Seite - 298 - in Differential Geometrical Theory of Statistics
Bild der Seite - 298 -
Text der Seite - 298 -
Entropy2016,18, 442
4.1. BasicBounds
Forapairofgivenm(x)andm′(x),weonlyneedtoderiveboundsofHα(m :m′) inEquation(31)
so thatLα(m :m′)≤Hα(m :m′)≤Uα(m :m′). Then theα-divergenceDα(m :m′) canbebounded
byalinear transformationof therange [Lα(m :m′),Uα(m :m′)]. In thefollowingwealwaysassume
without lossofgeneralityα≥1/2.OtherwisewecanboundDα(m :m′)byconsideringequivalently
theboundsofD1−α(m′ :m).
Recall that ineachelementaryslab Ir,wehave
max {
kw (r)p (r)(x),wδ(r)pδ(r)(x) }
≤m(x)≤ kwδ(r)pδ(r)(x). (32)
Notice thatkw (r)p (r)(x),wδ(r)pδ(r)(x), andkwδ(r)pδ(r)(x)areall singlecomponentdistributions
up toascalingcoefficient. Thegeneral thinking is toboundthemulti-componentmixturem(x)by
singlecomponentdistributions ineachelementary interval, so that the integral inEquation(31)canbe
computedinapiecewisemanner.
For theconvenienceofnotation,werewriteEquation(32)as
cν(r)pν(r)(x)≤m(x)≤ cδ(r)pδ(r)(x), (33)
where
cν(r)pν(r)(x) := kw (r)p (r)(x) or wδ(r)pδ(r)(x),
cδ(r)pδ(r)(x) := kwδ(r)pδ(r)(x). (34)
(35)
If1/2≤α<1, thenbothxα andx1−α aremonotonically increasingonR+. Thereforewehave
Aαν(r),ν′(r)(Ir)≤ ∫
Ir m(x)αm′(x)1−αdx≤Aαδ(r),δ′(r)(Ir), (36)
where
Aαi,j(I)= ∫
I (cipi(x)) α (
c′jp ′
j(x) )1−α
dx, (37)
and Idenotesan interval I=(a,b)⊂R. Theothercaseα>1 is similarbynoting thatxα andx1−α are
monotonically increasinganddecreasingonR+, respectively. Inconclusion,weobtain the following
boundsofHα(m :m′):
If 1/2≤α<1, Lα(m :m′)= ∑
r=1 Aαν(r),ν′(r)(Ir), Uα(m :m ′)= ∑
r=1 Aαδ(r),δ′(r)(Ir); (38)
ifα>1, Lα(m :m′)= ∑
r=1 Aαν(r),δ′(r)(Ir), Uα(m :m ′)= ∑
r=1 Aαδ(r),ν′(r)(Ir). (39)
The remaining problem is to compute the definite integral Aαi,j(I) in the above equations.
Here we assume all mixture components are in the same exponential family so that
pi(x)= p(x;θi)= h(x)exp (
θ i t(x)−F(θi) )
,whereh(x) isabasemeasure, t(x) isavectorofsufficient
statistics, andthe functionF isknownas thecumulantgeneratingfunction. Then it is straightforward
fromEquation(37) that
Aαi,j(I)= c α
i (c ′
j) 1−α ∫
I h(x)exp ((
αθi+(1−α)θ′j )
t(x)−αF(θi)−(1−α)F(θ′j) )
dx. (40)
298
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