Seite - 191 - in Differential Geometrical Theory of Statistics
Bild der Seite - 191 -
Text der Seite - 191 -
Entropy2016,18, 433
5.3. Theα-ConnetionsofChentsov
Still, nowadays, the information geometry deals with models (Θ,P) whose underlying
m-dimensionalmanifoldΘ isanopensubsetof theeuclideanspaceRm.Further theFisher information
g isassumedtoberegular,viz (Θ,g) isaRiemannianmanifold. In thispaper thisclassical information
geometry iscalledthe local informationgeometry. This“localnature”willbeexplainedinPartBof
thispaper.
At themomentweplanto investigateother topologicalpropertiesof the local statisticalmodels.
Let(Θ,P)beanm-dimensional local statisticalmodel forameasurableset(Ξ,Ω). ThemanifoldΘ
isadomain in theEuclideanspaceRm.The functionP isnonnegative. It isdefinedinΘ×Ξ.Werecall
therequirementsP is subject to.
(1) ∀ξ∈Ξ the function
Θ θ→P(θ,ξ)
is smooth.
(2) ∀θ∈Θ the triple
(Ξ,Ω,P(θ,−))
isaprobabilityspace.
(3) ∀θ,θ∗∈Θwithθ = θ∗ thereexistsξ∈Ξsuchthat
P(θ,ξ) =P(θ′,ξ).
(with therequirement (3) (Θ,P) is called identifiable.)
(4) Thedifferentiationdθ commuteswiththeintegration ∫
Ξ .TheFisher informationofamodel(Θ,P)
is thesymmetricbi-linear formgwhich isdefinedby
g(X,Y)(θ)= ∫
Ξ P(θ,ξ)[[dθlog(P)]⊗2(X,Y)](θ,ξ)dξ.
Heredθ stands for thedifferentiationwithrespect to theargumentθ∈Θ.
(5) TheFisher information ispositivedefinite.
Remark 3. The Fisher information g can be defined using any Koszul connection∇ according to the
following formula
g(X,Y)(θ)=− ∫
Ξ P(θ,ξ)[(∇2log(P))(X,Y)](θ,ξ)]dξ.
Therightmemberof the last equalitydoesnotdependonthe choice of theKoszul connection∇.
Fromnowon,we dealwith a generic statisticalmodel. Thismeans thatwe do not assume the Fisher
informationg isdefinite.
Letθ=(θ1,...,θm)beasystemofEuclideancoordinate functions inRm.Toeveryrealnumberα is assigned
the torsion freeKoszul connection∇α whoseChristoffel symbols in the coordinate (θj)are
Γαij,k= ∫
Ξ P(θ,ξ)[( ∂2log(P(θ,ξ))
∂θi∂θj + 1−α
2 ∂log(P(θ,ξ))
∂θi ∂log(P(θ,ξ))
∂θj ) ∂log(p(θ,ξ))
∂θk ]dξ.
Thisdefinitionagreeswithanyaffinecoordinate change.Weput∂i= ∂∂θi .Wehave
∇α∂i∂j=∑
k Γαij,k∂k.
191
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