Page - 298 - in Differential Geometrical Theory of Statistics
Image of the Page - 298 -
Text of the Page - 298 -
Entropy2016,18, 442
4.1. BasicBounds
Forapairofgivenm(x)andmā²(x),weonlyneedtoderiveboundsofHα(m :mā²) inEquation(31)
so thatLα(m :mā²)ā¤Hα(m :mā²)ā¤Uα(m :mā²). Then theα-divergenceDα(m :mā²) canbebounded
byalinear transformationof therange [Lα(m :mā²),Uα(m :mā²)]. In thefollowingwealwaysassume
without lossofgeneralityαā„1/2.OtherwisewecanboundDα(m :mā²)byconsideringequivalently
theboundsofD1āα(mā² :m).
Recall that ineachelementaryslab Ir,wehave
max {
kw (r)p (r)(x),wΓ(r)pΓ(r)(x) }
ā¤m(x)⤠kwĪ“(r)pĪ“(r)(x). (32)
Notice thatkw (r)p (r)(x),wΓ(r)pΓ(r)(x), andkwΓ(r)pΓ(r)(x)areall singlecomponentdistributions
up toascalingcoefļ¬cient. 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)ā¤m(x)⤠cĪ“(r)pĪ“(r)(x), (33)
where
cν(r)pν(r)(x) := kw (r)p (r)(x) or wΓ(r)pΓ(r)(x),
cΓ(r)pΓ(r)(x) := kwΓ(r)pΓ(r)(x). (34)
(35)
If1/2ā¤Ī±<1, thenbothxα andx1āα aremonotonically increasingonR+. Thereforewehave
Aαν(r),νā²(r)(Ir)⤠ā«
Ir m(x)αmā²(x)1āαdxā¤AαΓ(r),Ī“ā²(r)(Ir), (36)
where
Aαi,j(I)= ā«
I (cipi(x)) α (
cā²jp ā²
j(x) )1āα
dx, (37)
and Idenotesan interval I=(a,b)āR. Theothercaseα>1 is similarbynoting thatxα andx1āα are
monotonically increasinganddecreasingonR+, respectively. Inconclusion,weobtain the following
boundsofHα(m :mā²):
If 1/2ā¤Ī±<1, Lα(m :mā²)= ā
r=1 Aαν(r),νā²(r)(Ir), Uα(m :m ā²)= ā
r=1 AαΓ(r),Ī“ā²(r)(Ir); (38)
ifα>1, Lα(m :mā²)= ā
r=1 Aαν(r),Ī“ā²(r)(Ir), Uα(m :m ā²)= ā
r=1 AαΓ(r),νā²(r)(Ir). (39)
The remaining problem is to compute the deļ¬nite 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;Īøi)= h(x)exp (
Īø i t(x)āF(Īøi) )
,whereh(x) isabasemeasure, t(x) isavectorofsufļ¬cient
statistics, andthe functionF isknownas thecumulantgeneratingfunction. Then it is straightforward
fromEquation(37) that
Aαi,j(I)= c α
i (c ā²
j) 1āα ā«
I h(x)exp ((
αθi+(1āα)Īøā²j )
t(x)āαF(Īøi)ā(1āα)F(Īøā²j) )
dx. (40)
298
Differential Geometrical Theory of Statistics
- Title
- Differential Geometrical Theory of Statistics
- Authors
- FrƩdƩric 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