简介:In1980'sdifferentialgeometricmethodsaresuccessfullyusedtostudycurvedexpomentialfamiliesandnormalnonlinearregressionmodels.Thispaperpresentsanewgeometricstructuretostudymultinomialdistributionmodelswhichcontainasetofnonlinearparameters.Basedonthisgeometricstructure,thesuthorsstudyseveralasymptoticpropertiesforsequentialestimation.Thebias,thevarianceandtheinformationlossofthesequentialestimatesaregivenfromgeomentricviewpoint,andalimittheoremconnectedwiththeobservedandexpectedFisherinformationisobtainedintermsofcurvatvremeasures.Theresultsshowthatthesequentialestimationprocedncehassomebetterpropertieswhicharegenerallyimpossiblefornonsequentialestimationprocedures.