RでKLダイバージェンス

RでKLダイバージェンス（仮）
0を除かないと発散するので、除くのがポイント

for(i in 1:length(list1)){
list_name<-list1[i]

	v1<-qb[[list_name]]

	a1<-v1[qb$y==1]
	a2<-v1[qb$y==0]

	aa1<-table(a1)
	aa2<-table(a2)

	ll<-unique(c(names(aa1),names(aa2)))	
	aaa1<-NULL
	aaa2<-NULL
	for(i in 1:length(ll)){
		aaa1[i]<-aa1[ll[i]]
		aaa2[i]<-aa2[ll[i]]
	}
	aaa1<-as.vector(aaa1)
	aaa2<-as.vector(aaa2)
	aaa1[is.na(aaa1)]<-0
	aaa2[is.na(aaa2)]<-0

#aaa2[is.na(aaa2)]<-0

#m<-matrix(c(aaa1,aaa2),nrow=2)

	prob<-(aaa1+aaa2)/sum(aaa1+aaa2)
	aaa1<-aaa1[which(prob!=0)]
	aaa2<-aaa2[which(prob!=0)]
	prob<-prob[prob!=0]
	
	entropy1<--sum(log((aaa1^(aaa1/sum(aaa1)))/sum(aaa1)))
	entropy2<--sum(log((aaa2^(aaa2/sum(aaa2)))/sum(aaa2)))
	aaa2b<-aaa2[(aaa1==0 & aaa2==0 )| (aaa1!=0)]
	aaa1b<-aaa1[(aaa1==0 & aaa2==0) | (aaa1!=0)]
	

	v1<-log(((aaa2b/sum(aaa2b))^(aaa2b/sum(aaa2b))))
	v2<-log(((aaa1b/sum(aaa1b))^(aaa2b/sum(aaa2b))))
	info1<- sum(v1-v2)
	
	info2<-KL.empirical(aaa2b,aaa1b)

#chisq.test()
	cat(list_name,"\n")
	str<-chisq.test(aaa1,p=prob)
	cat(str$p.value,"\n")
	cat("ent1",entropy1,"\n")
	cat("ent2",entropy2,"\n")
	cat("info1",info1,"\n")	
	cat("info2",info2,"\n")