Seite - 83 - in Differential Geometrical Theory of Statistics
Bild der Seite - 83 -
Text der Seite - 83 -
Entropy2016,18, 386
Figure9.GenerationofKoszulelements fromCartan innerproduct.
This informationgeometryhasbeen intensivelystudiedforstructuredmatrices [151–166]andin
statistics [167]andis linkedto theseminalworkofSiegel [168]onsymmetricboundeddomains.
Wecanapply thisKoszulgeometry frameworkforconesofsymmetricpositivedefinitematrices.
Let the innerproduct 〈η,ξ〉=Tr(ηTξ) ,∀η,ξ∈Sym(n)givenbyCartan-Killingform,Ωbethesetof
symmetricpositivedefinitematrices isanopenconvexconeandisself-dualΩ∗=Ω.
〈η,ξ〉=Tr(ηTξ) ,∀η,ξ∈Sym(n)
ψΩ(β)=
Ω∗ e−〈β,ξ〉dξ=det(β)−n+12 ψΩ(Id)
ξˆ= ∂Φ(β)
∂β = ∂(−logψΩ(β))
∂β = n+1
2 β−1 (158)
pξˆ(ξ)= e −〈Θ−1(ξˆ),ξ〉+Φ(Θ−1(ξˆ)) =ψΩ(Id) · [
det (
αξˆ−1 )] ·e−Tr(αξˆ−1ξ)
withα= n+1
2 (159)
Wewill in the following illustrate informationgeometry formultivariateGaussiandensity [169]:
pξˆ(ξ)= 1
(2π)n/2det(R)1/2 e− 1
2(z−m)TR−1(z−m) (160)
Ifwedevelop:
1
2 (z−m)TR−1(z−m) = 1
2 [ zTR−1z−mTR−1z−zTR−1m+mTR−1m]
= 1
2 zTR−1z−mTR−1z+ 12mTR−1m (161)
Wecanwrite thedensityasaGibbsdensity:
pξˆ(ξ)= 1
(2π)n/2det(R)1/2e 1
2m TR−1m e−[−mTR−1z+12zTR−1z] = 1
Z e−〈ξ,β〉
ξ= [
z
zzT ]
andβ= ⎡⎣ −R−1m1
2 R−1 ⎤⎦=[ a
H ]
with 〈ξ,β〉= aTz+zTHz=Tr[zaT+HTzzT] (162)
83
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