Aunifyinginvestigationofinterior-pointmethodsforconvexprogrammingReport92-89D.denHertogF.JarreC.RoosT.TerlakyTechnischeUniversiteitDelftDelftUniversityofTechnologyFaculteitderTechnischeWiskundeenInformaticaFacultyofTechnicalMathematicsandInformaticsISSN0922-5641Copyrightc 1992bytheFacultyofTechnicalMathematicsandInformatics,Delft,TheNetherlands.NopartofthisJournalmaybereproducedinanyform,byprint,photoprint,microfilm,oranyothermeanswithoutpermissionfromtheFacultyofTechnicalMathematicsandInformatics,DelftUniversityofTechnology,TheNetherlands.CopiesofthesereportsmaybeobtainedfromthebureauoftheFacultyofTechnicalMathematicsandInformatics,Julianalaan132,2628BLDelft,phone+3115784568.AselectionofthesereportsisavailableinPostScriptformattheFaculty’sanonymousftp-site.Theyarelocatedinthedirectory/pub/publications/tech-reportsatftp.twi.tudelft.nlDELFTUNIVERSITYOFTECHNOLOGYREPORT92{89AUNIFYINGINVESTIGATIONOFINTERIORPOINTMETHODSFORCONVEXPROGRAMMINGD.denHertog,F.Jarre,C.Roos,T.TerlakyISSN0922{5641ReportsoftheFacultyofTechnicalMathematicsandInformatics92{89Delft1993iD.denHertog,C.RoosandT.Terlaky,FacultyofTechnicalMathematicsandInformatics,DelftUniversityofTechnology,P.O.Box5031,2600GADelft,TheNetherlands.e{mail:roos@dutiosa.twi.tudelft.nlterlaky@dutiosa.twi.tudelft.nlF.JarreInstitutf urAngewandteMathematikundStatistik,Universit atW urzburg,AmHubland,8700W urzburg,Germany.ThisworkiscompletedwiththesupportofaresearchgrantfromSHELL.ThefourthauthorisonleavefromtheE otv osUniversity,Budapest,andpartiallysupportedbyOTKANo.2116.Copyrightc 1993byFacultyofTechnicalMathematicsandInfor-matics,Delft,TheNetherlands.NopartofthisJournalmaybereproducedinanyform,byprint,photoprint,micro lmoranyothermeanswithoutwrittenpermis-sionfromFacultyofTechnicalMathematicsandInformatics,DelftUniversityofTechnology,TheNetherlands.iiAbstractIntherecentpastanumberofpaperswerewrittenthatpresentlowcomplexityinterior-pointmethodsfordi erentclassesofconvexprograms.Goalofthisarticleistoshowthatthelogarithmicbarrierfunctionassociatedwiththeseprogramsisself-concordant,andthattheanalysesofinterior-pointmethodsfortheseprogramscanthusbereducedtotheanalysisofinterior-pointmethodswithself-concordantbarrierfunctions.Keywords:interior-pointmethod,barrierfunction,dualgeometricprogramming,(extended)en-tropyprogramming,primalandduallp-programming,relativeLipschitzcondition,scaledLipschitzcondition,self-concordance.1IntroductionThee ciencyofabarriermethodforsolvingconvexprogramsstronglydependsonthepropertiesofthebarrierfunctionused.Akeypropertythatissu cienttoprovefastconvergenceforbarriermethodsisthepropertyofself-concordanceintroducedin[17].Thisconditionnotonlyallowsaproofofpolynomialconvergence,butnumericalexperimentsin[1,11,14]andothersfurtherindi-catethatnumericalalgorithmsbasedonself-concordantbarrierfunctionsareofpracticalinterestande ectivelyexploitthestructureoftheunderlyingproblem.Awell-knownbarrierfunctionforsolvingconvexprogramsisthelogarithmicbarrierfunction,introducedbyFrisch[4]andFiaccoandMcCormick[3].Todescribethelogarithmicbarrierfunctionmoreprecisely,wewill rstgiveageneralformfortheclassesofproblemsconsideredinthispaper:(CP)8:minf0(x)fi(x) 0;i=1; ;m;Ax=b;whereAisp nmatrixandbanp-dimensionalvector.Thelogarithmicbarrierfunctionforthisprogramisgivenby (x; )=f0(x) mXi=1ln( fi(x));where 0isthebarrierparameter.Weshowthatforseveralclassesofconvexproblemsforwhichinterior-pointmethodswerepresentedintheliteraturethelogarithmicbarrierfunctionisself-concordant.Theseclassesare:dualgeometricprogramming,(extended)entropyprogramming,primalandduallp-programming.Sincefordualgeometricprogrammingandduallp-programmingnocomplexityresultsareknownintheliterature,theseself-concordanceproofsenlargetheclassofproblemsforwhichpolynomialitycanbeproved.(In[12]onlyaconvergenceanalysisisgiven.)Moreover,weshowthatsomeothersmoothnessconditionsusedintheliterature(relativeLipschitzcondition[9,7],scaledLipschitzcondition[25,13],MonteiroandAdler’scondition[16])arealsocoveredbythisself-concordancecondition.Theseobservationsallowauni cationoftheanalysesofinterior-pointmethodsforanumberofconvexproblems.Thearticleisdividedinthreeparts.InSection2wegivethede nitionofself-concordanceandstatesomebasiclemmasaboutself-concordantfunctions.InSections3-6weproveself-concordancefortheclassesofproblemstreatedin[5,12,23],andinSection7weshowthatthesmoothnessconditionsusedin[7,9,13,16,25]implyself-concordanceofthebarrierfunction.2SomegeneralcompositionrulesLetus rstgivetheprecisede nitionofself-concordanceasgivenbyNesterovandNemirovsky[17]:De nitionofself-concordance:LetF0beanopenconvexsubsetofIRn.Afunction’:F0!IRiscalled -self-concordantonF0, 0,if’isthreetimescontinuouslydi erentiableinF0andifforallx2F0andh2IRnthefollowinginequalityholds:r3’(x)[h;h;h] 2 hTr2’(x)h 32;1wherer3’(x)[h;h;h]denotesthethirddi erentialof’atxandh.Intuitively,sincer3’describesthechangeinr2’,andsincer3’isboundedbyasuitablepowerofr2’,thisconditionimpliesthattherelativechangeofr2’isboundedby2 .Theassociatednormtomeasurethere