Page - 259 - in Differential Geometrical Theory of Statistics
Image of the Page - 259 -
Text of the Page - 259 -
Entropy2016,18, 277
Proof. Weprove (a). Let (φnk)kbeaconvergent subsequenceof (φk)kwhichconverges toφ∞. First,
φ∞∈Φ0,becauseΦ0 is closedandthesubsequence (φnk) isasequenceofelementsofΦ0 (provedin
Proposition1b).
Letusnowshowthat thesubsequence (φnk+1)alsoconverges toφ∞.Wesimplyhave:
‖φnk+1−φ∞‖ ≤ ‖φnk−φ∞‖+‖φnk+1−φnk‖.
Sinceφk+1−φk→0andφnk→φ∞,weconcludethatφnk+1→φ∞.
By definition of φnk+1, it verifies the infimum in recurrence (11), so that the gradient of the
optimizedfunction iszero:
∇Dˆϕ(pφnk+1,pφT)+∇Dψ(φnk+1,φnk)=0.
Using the continuity assumptionsA1andACof thegradients, one canpass to the limitwith
noproblem:
∇Dˆϕ(pφ∞,pφT)+∇Dψ(φ∞,φ∞)=0.
However, thegradient∇Dψ(φ∞,φ∞)=0because (recall thatψ′(1)=0) foranyφ∈Φ
∇Dψ(φ,φ)= n
∑
i=1 ∫
X ∇hi(x|φ)
hi(x|φ) ψ ′ ( hi(x|φ)
hi(x|φ) )
hi(x|φ)dx= n
∑
i=1 ∫
X ∇hi(x|φ)ψ′(1)dx,
which isequal tozerosinceψ′(1)=0. This implies that∇Dˆϕ(pφ∞,pφT)=0.
Weprove(b).Weuseagainthedefinitionofthearginf.Astheoptimizedfunctionisnotnecessarily
differentiableat thepointsof thesequence(φk)k, anecessaryconditionforφk+1 tobeaninfimumisthat
0belongs to thesubgradientof the functiononφk+1. SinceDψ(φ,φk) isassumedtobedifferentiable,
theoptimalitycondition is translated into:
−∇Dψ(φk+1,φk)∈∂Dˆϕ(pφk+1,pφT) ∀k.
Since Dˆϕ(pφ,pφT) is continuous, then its subgradient isoutersemicontinuous (see [13]Chapter8,
Proposition 7). We use the same arguments presented in (a) to conclude the existence of two
subsequences (φnk)k and (φnk+1)k which converge to the same limit φ∞. By definition of outer
semicontinuity,andsinceφnk+1→φ∞,wehave:
limsup
φnk+1→φ∞ ∂Dˆϕ(pφnk+1,pφT)⊂∂Dˆϕ(pφ∞,pφT). (12)
Wewant toprove that0∈ limsup
φnk+1→φ∞∂Dˆϕ(pφnk+1,pφT). Bydefinitionof the (outer) limsup
(see [13]Chapter4,Definition1orChapter5B):
limsup
φ→φ∞ ∂Dˆϕ(pφ,pφT)= {
u|∃φk→φ∞,∃uk→uwithuk∈∂Dˆϕ(pφk,pφT) }
.
Inourscenario,φ=φnk+1,φk=φnk+1,u=0anduk=∇1Dψ(φnk+1,φnk). Thecontinuityof∇1Dψ
withrespect tobothargumentsandthefact that the twosubsequencesφnk+1 andφnk converge to the
same limit, imply thatuk →∇1Dψ(φ∞,φ∞) = 0. Hence,u= 0∈ limsupφnk+1→φ∞∂Dˆϕ(pφnk+1,pφT).
By inclusion(12),wegetourresult:
0∈∂Dˆϕ(pφ∞,pφT).
Thisends theproof.
Theassumption{φk+1−φk}→0used inProposition2 isnoteasy tobecheckedunlessonehasa
close formulaofφk. Thefollowingpropositiongivesamethodtoprovesuchassumption. Thismethod
259
Differential Geometrical Theory of Statistics
- Title
- Differential Geometrical Theory of Statistics
- Authors
- Frédéric Barbaresco
- Frank Nielsen
- Editor
- MDPI
- Location
- Basel
- Date
- 2017
- Language
- English
- License
- CC BY-NC-ND 4.0
- ISBN
- 978-3-03842-425-3
- Size
- 17.0 x 24.4 cm
- Pages
- 476
- Keywords
- Entropy, Coding Theory, Maximum entropy, Information geometry, Computational Information Geometry, Hessian Geometry, Divergence Geometry, Information topology, Cohomology, Shape Space, Statistical physics, Thermodynamics
- Categories
- Naturwissenschaften Physik