Seite - 261 - in Differential Geometrical Theory of Statistics

Entropy2016,18, 277 If (φk)kdoesnotconverge. SinceΦ0 is compactand∀k,φk∈Φ0 (provedinProposition1), there exists a subsequence (φN0(k))k such thatφN0(k)→ φ˜. Let us take the subsequence (φN0(k)−1)k. This subsequencedoesnotnecessarilyconverge; it isstillcontainedinthecompactΦ0, sothatwecanextract a further subsequence (φN1◦N0(k)−1)kwhichconverges to, say, φ¯. Now, thesubsequence (φN1◦N0(k))k converges to φ˜, because it isasubsequenceof (φN0(k))k.Wehaveproveduntilnowtheexistenceof two convergentsubsequencesφN(k)−1 andφN(k)withaprioridifferent limits. Forsimplicityandwithout any lossofgenerality,wewill consider thesesubsequences tobeφk andφk+1, respectively. Conservingpreviousnotations, suppose thatφk+1→ φ˜andφk→ φ¯.Weuseagain inequality (13): Dˆ(pφk+1,pφT)+Dψ(φ k+1,φk)≤ Dˆ(pφk,pφT). Bytakingthelimitsof thetwopartsof theinequalityask tendstoinfinity,andusingthecontinuity of the twofunctions,wehave Dˆ(pφ˜,pφT)+Dψ(φ˜,φ¯)≤ Dˆ(pφ¯,pφT). Recall that underA1-2, the sequence ( Dˆϕ(pφk,pφT) ) k converges, so that it has the same limit for any subsequence, i.e., Dˆ(pφ˜,pφT) = Dˆ(pφ¯,pφT). We also use the fact that the distance-like functionDψ is nonnegative todeduce thatDψ(φ˜,φ¯) = 0. Looking closely at thedefinitionof this divergence(10),wegetthatif thesumiszero, theneachtermisalsozerosinceall termsarenonnegative. Thismeans that: ∀i∈{1,··· ,n}, ∫ X ψ ( hi(x|φ˜) hi(x|φ¯) ) hi(x|φ¯)dx=0. The integrandsarenonnegative functions, so theyvanishalmosteverywherewithrespect to the measuredxdefinedonthespaceof labels. ∀i∈{1,··· ,n}, ψ ( hi(x|φ˜) hi(x|φ¯) ) hi(x|φ¯)=0 dx−a.e. Theconditionaldensitieshiaresupposedtobepositive(whichcanbeensuredbyasuitablechoice of the initialpointφ0), i.e.,hi(x|φ¯)>0,dx−a.e.Hence,ψ ( hi(x|φ˜) hi(x|φ¯) ) =0,dx−a.e.Ontheotherhand,ψ is chosen inawaythatψ(z)=0 iffz=1. Therefore: ∀i∈{1,··· ,n}, hi(x|φ˜)=hi(x|φ¯) dx−a.e. (14) Sinceφk+1 is,bydefinition,an infimumofφ → Dˆ(pφ,pφT)+Dψ(φ,φk), thenthegradientof this function iszeroonφk+1. It results that: ∇Dˆ(pφk+1,pφT)+∇Dψ(φk+1,φk)=0, ∀k. Takingthe limitonk, andusingthecontinuityof thederivatives,weget that: ∇Dˆ(pφ˜,pφT)+∇Dψ(φ˜,φ¯)=0. (15) Letuswriteexplicitly thegradientof theseconddivergence: ∇Dψ(φ˜,φ¯)= n ∑ i=1 ∫ X ∇hi(x|φ˜) hi(x|φ¯) ψ ′ ( hi(x|φ˜) hi(x|φ¯) ) hi(x|φ¯). Weusenowthe identities (14), andthe fact thatψ′(1)=0, todeduce that: ∇Dψ(φ˜,φ¯)=0. 261