Seite - 196 - in Differential Geometrical Theory of Statistics
Bild der Seite - 196 -
Text der Seite - 196 -
Entropy2016,18, 433
The identities above showthat
∂2a
∂θi∂θj (θ0)=0
for allθ0∈U.Thereby the function
θ→ψ(θ)+ log(P(θ,ξ))
dependsaffinelyonθ1,...,θm. So there exists aRm+1-valued function
Ξ ξ→ (C(ξ),F1(ξ),...,Fm(ξ))∈Rm+1
such that
ψ(θ)+ log(P(θ,ξ))=C(ξ)+ m
∑
1 Fi(ξ)θi.
Infinalweget
P(θ,ξ)= exp(C(ξ)+∑Fi(ξ)θi−ψ(θ))
for all (θ,ξ)∈U×Ξ. So (Θ,P) is locallyanexponential family.
Since the exponential function is injective this local property of (Θ,P) is a global property, in other
words themodel is globally an exponentialmodel. In final assertion (1) implies assertion(2). This ends the
demonstrationof the theorem
SomeComments.
(i) Itmust be noticed that the demonstration above is independent of the rank of the Fisher information g.
Therefore, the theoremholds insingular statisticalmodels.
(ii) In regular statisticalmodels the theoremabove leads to thenotionof e-m-flatnessas in [18].
(iii)When the Fisher information g is semi-definite the dualistic relation is meaningless. However data
(Θ,g,∇,∇∗)mayberegardedasdatadependingonthe transversal structureof thedistributionKer(g).
(iv) In theanalytic category theFisher information isaRiemannian foliation. Therefore, both the information
geometry and the topology of information are transversal concepts. Thismay be called the transversal
geometryandthe transversal topologyofFisher-Riemannian foliations.
(v) The theoremabovedoesnot solve thequestionashowfar frombeinganexponential family is agivenmodel.
It only tellsus that exponential families areobjects of theHessiangeometry.
Theframeworkforaddressingthechallenge justmentionedis the theoryof invariants. That is the
purposeofa forthgoingwork. Somenewresultsareanouncedin theAppendixAtothispaper.
6.TheSimilarityStructureandtheHyperbolicity
Weconsideraduallyflatpair(M,g,∇,∇∗). Both(M,∇,g)and(M,g,∇∗)are locallyhyperbolic in
thesenseof [2]. So they locallysupport thegeometryofKoszul. That isaconsequenceof theclassical
LemmaofPoincare.
EverypointofMhasanopenneighborhoodUsupportingalocaldeRhamcloseddifferential1-forms
ω∈C1KV(A,R)
and
ω∗ ∈C1KV(A∗,R)
subject to the followingrequirements
g|U= δω,
g|U= δ∗ω∗.
196
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