Seite - 258 - in Differential Geometrical Theory of Statistics
Bild der Seite - 258 -
Text der Seite - 258 -
Entropy2016,18, 277
Westart theconvergencepropertiesbyprovingthat theobjective function Dˆϕ(pφ|pφT)decreases
alongside the the sequence (φk)k, and give a possible set of conditions for the existence of the
sequence (φk)k.
Proposition1. (a)Assumethat the sequence (φk)k iswell defined inΦ, then Dˆϕ(pφk+1|pφT)≤ Dˆϕ(pφk|pφT),
and (b)∀k,φk ∈Φ0. (c) AssumeA0 andA2 are verified, then the sequence (φk)k is defined and bounded.
Moreover, the sequence (Dˆϕ(pφk|pφT))k converges.
Proof. Weprove (a).Wehavebydefinitionof thearginf:
Dˆϕ(pφk+1,pφT)+Dψ(φ k+1,φk)≤ Dˆϕ(pφk,pφT)+Dψ(φk,φk).
Weuse the fact thatDψ(φk,φk)=0for theright-handsideandthatDψ(φk+1,φk)≥0for the left-hand
sideof theprevious inequality.Hence, Dˆϕ(pφk+1,pφT)≤ Dˆϕ(pφk,pφT).
Weprove (b)using thedecreasingpropertypreviouslyproved in (a). Wehaveby recurrence
∀k,Dˆϕ(pφk+1,pφT)≤ Dˆϕ(pφk,pφT)≤···≤ Dˆϕ(pφ0,pφT). Theresult followsdirectlybydefinitionofΦ0.
Weprove(c)byinductiononk. Fork=0,clearlyφ0 iswelldefinedsincewechoose it. Thechoice
of the initialpointφ0 of thesequencemayinfluence theconvergenceof thesequence. See theExample
of theGaussianmixture inSection4.2. Suppose, for some k≥ 0, thatφk exists. Weprove that the
infimumisattained inΦ0. Letφ∈Φbeanyvectoratwhich thevalueof theoptimizedfunctionhasa
value less than itsvalueatφk, i.e., Dˆϕ(pφ,pφT)+Dψ(φ,φ k)≤ Dˆϕ(pφk,pφT)+Dψ(φk,φk).Wehave:
Dˆϕ(pφ,pφT) ≤ Dˆϕ(pφ,pφT)+Dψ(φ,φk)
≤ Dˆϕ(pφk,pφT)+Dψ(φk,φk)
≤ Dˆϕ(pφk,pφT)
≤ Dˆϕ(pφ0,pφT).
Thefirst line followsfromthenonnegativityofDψ. As Dˆϕ(pφ,pφT)≤ Dˆϕ(pφ0,pφT), thenφ∈Φ0.
Thus, the infimumcanbe calculated for vectors inΦ0 instead ofΦ. SinceΦ0 is compact and the
optimizedfunction is lowersemicontinuous(thesumof twolowersemicontinuousfunctions), then
the infimumexistsandisattained inΦ0.Wemaynowdefineφk+1 tobeavectorwhosecorresponding
value isequal to the infimum.
Convergenceof the sequence (Dˆϕ(pφk,pφT))k comes fromthe fact that it isnon increasingand
bounded. It isnon increasingbyvirtueof (a). Boundedness comes fromthe lower semicontinuity
ofφ → Dˆϕ(pφ,pφT). Indeed,∀k,Dˆϕ(pφk,pφT)≥ infφ∈Φ0 Dˆϕ(pφ,pφT). The infimumofaproper lower
semicontinuous function on a compact set exists and is attained on this set. Hence, the quantity
infφ∈Φ0 Dˆϕ(pφ,pφT)existsandisfinite. Thisends theproof.
Compactness inpart (c) canbe replacedby inf-compactness of functionφ → Dˆϕ(pφ|pφT)and
continuityofDψ withrespect to itsfirstargument. Theconvergenceof thesequence (Dˆϕ(φk|φT))k is
an interestingproperty, since, ingeneral, there isno theoreticalguarantee,or it isdifficult toprove
that thewholesequence (φk)k converges. Itmayalsocontinuetofluctuatearoundaminimum.The
decreaseof the error criterion Dˆϕ(φk|φT)between two iterationshelpsusdecidewhen to stop the
iterativeprocedure.
Proposition2. SupposeA1verified,Φ0 is closedand{φk+1−φk}→0.
(a) IfACisverified, thenany limitpointof (φk)k is a stationarypointofφ → Dˆϕ(pφ|pφT);
(b) IfACisdropped, thenany limitpoint of (φk)k is a“generalized”stationarypoint ofφ → Dˆϕ(pφ|pφT),
i.e., zerobelongs to the subgradientofφ → Dˆϕ(pφ|pφT) calculatedat the limitpoint.
258
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