Seite - 307 - in Differential Geometrical Theory of Statistics

Entropy2016,18, 442 (anadditivelyweightedBregman–Voronoidiagram[49,50] for componentsbelonging to the same exponential family). However, it becomesmore complex to compute in the elementaryVoronoi cellsV, the functionsCi,j(V)andMi(V) (in1D, theVoronoicellsaresegments).Wemayobtainhybrid algorithmsbyapproximatingorestimatingthese functions. In2D, it is thuspossible toobtain lower andupperboundson theMutual Information [51] (MI)when the jointdistributionm(x,y) is a 2D mixtureofGaussians: I(M;M′)= ∫ m(x,y) log m(x,y) m(x)m′(y)dxdy. Indeed, themarginaldistributionsm(x)andm′(y)areunivariateGaussianmixtures. APythoncode implementing thosecomputational-geometricmethods for reproducible research isavailableonline [52]. Acknowledgments: Theauthorsgratefully thank the referees for their comments. Thisworkwascarriedout whileKeSunwasvisitingFrankNielsenatEcolePolytechnique,Palaiseau,France. AuthorContributions:FrankNielsenandKeSuncontributedto the theoretical resultsaswellas to thewritingof thearticle.KeSunimplementedthemethodsandperformedthenumericalexperiments.Allauthorshaveread andapprovedthefinalmanuscript. Conflictsof Interest:Theauthorsdeclarenoconflictof interest. AppendixA. TheKullback–LeiblerDivergenceofMixtureModelsIsNotAnalytic [6] Ideally,weaimatgettingafinite lengthclosed-formformula tocompute theKLdivergenceof finitemixturemodels.However, this isprovablymathematically intractable [6]becauseof the log-sum terminthe integral, asweshallprovebelow. Analytic expressions encompass closed-form formula and may include special functions (e.g., Gamma function) but do not allow to use limits or integrals. An analytic function f(x) is a C∞ function (infinitely differentiable) such that around any point x0 the k-order Taylor series Tk(x)=∑ki=0 f(i)(x0) i! (x−x0)i converges to f(x): limk→∞Tk(x) = f(x) for x belonging to a neighborhood Nr(x0) = {x : |x−x0| ≤ r} of x0, where r is called the radius of convergence. Theanalyticpropertyofa function isequivalent to thecondition that foreach k∈N, thereexistsa constant c suchthat ∣∣∣dk fdxk(x)∣∣∣≤ ck+1k!. Toprove that theKLofmixtures isnotanalytic (hencedoesnotadmitaclosed-formformula), we shall adapt the proof reported in [6] (in Japanese, we thank Professor Aoyagi for sending us his paper [6]). We shall prove that KL(p : q) is not analytic for p(x) = G(x;0,1) and q(x;w)=(1−w)G(x;0,1)+wG(x;1,1),wherew∈ (0,1), andG(x;μ,σ)= 1√ 2πσ exp(−(x−μ)22σ2 ) is the densityofaunivariateGaussianofmeanμandstandarddeviationσ. LetD(w)=KL(p(x) : q(x;w)) denote the KL divergence between these two mixtures (p has a single component and q has twocomponents). Wehave log p(x) q(x;w) = log exp ( −x22 ) (1−w)exp ( −x22 ) +wexp ( −(x−1)22 )=− log(1+w(ex−12−1)). (A1) Therefore dkD dwk = (−1)k k ∫ p(x)(ex− 1 2−1)dx. Let x0 be the root of the equation ex− 1 2 −1= ex2 so that for x≥ x0, wehave ex−12 −1≥ ex2 . It followsthat: ∣∣∣∣∣dkDdwk ∣∣∣∣∣≥ 1k ∫ ∞ x0 p(x)e kx 2 dx= 1 k e k2 8 Ak 307