Seite - 322 - in Differential Geometrical Theory of Statistics
Bild der Seite - 322 -
Text der Seite - 322 -
Entropy2017,19, 7
ThemetricgN is conformallyequivalent togMwithconformal factorZq(pθ)= ∫ Ω{p(x;θ)}qdx.
That is, gN(θ)=Zq(pθ)gM(θ). (Seealso [6]). Naudtsgavea furthergeneralizationofFishermetric
andheshowedaCramér–Raotypeboundtheorem[5].
6.ConcludingRemarks
In thispaper,we introducedasequenceofescortdistributions. Thenwegaverepresentationsof
Riemannianmetricsandcubic formsfromaviewpointof thesequenceofescortdistributions.
Inparticular,wecandefinethefollowing(0,2)-tensorfieldsonaq-exponential family. Forpθ ∈Sq,
setηi= ∂iψ(θ).
(1) Fromthestandardexpectation,weobtain
g(0)ij (θ) := Gij(θ) := ∫
Ω (∂i lnq pθ)(∂j lnq pθ)pθdx
= Ep[(Fi(x)−ηi)(Fj(x)−ηj)].
The tensorG isacovariancematrix.However,Gmaynotbe important inanomalousstatistics.
(2) Fromtheq-expectation,weobtain
g(1)ij (θ) := g M
ij (θ) = ∫
Ω (∂i lnq pθ)(∂j lnq pθ){pθ}qdx
= ∫
Ω (∂i lnq pθ)(∂j lnq pθ)Pq(x;θ)dx
= Eq,p[(Fi(x)−ηi)(Fj(x)−ηj)].
TheRiemannianmetricgM isaHessianmetric, andit is inducedfromtheβ-divergence (20).
(3) Fromtheexpectationwithrespect to thesecondescortdistribution,weobtain
g(2)ij (θ) := g F
ij(θ) = ∫
Ω (∂i lnq pθ)(∂j lnq pθ){pθ}2q−1dx
= 1
q ∫
Ω (∂i lnq pθ)(∂j lnq pθ)P˜q(x;θ)dx
= 1
q Eq,(2),p[(Fi(x)−ηi)(Fj(x)−ηj)].
gqij(θ) = Zq(p)
q gFij.
TheRiemannianmetric gF is a Fishermetric. Hence gF is invariant to the choiceof reference
measureonΩ, but it isnotaHessianmetric. Inaddition, gF is inducedfromtheα-divergence (14).
TheconformalRiemannianmetricgq isaq-Fishermetric. It isaHessianmetric, and it is inducedfrom
anormalizedTsallis relativeentropy(15).
WemaydefineaRiemannianmetricandacubic formfromhigherorderescortexpectations:
g(n)ij (θ) := ∫
Ω (∂i lnq pθ)(∂j lnq pθ)Pq,(n)(x;θ)dx,
C(n)ij (θ) := ∫
Ω (∂i lnq pθ)(∂j lnq pθ)(∂k lnq pθ)Pq,(n+1)(x;θ)dx.
Thenweobtainasequenceofstatisticalmanifoldstructures.
(Sq,g(1),C(1)) → (Sq,g(2),C(2)) → ··· → (Sq,g(n),C(n)) → ···
However, the geometric meaning of this sequence is not clear at this moment. Elucidating
geometricpropertiesof this sequence isa futureproblem.
322
Differential Geometrical Theory of Statistics
- Titel
- Differential Geometrical Theory of Statistics
- Autoren
- Frédéric Barbaresco
- Frank Nielsen
- Herausgeber
- MDPI
- Ort
- Basel
- Datum
- 2017
- Sprache
- englisch
- Lizenz
- CC BY-NC-ND 4.0
- ISBN
- 978-3-03842-425-3
- Abmessungen
- 17.0 x 24.4 cm
- Seiten
- 476
- Schlagwörter
- Entropy, Coding Theory, Maximum entropy, Information geometry, Computational Information Geometry, Hessian Geometry, Divergence Geometry, Information topology, Cohomology, Shape Space, Statistical physics, Thermodynamics
- Kategorien
- Naturwissenschaften Physik