedbyCONICET(PIP4262013-2015) 3PartiallysupportedbyCONICET(PIP1682014-2016) PreprintsubmittedtoElsevier January3,2017 ofB beingthedirectsumofrangesofA andB−A andtherangeofB∗beingthedirectsumof rangesofA∗ and B∗−A∗. Thustheminusorderisintrinsicallyalgebraicinnature. Thisplays anequivalentroletoaknowncharacterizationwhenA andB arematrices[25,31];thatA≤− B ifandonlyiftherankof B−A isthedifferenceoftherankof B andtherankofA. As a consequence, diverse concepts that have been developed for matrices and opera- tors are in fact manifestations of the minus order. These include, for example, the notions ofweaklybicomplementarymatricesdefinedduetoWerner[37],andquasidirectadditionof operatorsdefinedbyLešnjakandŠemrl[28]. Althoughinthesepaperstheminusorderdoes notappearexplicitly,thesenotionswhenappliedtooperatorsA and B areequivalenttosay- ing that A ≤− A+B. The minus order also lurks in the papers of Baksalary and Trenkler [9], Baksalary,SˇemrlandStyan[7],Mitra[29]andArias,CorachandMaestripieri[5]. The minus order can be weakened to what we call left and right minus orders. As with theminusorder,theseordersareeasilyderivedfromarangeadditivitycondition. Ithappens that they truly differ from the minus order only in the infinite dimensional setting. When A≤− B,wegivesomeapplicationstoformulasforgeneralizedinversesofsumsA+B interms ofgeneralizedinversesofAandB,andweshowthatcertainoptimizationproblemsinvolving theoperatorA+B canbedecoupledintoasystemofsimilarproblemsforA and B. Thepaperisorganizedasfollows.InSection2wecollectsomeusefulknownresultsabout rangeadditivity,whileinSection3,theminusorderisdefinedandtheconnectionwithrange additivity is made. Motivated by the concepts of the left and the right star orders, we define leftandtherightminusorderson L((cid:72)). Formatrices,theseareequivalenttotheminusor- der,withdifferencesonlyemergeintheinfinitedimensionalcontext. Proposition3.13char- acterizes the left minus order in terms of densely defined, though not necessarily bounded, projections. Additionally,theleftminus,therightminusandtheminusordersarecharacter- ized in terms of (densely defined) inner generalized inverses, generalizing a matricial result (see[30]). Finally,Section4isdevotedtoapplications.Webeginbyrelatingtheminuspartialorderto someformulasforreflexiveinnerinversesofthesumoftwooperators. Inparticular,wegive analternativeprooffortheFill-FishkindformulafortheMoore-Penroseinverseofasum,as found in [21] for matrices and extended to L((cid:72)) by Arias et al. [5]. We also apply the new characterizationoftheminusordertosystemsofequationsandleastsquaresproblems. We include a final remark about a possible generalization of the minus order involving densely definedprojectionswithclosedrange. 2. Preliminaries Throughout, ((cid:72),〈·,·〉) denotes a complex Hilbert space and L((cid:72)) the algebra of linear boundedoperatorson(cid:72),(cid:81)isthesubsetofL((cid:72))ofobliqueprojections,i.e.,(cid:81)={Q∈L((cid:72)): Q2=Q}and(cid:80) thesubsetof(cid:81) oforthogonalprojections,i.e.,(cid:80) ={P∈L((cid:72)):P2=P=P∗}. Given(cid:77) and(cid:78) twoclosedsubspacesof(cid:72),write(cid:77) +˙ (cid:78) forthedirectsumof(cid:77) and (cid:78) ,(cid:77) ⊕(cid:78) theorthogonalsumand(cid:77) (cid:9)(cid:78) =(cid:77) ∩((cid:77) ∩(cid:78) )⊥. If(cid:77) +˙ (cid:78) =(cid:72),theoblique projection with range (cid:77) and null space (cid:78) is P(cid:77)//(cid:78) and P(cid:77) = P(cid:77)//(cid:77)⊥ is the orthogonal projectiononto(cid:77). 2 ForA ∈L((cid:72)),R(A)standsfortherangeofA,N(A)foritsnullspaceandP forP . The A R(A) Moore-PenroseinverseofA isthe(denselydefined)operatorA†:R(A)⊕R(A)⊥→(cid:72),defined byA†|R(A)=(A|N(A)⊥)−1 andN(A†)=R(A)⊥. ItholdsthatA†∈L((cid:72))ifandonlyifA hasaclosed range. Given (cid:77) and (cid:78) two closed subspaces of (cid:72), the minimal angle between (cid:77) and (cid:78) is α ((cid:77),(cid:78) )∈[0,π/2],thecosineofwhichis 0 c ((cid:77),(cid:78) )=sup(cid:8)|(cid:10)ξ,η(cid:11)|: ξ∈(cid:77), (cid:107)ξ(cid:107)≤1, η∈(cid:78) , (cid:107)η(cid:107)≤1(cid:9) ∈[0,1]. 0 Whentheminimalanglebetween(cid:77) and(cid:78) isstrictlylessthat1,thenthesum(cid:77) +(cid:78) is closedanddirect,moreover,wehavethefollowing. Proposition2.1. Let (cid:77) and (cid:78) be two closed subspaces of (cid:72). The following statements are equivalent: (1) c ((cid:77),(cid:78) )<1; 0 (2) (cid:77) +˙ (cid:78) isclosed; (3) (cid:72) =(cid:77)⊥+(cid:78) ⊥. Foraproof,seeLemma2.11andTheorem2.12in[15]. ForA,B ∈L((cid:72)),italwaysholdsthatR(A+B)⊆R(A)+R(B). WesaythatA andB havethe rangeadditivitypropertyifR(A+B)=R(A)+R(B).Inthiscase,R(A)⊆R(A+B).Conversely,if R(A)⊆R(A+B)then,forx ∈(cid:72),Bx =(A+B)x−Ax ∈R(A+B).Wehaveprovedthefollowing. Lemma2.2([5,Proposition2.4]). ForA,B ∈L((cid:72)),R(A+B)=R(A)+R(B)ifandonlyifR(A)⊆ R(A+B). Operators having the range additivity property were characterized in [5, Theorem 2.10]. CloselyrelatedisthefollowingforoperatorsA,B ∈L((cid:72))satisfyingtheconditionR(A)∩R(B)= {0}. Proposition2.3([5, Theorem 2.10]). Consider A,B ∈ L((cid:72)) such that R(A)∩R(B)={0} then R(A+B)=R(A)+˙ R(B)ifandonlyif(cid:72) =N(A)+N(B). ThenextresultwillbeusefulincharacterizingtheminusorderinSection3(see[5,Propo- sition2.2]). Proposition2.4. ForA,B ∈L((cid:72))considerthefollowingstatements: (1) R(A∗)+˙ R(B∗)isclosed; (2) thereexistsQ∈(cid:81) suchthatA∗=Q(A∗+B∗); (3) N(A)+N(B)=(cid:72); 3 (4) R(A+B)=R(A)+R(B). Then(1)⇔(2)⇔(3)⇒(4). Theimplication(4)⇒(3)holdsifR(A)∩R(B)={0}. Foraproofof(1)⇔(2)see[2, Proposition4.13]. (1)⇔(3)wasstatedinProposition2.1. Theimplication(3)⇒(4)followsfromtheproofof[3,Proposition2.8]. (4)⇒(3)followsfrom Proposition2.3sinceR(A)∩R(B)={0}. 3. Theminusorder Differentdefinitionshavebeengivenfortheminus(partial)order. Foroperatorsweoffer onewhichequivalenttothoseappearingin[2]and[36]. Definition3.1. ForA,B ∈L((cid:72)),A≤− B ifthereexistP,Q∈(cid:81) suchthatA=PB andA∗=QB∗. Proofsthat≤− isapartialorderonL((cid:72))canbefoundin[2,Corollary4.14]and[36,Corol- lary 3]. It is easy to see that the ranges of P and Q can be fixed so that R(P) = R(A) and R(Q)=R(A∗). Fordetails,see[2,Proposition4.13]andthedefinitionofminusorderin[36]. Inthenextpropositionwecollectsomecharacterizationsoftheminusorderintermsof angleconditionsandsumofclosedsubspaces. Proposition3.2. ConsiderA,B ∈L((cid:72)). Thefollowingstatementsareequivalent: (1) A≤− B; (2) c (R(A),R(B−A))<1andc (R(A∗),R(B∗−A∗))<1; 0 0 (3) R(B)=R(A)+˙ R(B−A)andR(B∗)=R(A∗)+˙ R(B∗−A∗); (4) N(A)+N(B−A)=N(A∗)+N(B∗−A∗)=(cid:72); (5) thereexistsP∈(cid:81) suchthatA=PB andR(A)⊆R(B). Proof. The equivalences (1) ⇔ (2) ⇔ (4) follow applying the definition of the minus order andProposition2.1totheoperatorsA, B−A,A∗and B∗−A∗,seealso[2,Proposition4.13]. For (2) ⇔ (3), suppose that c (R(A),R(B−A)) < 1 and c (R(A∗),R(B∗−A∗)) < 1. Then 0 0 R(A) +˙ R(B−A) and R(A∗) +˙ R(B∗−A∗) are closed. In this case, R(B) ⊆ R(A) +˙ R(B−A). On the other hand, applying Proposition 2.4, there existsQ ∈ (cid:81) such that A∗ =QB∗. Then N(B∗) ⊆ N(A∗) and N(B∗) ⊆ N(B∗−A∗), or R(A) ⊆ R(B) and R(B−A) ⊆ R(B). Then R(B) = R(A)+˙ R(B−A). Similarly,R(B∗)=R(A∗)+˙ R(B∗−A∗). Seealso[36,Theorem2]. Conversely, if item 3 holds, then R(A) +˙ R(B−A) and R(A∗) +˙ R(B∗−A∗) are closed or equivalently, by Proposition2.1,item2holds. Next consider (1)⇔(5). If A ≤− B then A =PB = BQ∗ with P,Q ∈(cid:81), so that A =PB and R(A)⊆R(B). Conversely,supposeR(A)⊆R(B)andthereexistsP∈(cid:81) suchthatA=PB. Then byLemma2.2itholdsthatR(B)=R(A)+R(B −A). MoreoverR(A)∩R(B −A)={0}because R(A) ⊆ R(P) and R(B −A) ⊆ N(P), so that R(B) = R(A) +˙ R(B −A). In this case, (4) ⇒ (2) of Proposition 2.4 can be applied so that there existsQ ∈ (cid:81) such that A∗ =QB∗. Therefore A≤− B. 4 The following is a key result that will be useful on many occasions throughout the pa- per. Itgivesanewcharacterizationoftheminuspartialorderintermsoftherangeadditivity property,showingthattheminusorderhasanalgebraicnature. Theorem3.3. ConsiderA,B ∈L((cid:72)). Thenthefollowingassertionsareequivalent: (1) A≤− B; (2) R(B)=R(A)+˙ R(B−A)andR(B∗)=R(A∗)+˙ R(B∗−A∗). Proof. Suppose that A ≤− B. By Proposition 3.2, it follows that R(A) +˙ R(B−A) and R(A∗) +˙ R((B−A)∗)areclosed. Inparticular,R(A)∩R(B−A)=R(A∗)∩R(B∗−A∗)={0}. Also,itfollows fromProposition2.4thatR(B∗)=R(A∗)+R(B∗−A∗)andR(B)=R(A)+R(B −A). Therefore, R(B)=R(A)+˙ R(B−A)andR(B∗)=R(A∗)+˙ R(B∗−A∗). Conversely,supposethatR(B)=R(A)+˙ R(B−A)andR(B∗)=R(A∗)+˙ R(B∗−A∗).Applying (4)⇒(1)inProposition2.4,itfollowsthatR(A)+˙ R(B−A)andR(A∗)+˙ R(B∗−A∗)areclosed. Hence,byProposition3.2andProposition2.1,A≤− B. Let A ∈ L((cid:72)) for 1 ≤ i ≤ k. Lešnjak and Šemrl [28] give the following definition: the i k operatorA=(cid:80)A isthequasidirectsumiftherangeofA isthedirectsumoftherangesofthe i i=1 A sandtheclosureoftherangeofA isthedirectsumoftheclosuresoftherangesoftheA s. i i Thenextresultmayberestatedassayingthat B isthequasidirectsumofA and B−A ifand onlyifA≤− B. Corollary3.4. IfA,B ∈L((cid:72)),thefollowingconditionsareequivalent: (1) A≤− B; (2) R(B)=R(A)+˙ R(B−A)andR(B)=R(A)+˙ R(B−A). Proof. (1)⇒(2)followsfromProposition3.2andTheorem3.3. Fortheconverse,sinceR(A)+˙ R(B−A)isclosed,fromProposition2.4wehavethatR(B∗)= R(A∗)+R(B∗−A∗). Toseethatthissumisdirect,applyingProposition2.4againandusingthe fact that R(B)=R(A)+˙ R(B −A) we get that R(A∗)∩R(B∗−A∗)={0}. Thus R(B∗)=R(A∗)+˙ R(B∗−A∗),andsobyTheorem3.3,A≤− B. The next result shows the behavior of the minus order when the operators have closed ranges. Corollary3.5. ConsiderA,B ∈L((cid:72))suchthatA ≤− B. ThenR(B)isclosedifandonlyifR(A) andR(B−A)areclosed. Proof. If A ≤− B, then by Corollary 3.4, R(B) = R(A) +˙ R(B−A) and R(B) = R(A) +˙ R(B−A). IfR(B)isclosedthenR(A)+˙ R(B−A)=R(A)+˙ R(B−A). HenceR(A)=R(A)andR(B−A)= R(B −A). Infact, givenx ∈R(A), thenx ∈R(A)+˙ R(B −A), sothatthereexistx ∈R(A)and 1 x ∈R(B−A)suchthatx =x +x . Butx −x =x ∈R(A)∩R(B)={0},andsox =x ∈R(A); 2 1 2 1 2 1 thatis,R(A)=R(A). Similarly,R(B−A)=R(B−A). TheconversefollowsbyCorollary3.4. 5 3.1. Theleftandrightminusorders Inthissectionwedefinetheleft andrightminusordersandshowthattheyareageneral- izationoftheleftandrightstarorders. Aswewillsee,theseordersarereallyonlyinteresting oninfinitedimensionalspaces. Formatrices,theycoincidewiththeminusorder. Webeginanalyzingthepropertiesoftheleftandrightstarorders. Originally,Drazin[20] introduced the star order on semigroups with involutions, Baksalary and Mitra [6] defined the left and right star orders for complex matrices, and later, Antezana, Cano, Mosconi and Stojanoff[1]extendedthestarordertothealgebraofboundedoperatorsonaHilbertspace. SeealsoDolinarandMarovt[18],DengandWang[14]andDjikic´ [16]. Given A,B ∈ L((cid:72)), the star order, left star order and right star order are respectively de- finedby ∗ • A≤B ifandonlyifA∗A=A∗B andAA∗=BA∗, • A∗≤B ifandonlyifA∗A=A∗B andR(A)⊆R(B),and • A≤∗B ifandonlyifAA∗=BA∗andR(A∗)⊆R(B∗). ∗ IfA,B ∈L((cid:72)),thenA≤B ifandonlyifthereexistP,Q∈(cid:80) suchthatA=PB andA∗=QB∗ (see[1,Proposition2.3]or[18,Theorem5]). WecanalwaystakeP=PA andQ=PA∗. Thenextresultisastraightforwardconsequenceof[14,Theorem2.1].Weincludeasimple proof. Proposition3.6. LetA,B ∈L((cid:72)). IfA∗≤B thenA≤− B. Proof. If A ∗≤ B , then A∗A = A∗B, or equivalently A∗(A −B) = 0. Hence P (A −B) = 0, or A A = P B. Conversely, if A = P B, then A∗A = A∗B. So A ∗≤ B is equivalent to A = P B and A A A R(A)⊆R(B). ByProposition3.2(5),thisgivesA≤− B. Thefollowingresultscharacterizetheleftandrightstarordersintermsofanorthogonal rangeadditivityproperty. Proposition3.7. ForA,B ∈L((cid:72)),A∗≤B ifandonlyifR(B)=R(A)⊕R(B−A). Proof. FromtheproofofProposition3.6,A∗≤B ifandonlyifA=P B andR(A)⊆R(B). Thus A R(B)=R(A)⊕R(B−A)sinceR(B−A)⊆N(P )=R(A)⊥. A Conversely, if R(B) = R(A) ⊕ R(B −A), then R(A) ⊆ R(B) and R(B −A) ⊆ R(A)⊥, so that A=P B. HenceA∗≤B. A Corollary3.8. ForA,B ∈L((cid:72)),A≤∗B ifandonlyifR(B∗)=R(A∗)⊕R(B∗−A∗). The next characterization of the star order follows from the previous results (or alterna- tively,fromTheorem3.3). Corollary3.9. GivenA,B ∈L((cid:72)),thefollowingstatementsareequivalent: 6 ∗ (1) A≤B; (2) A∗≤B andA≤∗B; (3) R(B)=R(A)⊕R(B−A)andR(B∗)=R(A∗)⊕R(B∗−A∗). ∗ Proof. Obviously, A ∗≤ B and A ≤∗ B. On the other hand, if A ≤ B, then by the proof of Proposition3.6,A =PAB andA∗ =PA∗B∗. HenceR(A∗)⊆R(B∗)andR(A)⊆R(B). Thus(1)⇔ (2). Theequivalenceoftheseto(3)followsfromProposition3.7. Asageneralizationoftheleftandrightstarorders,wenowdefinetheleft andrightminus orders. Definition3.10. ForA,B ∈L((cid:72)), • A−≤B ifandonlyifR(B)=R(A)+˙ R(B−A),and • A≤−B ifandonlyifR(B∗)=R(A∗)+˙ R(B∗−A∗). Proposition3.11. Therelations−≤and≤−definepartialorders. Proof. Weonlygivetheprooffor−≤,sincetheprooffor≤−isidentical. First of all, −≤ is clearly reflexive. So consider A,B ∈ L((cid:72)) such that A −≤ B and B −≤A. Then R(B) = R(A) +˙ R(B −A) and R(A) = R(B) +˙ R(B −A). From the last equality R(B − A) ⊆ R(A). But R(B −A)∩R(A) = {0}, so that R(B −A) = {0}. Therefore A = B thus −≤ is antisymmetric. To prove −≤ is transitive, consider A,B,C ∈ L((cid:72)) such that A −≤ B and B −≤ C. Then R(B)=R(A)+˙ R(B−A)andR(C)=R(B)+˙ R(C−B). SinceR(A)⊆R(B)⊆R(C),byLemma2.2, R(C)=R(A)+R(C −A). ItremainstoshowthatR(A)∩R(C −A)={0}. SinceR(A)∩R(C −A)⊆ R(A)∩(R(C−B)+R(B−A))wecanwritex ∈R(A)asx =x +x ,x ∈R(C−B)andx ∈R(B−A). 1 2 1 2 Thenx −x =x ∈R(B)∩R(C −B)={0},andsox =x . Hencex ∈R(A)∩R(B−A)={0};that 2 1 2 isx =0andR(A)∩R(C −A)={0}. ThisimpliesthatR(C)=R(A)+˙ R(C −A),orequivalently, A−≤C,andso−≤istransitive. ThenextcorollaryisaconsequenceofTheorem3.3. Corollary3.12. ForA,B ∈L((cid:72)),A−≤B andA≤−B ifandonlyifA≤− B. It follows from Proposition 2.4 that A −≤ B if and only if A∗ =QB∗ forQ ∈ (cid:81) and R(A)∩ R(B−A)={0}.Thereisalsoacharacterizationoftheleftminusordersimilartothatoftheleft starorderasfoundintheproofofProposition3.6. Weleavetheobviousversionfortheright minusorderunstated. Proposition3.13. For A,B ∈ L((cid:72)), A −≤ B if and only if there exists a (possibly unbounded) denselydefinedprojectionP suchthatA=PB andR(A)⊆R(B). 7 Proof. IfA−≤B thenR(B)=R(A)+˙ R(B−A)sothatR(A)⊆R(B).DefineP=PR(A)//R(B−A)⊕N(B∗). Then P is a densely defined projection and it is easy to check that A = PB. Conversely, if A = PB for a densely defined projection and R(A) ⊆ R(B) then R(B) = R(A)+R(B −A) by Lemma2.2,andthesumisdirectsinceR(A)⊆R(P)andR(B−A)⊆N(P). Remark 3.14. Theminusordercanbeseenasastarorderafterapplyingsuitableweightsto theHilbertspacesinvolved. Recallthat,ifA,B ∈L((cid:72),(cid:75))aresuchthatA≤− B thenthereexist projectionsP ∈L((cid:75))andQ ∈L((cid:72))suchthatA =PB = BQ. TheoperatorsW =Q∗Q+(I − 1 Q∗)(I−Q)∈L((cid:72))andW =P∗P+(I−P∗)(I−P)∈L((cid:75))arepositiveandinvertible. Hencethe 2 innerproductsin(cid:72) and(cid:75) respectively, (cid:10)x,y(cid:11) =(cid:10)W x,y(cid:11), forx,y ∈(cid:72) and 〈z,w〉 =〈W z,w〉, forz,w ∈(cid:75) W1 1 W2 2 give rise to equivalent norms. With these new inner products, the projections P andQ are ∗ orthogonalin(cid:75) =((cid:75),〈·,·〉 )and(cid:72) =((cid:72),〈·,·〉 ),respectively,andsoA≤B. W2 W2 W1 W1 On the other hand, A −≤ B if and only if there exists a densely defined projection P such thatA=PB andR(A)⊆R(B).Inthiscase,itispossibletofindapositiveandinvertibleweight W on(cid:75) suchthatP issymmetricwithrespectto〈·,·〉 (orequivalentlyA∗≤B inL((cid:72),(cid:75) )) ifa2ndonlyifP admitsaboundedextensionP˜∈(cid:81) (orWe2quivalentlyA≤− B). W2 Hereisaproofofthelaststatement: supposethatthereexistsaweightW on(cid:75) positive 2 andinvertiblesuchthatP issymmetricwithrespectto〈·,·〉 . SinceP isa(denselydefined) W2 idempotent then (cid:68)(P) = R(P) +˙ N(P), where (cid:68)(P) is the domain of P. Moreover, given x ∈ R(P)andy ∈N(P)wehave(cid:10)x,y(cid:11) =(cid:10)Px,y(cid:11) =(cid:10)x,Py(cid:11) =0becauseP issymmetricwith W2 W2 W2 respectto〈·,·〉 andy ∈N(P). Hence(cid:68)(P)=R(P)⊕ N(P),andconsequently(cid:72) =R(P)⊕ W2 W2 W2 N(P), where the closures are taken with respect to 〈·,·〉 . Then P = P is a bounded W2 R(P)//N(P) extensionofP. Conversely,supposethatthereexistsP˜∈(cid:81)suchthatP⊆P˜,andletW =P˜∗P˜+(I−P˜)∗(I− 2 P˜), which is positive and invertible and satisfies W P˜ = P˜∗W . Finally, P is symmetric with 2 2 respectto〈·,·〉 . Infact,ifx,y ∈(cid:68)(P)then(cid:10)Px,y(cid:11) =(cid:172)P˜x,y(cid:182) =(cid:172)W P˜x,y(cid:182)=(cid:172)x,W P˜y(cid:182)= (cid:172)W x,P˜y(cid:182)=(cid:172)Wx2,P˜y(cid:182) =(cid:10)x,Py(cid:11) . W2 W2 2 2 2 W2 W2 Corollary3.15. LetA,B ∈L((cid:72))besuchthatA−≤B. ThenR(B∗)=R(A∗)+˙ R(B∗−A∗). Proof. FromProposition3.13(3),ifA−≤B,thenA=PB andN(B)⊆N(A)orR(A∗)⊆R(B∗)and inthesameway,R(B∗−A∗)⊆R(B∗). ThenbyProposition2.4(1),R(A∗)+˙ R(B∗−A∗)⊆R(B∗). Ontheotherhand,R(B∗)⊆R(A∗)+R(B∗−A∗)⊆R(A∗)+˙ R(B∗−A∗). Hence R(B∗)⊆R(A∗)+˙ R(B∗−A∗)⊆R(B∗). ButR(A∗)+˙ R(B∗−A∗)isclosedbyProposition2.4. Therefore,R(B∗)=R(A∗)+˙ R(B∗−A∗). Corollary3.16. LetA,B ∈L((cid:72))suchthatA−≤B. IfR(B)isclosedthenR(A)andR(B−A)are closedandA≤− B. 8 Proof. Since A −≤ B, by Corollary 3.15, R(B∗) = R(A∗) +˙ R(B∗−A∗). If R(B) is closed, then R(B∗)isclosedand R(A∗)+˙ R(B∗−A∗)=R(B∗)=R(B∗)⊆R(A∗)+˙ R(B∗−A∗). ThereforeR(A∗)+˙ R(B∗−A∗)=R(A∗)+˙ R(B∗−A∗).ThisimpliesthatR(A∗)=R(A∗)andR(B∗−A∗)= R(B∗−A∗)andA≤− B. Theabovecorollaryshowsthat,unliketheleft(right)starorder,theleft(right)minusorder coincideswiththeminusorderwhenappliedtomatrices.Howeverforoperatorstheseorders arenotthesame. Example3.17(Seealso[7]). LetA∈L((cid:72))beanoperatorsuchthatR(A)(cid:54)=R(A)andthatthere existsx ∈R(A)\R(A)whichisnotorthogonaltoN(A). Forexample,consider(cid:72) =l2((cid:78))the spaceofallsquare-summablesequences,operatorA definedasA :(xn)n∈(cid:78)(cid:55)→((1/n)xn+1)n∈(cid:78), and take x to be x = (1/n)n∈(cid:78). Define operator B as B = A+Px, where Px is the orthogonal projection onto the one-dimensional subspace spanned by {x}. Since N(A) (cid:54)⊆ N(P ), and x N(P )isofco-dimensionone,wehave(cid:72) =N(A)+N(P ),whichaccordingtoProposition2.4 x x showsthatA andP arerange-additive; thatis,R(A)+R(B −A)=R(B). WealsohaveR(A)∩ x R(B −A)={0}showingthatA −≤ B. Ontheotherhand,R(A)∩R(B−A)(cid:54)={0}soA ≤− B does nothold. ApplyingTheorem3.3itispossibletodefinetheminusorderintermsoftheinnergener- alizedinversesoftheoperatorsinvolved. ByaninnerinverseofanoperatorA ∈L((cid:72),(cid:75))we meanadenselydefinedoperatorA−:(cid:68)(A−)⊆(cid:75) →(cid:72) satisfyingR(A)⊆(cid:68)(A−)andAA−A=A. Proposition3.18. ForA,B ∈L((cid:72)),thefollowingconditionsareequivalent: (1) A−≤B; (2) there exists an inner inverse A− of A such that A−A =A−B and AA−x = BA−x for every x ∈(cid:68)(A−). Proof. SupposethatA −≤ B. If(cid:78) isacomplementofR(B),then(cid:72) =R(A)+˙ R(B−A)+˙ (cid:78) . From Proposition 2.4 we know that N(A)+N(B −A)=(cid:72); so if (cid:77) =N(B −A)(cid:9)N(A), then (cid:72) = N(A) +˙ (cid:77). Let A be the restriction of A to (cid:77), and define A− as A−1 on R(A), and 1 1 as the null operator on R(B −A) +˙ (cid:78) . Then A− is densely defined and the domain of A− is (cid:68)(A−)=R(B)+˙ (cid:78) .Inthiscase,(A−B)A−x =0foreveryx ∈(cid:68)(A−),becauseR(A−)⊆N(A−B). Ontheotherhand,sinceR(A−B)⊆(cid:68)(A−),wefindthatA−(A−B)=0sinceR(A−B)⊆N(A−). Fortheconverse,supposethatthereexistsaninnerinverseA− ofA suchthatA−A =A−B and AA−x = BA−x for every x ∈ (cid:68)(A−). In particular, if z ∈ (cid:72) then Az ∈ R(A) ⊆ D(A−), so thatAz =AA−Az =BA−Az. HenceR(A)⊆R(B),showingthatR(B)=R(A)+R(B−A). From A−A=A−B wehaveR(A−B)⊆N(A−),whileN(A−)∩R(A)={0},andsoR(B)=R(A)+˙ R(B−A). Therefore,A−≤B. Corollary3.19. ForA,B ∈L((cid:72)),thefollowingconditionsareequivalent: 9 (1) A≤− B; (2) thereexistinnerinversesA−ofA and(A∗)−ofA∗suchthat (i) A−A=A−B andAA−x =BA−x foreveryx ∈(cid:68)(A−), (ii) (A∗)−A∗=(A∗)−B∗andA∗(A∗)−x =B∗(A∗)−x foreveryx ∈(cid:68)((A∗)−). 4. Applications 4.1. GeneralizedinversesofA+B InthissectionwestatetheformulasforarbitraryreflexiveinversesofA+B intermsofthe inversesofAandB,whenA−≤A+B.Forthesakeofsimplicity,webeginbygivingtheformula for the Moore-Penrose inverse. Theorem 4.7 states the result in the most general form, and fromthistheoremmanyexistingresultsinthesubjectcanberecovered. IfA ≤− A+B thenA =P(A+B)forsomeP ∈(cid:81). UsingtheprojectionP wecanconstruct a projection E ∈ (cid:81) onto R(A+B) that will be useful in stating the formula for the Moore- PenroseinverseofA+B. Lemma4.1. LetA,B ∈L((cid:72))besuchthatA≤− A+B,andP∈(cid:81) besuchthatA=P(A+B). Set E =P P+P (I −P). A B Then E ∈ (cid:81) and R(E) = R(A+B). Moreover, E is selfadjoint if and only if P = P(cid:77)//(cid:78) where (cid:77) =R(A)⊕(cid:77) ,(cid:78) =R(B)⊕(cid:78) with(cid:77) and(cid:78) closedsubspacessuchthat(cid:77) ,(cid:78) ⊆N(A∗)∩ 1 1 1 1 1 1 N(B∗). Proof. If A = P(A+B) then R(A) ⊆ R(P) and R(B) ⊆ N(P). Therefore P P and P (I −P) are A B projections,withR(P P)=R(A)andR(P (I −P))=R(B). Moreover, A B P PP (I −P)=P (I −P)P P=0. A B B A Therefore E =P P+P (I −P)isaprojection. Also,R(A)=R(P P)=R(EP)⊆R(E). Applying A B A Lemma2.2,R(E)=R(A)+R(B)=R(A+B)becauseA≤− A+B. Finally,ifP =P(cid:77)//(cid:78) thenthereexistclosedsubspaces(cid:77)1,(cid:78)1 suchthat(cid:77) =R(A)⊕(cid:77)1 and (cid:78) = R(B) ⊕ (cid:78) . Hence P P = P , P (I −P) = P and E = 1 A R(A)//R(B)+˙(cid:78)1+˙(cid:77)1 B R(B)//R(A)+˙(cid:77)1+˙(cid:78)1 P . SinceA ≤− A+B,itfollowsthat E∗=E ifandonlyif(cid:77) +˙ (cid:78) =(R(A+B))⊥= R(A+B)//(cid:77)1+˙(cid:78)1 1 1 (R(A)+˙ R(B))⊥=N(A∗)∩N(B∗),orequivalently,(cid:77) and(cid:78) areincludedinN(A∗)∩N(B∗). 1 1 Definition4.2. LetA,B ∈L((cid:72))suchthatA ≤− A+B. ConsiderP,Q ∈(cid:81) suchthatA =P(A+ B)=(A+B)Q,thenP willbecalledoptimalforA andB ifE =P P+P (I−P)isselfadjoint. In A B asymmetricway,sinceA∗=Q∗(A∗+B∗),Q willbecalledoptimalforA and B ifQ∗ isoptimal forA∗and B∗,i.e.,F =PA∗Q∗+PB∗(I −Q∗)isselfadjoint. 10

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.