Seite - 255 - in Differential Geometrical Theory of Statistics

Entropy2016,18, 277 algorithm,andextendstheEMalgorithm.Ourconvergenceproofdemandssomeregularity(continuity anddifferentiability)of theestimateddivergencewithrespect to theparametervectorφ)which isnot simplycheckedusing(2). Recentresults in thebookofRockafellarandWets [13]providesufficient conditions to prove continuity anddifferentiability of supremal functions of the formof (2)with respect toφ. Differentiabilitywith respect toφ still remainsaveryhard task; therefore, our results covercaseswhentheobjective function isnotdifferentiable. Thepaperisorganizedasfollows: inSection2,wepresentthegeneralcontext.Wealsopresentthe derivationofouralgorithmfromtheEMalgorithmandpassingbyTseng’sgeneralization. InSection3, wepresent someconvergenceproperties.Wediscuss inSection4avariantof thealgorithmwitha theoreticalglobal infimum,andanexampleof the two-Gaussianmixturemodelandaconvergence proof of the EMalgorithm in the spirit of our approach. Finally, Section 5 contains simulations confirmingour claimabout the efficiencyand the robustnessof ourapproach in comparisonwith theMLE.Thealgorithmisalsoappliedto theso-calledminimumdensitypowerdivergence (MDPD) introducedby[14]. 2.ADescriptionof theAlgorithm 2.1.GeneralContextandNotations Let (X,Y) be a couple of random variables with joint probability density function f(x,y|φ) parametrizedbyavectorofparametersφ∈Φ⊂Rd. Let (X1,Y1),··· , (Xn,Yn)ben copiesof (X,Y) independentlyand identicallydistributed. Finally, let (x1,y1),··· ,(xn,yn)ben realizationsof then copiesof (X,Y). Thexisare theunobserveddata (labels)andtheyisare theobservations. Thevector of parametersφ is unknownandneeds tobe estimated. Theobserveddata yi are supposed tobe realnumbers,andthe labelsxibelongtoaspaceX notnecessarilyfiniteunlessmentionedotherwise. Themarginaldensityof theobserveddata isgivenby pφ(y)= ∫ f(x,y|φ)dx,wheredx is ameasure definedonthe label space (forexample, thecountingmeasure ifweworkwithmixturemodels). Foraparametrized function f withaparameter a,wewrite f(x|a). Weuse thenotationφk for sequenceswith the indexabove. ThederivativesofarealvaluedfunctionψdefinedonRaredenoted ψ′,ψ′′, etc.Wedenote∇f thegradientofa real function f definedonRd. Forageneric functionof two (vectorial)argumentsD(φ|θ), then∇1D(φ|θ)denotes thegradientwithrespect to thefirst (vectorial) variable. Finally, foranysetA,weuse int(A) todenote the interiorofA. 2.2. EMAlgorithmandTseng’sGeneralization TheEMalgorithmestimates theunknownparametervectorby(see [15]): φk+1=argmax Φ E [ log(f(X,Y|φ)) ∣∣∣Y=y,φk] , whereX=(X1,··· ,Xn),Y=(Y1,··· ,Yn)andy=(y1,··· ,yn). By independencebetweenthecouples (Xi,Yi)’s, theprevious iterationmaybewrittenas: φk+1 = argmax Φ n ∑ i=1 E [ log(f(Xi,Yi|φ)) ∣∣∣Yi=yi,φk] = argmax Φ n ∑ i=1 ∫ X log(f(x,yi|φ))hi(x|φk)dx, (6) where hi(x|φk) = f(x,yi|φ k) p φk(yi) is the conditional density of the labels (at step k) provided yi whichwe supposetobepositivedx−almosteverywhere. It iswell-knownthat theEMiterationscanberewritten asadifferencebetweenthe log-likelihoodandaKullback–Lieblerdistance-like function. Indeed, 255