摘要
AbriefsurveyofformerandrecentresultsonHuber'sminimaxapproachinrobuststatisticsisgiven.TheleastinformativedistributionsminimizingFisherinformationforlocationoverseveraldistributionclasseswithupper-boundedvariancesandsubrangesarewrittendown.TheseleastinformativedistributionsarequalitativelydifferentfromclassicalHuber'ssolutionandhavethefollowingcommonstructure:(i)withrelativelysmallvariancestheyareshort-tailed,inparticularnormal;(ii)withrelativelylargevariancestheyareheavytailed,inparticulartheLaplace;(iii)theyarecompromisewithrelativelymoderatevariances.TheseresultsallowtoraisetheefficiencyofminimaxrobustproceduresretaininghighstabilityascomparedtoclassicalHuber'sprocedureforcontaminatednormalpopulations.Inapplicationtosignaldetectionproblems,theproposedminimaxdetectionrulehasprovedtoberobustandclosetoHuber'sforheavy-taileddistributionsandmoreefficientthanHuber'sforshort-tailedonesbothinasymptoticsandonfinitesamples。
出版日期
2005年02月12日(中国期刊网平台首次上网日期,不代表论文的发表时间)