Consistency Analysis of Spectral Regularization Algorithms英文文献资料.docVIP

HindawiPublishingCorporation AbstractandAppliedAnalysis Volume2012,ArticleID436510,16pages doi:10.1155/2012/436510 ResearchArticle ConsistencyAnalysisof SpectralRegularization Algorithms YukuiZhuandHongweiSun ShandongProvincialKeyLaboratory ofNetworkbasedIntelligent Computing, SchoolofMathematic Science,University ofJinan,Jinan250022,China Correspondence shouldbeaddressedtoHongwei Sun,ss sunhw@ Received10January2012;Revised30March2012;Accepted 16April2012 Academic Editor:OndˇrejDoˇsly′ Copyright q2012Y.ZhuandH.Sun.Thisisanopenaccessarticledistributed undertheCreative Commons Attribution License, which permits unrestricted use,distribution, andreproduction in anymedium,providedtheoriginalworkisproperlycited. Weinvestigate theconsistency ofspectralregularization algorithms. Wegeneralize theusualde?- nitionofregularization functiontoenrichthecontentofspectralregularization algorithms. Under a more general prior condition, using re?ned error decompositions and techniques of operator normestimation, satisfactory errorboundsandlearningratesareproved. 1.Introduction Inthispaper,westudytheconsistencyanalysisofspectralregularizationalgorithmsinreg- ressionlearning. Y R.LetTheX,regressiondbeacompactlearningmetricaimsatspaceestimatingandρaorprobabilityapproximatingdistributiontheregressiononZXfunction×Ywith    fρx  ydρ y|x 1.1 Y through a set of samples z  {xi,yi}mi1 ∈ Zm drawn independently and identically ac- cordingtoρfromZ. Inlearningtheory,areproducingkernelHilbertspaceRKHSassociatedwithaMer- cerkernelKx,yisusuallytakenasthehypothesisspace.RecallthatafunctionK : X× X → RiscalledaMercerkernelifitiscontinuous,symmetric,andpositivesemide?nite. ThereproducingkernelHilbertspaceH isde?nedtobetheclosureofthelinearspanof K K :K·,x, x∈X.Thereproducingpropertytakestheform x  fxf,Kx K, ?f ∈HK, ?x∈X. 1.2  2 AbstractandAppliedAnalysis FortheMercerkernelKx,y,wedenotethat κmax Kx,x. 1.3 x∈X Our?rstcontributi